* Re: [PATCH libibverbs v3 1/3] Add new call ibv_cmd_create_ah_ex which supports extra parameters
From: Knut Omang @ 2016-09-20 14:05 UTC (permalink / raw)
To: Leon Romanovsky
Cc: Yishai Hadas, Doug Ledford, linux-rdma-u79uwXL29TY76Z2rM5mHXA,
Mukesh Kacker, Jason Gunthorpe
In-Reply-To: <20160920132945.GO26673-2ukJVAZIZ/Y@public.gmane.org>
On Tue, 2016-09-20 at 16:29 +0300, Leon Romanovsky wrote:
> On Tue, Sep 20, 2016 at 01:37:39PM +0200, Knut Omang wrote:
> > On Tue, 2016-09-20 at 14:17 +0300, Yishai Hadas wrote:
> > > On 9/20/2016 9:22 AM, Knut Omang wrote:
> > > >
> > > > The original call ibv_cmd_create_ah does not allow vendor specific request or
> > > > response parameters to be defined and passed through to kernel space.
> > > >
> > > > To keep backward compatibility, introduce a new extended call which accepts
> > > > cmd and resp parameters similar to other parts of the API.
> > > > Reimplement the old call to just use the new extended call.
> > > >
> > > > The new call is versioned in the shared library as IBVERBS_1.4.
> > > >
> > > > Signed-off-by: Knut Omang <knut.omang-QHcLZuEGTsvQT0dZR+AlfA@public.gmane.org>
> > > > Reviewed-by: Mukesh Kacker <mukesh.kacker-QHcLZuEGTsvQT0dZR+AlfA@public.gmane.org>
> > > > ---
> > > > include/infiniband/driver.h | 4 +++-
> > > > src/cmd.c | 50 +++++++++++++++++++++-----------------
> > > > src/libibverbs.map | 6 ++++-
> > > > 3 files changed, 37 insertions(+), 23 deletions(-)
> > > >
> > > > diff --git a/include/infiniband/driver.h b/include/infiniband/driver.h
> > > > index ea3dade..9356dc8 100644
> > > > --- a/include/infiniband/driver.h
> > > > +++ b/include/infiniband/driver.h
> > > > @@ -234,6 +234,10 @@ int ibv_cmd_post_srq_recv(struct ibv_srq *srq, struct ibv_recv_wr *wr,
> > > > struct ibv_recv_wr **bad_wr);
> > > > int ibv_cmd_create_ah(struct ibv_pd *pd, struct ibv_ah *ah,
> > > > struct ibv_ah_attr *attr);
> > > > +int ibv_cmd_create_ah_ex(struct ibv_pd *pd, struct ibv_ah *ah,
> > > > + struct ibv_ah_attr *attr,
> > > > + struct ibv_create_ah *cmd, size_t cmd_size,
> > > > + struct ibv_create_ah_resp *resp, size_t resp_size);
> > > > int ibv_cmd_destroy_ah(struct ibv_ah *ah);
> > > > int ibv_cmd_attach_mcast(struct ibv_qp *qp, const union ibv_gid *gid, uint16_t lid);
> > > > int ibv_cmd_detach_mcast(struct ibv_qp *qp, const union ibv_gid *gid, uint16_t lid);
> > > > diff --git a/src/cmd.c b/src/cmd.c
> > > > index 11f6509..58b9dcd 100644
> > > > --- a/src/cmd.c
> > > > +++ b/src/cmd.c
> > > > @@ -1493,38 +1493,44 @@ int ibv_cmd_post_srq_recv(struct ibv_srq *srq, struct ibv_recv_wr *wr,
> > > > return ret;
> > > > }
> > > >
> > > > -int ibv_cmd_create_ah(struct ibv_pd *pd, struct ibv_ah *ah,
> > > > - struct ibv_ah_attr *attr)
> > > > +int ibv_cmd_create_ah_ex(struct ibv_pd *pd, struct ibv_ah *ah, struct ibv_ah_attr *attr,
> > > > + struct ibv_create_ah *cmd, size_t cmd_size,
> > > > + struct ibv_create_ah_resp *resp, size_t resp_size)
> > > > {
> > > > - struct ibv_create_ah cmd;
> > > > - struct ibv_create_ah_resp resp;
> > > > -
> > > > - IBV_INIT_CMD_RESP(&cmd, sizeof cmd, CREATE_AH, &resp, sizeof resp);
> > > > - cmd.user_handle = (uintptr_t) ah;
> > > > - cmd.pd_handle = pd->handle;
> > > > - cmd.attr.dlid = attr->dlid;
> > > > - cmd.attr.sl = attr->sl;
> > > > - cmd.attr.src_path_bits = attr->src_path_bits;
> > > > - cmd.attr.static_rate = attr->static_rate;
> > > > - cmd.attr.is_global = attr->is_global;
> > > > - cmd.attr.port_num = attr->port_num;
> > > > - cmd.attr.grh.flow_label = attr->grh.flow_label;
> > > > - cmd.attr.grh.sgid_index = attr->grh.sgid_index;
> > > > - cmd.attr.grh.hop_limit = attr->grh.hop_limit;
> > > > - cmd.attr.grh.traffic_class = attr->grh.traffic_class;
> > > > - memcpy(cmd.attr.grh.dgid, attr->grh.dgid.raw, 16);
> > > > + IBV_INIT_CMD_RESP(cmd, cmd_size, CREATE_AH, resp, resp_size);
> > > > + cmd->user_handle = (uintptr_t) ah;
> > > > + cmd->pd_handle = pd->handle;
> > > > + cmd->attr.dlid = attr->dlid;
> > > > + cmd->attr.sl = attr->sl;
> > > > + cmd->attr.src_path_bits = attr->src_path_bits;
> > > > + cmd->attr.static_rate = attr->static_rate;
> > > > + cmd->attr.is_global = attr->is_global;
> > > > + cmd->attr.port_num = attr->port_num;
> > > > + cmd->attr.grh.flow_label = attr->grh.flow_label;
> > > > + cmd->attr.grh.sgid_index = attr->grh.sgid_index;
> > > > + cmd->attr.grh.hop_limit = attr->grh.hop_limit;
> > > > + cmd->attr.grh.traffic_class = attr->grh.traffic_class;
> > > > + memcpy(cmd->attr.grh.dgid, attr->grh.dgid.raw, 16);
> > > >
> > > > - if (write(pd->context->cmd_fd, &cmd, sizeof cmd) != sizeof cmd)
> > > > + if (write(pd->context->cmd_fd, cmd, cmd_size) != cmd_size)
> > > > return errno;
> > > >
> > > > - (void) VALGRIND_MAKE_MEM_DEFINED(&resp, sizeof resp);
> > > > + (void) VALGRIND_MAKE_MEM_DEFINED(resp, sizeof *resp);
> > > >
> > > > - ah->handle = resp.handle;
> > > > + ah->handle = resp->handle;
> > > > ah->context = pd->context;
> > > >
> > > > return 0;
> > > > }
> > > >
> > > > +int ibv_cmd_create_ah(struct ibv_pd *pd, struct ibv_ah *ah,
> > > > + struct ibv_ah_attr *attr)
> > > > +{
> > > > + struct ibv_create_ah cmd;
> > > > + struct ibv_create_ah_resp resp;
> > > > + return ibv_cmd_create_ah_ex(pd, ah, attr, &cmd, sizeof(cmd), &resp, sizeof(resp));
> > > > +}
> > > > +
> > > > int ibv_cmd_destroy_ah(struct ibv_ah *ah)
> > > > {
> > > > struct ibv_destroy_ah cmd;
> > > > diff --git a/src/libibverbs.map b/src/libibverbs.map
> > > > index 46744e5..ddba63a 100644
> > > > --- a/src/libibverbs.map
> > > > +++ b/src/libibverbs.map
> > > > @@ -118,7 +118,6 @@ IBVERBS_1.1 {
> > > > ibv_cmd_create_qp_ex2;
> > > > ibv_cmd_open_qp;
> > > > ibv_cmd_rereg_mr;
> > > > -
> > > > } IBVERBS_1.0;
> > > >
> > > > IBVERBS_1.3 {
> > > > @@ -130,3 +129,8 @@ IBVERBS_1.3 {
> > > > ibv_cmd_destroy_rwq_ind_table;
> > > > ibv_query_gid_type;
> > > > } IBVERBS_1.1;
> > > > +
> > > > +IBVERBS_1.4 {
> > > Any reason to add 1.4 instead of using 1.3 ? in my understanding the
> > > idea is that new symbols for next release will use a new ABI version,
> > > this is what 1.3 already supplied. Added Jason for his input as well.
> >
> > good point - Jason?
> >
> > > >
> > > > + global:
> > > > + ibv_cmd_create_ah_ex;
> > > When do you plan to contribute the driver code for that ? for now no
> > > upstream drivers uses this API, taking some code for future usage with
> > > no current usage is not the preferred way.
> >
> > The plan is ASAP, the kernel side has been approved for release by Oracle's
> > legal dep, but the user library is still stuck, and with an undesirable
> > license combination given Jason's recent consolidation efforts :-/
>
> Jason didn't change licenses. You still can release your driver under
> dual-license in similar way to other drivers.
The problem is that libsif is stuck under a BSD only license imposed by
the first set of lawyers involved, and to be compatible, we need it
to be dual license GPLv2 + BSD like the rest of the provider libraries.
Knut
> >
> > Knut
> >
> > > >
> > > > +} IBVERBS_1.3;
> > > >
> > --
> > To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
> > the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
> > More majordomo info at http://vger.kernel.org/majordomo-info.html
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
^ permalink raw reply
* Re: [PATCH] rdma_rxe: Ensure rdma_rxe module init occurs at correct time
From: Yonatan Cohen @ 2016-09-20 13:59 UTC (permalink / raw)
To: Stephen Bates, Doug Ledford
Cc: monis-VPRAkNaXOzVWk0Htik3J/w, linux-rdma-u79uwXL29TY76Z2rM5mHXA
In-Reply-To: <a3e5f61b65cdf92c192daef1da319718.squirrel-2RFepEojUI2m/tTznewQxVaTQe2KTcn/@public.gmane.org>
On 9/19/2016 2:59 AM, Stephen Bates wrote:
> On Fri, September 16, 2016 1:06 pm, Doug Ledford wrote:
>> On 9/15/2016 5:45 PM, Stephen Bates wrote:
>>
>>> When the Soft-RoCE module is built into the kernel there is a
>>> initialization issue when CONFIG_IPV6 is selected. Use late_initcall to
>>> ensure the rxe module is loaded later. This fix is based on one proposed
>>> by Logan Gunthorpe on a much older code base [1].
>>
>> Your commit message doesn't make sense. You start by saying that
>> soft-RoCE is compiled in, then say "the rxe module is loaded later". Well,
>> which is it? I'm not sure exactly the problem you are fixing. However, I
>> know that the series Leon posted for the rxe driver included a fix
>> specifically for IPv6 tunnel creation issues that result in an oops. I
>> suspect that patch might solve your problem. So, I'm going to drop this
>> patch. Once you get a chance to test with the other patch applied, if
>> your problem still exists, then please repost this patch with a clearer
>> commit message and we can add this as well.
>
> Doug
>
> Yes on reflection the commit message associated with this patch was not as
> clear as it could have been. Apologies. I will be more explicit on the v1.
> To answer your questions:
>
> 1. This patch should be applied on top of the three patches that came from
> Leon [1]. I will make that clear in v1.
>
> 2. My patch applies to the case where rmda_rxe is built-in and not when it
> is a module. Currently the module_init is called before IPv6 is setup and
> that results in a load failure of rxe.
>
> 3. When rdma_rxe is built as a module the problem is not observed because
> the rdma_rxe module is not probed until the system is up and running.
>
> 4. The change from module_init to late_initcall means that IPv6 has had
> enough time to come up befoe the rdma_rxe init is executed.
>
> I hope that clarifies things. I will get a v1 out shortly with an improved
> commit log.
>
> [1] http://marc.info/?l=linux-rdma&m=147307739626544&w=2
>
> Cheers
>
> Stephen Bates
>
Hi Stephen,
I reproduced the issue you've reported, and tested your fix as well and
it looks good to me.
You can add my reviewed-by tag to the future v1 patch.
Yonatan,
Thanks
>>
>>> [1] https://github.com/sbates130272/qemu-minimal/blob/rdma/patches/\
>>> 0001-Changes-needed-to-use-rxe-compiled-into-the-kernel.patch
>>>
>>>
>>> Signed-off-by: Stephen Bates <sbates-pv7U853sEMVWk0Htik3J/w@public.gmane.org>
Reviewed-By: Yonatan Cohen <yonatanc-VPRAkNaXOzVWk0Htik3J/w@public.gmane.org>
>>> ---
>>> drivers/infiniband/sw/rxe/rxe.c | 2 +- 1 file changed, 1 insertion(+), 1
>>> deletion(-)
>>>
>>> diff --git a/drivers/infiniband/sw/rxe/rxe.c
>>> b/drivers/infiniband/sw/rxe/rxe.c index 08e09d9..97bc74a 100644 ---
>>> a/drivers/infiniband/sw/rxe/rxe.c +++ b/drivers/infiniband/sw/rxe/rxe.c
>>> @@ -392,5 +392,5 @@ static void __exit rxe_module_exit(void)
>>> pr_info("rxe: unloaded\n");
>>> }
>>>
>>>
>>> -module_init(rxe_module_init);
>>> +late_initcall(rxe_module_init);
>>> module_exit(rxe_module_exit);
>>>
>>
>>
>> --
>> Doug Ledford <dledford-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>
>> GPG Key ID: 0E572FDD
>>
>>
>>
>
> --
> To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
> the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
> More majordomo info at http://vger.kernel.org/majordomo-info.html
>
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
^ permalink raw reply
* Re: [PATCH libibverbs v2 3/3] Provide remote XRC SRQ number in kernel post_send.
From: Knut Omang @ 2016-09-20 13:38 UTC (permalink / raw)
To: Leon Romanovsky
Cc: Doug Ledford, linux-rdma-u79uwXL29TY76Z2rM5mHXA, Mukesh Kacker
In-Reply-To: <20160920121741.GL26673-2ukJVAZIZ/Y@public.gmane.org>
On Tue, 2016-09-20 at 15:17 +0300, Leon Romanovsky wrote:
> On Tue, Sep 20, 2016 at 01:08:35PM +0200, Knut Omang wrote:
> > On Tue, 2016-09-20 at 14:01 +0300, Leon Romanovsky wrote:
> > > On Tue, Sep 20, 2016 at 12:43:30PM +0200, Knut Omang wrote:
> > > >
> > > > On Tue, 2016-09-20 at 13:18 +0300, Leon Romanovsky wrote:
> > > > >
> > > > > On Mon, Sep 19, 2016 at 11:12:36AM +0200, Knut Omang wrote:
> > > > > >
> > > > > >
> > > > > > On Mon, 2016-09-19 at 08:29 +0300, Leon Romanovsky wrote:
> > > > > > >
> > > > > > >
> > > > > > > On Sat, Sep 17, 2016 at 05:59:13AM +0200, Knut Omang wrote:
> > > > > > > >
> > > > > > > >
> > > > > > > >
> > > > > > > > Also proper end align the ibv_kern_send_wr struct.
> > > > > > > >
> > > > > > > > Requires a corresponding kernel change.
> > > > > > > >
> > > > > > > > Signed-off-by: Knut Omang <knut.omang-QHcLZuEGTsvQT0dZR+AlfA@public.gmane.org>
> > > > > > > > Reviewed-by: Mukesh Kacker <mukesh.kacker-QHcLZuEGTsvQT0dZR+AlfA@public.gmane.org>
> > > > > > > > ---
> > > > > > > > include/infiniband/kern-abi.h | 1 +
> > > > > > > > src/cmd.c | 2 ++
> > > > > > > > 2 files changed, 3 insertions(+), 0 deletions(-)
> > > > > > > >
> > > > > > > > diff --git a/include/infiniband/kern-abi.h b/include/infiniband/kern-abi.h
> > > > > > > > index 8bdeef5..7b1d310 100644
> > > > > > > > --- a/include/infiniband/kern-abi.h
> > > > > > > > +++ b/include/infiniband/kern-abi.h
> > > > > > > > @@ -800,6 +800,7 @@ struct ibv_kern_send_wr {
> > > > > > > > union {
> > > > > > > > struct {
> > > > > > > > __u32 remote_srqn;
> > > > > > > > + __u32 reserved;
> > > > > > > > } xrc;
> > > > > > > > } qp_type;
> > > > > > > > };
> > > > > > > > diff --git a/src/cmd.c b/src/cmd.c
> > > > > > > > index a418ee1..a4e2f75 100644
> > > > > > > > --- a/src/cmd.c
> > > > > > > > +++ b/src/cmd.c
> > > > > > > > @@ -1293,6 +1293,8 @@ int ibv_cmd_post_send(struct ibv_qp *ibqp, struct ibv_send_wr *wr,
> > > > > > > > tmp->wr.ud.remote_qpn = i->wr.ud.remote_qpn;
> > > > > > > > tmp->wr.ud.remote_qkey = i->wr.ud.remote_qkey;
> > > > > > > > } else {
> > > > > > > > + if (ibqp->qp_type == IBV_QPT_XRC_SEND)
> > > > > > > > + tmp->qp_type.xrc.remote_srqn = i->qp_type.xrc.remote_srqn;
> > > > > > > It will be checked for all QPTs and for all consumers. Any chances to
> > > > > > > optimize it?
> > > > > > All QPTs except UD, yes.
> > > > > >
> > > > > > The easiest is perhaps just to remove the test (and set the attribute in all cases), which will work fine
> > > > > > as long as the qp_type union is not used for anything else, but is it worth
> > > > > > the loss of intuitiveness/future potential correctness of the code?
> > > > > Let's concentrate on the present and as far as I see there are no kernel users who
> > > > > need this field xrc.remote_srqn field.
> > > > So I assume then that you are ok with this code, given my explanation?
> > > Partially,
> > >
> > > I'm fine with the claim about readability and the fact that
> > > this code is not in data-path relaxed me a little bit.
> >
> > Good!
> >
> > > I'm not fine with the part that this code doesn't do anything in kernel.
> >
> > I know, I have two big patch sets one kernel and one for the new "rdma consolidated"
> > in the queue that should fix that, just want to save a few versions of those
> > ~32000 lines of code by preparing the baseline :-)
>
> It won't save. We DO believe you that this this new addition to UAPI is really
> needed, but we NEED to see that it is unavoidable.
>
> The proper flow for submission is driver->kernel_core->libraries and not
> kernel_core->libraries->driver.
Ok, got it..
Knut
> Thanks
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
^ permalink raw reply
* Re: [PATCH v4 01/16] vmxnet3: Move PCI Id to pci_ids.h
From: Bjorn Helgaas @ 2016-09-20 13:37 UTC (permalink / raw)
To: Adit Ranadive
Cc: Yuval Shaia, dledford-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org,
linux-rdma-u79uwXL29TY76Z2rM5mHXA@public.gmane.org, pv-drivers,
netdev-u79uwXL29TY76Z2rM5mHXA@public.gmane.org,
linux-pci-u79uwXL29TY76Z2rM5mHXA@public.gmane.org,
Jorgen S. Hansen, Aditya Sarwade, George Zhang, Bryan Tan
In-Reply-To: <DM2PR0501MB84422A1029EE72FC727F960C5F10-SXUgMXdxDSdudfUQJyYm35LWcSx1hRipwIZJ9u9yWa8oOQlpcoRfSA@public.gmane.org>
On Wed, Sep 14, 2016 at 07:36:34PM +0000, Adit Ranadive wrote:
> On Wed, Sep 14, 2016 at 09:25:18 -0700, Yuval Shaia wrote:
> > On Wed, Sep 14, 2016 at 04:00:25PM +0000, Adit Ranadive wrote:
> > > On Wed, Sep 14, 2016 at 04:09:12 -0700, Yuval Shaia wrote:
> > > > Please update vmxnet3_drv.c accordingly.
> > >
> > > Any reason why? I don't think we need to. Vmxnet3 should just pick up
> > > the moved PCI device id from pci_ids.h file.
> >
> > So now you need to include it from vmxnet3_drv.c.
> > Same with pvrdma_main.c
>
> If you're asking me to include pci_ids.h in our drivers we already do that
> by including pci.h in both the drivers.
> pci.h already includes pci_ids.h -
> https://git.kernel.org/cgit/linux/kernel/git/torvalds/linux.git/tree/include/linux/pci.h#n35
>
> If that's going to change maybe someone from the PCI group can comment on.
I don't contemplate a change there. It's fine to get pci_ids.h via pci.h.
Almost all drivers do that today. If we do change it someday, we can
change this driver along with all the rest.
Bjorn
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
^ permalink raw reply
* RE: [RFC v2 00/11] QLogic RDMA Driver (qedr) RFC
From: Elior, Ariel @ 2016-09-20 13:33 UTC (permalink / raw)
To: Leon Romanovsky, Amrani, Ram
Cc: davem-fT/PcQaiUtIeIZ0/mPfg9Q@public.gmane.org,
dledford-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org,
Kalderon, Michal, Mintz, Yuval, Borundia, Rajesh,
linux-rdma-u79uwXL29TY76Z2rM5mHXA@public.gmane.org,
netdev-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
In-Reply-To: <20160920121151.GK26673-2ukJVAZIZ/Y@public.gmane.org>
> On Tue, Sep 20, 2016 at 01:35:52PM +0300, Ram Amrani wrote:
...
> > The series adds on top of RFC v1:
> > * a check for all drivers that IB_ACCESS_MW_BIND isn't set for ib_get_dma_mr
> > * relocation of qedr user API to include/rdma/uapi/providers/
> > * removal of qedr_devlist_local
> > * fixed error handling in qedr_alloc_resources()
> > * configuration of PBL in ib_map_mr_sg() driver implementation,
> > rather than post_send's IB_WR_REG_MR
> > * misc.: placed code in proper patch, fixed a few comments,
> > removed extra parentheses
> >
> > Thanks for everyone which pointed out problems in the driver.
> >
> > Any review/comment is appreciated.
>
> Very nice,
> Any reason why didn't you drop debug module parameter and decided to
> mimic already available kernel core functionality?
>
> You got technical explanations why it is bad idea to use it. If you need additional
> voices
> to support my claims, you will find them in thread about VERBOSE flag and responses
> from
> Doug, and Dennis.
>
> Thanks
Hi Leon,
The RFC cover letter lists what has been addressed. Debug printouts are not addressed in V2 as the discussion on that topic is not concluded (more thoughts from us on debug printouts incoming on the thread). There were many comments to V1 which are not relevant to debug printouts which are addressed by V2. We are requesting further comment, hence RFC V2. Rest assured, if it is the final opinion in the relevant discussion that pr_debug is the way to go, that's what we'll do.
Thanks,
Ariel
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
^ permalink raw reply
* Re: [PATCH libibverbs v3 1/3] Add new call ibv_cmd_create_ah_ex which supports extra parameters
From: Leon Romanovsky @ 2016-09-20 13:29 UTC (permalink / raw)
To: Knut Omang
Cc: Yishai Hadas, Doug Ledford, linux-rdma-u79uwXL29TY76Z2rM5mHXA,
Mukesh Kacker, Jason Gunthorpe
In-Reply-To: <1474371459.8837.32.camel-QHcLZuEGTsvQT0dZR+AlfA@public.gmane.org>
[-- Attachment #1: Type: text/plain, Size: 6820 bytes --]
On Tue, Sep 20, 2016 at 01:37:39PM +0200, Knut Omang wrote:
> On Tue, 2016-09-20 at 14:17 +0300, Yishai Hadas wrote:
> > On 9/20/2016 9:22 AM, Knut Omang wrote:
> > >
> > > The original call ibv_cmd_create_ah does not allow vendor specific request or
> > > response parameters to be defined and passed through to kernel space.
> > >
> > > To keep backward compatibility, introduce a new extended call which accepts
> > > cmd and resp parameters similar to other parts of the API.
> > > Reimplement the old call to just use the new extended call.
> > >
> > > The new call is versioned in the shared library as IBVERBS_1.4.
> > >
> > > Signed-off-by: Knut Omang <knut.omang-QHcLZuEGTsvQT0dZR+AlfA@public.gmane.org>
> > > Reviewed-by: Mukesh Kacker <mukesh.kacker-QHcLZuEGTsvQT0dZR+AlfA@public.gmane.org>
> > > ---
> > > include/infiniband/driver.h | 4 +++-
> > > src/cmd.c | 50 +++++++++++++++++++++-----------------
> > > src/libibverbs.map | 6 ++++-
> > > 3 files changed, 37 insertions(+), 23 deletions(-)
> > >
> > > diff --git a/include/infiniband/driver.h b/include/infiniband/driver.h
> > > index ea3dade..9356dc8 100644
> > > --- a/include/infiniband/driver.h
> > > +++ b/include/infiniband/driver.h
> > > @@ -234,6 +234,10 @@ int ibv_cmd_post_srq_recv(struct ibv_srq *srq, struct ibv_recv_wr *wr,
> > > struct ibv_recv_wr **bad_wr);
> > > int ibv_cmd_create_ah(struct ibv_pd *pd, struct ibv_ah *ah,
> > > struct ibv_ah_attr *attr);
> > > +int ibv_cmd_create_ah_ex(struct ibv_pd *pd, struct ibv_ah *ah,
> > > + struct ibv_ah_attr *attr,
> > > + struct ibv_create_ah *cmd, size_t cmd_size,
> > > + struct ibv_create_ah_resp *resp, size_t resp_size);
> > > int ibv_cmd_destroy_ah(struct ibv_ah *ah);
> > > int ibv_cmd_attach_mcast(struct ibv_qp *qp, const union ibv_gid *gid, uint16_t lid);
> > > int ibv_cmd_detach_mcast(struct ibv_qp *qp, const union ibv_gid *gid, uint16_t lid);
> > > diff --git a/src/cmd.c b/src/cmd.c
> > > index 11f6509..58b9dcd 100644
> > > --- a/src/cmd.c
> > > +++ b/src/cmd.c
> > > @@ -1493,38 +1493,44 @@ int ibv_cmd_post_srq_recv(struct ibv_srq *srq, struct ibv_recv_wr *wr,
> > > return ret;
> > > }
> > >
> > > -int ibv_cmd_create_ah(struct ibv_pd *pd, struct ibv_ah *ah,
> > > - struct ibv_ah_attr *attr)
> > > +int ibv_cmd_create_ah_ex(struct ibv_pd *pd, struct ibv_ah *ah, struct ibv_ah_attr *attr,
> > > + struct ibv_create_ah *cmd, size_t cmd_size,
> > > + struct ibv_create_ah_resp *resp, size_t resp_size)
> > > {
> > > - struct ibv_create_ah cmd;
> > > - struct ibv_create_ah_resp resp;
> > > -
> > > - IBV_INIT_CMD_RESP(&cmd, sizeof cmd, CREATE_AH, &resp, sizeof resp);
> > > - cmd.user_handle = (uintptr_t) ah;
> > > - cmd.pd_handle = pd->handle;
> > > - cmd.attr.dlid = attr->dlid;
> > > - cmd.attr.sl = attr->sl;
> > > - cmd.attr.src_path_bits = attr->src_path_bits;
> > > - cmd.attr.static_rate = attr->static_rate;
> > > - cmd.attr.is_global = attr->is_global;
> > > - cmd.attr.port_num = attr->port_num;
> > > - cmd.attr.grh.flow_label = attr->grh.flow_label;
> > > - cmd.attr.grh.sgid_index = attr->grh.sgid_index;
> > > - cmd.attr.grh.hop_limit = attr->grh.hop_limit;
> > > - cmd.attr.grh.traffic_class = attr->grh.traffic_class;
> > > - memcpy(cmd.attr.grh.dgid, attr->grh.dgid.raw, 16);
> > > + IBV_INIT_CMD_RESP(cmd, cmd_size, CREATE_AH, resp, resp_size);
> > > + cmd->user_handle = (uintptr_t) ah;
> > > + cmd->pd_handle = pd->handle;
> > > + cmd->attr.dlid = attr->dlid;
> > > + cmd->attr.sl = attr->sl;
> > > + cmd->attr.src_path_bits = attr->src_path_bits;
> > > + cmd->attr.static_rate = attr->static_rate;
> > > + cmd->attr.is_global = attr->is_global;
> > > + cmd->attr.port_num = attr->port_num;
> > > + cmd->attr.grh.flow_label = attr->grh.flow_label;
> > > + cmd->attr.grh.sgid_index = attr->grh.sgid_index;
> > > + cmd->attr.grh.hop_limit = attr->grh.hop_limit;
> > > + cmd->attr.grh.traffic_class = attr->grh.traffic_class;
> > > + memcpy(cmd->attr.grh.dgid, attr->grh.dgid.raw, 16);
> > >
> > > - if (write(pd->context->cmd_fd, &cmd, sizeof cmd) != sizeof cmd)
> > > + if (write(pd->context->cmd_fd, cmd, cmd_size) != cmd_size)
> > > return errno;
> > >
> > > - (void) VALGRIND_MAKE_MEM_DEFINED(&resp, sizeof resp);
> > > + (void) VALGRIND_MAKE_MEM_DEFINED(resp, sizeof *resp);
> > >
> > > - ah->handle = resp.handle;
> > > + ah->handle = resp->handle;
> > > ah->context = pd->context;
> > >
> > > return 0;
> > > }
> > >
> > > +int ibv_cmd_create_ah(struct ibv_pd *pd, struct ibv_ah *ah,
> > > + struct ibv_ah_attr *attr)
> > > +{
> > > + struct ibv_create_ah cmd;
> > > + struct ibv_create_ah_resp resp;
> > > + return ibv_cmd_create_ah_ex(pd, ah, attr, &cmd, sizeof(cmd), &resp, sizeof(resp));
> > > +}
> > > +
> > > int ibv_cmd_destroy_ah(struct ibv_ah *ah)
> > > {
> > > struct ibv_destroy_ah cmd;
> > > diff --git a/src/libibverbs.map b/src/libibverbs.map
> > > index 46744e5..ddba63a 100644
> > > --- a/src/libibverbs.map
> > > +++ b/src/libibverbs.map
> > > @@ -118,7 +118,6 @@ IBVERBS_1.1 {
> > > ibv_cmd_create_qp_ex2;
> > > ibv_cmd_open_qp;
> > > ibv_cmd_rereg_mr;
> > > -
> > > } IBVERBS_1.0;
> > >
> > > IBVERBS_1.3 {
> > > @@ -130,3 +129,8 @@ IBVERBS_1.3 {
> > > ibv_cmd_destroy_rwq_ind_table;
> > > ibv_query_gid_type;
> > > } IBVERBS_1.1;
> > > +
> > > +IBVERBS_1.4 {
> > Any reason to add 1.4 instead of using 1.3 ? in my understanding the
> > idea is that new symbols for next release will use a new ABI version,
> > this is what 1.3 already supplied. Added Jason for his input as well.
>
> good point - Jason?
>
> > >
> > > + global:
> > > + ibv_cmd_create_ah_ex;
> > When do you plan to contribute the driver code for that ? for now no
> > upstream drivers uses this API, taking some code for future usage with
> > no current usage is not the preferred way.
>
> The plan is ASAP, the kernel side has been approved for release by Oracle's
> legal dep, but the user library is still stuck, and with an undesirable
> license combination given Jason's recent consolidation efforts :-/
Jason didn't change licenses. You still can release your driver under
dual-license in similar way to other drivers.
>
> Knut
>
> > >
> > > +} IBVERBS_1.3;
> > >
> --
> To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
> the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
> More majordomo info at http://vger.kernel.org/majordomo-info.html
[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 819 bytes --]
^ permalink raw reply
* Re: VERBOSE flags
From: Leon Romanovsky @ 2016-09-20 12:30 UTC (permalink / raw)
To: Dalessandro, Dennis
Cc: linux-rdma-u79uwXL29TY76Z2rM5mHXA@public.gmane.org,
dledford-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org
In-Reply-To: <1474373496.12415.28.camel-ral2JQCrhuEAvxtiuMwx3w@public.gmane.org>
[-- Attachment #1: Type: text/plain, Size: 1458 bytes --]
On Tue, Sep 20, 2016 at 12:11:40PM +0000, Dalessandro, Dennis wrote:
> On Tue, 2016-09-20 at 13:56 +0300, Leon Romanovsky wrote:
> > > Yes we do have that, which illustrates that trace points can not do
> > > everything. I had hoped to deprecate that at some point.
> > > Regardless,
> > > sometimes you want additional code to be added or to do something
> > > slightly different. You can't do that with a tracepoint.
> >
> > Can you elaborate more on this topic? Which limitations did you see
> > in tracepoints?
>
> Just making up an example. Say you have a hash table. You want to dump
> the items. If you iterate over it and do a trace point, yeah the actual
> trace point has little to no perf impact, but you still walk the
> table.
>
> There are situations where you need to do more for debugging than just
> printing a message.
Thanks
>
> > In regards of other discussion in KS 2016, "tracepoints and ABI
> > stability warranties" [1]. Are you strict with this tracepoints ABI?
>
> Honestly we are not super strict about it, but do take care not to
> break/remove something that could be useful. It will be interesting to
> see what comes out of the discussion. Perhaps we do need to be more
> strict about them going forward.
I followed the discussion and it looks like that majority are against to
declare tracepoints as ABI. Otherwise, you will find yourself unable to
change the kernel code.
>
> -Denny
[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 819 bytes --]
^ permalink raw reply
* Re: [PATCH libibverbs v2 3/3] Provide remote XRC SRQ number in kernel post_send.
From: Leon Romanovsky @ 2016-09-20 12:17 UTC (permalink / raw)
To: Knut Omang; +Cc: Doug Ledford, linux-rdma-u79uwXL29TY76Z2rM5mHXA, Mukesh Kacker
In-Reply-To: <1474369715.8837.19.camel-QHcLZuEGTsvQT0dZR+AlfA@public.gmane.org>
[-- Attachment #1: Type: text/plain, Size: 3547 bytes --]
On Tue, Sep 20, 2016 at 01:08:35PM +0200, Knut Omang wrote:
> On Tue, 2016-09-20 at 14:01 +0300, Leon Romanovsky wrote:
> > On Tue, Sep 20, 2016 at 12:43:30PM +0200, Knut Omang wrote:
> > >
> > > On Tue, 2016-09-20 at 13:18 +0300, Leon Romanovsky wrote:
> > > >
> > > > On Mon, Sep 19, 2016 at 11:12:36AM +0200, Knut Omang wrote:
> > > > >
> > > > >
> > > > > On Mon, 2016-09-19 at 08:29 +0300, Leon Romanovsky wrote:
> > > > > >
> > > > > >
> > > > > > On Sat, Sep 17, 2016 at 05:59:13AM +0200, Knut Omang wrote:
> > > > > > >
> > > > > > >
> > > > > > >
> > > > > > > Also proper end align the ibv_kern_send_wr struct.
> > > > > > >
> > > > > > > Requires a corresponding kernel change.
> > > > > > >
> > > > > > > Signed-off-by: Knut Omang <knut.omang-QHcLZuEGTsvQT0dZR+AlfA@public.gmane.org>
> > > > > > > Reviewed-by: Mukesh Kacker <mukesh.kacker-QHcLZuEGTsvQT0dZR+AlfA@public.gmane.org>
> > > > > > > ---
> > > > > > > include/infiniband/kern-abi.h | 1 +
> > > > > > > src/cmd.c | 2 ++
> > > > > > > 2 files changed, 3 insertions(+), 0 deletions(-)
> > > > > > >
> > > > > > > diff --git a/include/infiniband/kern-abi.h b/include/infiniband/kern-abi.h
> > > > > > > index 8bdeef5..7b1d310 100644
> > > > > > > --- a/include/infiniband/kern-abi.h
> > > > > > > +++ b/include/infiniband/kern-abi.h
> > > > > > > @@ -800,6 +800,7 @@ struct ibv_kern_send_wr {
> > > > > > > union {
> > > > > > > struct {
> > > > > > > __u32 remote_srqn;
> > > > > > > + __u32 reserved;
> > > > > > > } xrc;
> > > > > > > } qp_type;
> > > > > > > };
> > > > > > > diff --git a/src/cmd.c b/src/cmd.c
> > > > > > > index a418ee1..a4e2f75 100644
> > > > > > > --- a/src/cmd.c
> > > > > > > +++ b/src/cmd.c
> > > > > > > @@ -1293,6 +1293,8 @@ int ibv_cmd_post_send(struct ibv_qp *ibqp, struct ibv_send_wr *wr,
> > > > > > > tmp->wr.ud.remote_qpn = i->wr.ud.remote_qpn;
> > > > > > > tmp->wr.ud.remote_qkey = i->wr.ud.remote_qkey;
> > > > > > > } else {
> > > > > > > + if (ibqp->qp_type == IBV_QPT_XRC_SEND)
> > > > > > > + tmp->qp_type.xrc.remote_srqn = i->qp_type.xrc.remote_srqn;
> > > > > > It will be checked for all QPTs and for all consumers. Any chances to
> > > > > > optimize it?
> > > > > All QPTs except UD, yes.
> > > > >
> > > > > The easiest is perhaps just to remove the test (and set the attribute in all cases), which will work fine
> > > > > as long as the qp_type union is not used for anything else, but is it worth
> > > > > the loss of intuitiveness/future potential correctness of the code?
> > > > Let's concentrate on the present and as far as I see there are no kernel users who
> > > > need this field xrc.remote_srqn field.
> > > So I assume then that you are ok with this code, given my explanation?
> > Partially,
> >
> > I'm fine with the claim about readability and the fact that
> > this code is not in data-path relaxed me a little bit.
>
> Good!
>
> > I'm not fine with the part that this code doesn't do anything in kernel.
>
> I know, I have two big patch sets one kernel and one for the new "rdma consolidated"
> in the queue that should fix that, just want to save a few versions of those
> ~32000 lines of code by preparing the baseline :-)
It won't save. We DO believe you that this this new addition to UAPI is really
needed, but we NEED to see that it is unavoidable.
The proper flow for submission is driver->kernel_core->libraries and not
kernel_core->libraries->driver.
Thanks
[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 819 bytes --]
^ permalink raw reply
* Re: [RFC v2 00/11] QLogic RDMA Driver (qedr) RFC
From: Leon Romanovsky @ 2016-09-20 12:11 UTC (permalink / raw)
To: Ram Amrani
Cc: davem-fT/PcQaiUtIeIZ0/mPfg9Q, dledford-H+wXaHxf7aLQT0dZR+AlfA,
Ariel.Elior-YGCgFSpz5w/QT0dZR+AlfA,
Michal.Kalderon-YGCgFSpz5w/QT0dZR+AlfA,
Yuval.Mintz-YGCgFSpz5w/QT0dZR+AlfA,
rajesh.borundia-YGCgFSpz5w/QT0dZR+AlfA,
linux-rdma-u79uwXL29TY76Z2rM5mHXA, netdev-u79uwXL29TY76Z2rM5mHXA
In-Reply-To: <1474367764-9555-1-git-send-email-Ram.Amrani-YGCgFSpz5w/QT0dZR+AlfA@public.gmane.org>
[-- Attachment #1: Type: text/plain, Size: 1476 bytes --]
On Tue, Sep 20, 2016 at 01:35:52PM +0300, Ram Amrani wrote:
> From: Ram amrani <Ram.Amrani-YGCgFSpz5w/QT0dZR+AlfA@public.gmane.org>
>
> This series introduces RoCE RDMA driver for the 579xx RDMA products by Qlogic.
> The RDMA support is added as an additional loadable module (qedr) over the
> Ethernet qede driver.
> The qedr module will support both RoCE and iWarp, although this series only
> adds RoCE support.
> The qed and qede drivers are enhanced with functionality required for RDMA
> support.
>
> Based on net-next 4c73195edbe3a5d7e14ea549bb261cf35c29f0cc.
>
> The series adds on top of RFC v1:
> * a check for all drivers that IB_ACCESS_MW_BIND isn't set for ib_get_dma_mr
> * relocation of qedr user API to include/rdma/uapi/providers/
> * removal of qedr_devlist_local
> * fixed error handling in qedr_alloc_resources()
> * configuration of PBL in ib_map_mr_sg() driver implementation,
> rather than post_send's IB_WR_REG_MR
> * misc.: placed code in proper patch, fixed a few comments,
> removed extra parentheses
>
> Thanks for everyone which pointed out problems in the driver.
>
> Any review/comment is appreciated.
Very nice,
Any reason why didn't you drop debug module parameter and decided to
mimic already available kernel core functionality?
You got technical explanations why it is bad idea to use it. If you need additional voices
to support my claims, you will find them in thread about VERBOSE flag and responses from
Doug, and Dennis.
Thanks
[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 819 bytes --]
^ permalink raw reply
* Re: VERBOSE flags
From: Dalessandro, Dennis @ 2016-09-20 12:11 UTC (permalink / raw)
To: leon-DgEjT+Ai2ygdnm+yROfE0A@public.gmane.org
Cc: linux-rdma-u79uwXL29TY76Z2rM5mHXA@public.gmane.org,
dledford-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org
In-Reply-To: <20160920105642.GI26673-2ukJVAZIZ/Y@public.gmane.org>
On Tue, 2016-09-20 at 13:56 +0300, Leon Romanovsky wrote:
> > Yes we do have that, which illustrates that trace points can not do
> > everything. I had hoped to deprecate that at some point.
> > Regardless,
> > sometimes you want additional code to be added or to do something
> > slightly different. You can't do that with a tracepoint.
>
> Can you elaborate more on this topic? Which limitations did you see
> in tracepoints?
Just making up an example. Say you have a hash table. You want to dump
the items. If you iterate over it and do a trace point, yeah the actual
trace point has little to no perf impact, but you still walk the
table.
There are situations where you need to do more for debugging than just
printing a message.
> In regards of other discussion in KS 2016, "tracepoints and ABI
> stability warranties" [1]. Are you strict with this tracepoints ABI?
Honestly we are not super strict about it, but do take care not to
break/remove something that could be useful. It will be interesting to
see what comes out of the discussion. Perhaps we do need to be more
strict about them going forward.
-Denny
^ permalink raw reply
* Re: [PATCH libibverbs v3 1/3] Add new call ibv_cmd_create_ah_ex which supports extra parameters
From: Knut Omang @ 2016-09-20 11:37 UTC (permalink / raw)
To: Yishai Hadas
Cc: Doug Ledford, linux-rdma-u79uwXL29TY76Z2rM5mHXA, Mukesh Kacker,
Jason Gunthorpe
In-Reply-To: <550041ef-b105-c079-068c-d96278ec34f1-LDSdmyG8hGV8YrgS2mwiifqBs+8SCbDb@public.gmane.org>
On Tue, 2016-09-20 at 14:17 +0300, Yishai Hadas wrote:
> On 9/20/2016 9:22 AM, Knut Omang wrote:
> >
> > The original call ibv_cmd_create_ah does not allow vendor specific request or
> > response parameters to be defined and passed through to kernel space.
> >
> > To keep backward compatibility, introduce a new extended call which accepts
> > cmd and resp parameters similar to other parts of the API.
> > Reimplement the old call to just use the new extended call.
> >
> > The new call is versioned in the shared library as IBVERBS_1.4.
> >
> > Signed-off-by: Knut Omang <knut.omang-QHcLZuEGTsvQT0dZR+AlfA@public.gmane.org>
> > Reviewed-by: Mukesh Kacker <mukesh.kacker-QHcLZuEGTsvQT0dZR+AlfA@public.gmane.org>
> > ---
> > include/infiniband/driver.h | 4 +++-
> > src/cmd.c | 50 +++++++++++++++++++++-----------------
> > src/libibverbs.map | 6 ++++-
> > 3 files changed, 37 insertions(+), 23 deletions(-)
> >
> > diff --git a/include/infiniband/driver.h b/include/infiniband/driver.h
> > index ea3dade..9356dc8 100644
> > --- a/include/infiniband/driver.h
> > +++ b/include/infiniband/driver.h
> > @@ -234,6 +234,10 @@ int ibv_cmd_post_srq_recv(struct ibv_srq *srq, struct ibv_recv_wr *wr,
> > struct ibv_recv_wr **bad_wr);
> > int ibv_cmd_create_ah(struct ibv_pd *pd, struct ibv_ah *ah,
> > struct ibv_ah_attr *attr);
> > +int ibv_cmd_create_ah_ex(struct ibv_pd *pd, struct ibv_ah *ah,
> > + struct ibv_ah_attr *attr,
> > + struct ibv_create_ah *cmd, size_t cmd_size,
> > + struct ibv_create_ah_resp *resp, size_t resp_size);
> > int ibv_cmd_destroy_ah(struct ibv_ah *ah);
> > int ibv_cmd_attach_mcast(struct ibv_qp *qp, const union ibv_gid *gid, uint16_t lid);
> > int ibv_cmd_detach_mcast(struct ibv_qp *qp, const union ibv_gid *gid, uint16_t lid);
> > diff --git a/src/cmd.c b/src/cmd.c
> > index 11f6509..58b9dcd 100644
> > --- a/src/cmd.c
> > +++ b/src/cmd.c
> > @@ -1493,38 +1493,44 @@ int ibv_cmd_post_srq_recv(struct ibv_srq *srq, struct ibv_recv_wr *wr,
> > return ret;
> > }
> >
> > -int ibv_cmd_create_ah(struct ibv_pd *pd, struct ibv_ah *ah,
> > - struct ibv_ah_attr *attr)
> > +int ibv_cmd_create_ah_ex(struct ibv_pd *pd, struct ibv_ah *ah, struct ibv_ah_attr *attr,
> > + struct ibv_create_ah *cmd, size_t cmd_size,
> > + struct ibv_create_ah_resp *resp, size_t resp_size)
> > {
> > - struct ibv_create_ah cmd;
> > - struct ibv_create_ah_resp resp;
> > -
> > - IBV_INIT_CMD_RESP(&cmd, sizeof cmd, CREATE_AH, &resp, sizeof resp);
> > - cmd.user_handle = (uintptr_t) ah;
> > - cmd.pd_handle = pd->handle;
> > - cmd.attr.dlid = attr->dlid;
> > - cmd.attr.sl = attr->sl;
> > - cmd.attr.src_path_bits = attr->src_path_bits;
> > - cmd.attr.static_rate = attr->static_rate;
> > - cmd.attr.is_global = attr->is_global;
> > - cmd.attr.port_num = attr->port_num;
> > - cmd.attr.grh.flow_label = attr->grh.flow_label;
> > - cmd.attr.grh.sgid_index = attr->grh.sgid_index;
> > - cmd.attr.grh.hop_limit = attr->grh.hop_limit;
> > - cmd.attr.grh.traffic_class = attr->grh.traffic_class;
> > - memcpy(cmd.attr.grh.dgid, attr->grh.dgid.raw, 16);
> > + IBV_INIT_CMD_RESP(cmd, cmd_size, CREATE_AH, resp, resp_size);
> > + cmd->user_handle = (uintptr_t) ah;
> > + cmd->pd_handle = pd->handle;
> > + cmd->attr.dlid = attr->dlid;
> > + cmd->attr.sl = attr->sl;
> > + cmd->attr.src_path_bits = attr->src_path_bits;
> > + cmd->attr.static_rate = attr->static_rate;
> > + cmd->attr.is_global = attr->is_global;
> > + cmd->attr.port_num = attr->port_num;
> > + cmd->attr.grh.flow_label = attr->grh.flow_label;
> > + cmd->attr.grh.sgid_index = attr->grh.sgid_index;
> > + cmd->attr.grh.hop_limit = attr->grh.hop_limit;
> > + cmd->attr.grh.traffic_class = attr->grh.traffic_class;
> > + memcpy(cmd->attr.grh.dgid, attr->grh.dgid.raw, 16);
> >
> > - if (write(pd->context->cmd_fd, &cmd, sizeof cmd) != sizeof cmd)
> > + if (write(pd->context->cmd_fd, cmd, cmd_size) != cmd_size)
> > return errno;
> >
> > - (void) VALGRIND_MAKE_MEM_DEFINED(&resp, sizeof resp);
> > + (void) VALGRIND_MAKE_MEM_DEFINED(resp, sizeof *resp);
> >
> > - ah->handle = resp.handle;
> > + ah->handle = resp->handle;
> > ah->context = pd->context;
> >
> > return 0;
> > }
> >
> > +int ibv_cmd_create_ah(struct ibv_pd *pd, struct ibv_ah *ah,
> > + struct ibv_ah_attr *attr)
> > +{
> > + struct ibv_create_ah cmd;
> > + struct ibv_create_ah_resp resp;
> > + return ibv_cmd_create_ah_ex(pd, ah, attr, &cmd, sizeof(cmd), &resp, sizeof(resp));
> > +}
> > +
> > int ibv_cmd_destroy_ah(struct ibv_ah *ah)
> > {
> > struct ibv_destroy_ah cmd;
> > diff --git a/src/libibverbs.map b/src/libibverbs.map
> > index 46744e5..ddba63a 100644
> > --- a/src/libibverbs.map
> > +++ b/src/libibverbs.map
> > @@ -118,7 +118,6 @@ IBVERBS_1.1 {
> > ibv_cmd_create_qp_ex2;
> > ibv_cmd_open_qp;
> > ibv_cmd_rereg_mr;
> > -
> > } IBVERBS_1.0;
> >
> > IBVERBS_1.3 {
> > @@ -130,3 +129,8 @@ IBVERBS_1.3 {
> > ibv_cmd_destroy_rwq_ind_table;
> > ibv_query_gid_type;
> > } IBVERBS_1.1;
> > +
> > +IBVERBS_1.4 {
> Any reason to add 1.4 instead of using 1.3 ? in my understanding the
> idea is that new symbols for next release will use a new ABI version,
> this is what 1.3 already supplied. Added Jason for his input as well.
good point - Jason?
> >
> > + global:
> > + ibv_cmd_create_ah_ex;
> When do you plan to contribute the driver code for that ? for now no
> upstream drivers uses this API, taking some code for future usage with
> no current usage is not the preferred way.
The plan is ASAP, the kernel side has been approved for release by Oracle's
legal dep, but the user library is still stuck, and with an undesirable
license combination given Jason's recent consolidation efforts :-/
Knut
> >
> > +} IBVERBS_1.3;
> >
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
^ permalink raw reply
* Re: [PATCH libibverbs v3 1/3] Add new call ibv_cmd_create_ah_ex which supports extra parameters
From: Yishai Hadas @ 2016-09-20 11:17 UTC (permalink / raw)
To: Knut Omang
Cc: Doug Ledford, linux-rdma-u79uwXL29TY76Z2rM5mHXA, Mukesh Kacker,
Jason Gunthorpe
In-Reply-To: <d551c3792e884443903c67212f505082cd53b135.1474352400.git-series.knut.omang-QHcLZuEGTsvQT0dZR+AlfA@public.gmane.org>
On 9/20/2016 9:22 AM, Knut Omang wrote:
> The original call ibv_cmd_create_ah does not allow vendor specific request or
> response parameters to be defined and passed through to kernel space.
>
> To keep backward compatibility, introduce a new extended call which accepts
> cmd and resp parameters similar to other parts of the API.
> Reimplement the old call to just use the new extended call.
>
> The new call is versioned in the shared library as IBVERBS_1.4.
>
> Signed-off-by: Knut Omang <knut.omang-QHcLZuEGTsvQT0dZR+AlfA@public.gmane.org>
> Reviewed-by: Mukesh Kacker <mukesh.kacker-QHcLZuEGTsvQT0dZR+AlfA@public.gmane.org>
> ---
> include/infiniband/driver.h | 4 +++-
> src/cmd.c | 50 +++++++++++++++++++++-----------------
> src/libibverbs.map | 6 ++++-
> 3 files changed, 37 insertions(+), 23 deletions(-)
>
> diff --git a/include/infiniband/driver.h b/include/infiniband/driver.h
> index ea3dade..9356dc8 100644
> --- a/include/infiniband/driver.h
> +++ b/include/infiniband/driver.h
> @@ -234,6 +234,10 @@ int ibv_cmd_post_srq_recv(struct ibv_srq *srq, struct ibv_recv_wr *wr,
> struct ibv_recv_wr **bad_wr);
> int ibv_cmd_create_ah(struct ibv_pd *pd, struct ibv_ah *ah,
> struct ibv_ah_attr *attr);
> +int ibv_cmd_create_ah_ex(struct ibv_pd *pd, struct ibv_ah *ah,
> + struct ibv_ah_attr *attr,
> + struct ibv_create_ah *cmd, size_t cmd_size,
> + struct ibv_create_ah_resp *resp, size_t resp_size);
> int ibv_cmd_destroy_ah(struct ibv_ah *ah);
> int ibv_cmd_attach_mcast(struct ibv_qp *qp, const union ibv_gid *gid, uint16_t lid);
> int ibv_cmd_detach_mcast(struct ibv_qp *qp, const union ibv_gid *gid, uint16_t lid);
> diff --git a/src/cmd.c b/src/cmd.c
> index 11f6509..58b9dcd 100644
> --- a/src/cmd.c
> +++ b/src/cmd.c
> @@ -1493,38 +1493,44 @@ int ibv_cmd_post_srq_recv(struct ibv_srq *srq, struct ibv_recv_wr *wr,
> return ret;
> }
>
> -int ibv_cmd_create_ah(struct ibv_pd *pd, struct ibv_ah *ah,
> - struct ibv_ah_attr *attr)
> +int ibv_cmd_create_ah_ex(struct ibv_pd *pd, struct ibv_ah *ah, struct ibv_ah_attr *attr,
> + struct ibv_create_ah *cmd, size_t cmd_size,
> + struct ibv_create_ah_resp *resp, size_t resp_size)
> {
> - struct ibv_create_ah cmd;
> - struct ibv_create_ah_resp resp;
> -
> - IBV_INIT_CMD_RESP(&cmd, sizeof cmd, CREATE_AH, &resp, sizeof resp);
> - cmd.user_handle = (uintptr_t) ah;
> - cmd.pd_handle = pd->handle;
> - cmd.attr.dlid = attr->dlid;
> - cmd.attr.sl = attr->sl;
> - cmd.attr.src_path_bits = attr->src_path_bits;
> - cmd.attr.static_rate = attr->static_rate;
> - cmd.attr.is_global = attr->is_global;
> - cmd.attr.port_num = attr->port_num;
> - cmd.attr.grh.flow_label = attr->grh.flow_label;
> - cmd.attr.grh.sgid_index = attr->grh.sgid_index;
> - cmd.attr.grh.hop_limit = attr->grh.hop_limit;
> - cmd.attr.grh.traffic_class = attr->grh.traffic_class;
> - memcpy(cmd.attr.grh.dgid, attr->grh.dgid.raw, 16);
> + IBV_INIT_CMD_RESP(cmd, cmd_size, CREATE_AH, resp, resp_size);
> + cmd->user_handle = (uintptr_t) ah;
> + cmd->pd_handle = pd->handle;
> + cmd->attr.dlid = attr->dlid;
> + cmd->attr.sl = attr->sl;
> + cmd->attr.src_path_bits = attr->src_path_bits;
> + cmd->attr.static_rate = attr->static_rate;
> + cmd->attr.is_global = attr->is_global;
> + cmd->attr.port_num = attr->port_num;
> + cmd->attr.grh.flow_label = attr->grh.flow_label;
> + cmd->attr.grh.sgid_index = attr->grh.sgid_index;
> + cmd->attr.grh.hop_limit = attr->grh.hop_limit;
> + cmd->attr.grh.traffic_class = attr->grh.traffic_class;
> + memcpy(cmd->attr.grh.dgid, attr->grh.dgid.raw, 16);
>
> - if (write(pd->context->cmd_fd, &cmd, sizeof cmd) != sizeof cmd)
> + if (write(pd->context->cmd_fd, cmd, cmd_size) != cmd_size)
> return errno;
>
> - (void) VALGRIND_MAKE_MEM_DEFINED(&resp, sizeof resp);
> + (void) VALGRIND_MAKE_MEM_DEFINED(resp, sizeof *resp);
>
> - ah->handle = resp.handle;
> + ah->handle = resp->handle;
> ah->context = pd->context;
>
> return 0;
> }
>
> +int ibv_cmd_create_ah(struct ibv_pd *pd, struct ibv_ah *ah,
> + struct ibv_ah_attr *attr)
> +{
> + struct ibv_create_ah cmd;
> + struct ibv_create_ah_resp resp;
> + return ibv_cmd_create_ah_ex(pd, ah, attr, &cmd, sizeof(cmd), &resp, sizeof(resp));
> +}
> +
> int ibv_cmd_destroy_ah(struct ibv_ah *ah)
> {
> struct ibv_destroy_ah cmd;
> diff --git a/src/libibverbs.map b/src/libibverbs.map
> index 46744e5..ddba63a 100644
> --- a/src/libibverbs.map
> +++ b/src/libibverbs.map
> @@ -118,7 +118,6 @@ IBVERBS_1.1 {
> ibv_cmd_create_qp_ex2;
> ibv_cmd_open_qp;
> ibv_cmd_rereg_mr;
> -
> } IBVERBS_1.0;
>
> IBVERBS_1.3 {
> @@ -130,3 +129,8 @@ IBVERBS_1.3 {
> ibv_cmd_destroy_rwq_ind_table;
> ibv_query_gid_type;
> } IBVERBS_1.1;
> +
> +IBVERBS_1.4 {
Any reason to add 1.4 instead of using 1.3 ? in my understanding the
idea is that new symbols for next release will use a new ABI version,
this is what 1.3 already supplied. Added Jason for his input as well.
> + global:
> + ibv_cmd_create_ah_ex;
When do you plan to contribute the driver code for that ? for now no
upstream drivers uses this API, taking some code for future usage with
no current usage is not the preferred way.
> +} IBVERBS_1.3;
>
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
^ permalink raw reply
* Re: [PATCH libibverbs v2 3/3] Provide remote XRC SRQ number in kernel post_send.
From: Knut Omang @ 2016-09-20 11:08 UTC (permalink / raw)
To: Leon Romanovsky
Cc: Doug Ledford, linux-rdma-u79uwXL29TY76Z2rM5mHXA, Mukesh Kacker
In-Reply-To: <20160920110115.GJ26673-2ukJVAZIZ/Y@public.gmane.org>
On Tue, 2016-09-20 at 14:01 +0300, Leon Romanovsky wrote:
> On Tue, Sep 20, 2016 at 12:43:30PM +0200, Knut Omang wrote:
> >
> > On Tue, 2016-09-20 at 13:18 +0300, Leon Romanovsky wrote:
> > >
> > > On Mon, Sep 19, 2016 at 11:12:36AM +0200, Knut Omang wrote:
> > > >
> > > >
> > > > On Mon, 2016-09-19 at 08:29 +0300, Leon Romanovsky wrote:
> > > > >
> > > > >
> > > > > On Sat, Sep 17, 2016 at 05:59:13AM +0200, Knut Omang wrote:
> > > > > >
> > > > > >
> > > > > >
> > > > > > Also proper end align the ibv_kern_send_wr struct.
> > > > > >
> > > > > > Requires a corresponding kernel change.
> > > > > >
> > > > > > Signed-off-by: Knut Omang <knut.omang-QHcLZuEGTsvQT0dZR+AlfA@public.gmane.org>
> > > > > > Reviewed-by: Mukesh Kacker <mukesh.kacker-QHcLZuEGTsvQT0dZR+AlfA@public.gmane.org>
> > > > > > ---
> > > > > > include/infiniband/kern-abi.h | 1 +
> > > > > > src/cmd.c | 2 ++
> > > > > > 2 files changed, 3 insertions(+), 0 deletions(-)
> > > > > >
> > > > > > diff --git a/include/infiniband/kern-abi.h b/include/infiniband/kern-abi.h
> > > > > > index 8bdeef5..7b1d310 100644
> > > > > > --- a/include/infiniband/kern-abi.h
> > > > > > +++ b/include/infiniband/kern-abi.h
> > > > > > @@ -800,6 +800,7 @@ struct ibv_kern_send_wr {
> > > > > > union {
> > > > > > struct {
> > > > > > __u32 remote_srqn;
> > > > > > + __u32 reserved;
> > > > > > } xrc;
> > > > > > } qp_type;
> > > > > > };
> > > > > > diff --git a/src/cmd.c b/src/cmd.c
> > > > > > index a418ee1..a4e2f75 100644
> > > > > > --- a/src/cmd.c
> > > > > > +++ b/src/cmd.c
> > > > > > @@ -1293,6 +1293,8 @@ int ibv_cmd_post_send(struct ibv_qp *ibqp, struct ibv_send_wr *wr,
> > > > > > tmp->wr.ud.remote_qpn = i->wr.ud.remote_qpn;
> > > > > > tmp->wr.ud.remote_qkey = i->wr.ud.remote_qkey;
> > > > > > } else {
> > > > > > + if (ibqp->qp_type == IBV_QPT_XRC_SEND)
> > > > > > + tmp->qp_type.xrc.remote_srqn = i->qp_type.xrc.remote_srqn;
> > > > > It will be checked for all QPTs and for all consumers. Any chances to
> > > > > optimize it?
> > > > All QPTs except UD, yes.
> > > >
> > > > The easiest is perhaps just to remove the test (and set the attribute in all cases), which will work fine
> > > > as long as the qp_type union is not used for anything else, but is it worth
> > > > the loss of intuitiveness/future potential correctness of the code?
> > > Let's concentrate on the present and as far as I see there are no kernel users who
> > > need this field xrc.remote_srqn field.
> > So I assume then that you are ok with this code, given my explanation?
> Partially,
>
> I'm fine with the claim about readability and the fact that
> this code is not in data-path relaxed me a little bit.
Good!
> I'm not fine with the part that this code doesn't do anything in kernel.
I know, I have two big patch sets one kernel and one for the new "rdma consolidated"
in the queue that should fix that, just want to save a few versions of those
~32000 lines of code by preparing the baseline :-)
Knut
>
> >
> >
> > Thanks,
> > Knut
> >
> > >
> > > >
> > > >
> > > >
> > > > Note that this is for user mode post_send implemented by the kernel, not the "fast path"
> > > > pure user mode.
> > > >
> > > > Knut
> > > >
> > > > >
> > > > >
> > > > > >
> > > > > >
> > > > > >
> > > > > > switch (i->opcode) {
> > > > > > case IBV_WR_RDMA_WRITE:
> > > > > > case IBV_WR_RDMA_WRITE_WITH_IMM:
> > > > > > --
> > > > > > git-series 0.8.10
> > > > > > --
> > > > > > To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
> > > > > > the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
> > > > > > More majordomo info at http://vger.kernel.org/majordomo-info.html
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
^ permalink raw reply
* Re: [PATCH v2 5/8] ib_uverbs: Add padding to end align ib_uverbs_reg_mr_resp
From: Knut Omang @ 2016-09-20 11:02 UTC (permalink / raw)
To: Yishai Hadas
Cc: Doug Ledford, linux-rdma, linux-kernel, Sean Hefty,
Hal Rosenstock
In-Reply-To: <f049462a-5cd0-10bf-320b-c5bc634e93b9@dev.mellanox.co.il>
On Tue, 2016-09-20 at 13:45 +0300, Yishai Hadas wrote:
> On 9/16/2016 9:31 PM, Knut Omang wrote:
> >
> > The ib_uverbs_reg_mr_resp structure was not 64 bit end aligned
> > as required by the protocol. This causes alignment issues
> > if a device specific driver needs to transfer extra response
> > arguments.
> >
> > Avoid breaking backward compatibility by improving the handling
> > of the length checking in ib_uverbs_reg_mr to consider the case
> > where the kernel has been updated, but user space still has
> > the old length without padding.
> >
> > Signed-off-by: Knut Omang <knut.omang@oracle.com>
> > ---
> > drivers/infiniband/core/uverbs_cmd.c | 10 ++++++----
> > include/uapi/rdma/ib_user_verbs.h | 1 +
> > 2 files changed, 7 insertions(+), 4 deletions(-)
> >
> > diff --git a/drivers/infiniband/core/uverbs_cmd.c b/drivers/infiniband/core/uverbs_cmd.c
> > index dbc5885..fa8a717 100644
> > --- a/drivers/infiniband/core/uverbs_cmd.c
> > +++ b/drivers/infiniband/core/uverbs_cmd.c
> > @@ -967,16 +967,18 @@ ssize_t ib_uverbs_reg_mr(struct ib_uverbs_file *file,
> > struct ib_pd *pd;
> > struct ib_mr *mr;
> > int ret;
> > + size_t resp_size = sizeof resp;
> >
> > - if (out_len < sizeof resp)
> > - return -ENOSPC;
> > + if (out_len < resp_size) {
> > + resp_size = out_len;
> > + }
> You don't preserve the minimum response size error checking in your code
> (i.e -ENOSPC). We can expect an error if outlen is less than the
> offset of rkey in ib_uverbs_reg_mr_resp.
I agree, the minimal size needs to be the minimal supported size,
in particular since the memory for cmd may contain arbitrary stack content.
Good catch!
> Please note that below call to copy_to_user will succeed as it copies
> now based on resp_size and won't return the basic mandatory output.
Get it - will fix,
Thanks,
Knut
>
> >
> > if (copy_from_user(&cmd, buf, sizeof cmd))
> > return -EFAULT;
> >
> > INIT_UDATA(&udata, buf + sizeof cmd,
> > (unsigned long) cmd.response + sizeof resp,
> > - in_len - sizeof cmd, out_len - sizeof resp);
> > + in_len - sizeof cmd, out_len - resp_size);
> >
> > if ((cmd.start & ~PAGE_MASK) != (cmd.hca_va & ~PAGE_MASK))
> > return -EINVAL;
> > @@ -1030,7 +1032,7 @@ ssize_t ib_uverbs_reg_mr(struct ib_uverbs_file *file,
> > resp.mr_handle = uobj->id;
> >
> > if (copy_to_user((void __user *) (unsigned long) cmd.response,
> > - &resp, sizeof resp)) {
> > + &resp, resp_size)) {
> > ret = -EFAULT;
> > goto err_copy;
> > }
> > diff --git a/include/uapi/rdma/ib_user_verbs.h b/include/uapi/rdma/ib_user_verbs.h
> > index 7f035f4..6b8c9c0 100644
> > --- a/include/uapi/rdma/ib_user_verbs.h
> > +++ b/include/uapi/rdma/ib_user_verbs.h
> > @@ -307,6 +307,7 @@ struct ib_uverbs_reg_mr_resp {
> > __u32 mr_handle;
> > __u32 lkey;
> > __u32 rkey;
> > + __u32 reserved;
> > };
> >
> > struct ib_uverbs_rereg_mr {
> >
^ permalink raw reply
* Re: [PATCH libibverbs v2 3/3] Provide remote XRC SRQ number in kernel post_send.
From: Leon Romanovsky @ 2016-09-20 11:01 UTC (permalink / raw)
To: Knut Omang; +Cc: Doug Ledford, linux-rdma-u79uwXL29TY76Z2rM5mHXA, Mukesh Kacker
In-Reply-To: <1474368210.8837.3.camel-QHcLZuEGTsvQT0dZR+AlfA@public.gmane.org>
[-- Attachment #1: Type: text/plain, Size: 3300 bytes --]
On Tue, Sep 20, 2016 at 12:43:30PM +0200, Knut Omang wrote:
> On Tue, 2016-09-20 at 13:18 +0300, Leon Romanovsky wrote:
> > On Mon, Sep 19, 2016 at 11:12:36AM +0200, Knut Omang wrote:
> > >
> > > On Mon, 2016-09-19 at 08:29 +0300, Leon Romanovsky wrote:
> > > >
> > > > On Sat, Sep 17, 2016 at 05:59:13AM +0200, Knut Omang wrote:
> > > > >
> > > > >
> > > > > Also proper end align the ibv_kern_send_wr struct.
> > > > >
> > > > > Requires a corresponding kernel change.
> > > > >
> > > > > Signed-off-by: Knut Omang <knut.omang-QHcLZuEGTsvQT0dZR+AlfA@public.gmane.org>
> > > > > Reviewed-by: Mukesh Kacker <mukesh.kacker-QHcLZuEGTsvQT0dZR+AlfA@public.gmane.org>
> > > > > ---
> > > > > include/infiniband/kern-abi.h | 1 +
> > > > > src/cmd.c | 2 ++
> > > > > 2 files changed, 3 insertions(+), 0 deletions(-)
> > > > >
> > > > > diff --git a/include/infiniband/kern-abi.h b/include/infiniband/kern-abi.h
> > > > > index 8bdeef5..7b1d310 100644
> > > > > --- a/include/infiniband/kern-abi.h
> > > > > +++ b/include/infiniband/kern-abi.h
> > > > > @@ -800,6 +800,7 @@ struct ibv_kern_send_wr {
> > > > > union {
> > > > > struct {
> > > > > __u32 remote_srqn;
> > > > > + __u32 reserved;
> > > > > } xrc;
> > > > > } qp_type;
> > > > > };
> > > > > diff --git a/src/cmd.c b/src/cmd.c
> > > > > index a418ee1..a4e2f75 100644
> > > > > --- a/src/cmd.c
> > > > > +++ b/src/cmd.c
> > > > > @@ -1293,6 +1293,8 @@ int ibv_cmd_post_send(struct ibv_qp *ibqp, struct ibv_send_wr *wr,
> > > > > tmp->wr.ud.remote_qpn = i->wr.ud.remote_qpn;
> > > > > tmp->wr.ud.remote_qkey = i->wr.ud.remote_qkey;
> > > > > } else {
> > > > > + if (ibqp->qp_type == IBV_QPT_XRC_SEND)
> > > > > + tmp->qp_type.xrc.remote_srqn = i->qp_type.xrc.remote_srqn;
> > > > It will be checked for all QPTs and for all consumers. Any chances to
> > > > optimize it?
> > > All QPTs except UD, yes.
> > >
> > > The easiest is perhaps just to remove the test (and set the attribute in all cases), which will work fine
> > > as long as the qp_type union is not used for anything else, but is it worth
> > > the loss of intuitiveness/future potential correctness of the code?
> > Let's concentrate on the present and as far as I see there are no kernel users who
> > need this field xrc.remote_srqn field.
>
> So I assume then that you are ok with this code, given my explanation?
Partially,
I'm fine with the claim about readability and the fact that
this code is not in data-path relaxed me a little bit.
I'm not fine with the part that this code doesn't do anything in kernel.
>
> Thanks,
> Knut
>
> > >
> > >
> > > Note that this is for user mode post_send implemented by the kernel, not the "fast path"
> > > pure user mode.
> > >
> > > Knut
> > >
> > > >
> > > > >
> > > > >
> > > > > switch (i->opcode) {
> > > > > case IBV_WR_RDMA_WRITE:
> > > > > case IBV_WR_RDMA_WRITE_WITH_IMM:
> > > > > --
> > > > > git-series 0.8.10
> > > > > --
> > > > > To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
> > > > > the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
> > > > > More majordomo info at http://vger.kernel.org/majordomo-info.html
[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 819 bytes --]
^ permalink raw reply
* Re: VERBOSE flags
From: Leon Romanovsky @ 2016-09-20 10:56 UTC (permalink / raw)
To: Dalessandro, Dennis
Cc: linux-rdma-u79uwXL29TY76Z2rM5mHXA@public.gmane.org,
dledford-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org
In-Reply-To: <1474292075.6520.28.camel-ral2JQCrhuEAvxtiuMwx3w@public.gmane.org>
[-- Attachment #1: Type: text/plain, Size: 2440 bytes --]
On Mon, Sep 19, 2016 at 01:34:39PM +0000, Dalessandro, Dennis wrote:
> On Mon, 2016-09-19 at 16:25 +0300, Leon Romanovsky wrote:
> > On Mon, Sep 19, 2016 at 12:29:43PM +0000, Dalessandro, Dennis wrote:
> > > On Mon, 2016-09-19 at 09:15 +0300, Leon Romanovsky wrote:
> > > > The function pointers will cause us to create two similar
> > > > functions
> > > > (debug/regular) for every important function. It will leave to
> > > > code
> > > > duplication and/or introducing of new code to fill additional
> > > > fields
> > > > in returns to debug functions, so they will be able to print it.
> > >
> > > I do sort of like Doug's idea. It means you can easily instrument
> > > the
> > > heck out of a routine and have it always available. It provides
> > > more
> > > flexibility in what you can do in a debug version.
> > >
> > > You do make a valid point though about duplicating code. Perhaps
> > > that
> > > can be alleviated to some extent just by refactoring and adding
> > > inlines, etc.
> >
> > At the end, we will invent tracepoints :)
> >
> > >
> > > > I wanted to propose to use tracepoints to provide debug
> > > > information
> > > > in
> > > > data paths. The tracepoints infrastructure uses hooks in the code
> > > > with
> > > > minimal impact on performance.
> > >
> > > This is so far the path we have chosen for hfi1. We have a number
> > > of
> > > tracepoints in various parts of the code. Again it's available on
> > > demand without special kernels or boot/driver options.
> >
> > CONFIG_SDMA_VERBOSITY ????
>
> Yes we do have that, which illustrates that trace points can not do
> everything. I had hoped to deprecate that at some point. Regardless,
> sometimes you want additional code to be added or to do something
> slightly different. You can't do that with a tracepoint.
Can you elaborate more on this topic? Which limitations did you see
in tracepoints?
In regards of other discussion in KS 2016, "tracepoints and ABI
stability warranties" [1]. Are you strict with this tracepoints ABI?
>
> You can see we use tracepoints pretty extensively and that is our main
> vehicle for debugging. Here is where all the definitions are:
>
> $ ls trace_*
> trace_ctxts.h trace_dbg.h trace_ibhdrs.h trace_misc.h trace_rc.h t
> race_rx.h trace_tx.h
+1 :)
[1] https://lists.linuxfoundation.org/pipermail/ksummit-discuss/2016-September/003850.html
[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 819 bytes --]
^ permalink raw reply
* Re: [PATCH v2 5/8] ib_uverbs: Add padding to end align ib_uverbs_reg_mr_resp
From: Yishai Hadas @ 2016-09-20 10:45 UTC (permalink / raw)
To: Knut Omang
Cc: Doug Ledford, linux-rdma, linux-kernel, Sean Hefty,
Hal Rosenstock, Matan Barak, Sagi Grimberg, Yishai Hadas,
Leon Romanovsky, Majd Dibbiny, Eran Ben Elisha, Or Gerlitz
In-Reply-To: <aafa787a72e9df9a40b292fb33833e67ffb46c6a.1474049924.git-series.knut.omang@oracle.com>
On 9/16/2016 9:31 PM, Knut Omang wrote:
> The ib_uverbs_reg_mr_resp structure was not 64 bit end aligned
> as required by the protocol. This causes alignment issues
> if a device specific driver needs to transfer extra response
> arguments.
>
> Avoid breaking backward compatibility by improving the handling
> of the length checking in ib_uverbs_reg_mr to consider the case
> where the kernel has been updated, but user space still has
> the old length without padding.
>
> Signed-off-by: Knut Omang <knut.omang@oracle.com>
> ---
> drivers/infiniband/core/uverbs_cmd.c | 10 ++++++----
> include/uapi/rdma/ib_user_verbs.h | 1 +
> 2 files changed, 7 insertions(+), 4 deletions(-)
>
> diff --git a/drivers/infiniband/core/uverbs_cmd.c b/drivers/infiniband/core/uverbs_cmd.c
> index dbc5885..fa8a717 100644
> --- a/drivers/infiniband/core/uverbs_cmd.c
> +++ b/drivers/infiniband/core/uverbs_cmd.c
> @@ -967,16 +967,18 @@ ssize_t ib_uverbs_reg_mr(struct ib_uverbs_file *file,
> struct ib_pd *pd;
> struct ib_mr *mr;
> int ret;
> + size_t resp_size = sizeof resp;
>
> - if (out_len < sizeof resp)
> - return -ENOSPC;
> + if (out_len < resp_size) {
> + resp_size = out_len;
> + }
You don't preserve the minimum response size error checking in your code
(i.e -ENOSPC). We can expect an error if outlen is less than the
offset of rkey in ib_uverbs_reg_mr_resp.
Please note that below call to copy_to_user will succeed as it copies
now based on resp_size and won't return the basic mandatory output.
> if (copy_from_user(&cmd, buf, sizeof cmd))
> return -EFAULT;
>
> INIT_UDATA(&udata, buf + sizeof cmd,
> (unsigned long) cmd.response + sizeof resp,
> - in_len - sizeof cmd, out_len - sizeof resp);
> + in_len - sizeof cmd, out_len - resp_size);
>
> if ((cmd.start & ~PAGE_MASK) != (cmd.hca_va & ~PAGE_MASK))
> return -EINVAL;
> @@ -1030,7 +1032,7 @@ ssize_t ib_uverbs_reg_mr(struct ib_uverbs_file *file,
> resp.mr_handle = uobj->id;
>
> if (copy_to_user((void __user *) (unsigned long) cmd.response,
> - &resp, sizeof resp)) {
> + &resp, resp_size)) {
> ret = -EFAULT;
> goto err_copy;
> }
> diff --git a/include/uapi/rdma/ib_user_verbs.h b/include/uapi/rdma/ib_user_verbs.h
> index 7f035f4..6b8c9c0 100644
> --- a/include/uapi/rdma/ib_user_verbs.h
> +++ b/include/uapi/rdma/ib_user_verbs.h
> @@ -307,6 +307,7 @@ struct ib_uverbs_reg_mr_resp {
> __u32 mr_handle;
> __u32 lkey;
> __u32 rkey;
> + __u32 reserved;
> };
>
> struct ib_uverbs_rereg_mr {
>
^ permalink raw reply
* Re: [PATCH libibverbs v2 3/3] Provide remote XRC SRQ number in kernel post_send.
From: Knut Omang @ 2016-09-20 10:43 UTC (permalink / raw)
To: Leon Romanovsky
Cc: Doug Ledford, linux-rdma-u79uwXL29TY76Z2rM5mHXA, Mukesh Kacker
In-Reply-To: <20160920101839.GF26673-2ukJVAZIZ/Y@public.gmane.org>
On Tue, 2016-09-20 at 13:18 +0300, Leon Romanovsky wrote:
> On Mon, Sep 19, 2016 at 11:12:36AM +0200, Knut Omang wrote:
> >
> > On Mon, 2016-09-19 at 08:29 +0300, Leon Romanovsky wrote:
> > >
> > > On Sat, Sep 17, 2016 at 05:59:13AM +0200, Knut Omang wrote:
> > > >
> > > >
> > > > Also proper end align the ibv_kern_send_wr struct.
> > > >
> > > > Requires a corresponding kernel change.
> > > >
> > > > Signed-off-by: Knut Omang <knut.omang-QHcLZuEGTsvQT0dZR+AlfA@public.gmane.org>
> > > > Reviewed-by: Mukesh Kacker <mukesh.kacker-QHcLZuEGTsvQT0dZR+AlfA@public.gmane.org>
> > > > ---
> > > > include/infiniband/kern-abi.h | 1 +
> > > > src/cmd.c | 2 ++
> > > > 2 files changed, 3 insertions(+), 0 deletions(-)
> > > >
> > > > diff --git a/include/infiniband/kern-abi.h b/include/infiniband/kern-abi.h
> > > > index 8bdeef5..7b1d310 100644
> > > > --- a/include/infiniband/kern-abi.h
> > > > +++ b/include/infiniband/kern-abi.h
> > > > @@ -800,6 +800,7 @@ struct ibv_kern_send_wr {
> > > > union {
> > > > struct {
> > > > __u32 remote_srqn;
> > > > + __u32 reserved;
> > > > } xrc;
> > > > } qp_type;
> > > > };
> > > > diff --git a/src/cmd.c b/src/cmd.c
> > > > index a418ee1..a4e2f75 100644
> > > > --- a/src/cmd.c
> > > > +++ b/src/cmd.c
> > > > @@ -1293,6 +1293,8 @@ int ibv_cmd_post_send(struct ibv_qp *ibqp, struct ibv_send_wr *wr,
> > > > tmp->wr.ud.remote_qpn = i->wr.ud.remote_qpn;
> > > > tmp->wr.ud.remote_qkey = i->wr.ud.remote_qkey;
> > > > } else {
> > > > + if (ibqp->qp_type == IBV_QPT_XRC_SEND)
> > > > + tmp->qp_type.xrc.remote_srqn = i->qp_type.xrc.remote_srqn;
> > > It will be checked for all QPTs and for all consumers. Any chances to
> > > optimize it?
> > All QPTs except UD, yes.
> >
> > The easiest is perhaps just to remove the test (and set the attribute in all cases), which will work fine
> > as long as the qp_type union is not used for anything else, but is it worth
> > the loss of intuitiveness/future potential correctness of the code?
> Let's concentrate on the present and as far as I see there are no kernel users who
> need this field xrc.remote_srqn field.
So I assume then that you are ok with this code, given my explanation?
Thanks,
Knut
> >
> >
> > Note that this is for user mode post_send implemented by the kernel, not the "fast path"
> > pure user mode.
> >
> > Knut
> >
> > >
> > > >
> > > >
> > > > switch (i->opcode) {
> > > > case IBV_WR_RDMA_WRITE:
> > > > case IBV_WR_RDMA_WRITE_WITH_IMM:
> > > > --
> > > > git-series 0.8.10
> > > > --
> > > > To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
> > > > the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
> > > > More majordomo info at http://vger.kernel.org/majordomo-info.html
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
^ permalink raw reply
* Re: VERBOSE flags
From: Leon Romanovsky @ 2016-09-20 10:42 UTC (permalink / raw)
To: Doug Ledford; +Cc: linux-rdma
In-Reply-To: <02761057-02de-d3d1-13d9-ca64e4ee9556-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>
[-- Attachment #1: Type: text/plain, Size: 2572 bytes --]
On Mon, Sep 19, 2016 at 12:03:34PM -0400, Doug Ledford wrote:
> On 9/19/2016 2:15 AM, Leon Romanovsky wrote:
> > On Sun, Sep 18, 2016 at 04:35:37PM -0400, Doug Ledford wrote:
>
> > Doug,
> > The function pointers will cause us to create two similar functions
> > (debug/regular) for every important function. It will leave to code
> > duplication and/or introducing of new code to fill additional fields
> > in returns to debug functions, so they will be able to print it.
>
> That's situationally specific. The original patch you posted that I
> said no to is an example of where this wouldn't have been true.
> Basically it all boils down to whether or not your debug code can be a
> wrapper of your existing function or if it needs to be embedded in the
> middle of it. If the latter, then you are more likely to have code
> duplication as you point out. But, then refactoring code can help with
> that problem too.
That specific patch was sent to ensure that if(0) {...} code isn't
removed by mistake.
>
> > I wanted to propose to use tracepoints to provide debug information in
> > data paths. The tracepoints infrastructure uses hooks in the code with
> > minimal impact on performance.
>
> Yes, but one of the things you said before was "zero impact on normal
> operations", but with tracepoints you just said "minimal impact". So it
> sounds like tracepoints don't meet your own initial requirements. I'm
> not against them, but utilizing the entry points allows for achievment
> of the "zero impact" requirement. And as Denny pointed out, some things
> can't be done with tracepoints. So, the final solution might be a mix
> of different things including different entry points, tracepoints, and
> some core infrastructure.
You are right and my initial requirement which is fine in theory (zero
impact) not so good in real life (precompile e.t.c).
Let's start from the tracepoints and see what elase we need to add to core.
>
> >
> >> Then we can drop the CONFIG_*_DEBUG options out of all of the
> >> drivers. That would be my preference. I don't want any compile time
> >> options because you never will be running a debug kernel when you need
> >> it. Per device, run time selectable so that you get 0 overhead in the
> >> normal case is what I would like to see.
> >
> > Great,
> > We are aligned on this point, so do I. I see no reason in special compiled
> > kernel or global module parameter to print debug prints for whole driver.
>
>
>
> --
> Doug Ledford <dledford-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>
> GPG Key ID: 0E572FDD
>
[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 819 bytes --]
^ permalink raw reply
* [RFC v2 12/12] IB/core: add a protection to ib_get_dma_mr
From: Ram Amrani @ 2016-09-20 10:36 UTC (permalink / raw)
To: davem, dledford
Cc: Ariel.Elior, Michal.Kalderon, Yuval.Mintz, rajesh.borundia,
linux-rdma, netdev, Ram amrani
In-Reply-To: <1474367764-9555-1-git-send-email-Ram.Amrani@cavium.com>
From: Ram amrani <Ram.Amrani@cavium.com>
Add a check that prevents ib_get_dma_mr from executing if
IB_ACCESS_MW_BIND is set.
Signed-off-by: Ram Amrani <Ram.Amrani@cavium.com>
---
drivers/infiniband/core/verbs.c | 3 +++
1 file changed, 3 insertions(+)
diff --git a/drivers/infiniband/core/verbs.c b/drivers/infiniband/core/verbs.c
index f2b776e..8cc6f79 100644
--- a/drivers/infiniband/core/verbs.c
+++ b/drivers/infiniband/core/verbs.c
@@ -1400,6 +1400,9 @@ struct ib_mr *ib_get_dma_mr(struct ib_pd *pd, int mr_access_flags)
if (err)
return ERR_PTR(err);
+ if (mr_access_flags & IB_ACCESS_MW_BIND)
+ return ERR_PTR(-EINVAL);
+
mr = pd->device->get_dma_mr(pd, mr_access_flags);
if (!IS_ERR(mr)) {
--
1.8.3.1
^ permalink raw reply related
* [RFC v2 11/12] qedr: Add events support and register IB device
From: Ram Amrani @ 2016-09-20 10:36 UTC (permalink / raw)
To: davem, dledford
Cc: Ariel.Elior, Michal.Kalderon, Yuval.Mintz, rajesh.borundia,
linux-rdma, netdev, Ram amrani
In-Reply-To: <1474367764-9555-1-git-send-email-Ram.Amrani@cavium.com>
From: Ram amrani <Ram.Amrani@cavium.com>
Add error handling support.
Register ib device with ib stack.
Signed-off-by: Rajesh Borundia <rajesh.borundia@cavium.com>
Signed-off-by: Ram Amrani <Ram.Amrani@cavium.com>
---
drivers/infiniband/hw/qedr/main.c | 107 ++++++++++++++++++++++++++++++++++++-
drivers/infiniband/hw/qedr/verbs.c | 37 +++++++++++++
drivers/infiniband/hw/qedr/verbs.h | 9 ++++
include/linux/qed/common_hsi.h | 1 +
4 files changed, 152 insertions(+), 2 deletions(-)
diff --git a/drivers/infiniband/hw/qedr/main.c b/drivers/infiniband/hw/qedr/main.c
index d3555bf..b43c4ad 100644
--- a/drivers/infiniband/hw/qedr/main.c
+++ b/drivers/infiniband/hw/qedr/main.c
@@ -83,10 +83,25 @@ static void qedr_get_dev_fw_str(struct ib_device *ibdev, char *str,
(fw_ver >> 8) & 0xFF, fw_ver & 0xFF);
}
+static struct net_device *qedr_get_netdev(struct ib_device *dev, u8 port_num)
+{
+ struct qedr_dev *qdev;
+
+ qdev = get_qedr_dev(dev);
+ dev_hold(qdev->ndev);
+
+ /* The HW vendor's device driver must guarantee
+ * that this function returns NULL before the net device reaches
+ * NETDEV_UNREGISTER_FINAL state.
+ */
+ return qdev->ndev;
+}
+
static int qedr_register_device(struct qedr_dev *dev)
{
strlcpy(dev->ibdev.name, "qedr%d", IB_DEVICE_NAME_MAX);
+ dev->ibdev.node_guid = dev->attr.node_guid;
memcpy(dev->ibdev.node_desc, QEDR_NODE_DESC, sizeof(QEDR_NODE_DESC));
dev->ibdev.owner = THIS_MODULE;
dev->ibdev.uverbs_abi_ver = QEDR_ABI_VERSION;
@@ -153,11 +168,15 @@ static int qedr_register_device(struct qedr_dev *dev)
dev->ibdev.post_send = qedr_post_send;
dev->ibdev.post_recv = qedr_post_recv;
+ dev->ibdev.process_mad = qedr_process_mad;
+ dev->ibdev.get_port_immutable = qedr_port_immutable;
+ dev->ibdev.get_netdev = qedr_get_netdev;
+
dev->ibdev.dma_device = &dev->pdev->dev;
dev->ibdev.get_link_layer = qedr_link_layer;
dev->ibdev.get_dev_fw_str = qedr_get_dev_fw_str;
- return 0;
+ return ib_register_device(&dev->ibdev, NULL);
}
/* This function allocates fast-path status block memory */
@@ -542,6 +561,86 @@ static int qedr_set_device_attr(struct qedr_dev *dev)
return 0;
}
+void qedr_affiliated_event(void *context, u8 e_code, void *fw_handle)
+{
+#define EVENT_TYPE_NOT_DEFINED 0
+#define EVENT_TYPE_CQ 1
+#define EVENT_TYPE_QP 2
+ struct qedr_dev *dev = (struct qedr_dev *)context;
+ union event_ring_data *data = fw_handle;
+ u64 roce_handle64 = ((u64)data->roce_handle.hi << 32) +
+ data->roce_handle.lo;
+ u8 event_type = EVENT_TYPE_NOT_DEFINED;
+ struct ib_event event;
+ struct ib_cq *ibcq;
+ struct ib_qp *ibqp;
+ struct qedr_cq *cq;
+ struct qedr_qp *qp;
+
+ switch (e_code) {
+ case ROCE_ASYNC_EVENT_CQ_OVERFLOW_ERR:
+ event.event = IB_EVENT_CQ_ERR;
+ event_type = EVENT_TYPE_CQ;
+ break;
+ case ROCE_ASYNC_EVENT_SQ_DRAINED:
+ event.event = IB_EVENT_SQ_DRAINED;
+ event_type = EVENT_TYPE_QP;
+ break;
+ case ROCE_ASYNC_EVENT_QP_CATASTROPHIC_ERR:
+ event.event = IB_EVENT_QP_FATAL;
+ event_type = EVENT_TYPE_QP;
+ break;
+ case ROCE_ASYNC_EVENT_LOCAL_INVALID_REQUEST_ERR:
+ event.event = IB_EVENT_QP_REQ_ERR;
+ event_type = EVENT_TYPE_QP;
+ break;
+ case ROCE_ASYNC_EVENT_LOCAL_ACCESS_ERR:
+ event.event = IB_EVENT_QP_ACCESS_ERR;
+ event_type = EVENT_TYPE_QP;
+ break;
+ default:
+ DP_ERR(dev, "unsupported event %d on handle=%llx\n", e_code,
+ roce_handle64);
+ }
+
+ switch (event_type) {
+ case EVENT_TYPE_CQ:
+ cq = (struct qedr_cq *)(uintptr_t)roce_handle64;
+ if (cq) {
+ ibcq = &cq->ibcq;
+ if (ibcq->event_handler) {
+ event.device = ibcq->device;
+ event.element.cq = ibcq;
+ ibcq->event_handler(&event, ibcq->cq_context);
+ }
+ } else {
+ WARN(1,
+ "Error: CQ event with NULL pointer ibcq. Handle=%llx\n",
+ roce_handle64);
+ }
+ DP_ERR(dev, "CQ event %d on hanlde %p\n", e_code, cq);
+ break;
+ case EVENT_TYPE_QP:
+ qp = (struct qedr_qp *)(uintptr_t)roce_handle64;
+ if (qp) {
+ ibqp = &qp->ibqp;
+ if (ibqp->event_handler) {
+ event.device = ibqp->device;
+ event.element.qp = ibqp;
+ ibqp->event_handler(&event, ibqp->qp_context);
+ }
+ } else {
+ WARN(1,
+ "Error: QP event with NULL pointer ibqp. Handle=%llx\n",
+ roce_handle64);
+ }
+ DP_ERR(dev, "QP event %d on hanlde %p\n", e_code, qp);
+ break;
+ default:
+ break;
+ }
+}
+
static int qedr_init_hw(struct qedr_dev *dev)
{
struct qed_rdma_add_user_out_params out_params;
@@ -570,6 +669,7 @@ static int qedr_init_hw(struct qedr_dev *dev)
cur_pbl->pbl_ptr = (u64)p_phys_table;
}
+ events.affiliated_event = qedr_affiliated_event;
events.context = dev;
in_params->events = &events;
@@ -669,11 +769,13 @@ static struct qedr_dev *qedr_add(struct qed_dev *cdev, struct pci_dev *pdev,
for (i = 0; i < ARRAY_SIZE(qedr_attributes); i++)
if (device_create_file(&dev->ibdev.dev, qedr_attributes[i]))
- goto reg_err;
+ goto sysfs_err;
DP_VERBOSE(dev, QEDR_MSG_INIT, "qedr driver loaded successfully\n");
return dev;
+sysfs_err:
+ ib_unregister_device(&dev->ibdev);
reg_err:
qedr_sync_free_irqs(dev);
irq_err:
@@ -693,6 +795,7 @@ static void qedr_remove(struct qedr_dev *dev)
* of the registered clients.
*/
qedr_remove_sysfiles(dev);
+ ib_unregister_device(&dev->ibdev);
qedr_stop_hw(dev);
qedr_sync_free_irqs(dev);
diff --git a/drivers/infiniband/hw/qedr/verbs.c b/drivers/infiniband/hw/qedr/verbs.c
index 5c5fdd1..6ccf093 100644
--- a/drivers/infiniband/hw/qedr/verbs.c
+++ b/drivers/infiniband/hw/qedr/verbs.c
@@ -3484,3 +3484,40 @@ int qedr_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc)
spin_unlock_irqrestore(&cq->cq_lock, flags);
return done;
}
+
+int qedr_process_mad(struct ib_device *ibdev, int process_mad_flags,
+ u8 port_num,
+ const struct ib_wc *in_wc,
+ const struct ib_grh *in_grh,
+ const struct ib_mad_hdr *mad_hdr,
+ size_t in_mad_size, struct ib_mad_hdr *out_mad,
+ size_t *out_mad_size, u16 *out_mad_pkey_index)
+{
+ struct qedr_dev *dev = get_qedr_dev(ibdev);
+
+ DP_VERBOSE(dev, QEDR_MSG_GSI,
+ "QEDR_PROCESS_MAD in_mad %x %x %x %x %x %x %x %x\n",
+ mad_hdr->attr_id, mad_hdr->base_version, mad_hdr->attr_mod,
+ mad_hdr->class_specific, mad_hdr->class_version,
+ mad_hdr->method, mad_hdr->mgmt_class, mad_hdr->status);
+ return IB_MAD_RESULT_SUCCESS;
+}
+
+int qedr_port_immutable(struct ib_device *ibdev, u8 port_num,
+ struct ib_port_immutable *immutable)
+{
+ struct ib_port_attr attr;
+ int err;
+
+ err = qedr_query_port(ibdev, port_num, &attr);
+ if (err)
+ return err;
+
+ immutable->pkey_tbl_len = attr.pkey_tbl_len;
+ immutable->gid_tbl_len = attr.gid_tbl_len;
+ immutable->core_cap_flags = RDMA_CORE_PORT_IBA_ROCE |
+ RDMA_CORE_PORT_IBA_ROCE_UDP_ENCAP;
+ immutable->max_mad_size = IB_MGMT_MAD_SIZE;
+
+ return 0;
+}
diff --git a/drivers/infiniband/hw/qedr/verbs.h b/drivers/infiniband/hw/qedr/verbs.h
index 9eb2722b..da4535e 100644
--- a/drivers/infiniband/hw/qedr/verbs.h
+++ b/drivers/infiniband/hw/qedr/verbs.h
@@ -63,4 +63,13 @@ int qedr_post_send(struct ib_qp *, struct ib_send_wr *,
struct ib_send_wr **bad_wr);
int qedr_post_recv(struct ib_qp *, struct ib_recv_wr *,
struct ib_recv_wr **bad_wr);
+int qedr_process_mad(struct ib_device *ibdev, int process_mad_flags,
+ u8 port_num, const struct ib_wc *in_wc,
+ const struct ib_grh *in_grh,
+ const struct ib_mad_hdr *in_mad,
+ size_t in_mad_size, struct ib_mad_hdr *out_mad,
+ size_t *out_mad_size, u16 *out_mad_pkey_index);
+
+int qedr_port_immutable(struct ib_device *ibdev, u8 port_num,
+ struct ib_port_immutable *immutable);
#endif
diff --git a/include/linux/qed/common_hsi.h b/include/linux/qed/common_hsi.h
index 1902763..734deb0 100644
--- a/include/linux/qed/common_hsi.h
+++ b/include/linux/qed/common_hsi.h
@@ -674,6 +674,7 @@ union event_ring_data {
struct iscsi_eqe_data iscsi_info;
struct malicious_vf_eqe_data malicious_vf;
struct initial_cleanup_eqe_data vf_init_cleanup;
+ struct regpair roce_handle;
};
/* Event Ring Entry */
--
1.8.3.1
^ permalink raw reply related
* [RFC v2 10/12] qedr: Add GSI support
From: Ram Amrani @ 2016-09-20 10:36 UTC (permalink / raw)
To: davem-fT/PcQaiUtIeIZ0/mPfg9Q, dledford-H+wXaHxf7aLQT0dZR+AlfA
Cc: Ariel.Elior-YGCgFSpz5w/QT0dZR+AlfA,
Michal.Kalderon-YGCgFSpz5w/QT0dZR+AlfA,
Yuval.Mintz-YGCgFSpz5w/QT0dZR+AlfA,
rajesh.borundia-YGCgFSpz5w/QT0dZR+AlfA,
linux-rdma-u79uwXL29TY76Z2rM5mHXA, netdev-u79uwXL29TY76Z2rM5mHXA,
Ram amrani
In-Reply-To: <1474367764-9555-1-git-send-email-Ram.Amrani-YGCgFSpz5w/QT0dZR+AlfA@public.gmane.org>
From: Ram amrani <Ram.Amrani-YGCgFSpz5w/QT0dZR+AlfA@public.gmane.org>
Add support for GSI over light L2.
Signed-off-by: Rajesh Borundia <rajesh.borundia-YGCgFSpz5w/QT0dZR+AlfA@public.gmane.org>
Signed-off-by: Ram Amrani <Ram.Amrani-YGCgFSpz5w/QT0dZR+AlfA@public.gmane.org>
---
drivers/infiniband/hw/qedr/Makefile | 2 +-
drivers/infiniband/hw/qedr/main.c | 3 +
drivers/infiniband/hw/qedr/qedr.h | 15 +
drivers/infiniband/hw/qedr/qedr_cm.c | 626 +++++++++++++++++++++++++++++
drivers/infiniband/hw/qedr/qedr_cm.h | 21 +
drivers/infiniband/hw/qedr/verbs.c | 39 ++
drivers/net/ethernet/qlogic/qed/qed_ll2.c | 97 ++++-
drivers/net/ethernet/qlogic/qed/qed_ll2.h | 22 +-
drivers/net/ethernet/qlogic/qed/qed_roce.c | 66 +++
9 files changed, 878 insertions(+), 13 deletions(-)
create mode 100644 drivers/infiniband/hw/qedr/qedr_cm.c
diff --git a/drivers/infiniband/hw/qedr/Makefile b/drivers/infiniband/hw/qedr/Makefile
index b10f2b1..ba7067c 100644
--- a/drivers/infiniband/hw/qedr/Makefile
+++ b/drivers/infiniband/hw/qedr/Makefile
@@ -1,3 +1,3 @@
obj-$(CONFIG_INFINIBAND_QEDR) := qedr.o
-qedr-y := main.o verbs.o
+qedr-y := main.o verbs.o qedr_cm.o
diff --git a/drivers/infiniband/hw/qedr/main.c b/drivers/infiniband/hw/qedr/main.c
index fd602ad..d3555bf 100644
--- a/drivers/infiniband/hw/qedr/main.c
+++ b/drivers/infiniband/hw/qedr/main.c
@@ -140,6 +140,9 @@ static int qedr_register_device(struct qedr_dev *dev)
dev->ibdev.query_pkey = qedr_query_pkey;
+ dev->ibdev.create_ah = qedr_create_ah;
+ dev->ibdev.destroy_ah = qedr_destroy_ah;
+
dev->ibdev.get_dma_mr = qedr_get_dma_mr;
dev->ibdev.dereg_mr = qedr_dereg_mr;
dev->ibdev.reg_user_mr = qedr_reg_user_mr;
diff --git a/drivers/infiniband/hw/qedr/qedr.h b/drivers/infiniband/hw/qedr/qedr.h
index 18af842..37a7740 100644
--- a/drivers/infiniband/hw/qedr/qedr.h
+++ b/drivers/infiniband/hw/qedr/qedr.h
@@ -53,6 +53,7 @@ enum DP_QEDR_MODULE {
QEDR_MSG_SQ = 0x80000,
QEDR_MSG_QP = (QEDR_MSG_SQ | QEDR_MSG_RQ),
QEDR_MSG_MR = 0x100000,
+ QEDR_MSG_GSI = 0x200000,
QEDR_MSG_MISC = 0x400000,
};
@@ -146,6 +147,10 @@ struct qedr_dev {
u8 num_hwfns;
uint wq_multiplier;
u8 gsi_ll2_mac_address[ETH_ALEN];
+ int gsi_qp_created;
+ struct qedr_cq *gsi_sqcq;
+ struct qedr_cq *gsi_rqcq;
+ struct qedr_qp *gsi_qp;
};
#define QEDR_MAX_SQ_PBL (0x8000)
@@ -244,6 +249,9 @@ struct qedr_cq {
u16 icid;
+ /* Lock to protect completion handler */
+ spinlock_t comp_handler_lock;
+
/* Lock to protect multiplem CQ's */
spinlock_t cq_lock;
u8 arm_flags;
@@ -290,6 +298,7 @@ struct qedr_qp_hwq_info {
u16 prod;
u16 cons;
u16 wqe_cons;
+ u16 gsi_cons;
u16 max_wr;
/* DB */
@@ -364,6 +373,7 @@ struct qedr_qp {
struct ib_sge sg_list[RDMA_MAX_SGE_PER_RQ_WQE];
u8 wqe_size;
+ u8 smac[ETH_ALEN];
u16 vlan_id;
int rc;
} *rqe_wr_id;
@@ -472,6 +482,11 @@ static inline struct qedr_qp *get_qedr_qp(struct ib_qp *ibqp)
return container_of(ibqp, struct qedr_qp, ibqp);
}
+static inline struct qedr_ah *get_qedr_ah(struct ib_ah *ibah)
+{
+ return container_of(ibah, struct qedr_ah, ibah);
+}
+
static inline struct qedr_mr *get_qedr_mr(struct ib_mr *ibmr)
{
return container_of(ibmr, struct qedr_mr, ibmr);
diff --git a/drivers/infiniband/hw/qedr/qedr_cm.c b/drivers/infiniband/hw/qedr/qedr_cm.c
new file mode 100644
index 0000000..8ba6240
--- /dev/null
+++ b/drivers/infiniband/hw/qedr/qedr_cm.c
@@ -0,0 +1,626 @@
+/* QLogic qed NIC Driver
+ * Copyright (c) 2015 QLogic Corporation
+ *
+ * This software is available under the terms of the GNU General Public License
+ * (GPL) Version 2, available from the file COPYING in the main directory of
+ * this source tree.
+ */
+#include <linux/dma-mapping.h>
+#include <linux/crc32.h>
+#include <linux/iommu.h>
+#include <net/ip.h>
+#include <net/ipv6.h>
+#include <net/udp.h>
+
+#include <rdma/ib_verbs.h>
+#include <rdma/ib_user_verbs.h>
+#include <rdma/iw_cm.h>
+#include <rdma/ib_umem.h>
+#include <rdma/ib_addr.h>
+#include <rdma/ib_cache.h>
+
+#include "qedr_hsi.h"
+#include <linux/qed/qed_if.h>
+#include <linux/qed/qed_roce_if.h>
+#include "qedr.h"
+#include "qedr_hsi.h"
+#include "verbs.h"
+#include <uapi/rdma/providers/qedr-abi.h>
+#include "qedr_hsi.h"
+#include "qedr_cm.h"
+
+void qedr_inc_sw_gsi_cons(struct qedr_qp_hwq_info *info)
+{
+ info->gsi_cons = (info->gsi_cons + 1) % info->max_wr;
+}
+
+void qedr_store_gsi_qp_cq(struct qedr_dev *dev, struct qedr_qp *qp,
+ struct ib_qp_init_attr *attrs)
+{
+ dev->gsi_qp_created = 1;
+ dev->gsi_sqcq = get_qedr_cq(attrs->send_cq);
+ dev->gsi_rqcq = get_qedr_cq(attrs->recv_cq);
+ dev->gsi_qp = qp;
+}
+
+void qedr_ll2_tx_cb(void *_qdev, struct qed_roce_ll2_packet *pkt)
+{
+ struct qedr_dev *dev = (struct qedr_dev *)_qdev;
+ struct qedr_cq *cq = dev->gsi_sqcq;
+ struct qedr_qp *qp = dev->gsi_qp;
+ unsigned long flags;
+
+ DP_VERBOSE(dev, QEDR_MSG_GSI,
+ "LL2 TX CB: gsi_sqcq=%p, gsi_rqcq=%p, gsi_cons=%d, ibcq_comp=%s\n",
+ dev->gsi_sqcq, dev->gsi_rqcq, qp->sq.gsi_cons,
+ cq->ibcq.comp_handler ? "Yes" : "No");
+
+ dma_free_coherent(&dev->pdev->dev, pkt->header.len, pkt->header.vaddr,
+ pkt->header.baddr);
+ kfree(pkt);
+
+ spin_lock_irqsave(&qp->q_lock, flags);
+ qedr_inc_sw_gsi_cons(&qp->sq);
+ spin_unlock_irqrestore(&qp->q_lock, flags);
+
+ if (cq->ibcq.comp_handler) {
+ spin_lock_irqsave(&cq->comp_handler_lock, flags);
+ (*cq->ibcq.comp_handler) (&cq->ibcq, cq->ibcq.cq_context);
+ spin_unlock_irqrestore(&cq->comp_handler_lock, flags);
+ }
+}
+
+static inline void qedr_print_hex_dump(const char *str, void *buf, size_t size)
+{
+ print_hex_dump(KERN_INFO, str, DUMP_PREFIX_OFFSET, 32, 1, buf,
+ size, false);
+}
+
+void qedr_ll2_rx_cb(void *_dev, struct qed_roce_ll2_packet *pkt,
+ struct qed_roce_ll2_rx_params *params)
+{
+ struct qedr_dev *dev = (struct qedr_dev *)_dev;
+ struct qedr_cq *cq = dev->gsi_rqcq;
+ struct qedr_qp *qp = dev->gsi_qp;
+ unsigned long flags;
+
+ /* CM packet dump - supported unless IOMMU is enabled */
+ if (unlikely((dev->dp_level <= QED_LEVEL_VERBOSE) &&
+ (dev->dp_module & QEDR_MSG_GSI) &&
+ (!iommu_present(&pci_bus_type))))
+ qedr_print_hex_dump("RX CM ",
+ phys_to_virt(pkt->payload[0].baddr),
+ pkt->payload[0].len + IB_GRH_BYTES);
+
+ spin_lock_irqsave(&qp->q_lock, flags);
+
+ qp->rqe_wr_id[qp->rq.gsi_cons].rc = params->rc;
+ qp->rqe_wr_id[qp->rq.gsi_cons].vlan_id = params->vlan_id;
+ qp->rqe_wr_id[qp->rq.gsi_cons].sg_list[0].length = pkt->payload[0].len;
+ ether_addr_copy(qp->rqe_wr_id[qp->rq.gsi_cons].smac, params->smac);
+
+ qedr_inc_sw_gsi_cons(&qp->rq);
+
+ spin_unlock_irqrestore(&qp->q_lock, flags);
+
+ if (cq->ibcq.comp_handler) {
+ spin_lock_irqsave(&cq->comp_handler_lock, flags);
+ (*cq->ibcq.comp_handler) (&cq->ibcq, cq->ibcq.cq_context);
+ spin_unlock_irqrestore(&cq->comp_handler_lock, flags);
+ }
+}
+
+static void qedr_destroy_gsi_cq(struct qedr_dev *dev,
+ struct ib_qp_init_attr *attrs)
+{
+ struct qed_rdma_destroy_cq_in_params iparams;
+ struct qed_rdma_destroy_cq_out_params oparams;
+ struct qedr_cq *cq;
+
+ cq = get_qedr_cq(attrs->send_cq);
+ iparams.icid = cq->icid;
+ dev->ops->rdma_destroy_cq(dev->rdma_ctx, &iparams, &oparams);
+ dev->ops->common->chain_free(dev->cdev, &cq->pbl);
+
+ cq = get_qedr_cq(attrs->recv_cq);
+ /* if a dedicated recv_cq was used, delete it too */
+ if (iparams.icid != cq->icid) {
+ iparams.icid = cq->icid;
+ dev->ops->rdma_destroy_cq(dev->rdma_ctx, &iparams, &oparams);
+ dev->ops->common->chain_free(dev->cdev, &cq->pbl);
+ }
+}
+
+static inline int qedr_check_gsi_qp_attrs(struct qedr_dev *dev,
+ struct ib_qp_init_attr *attrs)
+{
+ if (attrs->cap.max_recv_sge > QEDR_GSI_MAX_RECV_SGE) {
+ DP_ERR(dev,
+ " create gsi qp: failed. max_recv_sge is larger the max %d>%d\n",
+ attrs->cap.max_recv_sge, QEDR_GSI_MAX_RECV_SGE);
+ return -EINVAL;
+ }
+
+ if (attrs->cap.max_recv_wr > QEDR_GSI_MAX_RECV_WR) {
+ DP_ERR(dev,
+ " create gsi qp: failed. max_recv_wr is too large %d>%d\n",
+ attrs->cap.max_recv_wr, QEDR_GSI_MAX_RECV_WR);
+ return -EINVAL;
+ }
+
+ if (attrs->cap.max_send_wr > QEDR_GSI_MAX_SEND_WR) {
+ DP_ERR(dev,
+ " create gsi qp: failed. max_send_wr is too large %d>%d\n",
+ attrs->cap.max_send_wr, QEDR_GSI_MAX_SEND_WR);
+ return -EINVAL;
+ }
+
+ return 0;
+}
+
+struct ib_qp *qedr_create_gsi_qp(struct qedr_dev *dev,
+ struct ib_qp_init_attr *attrs,
+ struct qedr_qp *qp)
+{
+ struct qed_roce_ll2_params ll2_params;
+ int rc;
+
+ rc = qedr_check_gsi_qp_attrs(dev, attrs);
+ if (rc)
+ return ERR_PTR(rc);
+
+ /* configure and start LL2 */
+ memset(&ll2_params, 0, sizeof(ll2_params));
+ ll2_params.max_tx_buffers = attrs->cap.max_send_wr;
+ ll2_params.max_rx_buffers = attrs->cap.max_recv_wr;
+ ll2_params.cbs.tx_cb = qedr_ll2_tx_cb;
+ ll2_params.cbs.rx_cb = qedr_ll2_rx_cb;
+ ll2_params.cb_cookie = (void *)dev;
+ ll2_params.mtu = dev->ndev->mtu;
+ ether_addr_copy(ll2_params.mac_address, dev->ndev->dev_addr);
+ rc = dev->ops->roce_ll2_start(dev->cdev, &ll2_params);
+ if (rc) {
+ DP_ERR(dev, "create gsi qp: failed on ll2 start. rc=%d\n", rc);
+ return ERR_PTR(rc);
+ }
+
+ /* create QP */
+ qp->ibqp.qp_num = 1;
+ qp->rq.max_wr = attrs->cap.max_recv_wr;
+ qp->sq.max_wr = attrs->cap.max_send_wr;
+
+ qp->rqe_wr_id = kcalloc(qp->rq.max_wr, sizeof(*qp->rqe_wr_id),
+ GFP_KERNEL);
+ if (!qp->rqe_wr_id) {
+ DP_ERR(dev, "create gsi qp: failed on rqe_wr_id allocation\n");
+ goto err;
+ }
+ qp->wqe_wr_id = kcalloc(qp->sq.max_wr, sizeof(*qp->wqe_wr_id),
+ GFP_KERNEL);
+ if (!qp->wqe_wr_id) {
+ DP_ERR(dev, "create gsi qp: failed on wqe_wr_id allocation\n");
+ goto err;
+ }
+
+ qedr_store_gsi_qp_cq(dev, qp, attrs);
+ ether_addr_copy(dev->gsi_ll2_mac_address, dev->ndev->dev_addr);
+
+ /* the GSI CQ is handled by the driver so remove it from the FW */
+ qedr_destroy_gsi_cq(dev, attrs);
+ dev->gsi_rqcq->cq_type = QEDR_CQ_TYPE_GSI;
+ dev->gsi_rqcq->cq_type = QEDR_CQ_TYPE_GSI;
+
+ DP_VERBOSE(dev, QEDR_MSG_GSI, "created GSI QP %p\n", qp);
+
+ return &qp->ibqp;
+
+err:
+ kfree(qp->rqe_wr_id);
+
+ rc = dev->ops->roce_ll2_stop(dev->cdev);
+ if (rc)
+ DP_ERR(dev, "create gsi qp: failed destroy on create\n");
+
+ return ERR_PTR(-ENOMEM);
+}
+
+int qedr_destroy_gsi_qp(struct qedr_dev *dev)
+{
+ int rc;
+
+ rc = dev->ops->roce_ll2_stop(dev->cdev);
+ if (rc)
+ DP_ERR(dev, "destroy gsi qp: failed (rc=%d)\n", rc);
+ else
+ DP_VERBOSE(dev, (QEDR_MSG_QP | QEDR_MSG_GSI),
+ "destroy gsi qp: success\n");
+
+ return rc;
+}
+
+#define QEDR_MAX_UD_HEADER_SIZE (100)
+#define QEDR_GSI_QPN (1)
+static inline int qedr_gsi_build_header(struct qedr_dev *dev,
+ struct qedr_qp *qp,
+ struct ib_send_wr *swr,
+ struct ib_ud_header *udh,
+ int *roce_mode)
+{
+ bool has_vlan = false, has_grh_ipv6 = true;
+ struct ib_ah_attr *ah_attr = &get_qedr_ah(ud_wr(swr)->ah)->attr;
+ struct ib_global_route *grh = &ah_attr->grh;
+ union ib_gid sgid;
+ int send_size = 0;
+ u16 vlan_id = 0;
+ u16 ether_type;
+ struct ib_gid_attr sgid_attr;
+ int rc;
+ int ip_ver = 0;
+
+ bool has_udp = false;
+ int i;
+
+ send_size = 0;
+ for (i = 0; i < swr->num_sge; ++i)
+ send_size += swr->sg_list[i].length;
+
+ rc = ib_get_cached_gid(qp->ibqp.device, ah_attr->port_num,
+ grh->sgid_index, &sgid, &sgid_attr);
+ if (rc) {
+ DP_ERR(dev,
+ "gsi post send: failed to get cached GID (port=%d, ix=%d)\n",
+ ah_attr->port_num, grh->sgid_index);
+ return rc;
+ }
+
+ vlan_id = rdma_vlan_dev_vlan_id(sgid_attr.ndev);
+ if (vlan_id < VLAN_CFI_MASK)
+ has_vlan = true;
+ if (sgid_attr.ndev)
+ dev_put(sgid_attr.ndev);
+
+ if (!memcmp(&sgid, &zgid, sizeof(sgid))) {
+ DP_ERR(dev, "gsi post send: GID not found GID index %d\n",
+ ah_attr->grh.sgid_index);
+ return -ENOENT;
+ }
+
+ has_udp = (sgid_attr.gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP);
+ if (!has_udp) {
+ /* RoCE v1 */
+ ether_type = ETH_P_ROCE;
+ *roce_mode = ROCE_V1;
+ } else if (ipv6_addr_v4mapped((struct in6_addr *)&sgid)) {
+ /* RoCE v2 IPv4 */
+ ip_ver = 4;
+ ether_type = ETH_P_IP;
+ has_grh_ipv6 = false;
+ *roce_mode = ROCE_V2_IPV4;
+ } else {
+ /* RoCE v2 IPv6 */
+ ip_ver = 6;
+ ether_type = ETH_P_IPV6;
+ *roce_mode = ROCE_V2_IPV6;
+ }
+
+ rc = ib_ud_header_init(send_size, false, true, has_vlan,
+ has_grh_ipv6, ip_ver, has_udp, 0, udh);
+ if (rc) {
+ DP_ERR(dev, "gsi post send: failed to init header\n");
+ return rc;
+ }
+
+ /* ENET + VLAN headers */
+ ether_addr_copy(udh->eth.dmac_h, ah_attr->dmac);
+ ether_addr_copy(udh->eth.smac_h, dev->ndev->dev_addr);
+ if (has_vlan) {
+ udh->eth.type = htons(ETH_P_8021Q);
+ udh->vlan.tag = htons(vlan_id);
+ udh->vlan.type = htons(ether_type);
+ } else {
+ udh->eth.type = htons(ether_type);
+ }
+
+ /* BTH */
+ udh->bth.solicited_event = !!(swr->send_flags & IB_SEND_SOLICITED);
+ udh->bth.pkey = QEDR_ROCE_PKEY_DEFAULT;
+ udh->bth.destination_qpn = htonl(ud_wr(swr)->remote_qpn);
+ udh->bth.psn = htonl((qp->sq_psn++) & ((1 << 24) - 1));
+ udh->bth.opcode = IB_OPCODE_UD_SEND_ONLY;
+
+ /* DETH */
+ udh->deth.qkey = htonl(0x80010000);
+ udh->deth.source_qpn = htonl(QEDR_GSI_QPN);
+
+ if (has_grh_ipv6) {
+ /* GRH / IPv6 header */
+ udh->grh.traffic_class = grh->traffic_class;
+ udh->grh.flow_label = grh->flow_label;
+ udh->grh.hop_limit = grh->hop_limit;
+ udh->grh.destination_gid = grh->dgid;
+ memcpy(&udh->grh.source_gid.raw, &sgid.raw,
+ sizeof(udh->grh.source_gid.raw));
+ } else {
+ /* IPv4 header */
+ u32 ipv4_addr;
+
+ udh->ip4.protocol = IPPROTO_UDP;
+ udh->ip4.tos = htonl(ah_attr->grh.flow_label);
+ udh->ip4.frag_off = htons(IP_DF);
+ udh->ip4.ttl = ah_attr->grh.hop_limit;
+
+ ipv4_addr = qedr_get_ipv4_from_gid(sgid.raw);
+ udh->ip4.saddr = ipv4_addr;
+ ipv4_addr = qedr_get_ipv4_from_gid(ah_attr->grh.dgid.raw);
+ udh->ip4.daddr = ipv4_addr;
+ /* note: checksum is calculated by the device */
+ }
+
+ /* UDP */
+ if (has_udp) {
+ udh->udp.sport = htons(QEDR_ROCE_V2_UDP_SPORT);
+ udh->udp.dport = htons(ROCE_V2_UDP_DPORT);
+ udh->udp.csum = 0;
+ /* UDP length is untouched hence is zero */
+ }
+ return 0;
+}
+
+static inline int qedr_gsi_build_packet(struct qedr_dev *dev,
+ struct qedr_qp *qp,
+ struct ib_send_wr *swr,
+ struct qed_roce_ll2_packet **p_packet)
+{
+ u8 ud_header_buffer[QEDR_MAX_UD_HEADER_SIZE];
+ struct qed_roce_ll2_packet *packet;
+ struct pci_dev *pdev = dev->pdev;
+ int roce_mode, header_size;
+ struct ib_ud_header udh;
+ int i, rc;
+
+ *p_packet = NULL;
+
+ rc = qedr_gsi_build_header(dev, qp, swr, &udh, &roce_mode);
+ if (rc)
+ return rc;
+
+ header_size = ib_ud_header_pack(&udh, &ud_header_buffer);
+
+ packet = kzalloc(sizeof(*packet), GFP_ATOMIC);
+ if (!packet) {
+ DP_ERR(dev, "gsi post send: failed to allocate packet\n");
+ return -ENOMEM;
+ }
+
+ packet->header.vaddr = dma_alloc_coherent(&pdev->dev, header_size,
+ &packet->header.baddr,
+ GFP_ATOMIC);
+ if (!packet->header.vaddr) {
+ DP_ERR(dev,
+ "cm header build: failed to allocate buffer for header\n");
+ kfree(packet);
+ return -ENOMEM;
+ }
+
+ if (ether_addr_equal(udh.eth.smac_h, udh.eth.dmac_h))
+ packet->tx_dest = QED_ROCE_LL2_TX_DEST_NW;
+ else
+ packet->tx_dest = QED_ROCE_LL2_TX_DEST_LB;
+
+ packet->roce_mode = roce_mode;
+ memcpy(packet->header.vaddr, ud_header_buffer, header_size);
+ packet->header.len = header_size;
+ packet->n_seg = swr->num_sge;
+ for (i = 0; i < packet->n_seg; i++) {
+ packet->payload[i].baddr = swr->sg_list[i].addr;
+ packet->payload[i].len = swr->sg_list[i].length;
+ }
+
+ *p_packet = packet;
+
+ return 0;
+}
+
+int qedr_gsi_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
+ struct ib_send_wr **bad_wr)
+{
+ struct qed_roce_ll2_packet *pkt = NULL;
+ struct qedr_qp *qp = get_qedr_qp(ibqp);
+ struct qed_roce_ll2_tx_params params;
+ struct qedr_dev *dev = qp->dev;
+ unsigned long flags;
+ int rc;
+
+ if (qp->state != QED_ROCE_QP_STATE_RTS) {
+ *bad_wr = wr;
+ DP_ERR(dev,
+ "gsi post recv: failed to post rx buffer. state is %d and not QED_ROCE_QP_STATE_RTS\n",
+ qp->state);
+ return -EINVAL;
+ }
+
+ if (wr->num_sge > RDMA_MAX_SGE_PER_SQ_WQE) {
+ DP_ERR(dev, "gsi post send: num_sge is too large (%d>%d)\n",
+ wr->num_sge, RDMA_MAX_SGE_PER_SQ_WQE);
+ rc = -EINVAL;
+ goto err;
+ }
+
+ if (wr->opcode != IB_WR_SEND) {
+ DP_ERR(dev,
+ "gsi post send: failed due to unsupported opcode %d\n",
+ wr->opcode);
+ rc = -EINVAL;
+ goto err;
+ }
+
+ memset(¶ms, 0, sizeof(params));
+
+ spin_lock_irqsave(&qp->q_lock, flags);
+
+ rc = qedr_gsi_build_packet(dev, qp, wr, &pkt);
+ if (rc) {
+ spin_unlock_irqrestore(&qp->q_lock, flags);
+ goto err;
+ }
+
+ if (unlikely((dev->dp_level <= QED_LEVEL_VERBOSE) &&
+ (dev->dp_module & QEDR_MSG_GSI)))
+ qedr_print_hex_dump("TX CM Header ", pkt->header.vaddr,
+ pkt->header.len);
+
+ rc = dev->ops->roce_ll2_tx(dev->cdev, pkt, ¶ms);
+ if (!rc) {
+ qp->wqe_wr_id[qp->sq.prod].wr_id = wr->wr_id;
+ qedr_inc_sw_prod(&qp->sq);
+ DP_VERBOSE(qp->dev, QEDR_MSG_GSI,
+ "gsi post send: opcode=%d, in_irq=%ld, irqs_disabled=%d, wr_id=%llx\n",
+ wr->opcode, in_irq(), irqs_disabled(), wr->wr_id);
+ } else {
+ if (rc == QED_ROCE_TX_HEAD_FAILURE) {
+ /* TX failed while posting header - release resources */
+ dma_free_coherent(&dev->pdev->dev, pkt->header.len,
+ pkt->header.vaddr, pkt->header.baddr);
+ kfree(pkt);
+ } else if (rc == QED_ROCE_TX_FRAG_FAILURE) {
+ /* NTD since TX failed while posting a fragment. We will
+ * release the resources on TX callback
+ */
+ }
+
+ DP_ERR(dev, "gsi post send: failed to transmit (rc=%d)\n", rc);
+ rc = -EAGAIN;
+ *bad_wr = wr;
+ }
+
+ spin_unlock_irqrestore(&qp->q_lock, flags);
+
+ if (wr->next) {
+ DP_ERR(dev,
+ "gsi post send: failed second WR. Only one WR may be passed at a time\n");
+ *bad_wr = wr->next;
+ rc = -EINVAL;
+ }
+
+ return rc;
+
+err:
+ *bad_wr = wr;
+ return rc;
+}
+
+int qedr_gsi_post_recv(struct ib_qp *ibqp, struct ib_recv_wr *wr,
+ struct ib_recv_wr **bad_wr)
+{
+ struct qedr_dev *dev = get_qedr_dev(ibqp->device);
+ struct qedr_qp *qp = get_qedr_qp(ibqp);
+ struct qed_roce_ll2_buffer buf;
+ unsigned long flags;
+ int status = 0;
+ int rc;
+
+ if ((qp->state != QED_ROCE_QP_STATE_RTR) &&
+ (qp->state != QED_ROCE_QP_STATE_RTS)) {
+ *bad_wr = wr;
+ DP_ERR(dev,
+ "gsi post recv: failed to post rx buffer. state is %d and not QED_ROCE_QP_STATE_RTR/S\n",
+ qp->state);
+ return -EINVAL;
+ }
+
+ memset(&buf, 0, sizeof(buf));
+
+ spin_lock_irqsave(&qp->q_lock, flags);
+
+ while (wr) {
+ if (wr->num_sge > QEDR_GSI_MAX_RECV_SGE) {
+ DP_ERR(dev,
+ "gsi post recv: failed to post rx buffer. too many sges %d>%d\n",
+ wr->num_sge, QEDR_GSI_MAX_RECV_SGE);
+ goto err;
+ }
+
+ buf.baddr = wr->sg_list[0].addr;
+ buf.len = wr->sg_list[0].length;
+
+ rc = dev->ops->roce_ll2_post_rx_buffer(dev->cdev, &buf, 0, 1);
+ if (rc) {
+ DP_ERR(dev,
+ "gsi post recv: failed to post rx buffer (rc=%d)\n",
+ rc);
+ goto err;
+ }
+
+ memset(&qp->rqe_wr_id[qp->rq.prod], 0,
+ sizeof(qp->rqe_wr_id[qp->rq.prod]));
+ qp->rqe_wr_id[qp->rq.prod].sg_list[0] = wr->sg_list[0];
+ qp->rqe_wr_id[qp->rq.prod].wr_id = wr->wr_id;
+
+ qedr_inc_sw_prod(&qp->rq);
+
+ wr = wr->next;
+ }
+
+ spin_unlock_irqrestore(&qp->q_lock, flags);
+
+ return status;
+err:
+ spin_unlock_irqrestore(&qp->q_lock, flags);
+ *bad_wr = wr;
+ return -ENOMEM;
+}
+
+int qedr_gsi_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc)
+{
+ struct qedr_dev *dev = get_qedr_dev(ibcq->device);
+ struct qedr_cq *cq = get_qedr_cq(ibcq);
+ struct qedr_qp *qp = dev->gsi_qp;
+ unsigned long flags;
+ int i = 0;
+
+ spin_lock_irqsave(&cq->cq_lock, flags);
+
+ while (i < num_entries && qp->rq.cons != qp->rq.gsi_cons) {
+ memset(&wc[i], 0, sizeof(*wc));
+
+ wc[i].qp = &qp->ibqp;
+ wc[i].wr_id = qp->rqe_wr_id[qp->rq.cons].wr_id;
+ wc[i].opcode = IB_WC_RECV;
+ wc[i].pkey_index = 0;
+ wc[i].status = (qp->rqe_wr_id[qp->rq.cons].rc) ?
+ IB_WC_GENERAL_ERR : IB_WC_SUCCESS;
+ /* 0 - currently only one recv sg is supported */
+ wc[i].byte_len = qp->rqe_wr_id[qp->rq.cons].sg_list[0].length;
+ wc[i].wc_flags |= IB_WC_GRH | IB_WC_IP_CSUM_OK;
+ ether_addr_copy(wc[i].smac, qp->rqe_wr_id[qp->rq.cons].smac);
+ wc[i].wc_flags |= IB_WC_WITH_SMAC;
+ if (qp->rqe_wr_id[qp->rq.cons].vlan_id) {
+ wc[i].wc_flags |= IB_WC_WITH_VLAN;
+ wc[i].vlan_id = qp->rqe_wr_id[qp->rq.cons].vlan_id;
+ }
+
+ qedr_inc_sw_cons(&qp->rq);
+ i++;
+ }
+
+ while (i < num_entries && qp->sq.cons != qp->sq.gsi_cons) {
+ memset(&wc[i], 0, sizeof(*wc));
+
+ wc[i].qp = &qp->ibqp;
+ wc[i].wr_id = qp->wqe_wr_id[qp->sq.cons].wr_id;
+ wc[i].opcode = IB_WC_SEND;
+ wc[i].status = IB_WC_SUCCESS;
+
+ qedr_inc_sw_cons(&qp->sq);
+ i++;
+ }
+
+ spin_unlock_irqrestore(&cq->cq_lock, flags);
+
+ DP_VERBOSE(dev, (QEDR_MSG_CQ | QEDR_MSG_GSI),
+ "gsi poll_cq: requested entries=%d, actual=%d, qp->rq.cons=%d, qp->rq.gsi_cons=%x, qp->sq.cons=%d, qp->sq.gsi_cons=%d, qp_num=%d\n",
+ num_entries, i, qp->rq.cons, qp->rq.gsi_cons, qp->sq.cons,
+ qp->sq.gsi_cons, qp->ibqp.qp_num);
+
+ return i;
+}
diff --git a/drivers/infiniband/hw/qedr/qedr_cm.h b/drivers/infiniband/hw/qedr/qedr_cm.h
index b8a8b76..9ba6e15 100644
--- a/drivers/infiniband/hw/qedr/qedr_cm.h
+++ b/drivers/infiniband/hw/qedr/qedr_cm.h
@@ -32,9 +32,30 @@
#ifndef LINUX_QEDR_CM_H_
#define LINUX_QEDR_CM_H_
+#define QEDR_GSI_MAX_RECV_WR (4096)
+#define QEDR_GSI_MAX_SEND_WR (4096)
+
+#define QEDR_GSI_MAX_RECV_SGE (1) /* LL2 FW limitation */
+
+#define ETH_P_ROCE (0x8915)
+#define QEDR_ROCE_V2_UDP_SPORT (0000)
+
static inline u32 qedr_get_ipv4_from_gid(u8 *gid)
{
return *(u32 *)(void *)&gid[12];
}
+/* RDMA CM */
+int qedr_gsi_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc);
+int qedr_gsi_post_recv(struct ib_qp *ibqp, struct ib_recv_wr *wr,
+ struct ib_recv_wr **bad_wr);
+int qedr_gsi_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
+ struct ib_send_wr **bad_wr);
+struct ib_qp *qedr_create_gsi_qp(struct qedr_dev *dev,
+ struct ib_qp_init_attr *attrs,
+ struct qedr_qp *qp);
+void qedr_store_gsi_qp_cq(struct qedr_dev *dev,
+ struct qedr_qp *qp, struct ib_qp_init_attr *attrs);
+int qedr_destroy_gsi_qp(struct qedr_dev *dev);
+void qedr_inc_sw_gsi_cons(struct qedr_qp_hwq_info *info);
#endif
diff --git a/drivers/infiniband/hw/qedr/verbs.c b/drivers/infiniband/hw/qedr/verbs.c
index 30b71e3..5c5fdd1 100644
--- a/drivers/infiniband/hw/qedr/verbs.c
+++ b/drivers/infiniband/hw/qedr/verbs.c
@@ -1472,6 +1472,15 @@ struct ib_qp *qedr_create_qp(struct ib_pd *ibpd,
qedr_set_qp_init_params(dev, qp, pd, attrs);
+ if (attrs->qp_type == IB_QPT_GSI) {
+ if (udata) {
+ DP_ERR(dev,
+ "create qp: unexpected udata when creating GSI QP\n");
+ goto err0;
+ }
+ return qedr_create_gsi_qp(dev, attrs, qp);
+ }
+
memset(&in_params, 0, sizeof(in_params));
if (udata) {
@@ -2037,6 +2046,8 @@ int qedr_destroy_qp(struct ib_qp *ibqp)
rc = dev->ops->rdma_destroy_qp(dev->rdma_ctx, qp->qed_qp);
if (rc)
return rc;
+ } else {
+ qedr_destroy_gsi_qp(dev);
}
if (ibqp->uobject && ibqp->uobject->context) {
@@ -2052,6 +2063,23 @@ int qedr_destroy_qp(struct ib_qp *ibqp)
return rc;
}
+struct ib_ah *qedr_create_ah(struct ib_pd *ibpd, struct ib_ah_attr *attr)
+{
+ struct qedr_ah *ah = kzalloc(sizeof(*ah), GFP_ATOMIC);
+
+ ah->attr = *attr;
+
+ return &ah->ibah;
+}
+
+int qedr_destroy_ah(struct ib_ah *ibah)
+{
+ struct qedr_ah *ah = get_qedr_ah(ibah);
+
+ kfree(ah);
+ return 0;
+}
+
static void free_mr_info(struct qedr_dev *dev, struct mr_info *info)
{
struct qedr_pbl *pbl, *tmp;
@@ -2903,6 +2931,10 @@ int qedr_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
*bad_wr = NULL;
+ if (qp->qp_type == IB_QPT_GSI) {
+ return qedr_gsi_post_send(ibqp, wr, bad_wr);
+ }
+
spin_lock_irqsave(&qp->q_lock, flags);
if ((qp->state == QED_ROCE_QP_STATE_RESET) ||
@@ -2959,6 +2991,10 @@ int qedr_post_recv(struct ib_qp *ibqp, struct ib_recv_wr *wr,
unsigned long flags;
int status = 0;
+ if (qp->qp_type == IB_QPT_GSI) {
+ return qedr_gsi_post_recv(ibqp, wr, bad_wr);
+ }
+
spin_lock_irqsave(&qp->q_lock, flags);
if ((qp->state == QED_ROCE_QP_STATE_RESET) ||
@@ -3394,6 +3430,9 @@ int qedr_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc)
int update = 0;
int done = 0;
+ if (cq->cq_type == QEDR_CQ_TYPE_GSI)
+ return qedr_gsi_poll_cq(ibcq, num_entries, wc);
+
spin_lock_irqsave(&cq->cq_lock, flags);
old_cons = qed_chain_get_cons_idx_u32(&cq->pbl);
while (num_entries && is_valid_cqe(cq, cqe)) {
diff --git a/drivers/net/ethernet/qlogic/qed/qed_ll2.c b/drivers/net/ethernet/qlogic/qed/qed_ll2.c
index 4ea6a07..f907569 100644
--- a/drivers/net/ethernet/qlogic/qed/qed_ll2.c
+++ b/drivers/net/ethernet/qlogic/qed/qed_ll2.c
@@ -278,6 +278,7 @@ static void qed_ll2_txq_flush(struct qed_hwfn *p_hwfn, u8 connection_handle)
struct qed_ll2_tx_packet *p_pkt = NULL;
struct qed_ll2_info *p_ll2_conn;
struct qed_ll2_tx_queue *p_tx;
+ dma_addr_t tx_frag;
p_ll2_conn = qed_ll2_handle_sanity_inactive(p_hwfn, connection_handle);
if (!p_ll2_conn)
@@ -294,15 +295,25 @@ static void qed_ll2_txq_flush(struct qed_hwfn *p_hwfn, u8 connection_handle)
list_del(&p_pkt->list_entry);
b_last_packet = list_empty(&p_tx->active_descq);
list_add_tail(&p_pkt->list_entry, &p_tx->free_descq);
+
p_tx->cur_completing_packet = *p_pkt;
p_tx->cur_completing_bd_idx = 1;
b_last_frag = p_tx->cur_completing_bd_idx == p_pkt->bd_used;
-
- qed_ll2b_complete_tx_packet(p_hwfn, p_ll2_conn->my_id,
- p_pkt->cookie,
- p_pkt->bds_set[0].tx_frag,
- b_last_frag,
- b_last_packet);
+ tx_frag = p_pkt->bds_set[0].tx_frag;
+ if (p_ll2_conn->gsi_enable)
+ qed_ll2b_release_tx_gsi_packet(p_hwfn,
+ p_ll2_conn->my_id,
+ p_pkt->cookie,
+ tx_frag,
+ b_last_frag,
+ b_last_packet);
+ else
+ qed_ll2b_complete_tx_packet(p_hwfn,
+ p_ll2_conn->my_id,
+ p_pkt->cookie,
+ tx_frag,
+ b_last_frag,
+ b_last_packet);
}
}
@@ -314,6 +325,7 @@ static int qed_ll2_txq_completion(struct qed_hwfn *p_hwfn, void *p_cookie)
struct qed_ll2_tx_packet *p_pkt;
bool b_last_frag = false;
unsigned long flags;
+ dma_addr_t tx_frag;
int rc = -EINVAL;
spin_lock_irqsave(&p_tx->lock, flags);
@@ -354,11 +366,19 @@ static int qed_ll2_txq_completion(struct qed_hwfn *p_hwfn, void *p_cookie)
list_add_tail(&p_pkt->list_entry, &p_tx->free_descq);
spin_unlock_irqrestore(&p_tx->lock, flags);
- qed_ll2b_complete_tx_packet(p_hwfn,
- p_ll2_conn->my_id,
- p_pkt->cookie,
- p_pkt->bds_set[0].tx_frag,
- b_last_frag, !num_bds);
+ tx_frag = p_pkt->bds_set[0].tx_frag;
+ if (p_ll2_conn->gsi_enable)
+ qed_ll2b_complete_tx_gsi_packet(p_hwfn,
+ p_ll2_conn->my_id,
+ p_pkt->cookie,
+ tx_frag,
+ b_last_frag, !num_bds);
+ else
+ qed_ll2b_complete_tx_packet(p_hwfn,
+ p_ll2_conn->my_id,
+ p_pkt->cookie,
+ tx_frag,
+ b_last_frag, !num_bds);
spin_lock_irqsave(&p_tx->lock, flags);
}
@@ -369,6 +389,54 @@ out:
return rc;
}
+static int
+qed_ll2_rxq_completion_gsi(struct qed_hwfn *p_hwfn,
+ struct qed_ll2_info *p_ll2_info,
+ union core_rx_cqe_union *p_cqe,
+ unsigned long lock_flags, bool b_last_cqe)
+{
+ struct qed_ll2_rx_queue *p_rx = &p_ll2_info->rx_queue;
+ struct qed_ll2_rx_packet *p_pkt = NULL;
+ u16 packet_length, parse_flags, vlan;
+ u32 src_mac_addrhi;
+ u16 src_mac_addrlo;
+
+ if (!list_empty(&p_rx->active_descq))
+ p_pkt = list_first_entry(&p_rx->active_descq,
+ struct qed_ll2_rx_packet, list_entry);
+ if (!p_pkt) {
+ DP_NOTICE(p_hwfn,
+ "GSI Rx completion but active_descq is empty\n");
+ return -EIO;
+ }
+
+ list_del(&p_pkt->list_entry);
+ parse_flags = le16_to_cpu(p_cqe->rx_cqe_gsi.parse_flags.flags);
+ packet_length = le16_to_cpu(p_cqe->rx_cqe_gsi.data_length);
+ vlan = le16_to_cpu(p_cqe->rx_cqe_gsi.vlan);
+ src_mac_addrhi = le32_to_cpu(p_cqe->rx_cqe_gsi.src_mac_addrhi);
+ src_mac_addrlo = le16_to_cpu(p_cqe->rx_cqe_gsi.src_mac_addrlo);
+ if (qed_chain_consume(&p_rx->rxq_chain) != p_pkt->rxq_bd)
+ DP_NOTICE(p_hwfn,
+ "Mismatch between active_descq and the LL2 Rx chain\n");
+ list_add_tail(&p_pkt->list_entry, &p_rx->free_descq);
+
+ spin_unlock_irqrestore(&p_rx->lock, lock_flags);
+ qed_ll2b_complete_rx_gsi_packet(p_hwfn,
+ p_ll2_info->my_id,
+ p_pkt->cookie,
+ p_pkt->rx_buf_addr,
+ packet_length,
+ p_cqe->rx_cqe_gsi.data_length_error,
+ parse_flags,
+ vlan,
+ src_mac_addrhi,
+ src_mac_addrlo, b_last_cqe);
+ spin_lock_irqsave(&p_rx->lock, lock_flags);
+
+ return 0;
+}
+
static int qed_ll2_rxq_completion_reg(struct qed_hwfn *p_hwfn,
struct qed_ll2_info *p_ll2_conn,
union core_rx_cqe_union *p_cqe,
@@ -430,6 +498,10 @@ static int qed_ll2_rxq_completion(struct qed_hwfn *p_hwfn, void *cookie)
DP_NOTICE(p_hwfn, "LL2 - unexpected Rx CQE slowpath\n");
rc = -EINVAL;
break;
+ case CORE_RX_CQE_TYPE_GSI_OFFLOAD:
+ rc = qed_ll2_rxq_completion_gsi(p_hwfn, p_ll2_conn,
+ cqe, flags, b_last_cqe);
+ break;
case CORE_RX_CQE_TYPE_REGULAR:
rc = qed_ll2_rxq_completion_reg(p_hwfn, p_ll2_conn,
cqe, flags, b_last_cqe);
@@ -527,6 +599,7 @@ static int qed_sp_ll2_rx_queue_start(struct qed_hwfn *p_hwfn,
}
p_ramrod->action_on_error.error_type = action_on_error;
+ p_ramrod->gsi_offload_flag = p_ll2_conn->gsi_enable;
return qed_spq_post(p_hwfn, p_ent, NULL);
}
@@ -589,6 +662,7 @@ static int qed_sp_ll2_tx_queue_start(struct qed_hwfn *p_hwfn,
DP_NOTICE(p_hwfn, "Unknown connection type: %d\n", conn_type);
}
+ p_ramrod->gsi_offload_flag = p_ll2_conn->gsi_enable;
return qed_spq_post(p_hwfn, p_ent, NULL);
}
@@ -775,6 +849,7 @@ int qed_ll2_acquire_connection(struct qed_hwfn *p_hwfn,
p_ll2_info->tx_dest = p_params->tx_dest;
p_ll2_info->ai_err_packet_too_big = p_params->ai_err_packet_too_big;
p_ll2_info->ai_err_no_buf = p_params->ai_err_no_buf;
+ p_ll2_info->gsi_enable = p_params->gsi_enable;
rc = qed_ll2_acquire_connection_rx(p_hwfn, p_ll2_info, rx_num_desc);
if (rc)
diff --git a/drivers/net/ethernet/qlogic/qed/qed_ll2.h b/drivers/net/ethernet/qlogic/qed/qed_ll2.h
index b3783fd..80a5dc2 100644
--- a/drivers/net/ethernet/qlogic/qed/qed_ll2.h
+++ b/drivers/net/ethernet/qlogic/qed/qed_ll2.h
@@ -125,6 +125,7 @@ struct qed_ll2_info {
u8 tx_stats_en;
struct qed_ll2_rx_queue rx_queue;
struct qed_ll2_tx_queue tx_queue;
+ u8 gsi_enable;
};
/**
@@ -292,5 +293,24 @@ void qed_ll2_setup(struct qed_hwfn *p_hwfn,
*/
void qed_ll2_free(struct qed_hwfn *p_hwfn,
struct qed_ll2_info *p_ll2_connections);
-
+void qed_ll2b_complete_rx_gsi_packet(struct qed_hwfn *p_hwfn,
+ u8 connection_handle,
+ void *cookie,
+ dma_addr_t rx_buf_addr,
+ u16 data_length,
+ u8 data_length_error,
+ u16 parse_flags,
+ u16 vlan,
+ u32 src_mac_addr_hi,
+ u16 src_mac_addr_lo, bool b_last_packet);
+void qed_ll2b_complete_tx_gsi_packet(struct qed_hwfn *p_hwfn,
+ u8 connection_handle,
+ void *cookie,
+ dma_addr_t first_frag_addr,
+ bool b_last_fragment, bool b_last_packet);
+void qed_ll2b_release_tx_gsi_packet(struct qed_hwfn *p_hwfn,
+ u8 connection_handle,
+ void *cookie,
+ dma_addr_t first_frag_addr,
+ bool b_last_fragment, bool b_last_packet);
#endif
diff --git a/drivers/net/ethernet/qlogic/qed/qed_roce.c b/drivers/net/ethernet/qlogic/qed/qed_roce.c
index b4f851b..83d841d 100644
--- a/drivers/net/ethernet/qlogic/qed/qed_roce.c
+++ b/drivers/net/ethernet/qlogic/qed/qed_roce.c
@@ -2633,6 +2633,71 @@ void qed_rdma_remove_user(void *rdma_cxt, u16 dpi)
spin_unlock_bh(&p_hwfn->p_rdma_info->lock);
}
+void qed_ll2b_complete_tx_gsi_packet(struct qed_hwfn *p_hwfn,
+ u8 connection_handle,
+ void *cookie,
+ dma_addr_t first_frag_addr,
+ bool b_last_fragment, bool b_last_packet)
+{
+ struct qed_roce_ll2_packet *packet = cookie;
+ struct qed_roce_ll2_info *roce_ll2 = p_hwfn->ll2;
+
+ roce_ll2->cbs.tx_cb(roce_ll2->cb_cookie, packet);
+}
+
+void qed_ll2b_release_tx_gsi_packet(struct qed_hwfn *p_hwfn,
+ u8 connection_handle,
+ void *cookie,
+ dma_addr_t first_frag_addr,
+ bool b_last_fragment, bool b_last_packet)
+{
+ qed_ll2b_complete_tx_gsi_packet(p_hwfn, connection_handle,
+ cookie, first_frag_addr,
+ b_last_fragment, b_last_packet);
+}
+
+void qed_ll2b_complete_rx_gsi_packet(struct qed_hwfn *p_hwfn,
+ u8 connection_handle,
+ void *cookie,
+ dma_addr_t rx_buf_addr,
+ u16 data_length,
+ u8 data_length_error,
+ u16 parse_flags,
+ u16 vlan,
+ u32 src_mac_addr_hi,
+ u16 src_mac_addr_lo, bool b_last_packet)
+{
+ struct qed_roce_ll2_info *roce_ll2 = p_hwfn->ll2;
+ struct qed_roce_ll2_rx_params params;
+ struct qed_dev *cdev = p_hwfn->cdev;
+ struct qed_roce_ll2_packet pkt;
+
+ DP_VERBOSE(cdev,
+ QED_MSG_LL2,
+ "roce ll2 rx complete: bus_addr=%p, len=%d, data_len_err=%d\n",
+ (void *)(uintptr_t)rx_buf_addr,
+ data_length, data_length_error);
+
+ memset(&pkt, 0, sizeof(pkt));
+ pkt.n_seg = 1;
+ pkt.payload[0].baddr = rx_buf_addr;
+ pkt.payload[0].len = data_length;
+
+ memset(¶ms, 0, sizeof(params));
+ params.vlan_id = vlan;
+ *((u32 *)¶ms.smac[0]) = ntohl(src_mac_addr_hi);
+ *((u16 *)¶ms.smac[4]) = ntohs(src_mac_addr_lo);
+
+ if (data_length_error) {
+ DP_ERR(cdev,
+ "roce ll2 rx complete: data length error %d, length=%d\n",
+ data_length_error, data_length);
+ params.rc = -EINVAL;
+ }
+
+ roce_ll2->cbs.rx_cb(roce_ll2->cb_cookie, &pkt, ¶ms);
+}
+
static int qed_roce_ll2_set_mac_filter(struct qed_dev *cdev,
u8 *old_mac_address,
u8 *new_mac_address)
@@ -2717,6 +2782,7 @@ static int qed_roce_ll2_start(struct qed_dev *cdev,
ll2_params.tx_dest = CORE_TX_DEST_NW;
ll2_params.ai_err_packet_too_big = LL2_DROP_PACKET;
ll2_params.ai_err_no_buf = LL2_DROP_PACKET;
+ ll2_params.gsi_enable = true;
rc = qed_ll2_acquire_connection(QED_LEADING_HWFN(cdev), &ll2_params,
params->max_rx_buffers,
--
1.8.3.1
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
^ permalink raw reply related
* [RFC v2 09/12] qedr: Add LL2 RoCE interface
From: Ram Amrani @ 2016-09-20 10:36 UTC (permalink / raw)
To: davem-fT/PcQaiUtIeIZ0/mPfg9Q, dledford-H+wXaHxf7aLQT0dZR+AlfA
Cc: Ariel.Elior-YGCgFSpz5w/QT0dZR+AlfA,
Michal.Kalderon-YGCgFSpz5w/QT0dZR+AlfA,
Yuval.Mintz-YGCgFSpz5w/QT0dZR+AlfA,
rajesh.borundia-YGCgFSpz5w/QT0dZR+AlfA,
linux-rdma-u79uwXL29TY76Z2rM5mHXA, netdev-u79uwXL29TY76Z2rM5mHXA,
Ram amrani
In-Reply-To: <1474367764-9555-1-git-send-email-Ram.Amrani-YGCgFSpz5w/QT0dZR+AlfA@public.gmane.org>
From: Ram amrani <Ram.Amrani-YGCgFSpz5w/QT0dZR+AlfA@public.gmane.org>
Add light L2 interface for RoCE.
Signed-off-by: Rajesh Borundia <rajesh.borundia-YGCgFSpz5w/QT0dZR+AlfA@public.gmane.org>
Signed-off-by: Ram Amrani <Ram.Amrani-YGCgFSpz5w/QT0dZR+AlfA@public.gmane.org>
---
drivers/infiniband/hw/qedr/main.c | 34 +++-
drivers/infiniband/hw/qedr/qedr.h | 2 +-
drivers/net/ethernet/qlogic/qed/qed.h | 1 +
drivers/net/ethernet/qlogic/qed/qed_hsi.h | 3 +
drivers/net/ethernet/qlogic/qed/qed_ll2.c | 21 ++-
drivers/net/ethernet/qlogic/qed/qed_ll2.h | 7 +
drivers/net/ethernet/qlogic/qed/qed_roce.c | 241 +++++++++++++++++++++++++++++
drivers/net/ethernet/qlogic/qed/qed_roce.h | 1 +
include/linux/qed/qed_roce_if.h | 79 ++++++++++
9 files changed, 385 insertions(+), 4 deletions(-)
diff --git a/drivers/infiniband/hw/qedr/main.c b/drivers/infiniband/hw/qedr/main.c
index c54166e..fd602ad 100644
--- a/drivers/infiniband/hw/qedr/main.c
+++ b/drivers/infiniband/hw/qedr/main.c
@@ -710,6 +710,38 @@ static void qedr_shutdown(struct qedr_dev *dev)
qedr_remove(dev);
}
+static void qedr_mac_address_change(struct qedr_dev *dev)
+{
+ union ib_gid *sgid = &dev->sgid_tbl[0];
+ u8 guid[8], mac_addr[6];
+ int rc;
+
+ /* Update SGID */
+ ether_addr_copy(&mac_addr[0], dev->ndev->dev_addr);
+ guid[0] = mac_addr[0] ^ 2;
+ guid[1] = mac_addr[1];
+ guid[2] = mac_addr[2];
+ guid[3] = 0xff;
+ guid[4] = 0xfe;
+ guid[5] = mac_addr[3];
+ guid[6] = mac_addr[4];
+ guid[7] = mac_addr[5];
+ sgid->global.subnet_prefix = cpu_to_be64(0xfe80000000000000LL);
+ memcpy(&sgid->raw[8], guid, sizeof(guid));
+
+ /* Update LL2 */
+ rc = dev->ops->roce_ll2_set_mac_filter(dev->cdev,
+ dev->gsi_ll2_mac_address,
+ dev->ndev->dev_addr);
+
+ ether_addr_copy(dev->gsi_ll2_mac_address, dev->ndev->dev_addr);
+
+ qedr_ib_dispatch_event(dev, 1, IB_EVENT_GID_CHANGE);
+
+ if (rc)
+ DP_ERR(dev, "Error updating mac filter\n");
+}
+
/* event handling via NIC driver ensures that all the NIC specific
* initialization done before RoCE driver notifies
* event to stack.
@@ -727,7 +759,7 @@ static void qedr_notify(struct qedr_dev *dev, enum qede_roce_event event)
qedr_shutdown(dev);
break;
case QEDE_CHANGE_ADDR:
- qedr_ib_dispatch_event(dev, 1, IB_EVENT_GID_CHANGE);
+ qedr_mac_address_change(dev);
break;
default:
pr_err("Event not supported\n");
diff --git a/drivers/infiniband/hw/qedr/qedr.h b/drivers/infiniband/hw/qedr/qedr.h
index 3d6af4a..18af842 100644
--- a/drivers/infiniband/hw/qedr/qedr.h
+++ b/drivers/infiniband/hw/qedr/qedr.h
@@ -145,7 +145,7 @@ struct qedr_dev {
u8 dp_level;
u8 num_hwfns;
uint wq_multiplier;
-
+ u8 gsi_ll2_mac_address[ETH_ALEN];
};
#define QEDR_MAX_SQ_PBL (0x8000)
diff --git a/drivers/net/ethernet/qlogic/qed/qed.h b/drivers/net/ethernet/qlogic/qed/qed.h
index a92aa80..3387f47 100644
--- a/drivers/net/ethernet/qlogic/qed/qed.h
+++ b/drivers/net/ethernet/qlogic/qed/qed.h
@@ -429,6 +429,7 @@ struct qed_hwfn {
#endif
struct z_stream_s *stream;
+ struct qed_roce_ll2_info *ll2;
};
struct pci_params {
diff --git a/drivers/net/ethernet/qlogic/qed/qed_hsi.h b/drivers/net/ethernet/qlogic/qed/qed_hsi.h
index 2777d5b..72eee29 100644
--- a/drivers/net/ethernet/qlogic/qed/qed_hsi.h
+++ b/drivers/net/ethernet/qlogic/qed/qed_hsi.h
@@ -727,6 +727,9 @@ struct core_tx_bd_flags {
#define CORE_TX_BD_FLAGS_L4_PROTOCOL_SHIFT 6
#define CORE_TX_BD_FLAGS_L4_PSEUDO_CSUM_MODE_MASK 0x1
#define CORE_TX_BD_FLAGS_L4_PSEUDO_CSUM_MODE_SHIFT 7
+#define CORE_TX_BD_FLAGS_ROCE_FLAV_MASK 0x1
+#define CORE_TX_BD_FLAGS_ROCE_FLAV_SHIFT 12
+
};
struct core_tx_bd {
diff --git a/drivers/net/ethernet/qlogic/qed/qed_ll2.c b/drivers/net/ethernet/qlogic/qed/qed_ll2.c
index 1a96a2d..4ea6a07 100644
--- a/drivers/net/ethernet/qlogic/qed/qed_ll2.c
+++ b/drivers/net/ethernet/qlogic/qed/qed_ll2.c
@@ -301,7 +301,8 @@ static void qed_ll2_txq_flush(struct qed_hwfn *p_hwfn, u8 connection_handle)
qed_ll2b_complete_tx_packet(p_hwfn, p_ll2_conn->my_id,
p_pkt->cookie,
p_pkt->bds_set[0].tx_frag,
- b_last_frag, b_last_packet);
+ b_last_frag,
+ b_last_packet);
}
}
@@ -1025,6 +1026,7 @@ static void qed_ll2_prepare_tx_packet_set_bd(struct qed_hwfn *p_hwfn,
u16 vlan,
u8 bd_flags,
u16 l4_hdr_offset_w,
+ enum core_roce_flavor_type type,
dma_addr_t first_frag,
u16 first_frag_len)
{
@@ -1045,6 +1047,9 @@ static void qed_ll2_prepare_tx_packet_set_bd(struct qed_hwfn *p_hwfn,
DMA_REGPAIR_LE(start_bd->addr, first_frag);
start_bd->nbytes = cpu_to_le16(first_frag_len);
+ SET_FIELD(start_bd->bd_flags.as_bitfield, CORE_TX_BD_FLAGS_ROCE_FLAV,
+ type);
+
DP_VERBOSE(p_hwfn,
(NETIF_MSG_TX_QUEUED | QED_MSG_LL2),
"LL2 [q 0x%02x cid 0x%08x type 0x%08x] Tx Producer at [0x%04x] - set with a %04x bytes %02x BDs buffer at %08x:%08x\n",
@@ -1136,11 +1141,13 @@ int qed_ll2_prepare_tx_packet(struct qed_hwfn *p_hwfn,
u16 vlan,
u8 bd_flags,
u16 l4_hdr_offset_w,
+ enum qed_ll2_roce_flavor_type qed_roce_flavor,
dma_addr_t first_frag,
u16 first_frag_len, void *cookie, u8 notify_fw)
{
struct qed_ll2_tx_packet *p_curp = NULL;
struct qed_ll2_info *p_ll2_conn = NULL;
+ enum core_roce_flavor_type roce_flavor;
struct qed_ll2_tx_queue *p_tx;
struct qed_chain *p_tx_chain;
unsigned long flags;
@@ -1173,6 +1180,15 @@ int qed_ll2_prepare_tx_packet(struct qed_hwfn *p_hwfn,
goto out;
}
+ if (qed_roce_flavor == QED_LL2_ROCE) {
+ roce_flavor = CORE_ROCE;
+ } else if (qed_roce_flavor == QED_LL2_RROCE) {
+ roce_flavor = CORE_RROCE;
+ } else {
+ rc = -EINVAL;
+ goto out;
+ }
+
/* Prepare packet and BD, and perhaps send a doorbell to FW */
qed_ll2_prepare_tx_packet_set(p_hwfn, p_tx, p_curp,
num_of_bds, first_frag,
@@ -1180,6 +1196,7 @@ int qed_ll2_prepare_tx_packet(struct qed_hwfn *p_hwfn,
qed_ll2_prepare_tx_packet_set_bd(p_hwfn, p_ll2_conn, p_curp,
num_of_bds, CORE_TX_DEST_NW,
vlan, bd_flags, l4_hdr_offset_w,
+ roce_flavor,
first_frag, first_frag_len);
qed_ll2_tx_packet_notify(p_hwfn, p_ll2_conn);
@@ -1624,7 +1641,7 @@ static int qed_ll2_start_xmit(struct qed_dev *cdev, struct sk_buff *skb)
rc = qed_ll2_prepare_tx_packet(QED_LEADING_HWFN(cdev),
cdev->ll2->handle,
1 + skb_shinfo(skb)->nr_frags,
- vlan, flags, 0, mapping,
+ vlan, flags, 0, mapping, 0,
skb->len, skb, 1);
if (rc)
goto err;
diff --git a/drivers/net/ethernet/qlogic/qed/qed_ll2.h b/drivers/net/ethernet/qlogic/qed/qed_ll2.h
index a037c48..b3783fd 100644
--- a/drivers/net/ethernet/qlogic/qed/qed_ll2.h
+++ b/drivers/net/ethernet/qlogic/qed/qed_ll2.h
@@ -24,6 +24,12 @@
#define QED_MAX_NUM_OF_LL2_CONNECTIONS (4)
+enum qed_ll2_roce_flavor_type {
+ QED_LL2_ROCE,
+ QED_LL2_RROCE,
+ MAX_QED_LL2_ROCE_FLAVOR_TYPE
+};
+
enum qed_ll2_conn_type {
QED_LL2_TYPE_RESERVED,
QED_LL2_TYPE_ISCSI,
@@ -199,6 +205,7 @@ int qed_ll2_prepare_tx_packet(struct qed_hwfn *p_hwfn,
u16 vlan,
u8 bd_flags,
u16 l4_hdr_offset_w,
+ enum qed_ll2_roce_flavor_type qed_roce_flavor,
dma_addr_t first_frag,
u16 first_frag_len, void *cookie, u8 notify_fw);
diff --git a/drivers/net/ethernet/qlogic/qed/qed_roce.c b/drivers/net/ethernet/qlogic/qed/qed_roce.c
index 04e33ee..b4f851b 100644
--- a/drivers/net/ethernet/qlogic/qed/qed_roce.c
+++ b/drivers/net/ethernet/qlogic/qed/qed_roce.c
@@ -64,6 +64,7 @@
#include "qed_reg_addr.h"
#include "qed_sp.h"
#include "qed_roce.h"
+#include "qed_ll2.h"
void qed_async_roce_event(struct qed_hwfn *p_hwfn,
struct event_ring_entry *p_eqe)
@@ -2632,6 +2633,240 @@ void qed_rdma_remove_user(void *rdma_cxt, u16 dpi)
spin_unlock_bh(&p_hwfn->p_rdma_info->lock);
}
+static int qed_roce_ll2_set_mac_filter(struct qed_dev *cdev,
+ u8 *old_mac_address,
+ u8 *new_mac_address)
+{
+ struct qed_hwfn *hwfn = QED_LEADING_HWFN(cdev);
+ struct qed_ptt *p_ptt;
+ int rc = 0;
+
+ if (!hwfn->ll2 || hwfn->ll2->handle == QED_LL2_UNUSED_HANDLE) {
+ DP_ERR(cdev,
+ "qed roce mac filter failed - roce_info/ll2 NULL\n");
+ return -EINVAL;
+ }
+
+ p_ptt = qed_ptt_acquire(QED_LEADING_HWFN(cdev));
+ if (!p_ptt) {
+ DP_ERR(cdev,
+ "qed roce ll2 mac filter set: failed to acquire PTT\n");
+ return -EINVAL;
+ }
+
+ mutex_lock(&hwfn->ll2->lock);
+ if (old_mac_address)
+ qed_llh_remove_mac_filter(QED_LEADING_HWFN(cdev), p_ptt,
+ old_mac_address);
+ if (new_mac_address)
+ rc = qed_llh_add_mac_filter(QED_LEADING_HWFN(cdev), p_ptt,
+ new_mac_address);
+ mutex_unlock(&hwfn->ll2->lock);
+
+ qed_ptt_release(QED_LEADING_HWFN(cdev), p_ptt);
+
+ if (rc)
+ DP_ERR(cdev,
+ "qed roce ll2 mac filter set: failed to add mac filter\n");
+
+ return rc;
+}
+
+static int qed_roce_ll2_start(struct qed_dev *cdev,
+ struct qed_roce_ll2_params *params)
+{
+ struct qed_hwfn *hwfn = QED_LEADING_HWFN(cdev);
+ struct qed_roce_ll2_info *roce_ll2;
+ struct qed_ll2_info ll2_params;
+ int rc;
+
+ if (!params) {
+ DP_ERR(cdev, "qed roce ll2 start: failed due to NULL params\n");
+ return -EINVAL;
+ }
+ if (!params->cbs.tx_cb || !params->cbs.rx_cb) {
+ DP_ERR(cdev,
+ "qed roce ll2 start: failed due to NULL tx/rx. tx_cb=%p, rx_cb=%p\n",
+ params->cbs.tx_cb, params->cbs.rx_cb);
+ return -EINVAL;
+ }
+ if (!is_valid_ether_addr(params->mac_address)) {
+ DP_ERR(cdev,
+ "qed roce ll2 start: failed due to invalid Ethernet address %pM\n",
+ params->mac_address);
+ return -EINVAL;
+ }
+
+ /* Initialize */
+ roce_ll2 = kzalloc(sizeof(*roce_ll2), GFP_ATOMIC);
+ if (!roce_ll2) {
+ DP_ERR(cdev, "qed roce ll2 start: failed memory allocation\n");
+ return -ENOMEM;
+ }
+ memset(roce_ll2, 0, sizeof(*roce_ll2));
+ roce_ll2->handle = QED_LL2_UNUSED_HANDLE;
+ roce_ll2->cbs = params->cbs;
+ roce_ll2->cb_cookie = params->cb_cookie;
+ mutex_init(&roce_ll2->lock);
+
+ memset(&ll2_params, 0, sizeof(ll2_params));
+ ll2_params.conn_type = QED_LL2_TYPE_ROCE;
+ ll2_params.mtu = params->mtu;
+ ll2_params.rx_drop_ttl0_flg = true;
+ ll2_params.rx_vlan_removal_en = false;
+ ll2_params.tx_dest = CORE_TX_DEST_NW;
+ ll2_params.ai_err_packet_too_big = LL2_DROP_PACKET;
+ ll2_params.ai_err_no_buf = LL2_DROP_PACKET;
+
+ rc = qed_ll2_acquire_connection(QED_LEADING_HWFN(cdev), &ll2_params,
+ params->max_rx_buffers,
+ params->max_tx_buffers,
+ &roce_ll2->handle);
+ if (rc) {
+ DP_ERR(cdev,
+ "qed roce ll2 start: failed to acquire LL2 connection (rc=%d)\n",
+ rc);
+ goto err;
+ }
+
+ rc = qed_ll2_establish_connection(QED_LEADING_HWFN(cdev),
+ roce_ll2->handle);
+ if (rc) {
+ DP_ERR(cdev,
+ "qed roce ll2 start: failed to establish LL2 connection (rc=%d)\n",
+ rc);
+ goto err1;
+ }
+
+ hwfn->ll2 = roce_ll2;
+
+ rc = qed_roce_ll2_set_mac_filter(cdev, NULL, params->mac_address);
+ if (rc) {
+ hwfn->ll2 = NULL;
+ goto err2;
+ }
+ ether_addr_copy(roce_ll2->mac_address, params->mac_address);
+
+ return 0;
+
+err2:
+ qed_ll2_terminate_connection(QED_LEADING_HWFN(cdev), roce_ll2->handle);
+err1:
+ qed_ll2_release_connection(QED_LEADING_HWFN(cdev), roce_ll2->handle);
+err:
+ kfree(roce_ll2);
+ return rc;
+}
+
+static int qed_roce_ll2_stop(struct qed_dev *cdev)
+{
+ struct qed_hwfn *hwfn = QED_LEADING_HWFN(cdev);
+ struct qed_roce_ll2_info *roce_ll2 = hwfn->ll2;
+ int rc;
+
+ if (!cdev) {
+ DP_ERR(cdev, "qed roce ll2 stop: invalid cdev\n");
+ return -EINVAL;
+ }
+
+ if (roce_ll2->handle == QED_LL2_UNUSED_HANDLE) {
+ DP_ERR(cdev, "qed roce ll2 stop: cannot stop an unused LL2\n");
+ return -EINVAL;
+ }
+
+ /* remove LL2 MAC address filter */
+ rc = qed_roce_ll2_set_mac_filter(cdev, roce_ll2->mac_address, NULL);
+ eth_zero_addr(roce_ll2->mac_address);
+
+ rc = qed_ll2_terminate_connection(QED_LEADING_HWFN(cdev),
+ roce_ll2->handle);
+ if (rc)
+ DP_ERR(cdev,
+ "qed roce ll2 stop: failed to terminate LL2 connection (rc=%d)\n",
+ rc);
+
+ qed_ll2_release_connection(QED_LEADING_HWFN(cdev), roce_ll2->handle);
+
+ roce_ll2->handle = QED_LL2_UNUSED_HANDLE;
+
+ kfree(roce_ll2);
+
+ return rc;
+}
+
+static int qed_roce_ll2_tx(struct qed_dev *cdev,
+ struct qed_roce_ll2_packet *pkt,
+ struct qed_roce_ll2_tx_params *params)
+{
+ struct qed_hwfn *hwfn = QED_LEADING_HWFN(cdev);
+ struct qed_roce_ll2_info *roce_ll2 = hwfn->ll2;
+ enum qed_ll2_roce_flavor_type qed_roce_flavor;
+ u8 flags = 0;
+ int rc;
+ int i;
+
+ if (!cdev || !pkt || !params) {
+ DP_ERR(cdev,
+ "roce ll2 tx: failed tx because one of the following is NULL - drv=%p, pkt=%p, params=%p\n",
+ cdev, pkt, params);
+ return -EINVAL;
+ }
+
+ qed_roce_flavor = (pkt->roce_mode == ROCE_V1) ? QED_LL2_ROCE
+ : QED_LL2_RROCE;
+
+ if (pkt->roce_mode == ROCE_V2_IPV4)
+ flags |= BIT(CORE_TX_BD_FLAGS_IP_CSUM_SHIFT);
+
+ /* Tx header */
+ rc = qed_ll2_prepare_tx_packet(QED_LEADING_HWFN(cdev), roce_ll2->handle,
+ 1 + pkt->n_seg, 0, flags, 0,
+ qed_roce_flavor, pkt->header.baddr,
+ pkt->header.len, pkt, 1);
+ if (rc) {
+ DP_ERR(cdev, "roce ll2 tx: header failed (rc=%d)\n", rc);
+ return QED_ROCE_TX_HEAD_FAILURE;
+ }
+
+ /* Tx payload */
+ for (i = 0; i < pkt->n_seg; i++) {
+ rc = qed_ll2_set_fragment_of_tx_packet(QED_LEADING_HWFN(cdev),
+ roce_ll2->handle,
+ pkt->payload[i].baddr,
+ pkt->payload[i].len);
+ if (rc) {
+ /* If failed not much to do here, partial packet has
+ * been posted * we can't free memory, will need to wait
+ * for completion
+ */
+ DP_ERR(cdev,
+ "roce ll2 tx: payload failed (rc=%d)\n", rc);
+ return QED_ROCE_TX_FRAG_FAILURE;
+ }
+ }
+
+ return 0;
+}
+
+static int qed_roce_ll2_post_rx_buffer(struct qed_dev *cdev,
+ struct qed_roce_ll2_buffer *buf,
+ u64 cookie, u8 notify_fw)
+{
+ return qed_ll2_post_rx_buffer(QED_LEADING_HWFN(cdev),
+ QED_LEADING_HWFN(cdev)->ll2->handle,
+ buf->baddr, buf->len,
+ (void *)(uintptr_t)cookie, notify_fw);
+}
+
+static int qed_roce_ll2_stats(struct qed_dev *cdev, struct qed_ll2_stats *stats)
+{
+ struct qed_hwfn *hwfn = QED_LEADING_HWFN(cdev);
+ struct qed_roce_ll2_info *roce_ll2 = hwfn->ll2;
+
+ return qed_ll2_get_stats(QED_LEADING_HWFN(cdev),
+ roce_ll2->handle, stats);
+}
+
static const struct qed_rdma_ops qed_rdma_ops_pass = {
.common = &qed_common_ops_pass,
.fill_dev_info = &qed_fill_rdma_dev_info,
@@ -2659,6 +2894,12 @@ static const struct qed_rdma_ops qed_rdma_ops_pass = {
.rdma_free_tid = &qed_rdma_free_tid,
.rdma_register_tid = &qed_rdma_register_tid,
.rdma_deregister_tid = &qed_rdma_deregister_tid,
+ .roce_ll2_start = &qed_roce_ll2_start,
+ .roce_ll2_stop = &qed_roce_ll2_stop,
+ .roce_ll2_tx = &qed_roce_ll2_tx,
+ .roce_ll2_post_rx_buffer = &qed_roce_ll2_post_rx_buffer,
+ .roce_ll2_set_mac_filter = &qed_roce_ll2_set_mac_filter,
+ .roce_ll2_stats = &qed_roce_ll2_stats,
};
const struct qed_rdma_ops *qed_get_rdma_ops()
diff --git a/drivers/net/ethernet/qlogic/qed/qed_roce.h b/drivers/net/ethernet/qlogic/qed/qed_roce.h
index 71b37fc..f688a47 100644
--- a/drivers/net/ethernet/qlogic/qed/qed_roce.h
+++ b/drivers/net/ethernet/qlogic/qed/qed_roce.h
@@ -42,6 +42,7 @@
#include "qed.h"
#include "qed_dev_api.h"
#include "qed_hsi.h"
+#include "qed_ll2.h"
#define QED_RDMA_MAX_FMR (RDMA_MAX_TIDS)
#define QED_RDMA_MAX_P_KEY (1)
diff --git a/include/linux/qed/qed_roce_if.h b/include/linux/qed/qed_roce_if.h
index 0b6df6e..53047d3 100644
--- a/include/linux/qed/qed_roce_if.h
+++ b/include/linux/qed/qed_roce_if.h
@@ -39,6 +39,16 @@
#include <linux/slab.h>
#include <linux/qed/qed_if.h>
#include <linux/qed/qed_ll2_if.h>
+#include <linux/qed/rdma_common.h>
+
+enum qed_roce_ll2_tx_dest {
+ /* Light L2 TX Destination to the Network */
+ QED_ROCE_LL2_TX_DEST_NW,
+
+ /* Light L2 TX Destination to the Loopback */
+ QED_ROCE_LL2_TX_DEST_LB,
+ QED_ROCE_LL2_TX_DEST_MAX
+};
#define QED_RDMA_MAX_CNQ_SIZE (0xFFFF)
@@ -461,6 +471,61 @@ struct qed_rdma_counters_out_params {
#define QED_ROCE_TX_HEAD_FAILURE (1)
#define QED_ROCE_TX_FRAG_FAILURE (2)
+struct qed_roce_ll2_header {
+ void *vaddr;
+ dma_addr_t baddr;
+ size_t len;
+};
+
+struct qed_roce_ll2_buffer {
+ dma_addr_t baddr;
+ size_t len;
+};
+
+struct qed_roce_ll2_packet {
+ struct qed_roce_ll2_header header;
+ int n_seg;
+ struct qed_roce_ll2_buffer payload[RDMA_MAX_SGE_PER_SQ_WQE];
+ int roce_mode;
+ enum qed_roce_ll2_tx_dest tx_dest;
+};
+
+struct qed_roce_ll2_tx_params {
+ int reserved;
+};
+
+struct qed_roce_ll2_rx_params {
+ u16 vlan_id;
+ u8 smac[ETH_ALEN];
+ int rc;
+};
+
+struct qed_roce_ll2_cbs {
+ void (*tx_cb)(void *pdev, struct qed_roce_ll2_packet *pkt);
+
+ void (*rx_cb)(void *pdev, struct qed_roce_ll2_packet *pkt,
+ struct qed_roce_ll2_rx_params *params);
+};
+
+struct qed_roce_ll2_params {
+ u16 max_rx_buffers;
+ u16 max_tx_buffers;
+ u16 mtu;
+ u8 mac_address[ETH_ALEN];
+ struct qed_roce_ll2_cbs cbs;
+ void *cb_cookie;
+};
+
+struct qed_roce_ll2_info {
+ u8 handle;
+ struct qed_roce_ll2_cbs cbs;
+ u8 mac_address[ETH_ALEN];
+ void *cb_cookie;
+
+ /* Lock to protect ll2 */
+ struct mutex lock;
+};
+
enum qed_rdma_type {
QED_RDMA_TYPE_ROCE,
};
@@ -518,6 +583,20 @@ struct qed_rdma_ops {
int (*rdma_deregister_tid)(void *rdma_cxt, u32 itid);
int (*rdma_alloc_tid)(void *rdma_cxt, u32 *itid);
void (*rdma_free_tid)(void *rdma_cxt, u32 itid);
+ int (*roce_ll2_start)(struct qed_dev *cdev,
+ struct qed_roce_ll2_params *params);
+ int (*roce_ll2_stop)(struct qed_dev *cdev);
+ int (*roce_ll2_tx)(struct qed_dev *cdev,
+ struct qed_roce_ll2_packet *packet,
+ struct qed_roce_ll2_tx_params *params);
+ int (*roce_ll2_post_rx_buffer)(struct qed_dev *cdev,
+ struct qed_roce_ll2_buffer *buf,
+ u64 cookie, u8 notify_fw);
+ int (*roce_ll2_set_mac_filter)(struct qed_dev *cdev,
+ u8 *old_mac_address,
+ u8 *new_mac_address);
+ int (*roce_ll2_stats)(struct qed_dev *cdev,
+ struct qed_ll2_stats *stats);
};
const struct qed_rdma_ops *qed_get_rdma_ops(void);
--
1.8.3.1
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
^ permalink raw reply related
* [RFC v2 08/12] qedr: Add support for data path
From: Ram Amrani @ 2016-09-20 10:36 UTC (permalink / raw)
To: davem, dledford
Cc: Ariel.Elior, Michal.Kalderon, Yuval.Mintz, rajesh.borundia,
linux-rdma, netdev, Ram amrani
In-Reply-To: <1474367764-9555-1-git-send-email-Ram.Amrani@cavium.com>
From: Ram amrani <Ram.Amrani@cavium.com>
Implement fastpath verbs like ib_send_post, ib_post_recv and ib_poll_cq.
Signed-off-by: Rajesh Borundia <rajesh.borundia@cavium.com>
Signed-off-by: Ram Amrani <Ram.Amrani@cavium.com>
---
drivers/infiniband/hw/qedr/main.c | 9 +-
drivers/infiniband/hw/qedr/qedr.h | 19 +
drivers/infiniband/hw/qedr/qedr_hsi_rdma.h | 562 +++++++++++++++
drivers/infiniband/hw/qedr/verbs.c | 1029 ++++++++++++++++++++++++++++
drivers/infiniband/hw/qedr/verbs.h | 5 +
5 files changed, 1623 insertions(+), 1 deletion(-)
diff --git a/drivers/infiniband/hw/qedr/main.c b/drivers/infiniband/hw/qedr/main.c
index 66d33fa..c54166e 100644
--- a/drivers/infiniband/hw/qedr/main.c
+++ b/drivers/infiniband/hw/qedr/main.c
@@ -105,7 +105,10 @@ static int qedr_register_device(struct qedr_dev *dev)
QEDR_UVERBS(QUERY_QP) |
QEDR_UVERBS(DESTROY_QP) |
QEDR_UVERBS(REG_MR) |
- QEDR_UVERBS(DEREG_MR);
+ QEDR_UVERBS(DEREG_MR) |
+ QEDR_UVERBS(POLL_CQ) |
+ QEDR_UVERBS(POST_SEND) |
+ QEDR_UVERBS(POST_RECV);
dev->ibdev.phys_port_cnt = 1;
dev->ibdev.num_comp_vectors = dev->num_cnq;
@@ -143,6 +146,10 @@ static int qedr_register_device(struct qedr_dev *dev)
dev->ibdev.alloc_mr = qedr_alloc_mr;
dev->ibdev.map_mr_sg = qedr_map_mr_sg;
+ dev->ibdev.poll_cq = qedr_poll_cq;
+ dev->ibdev.post_send = qedr_post_send;
+ dev->ibdev.post_recv = qedr_post_recv;
+
dev->ibdev.dma_device = &dev->pdev->dev;
dev->ibdev.get_link_layer = qedr_link_layer;
dev->ibdev.get_dev_fw_str = qedr_get_dev_fw_str;
diff --git a/drivers/infiniband/hw/qedr/qedr.h b/drivers/infiniband/hw/qedr/qedr.h
index c16844b..3d6af4a 100644
--- a/drivers/infiniband/hw/qedr/qedr.h
+++ b/drivers/infiniband/hw/qedr/qedr.h
@@ -408,6 +408,25 @@ struct qedr_mr {
u32 npages;
};
+#define SET_FIELD2(value, name, flag) ((value) |= ((flag) << (name ## _SHIFT)))
+
+#define QEDR_RESP_IMM (RDMA_CQE_RESPONDER_IMM_FLG_MASK << \
+ RDMA_CQE_RESPONDER_IMM_FLG_SHIFT)
+#define QEDR_RESP_RDMA (RDMA_CQE_RESPONDER_RDMA_FLG_MASK << \
+ RDMA_CQE_RESPONDER_RDMA_FLG_SHIFT)
+#define QEDR_RESP_RDMA_IMM (QEDR_RESP_IMM | QEDR_RESP_RDMA)
+
+static inline void qedr_inc_sw_cons(struct qedr_qp_hwq_info *info)
+{
+ info->cons = (info->cons + 1) % info->max_wr;
+ info->wqe_cons++;
+}
+
+static inline void qedr_inc_sw_prod(struct qedr_qp_hwq_info *info)
+{
+ info->prod = (info->prod + 1) % info->max_wr;
+}
+
static inline int qedr_get_dmac(struct qedr_dev *dev,
struct ib_ah_attr *ah_attr, u8 *mac_addr)
{
diff --git a/drivers/infiniband/hw/qedr/qedr_hsi_rdma.h b/drivers/infiniband/hw/qedr/qedr_hsi_rdma.h
index 4770559..5c98d20 100644
--- a/drivers/infiniband/hw/qedr/qedr_hsi_rdma.h
+++ b/drivers/infiniband/hw/qedr/qedr_hsi_rdma.h
@@ -150,6 +150,12 @@ struct rdma_rq_sge {
struct regpair addr;
__le32 length;
__le32 flags;
+#define RDMA_RQ_SGE_L_KEY_MASK 0x3FFFFFF
+#define RDMA_RQ_SGE_L_KEY_SHIFT 0
+#define RDMA_RQ_SGE_NUM_SGES_MASK 0x7
+#define RDMA_RQ_SGE_NUM_SGES_SHIFT 26
+#define RDMA_RQ_SGE_RESERVED0_MASK 0x7
+#define RDMA_RQ_SGE_RESERVED0_SHIFT 29
};
struct rdma_srq_sge {
@@ -183,4 +189,560 @@ struct rdma_pwm_val32_data {
__le32 value;
};
+/* DIF Block size options */
+enum rdma_dif_block_size {
+ RDMA_DIF_BLOCK_512 = 0,
+ RDMA_DIF_BLOCK_4096 = 1,
+ MAX_RDMA_DIF_BLOCK_SIZE
+};
+
+/* DIF CRC initial value */
+enum rdma_dif_crc_seed {
+ RDMA_DIF_CRC_SEED_0000 = 0,
+ RDMA_DIF_CRC_SEED_FFFF = 1,
+ MAX_RDMA_DIF_CRC_SEED
+};
+
+/* RDMA DIF Error Result Structure */
+struct rdma_dif_error_result {
+ __le32 error_intervals;
+ __le32 dif_error_1st_interval;
+ u8 flags;
+#define RDMA_DIF_ERROR_RESULT_DIF_ERROR_TYPE_CRC_MASK 0x1
+#define RDMA_DIF_ERROR_RESULT_DIF_ERROR_TYPE_CRC_SHIFT 0
+#define RDMA_DIF_ERROR_RESULT_DIF_ERROR_TYPE_APP_TAG_MASK 0x1
+#define RDMA_DIF_ERROR_RESULT_DIF_ERROR_TYPE_APP_TAG_SHIFT 1
+#define RDMA_DIF_ERROR_RESULT_DIF_ERROR_TYPE_REF_TAG_MASK 0x1
+#define RDMA_DIF_ERROR_RESULT_DIF_ERROR_TYPE_REF_TAG_SHIFT 2
+#define RDMA_DIF_ERROR_RESULT_RESERVED0_MASK 0xF
+#define RDMA_DIF_ERROR_RESULT_RESERVED0_SHIFT 3
+#define RDMA_DIF_ERROR_RESULT_TOGGLE_BIT_MASK 0x1
+#define RDMA_DIF_ERROR_RESULT_TOGGLE_BIT_SHIFT 7
+ u8 reserved1[55];
+};
+
+/* DIF IO direction */
+enum rdma_dif_io_direction_flg {
+ RDMA_DIF_DIR_RX = 0,
+ RDMA_DIF_DIR_TX = 1,
+ MAX_RDMA_DIF_IO_DIRECTION_FLG
+};
+
+/* RDMA DIF Runt Result Structure */
+struct rdma_dif_runt_result {
+ __le16 guard_tag;
+ __le16 reserved[3];
+};
+
+/* Memory window type enumeration */
+enum rdma_mw_type {
+ RDMA_MW_TYPE_1,
+ RDMA_MW_TYPE_2A,
+ MAX_RDMA_MW_TYPE
+};
+
+struct rdma_sq_atomic_wqe {
+ __le32 reserved1;
+ __le32 length;
+ __le32 xrc_srq;
+ u8 req_type;
+ u8 flags;
+#define RDMA_SQ_ATOMIC_WQE_COMP_FLG_MASK 0x1
+#define RDMA_SQ_ATOMIC_WQE_COMP_FLG_SHIFT 0
+#define RDMA_SQ_ATOMIC_WQE_RD_FENCE_FLG_MASK 0x1
+#define RDMA_SQ_ATOMIC_WQE_RD_FENCE_FLG_SHIFT 1
+#define RDMA_SQ_ATOMIC_WQE_INV_FENCE_FLG_MASK 0x1
+#define RDMA_SQ_ATOMIC_WQE_INV_FENCE_FLG_SHIFT 2
+#define RDMA_SQ_ATOMIC_WQE_SE_FLG_MASK 0x1
+#define RDMA_SQ_ATOMIC_WQE_SE_FLG_SHIFT 3
+#define RDMA_SQ_ATOMIC_WQE_INLINE_FLG_MASK 0x1
+#define RDMA_SQ_ATOMIC_WQE_INLINE_FLG_SHIFT 4
+#define RDMA_SQ_ATOMIC_WQE_DIF_ON_HOST_FLG_MASK 0x1
+#define RDMA_SQ_ATOMIC_WQE_DIF_ON_HOST_FLG_SHIFT 5
+#define RDMA_SQ_ATOMIC_WQE_RESERVED0_MASK 0x3
+#define RDMA_SQ_ATOMIC_WQE_RESERVED0_SHIFT 6
+ u8 wqe_size;
+ u8 prev_wqe_size;
+ struct regpair remote_va;
+ __le32 r_key;
+ __le32 reserved2;
+ struct regpair cmp_data;
+ struct regpair swap_data;
+};
+
+/* First element (16 bytes) of atomic wqe */
+struct rdma_sq_atomic_wqe_1st {
+ __le32 reserved1;
+ __le32 length;
+ __le32 xrc_srq;
+ u8 req_type;
+ u8 flags;
+#define RDMA_SQ_ATOMIC_WQE_1ST_COMP_FLG_MASK 0x1
+#define RDMA_SQ_ATOMIC_WQE_1ST_COMP_FLG_SHIFT 0
+#define RDMA_SQ_ATOMIC_WQE_1ST_RD_FENCE_FLG_MASK 0x1
+#define RDMA_SQ_ATOMIC_WQE_1ST_RD_FENCE_FLG_SHIFT 1
+#define RDMA_SQ_ATOMIC_WQE_1ST_INV_FENCE_FLG_MASK 0x1
+#define RDMA_SQ_ATOMIC_WQE_1ST_INV_FENCE_FLG_SHIFT 2
+#define RDMA_SQ_ATOMIC_WQE_1ST_SE_FLG_MASK 0x1
+#define RDMA_SQ_ATOMIC_WQE_1ST_SE_FLG_SHIFT 3
+#define RDMA_SQ_ATOMIC_WQE_1ST_INLINE_FLG_MASK 0x1
+#define RDMA_SQ_ATOMIC_WQE_1ST_INLINE_FLG_SHIFT 4
+#define RDMA_SQ_ATOMIC_WQE_1ST_RESERVED0_MASK 0x7
+#define RDMA_SQ_ATOMIC_WQE_1ST_RESERVED0_SHIFT 5
+ u8 wqe_size;
+ u8 prev_wqe_size;
+};
+
+/* Second element (16 bytes) of atomic wqe */
+struct rdma_sq_atomic_wqe_2nd {
+ struct regpair remote_va;
+ __le32 r_key;
+ __le32 reserved2;
+};
+
+/* Third element (16 bytes) of atomic wqe */
+struct rdma_sq_atomic_wqe_3rd {
+ struct regpair cmp_data;
+ struct regpair swap_data;
+};
+
+struct rdma_sq_bind_wqe {
+ struct regpair addr;
+ __le32 l_key;
+ u8 req_type;
+ u8 flags;
+#define RDMA_SQ_BIND_WQE_COMP_FLG_MASK 0x1
+#define RDMA_SQ_BIND_WQE_COMP_FLG_SHIFT 0
+#define RDMA_SQ_BIND_WQE_RD_FENCE_FLG_MASK 0x1
+#define RDMA_SQ_BIND_WQE_RD_FENCE_FLG_SHIFT 1
+#define RDMA_SQ_BIND_WQE_INV_FENCE_FLG_MASK 0x1
+#define RDMA_SQ_BIND_WQE_INV_FENCE_FLG_SHIFT 2
+#define RDMA_SQ_BIND_WQE_SE_FLG_MASK 0x1
+#define RDMA_SQ_BIND_WQE_SE_FLG_SHIFT 3
+#define RDMA_SQ_BIND_WQE_INLINE_FLG_MASK 0x1
+#define RDMA_SQ_BIND_WQE_INLINE_FLG_SHIFT 4
+#define RDMA_SQ_BIND_WQE_RESERVED0_MASK 0x7
+#define RDMA_SQ_BIND_WQE_RESERVED0_SHIFT 5
+ u8 wqe_size;
+ u8 prev_wqe_size;
+ u8 bind_ctrl;
+#define RDMA_SQ_BIND_WQE_ZERO_BASED_MASK 0x1
+#define RDMA_SQ_BIND_WQE_ZERO_BASED_SHIFT 0
+#define RDMA_SQ_BIND_WQE_MW_TYPE_MASK 0x1
+#define RDMA_SQ_BIND_WQE_MW_TYPE_SHIFT 1
+#define RDMA_SQ_BIND_WQE_RESERVED1_MASK 0x3F
+#define RDMA_SQ_BIND_WQE_RESERVED1_SHIFT 2
+ u8 access_ctrl;
+#define RDMA_SQ_BIND_WQE_REMOTE_READ_MASK 0x1
+#define RDMA_SQ_BIND_WQE_REMOTE_READ_SHIFT 0
+#define RDMA_SQ_BIND_WQE_REMOTE_WRITE_MASK 0x1
+#define RDMA_SQ_BIND_WQE_REMOTE_WRITE_SHIFT 1
+#define RDMA_SQ_BIND_WQE_ENABLE_ATOMIC_MASK 0x1
+#define RDMA_SQ_BIND_WQE_ENABLE_ATOMIC_SHIFT 2
+#define RDMA_SQ_BIND_WQE_LOCAL_READ_MASK 0x1
+#define RDMA_SQ_BIND_WQE_LOCAL_READ_SHIFT 3
+#define RDMA_SQ_BIND_WQE_LOCAL_WRITE_MASK 0x1
+#define RDMA_SQ_BIND_WQE_LOCAL_WRITE_SHIFT 4
+#define RDMA_SQ_BIND_WQE_RESERVED2_MASK 0x7
+#define RDMA_SQ_BIND_WQE_RESERVED2_SHIFT 5
+ u8 reserved3;
+ u8 length_hi;
+ __le32 length_lo;
+ __le32 parent_l_key;
+ __le32 reserved4;
+};
+
+/* First element (16 bytes) of bind wqe */
+struct rdma_sq_bind_wqe_1st {
+ struct regpair addr;
+ __le32 l_key;
+ u8 req_type;
+ u8 flags;
+#define RDMA_SQ_BIND_WQE_1ST_COMP_FLG_MASK 0x1
+#define RDMA_SQ_BIND_WQE_1ST_COMP_FLG_SHIFT 0
+#define RDMA_SQ_BIND_WQE_1ST_RD_FENCE_FLG_MASK 0x1
+#define RDMA_SQ_BIND_WQE_1ST_RD_FENCE_FLG_SHIFT 1
+#define RDMA_SQ_BIND_WQE_1ST_INV_FENCE_FLG_MASK 0x1
+#define RDMA_SQ_BIND_WQE_1ST_INV_FENCE_FLG_SHIFT 2
+#define RDMA_SQ_BIND_WQE_1ST_SE_FLG_MASK 0x1
+#define RDMA_SQ_BIND_WQE_1ST_SE_FLG_SHIFT 3
+#define RDMA_SQ_BIND_WQE_1ST_INLINE_FLG_MASK 0x1
+#define RDMA_SQ_BIND_WQE_1ST_INLINE_FLG_SHIFT 4
+#define RDMA_SQ_BIND_WQE_1ST_RESERVED0_MASK 0x7
+#define RDMA_SQ_BIND_WQE_1ST_RESERVED0_SHIFT 5
+ u8 wqe_size;
+ u8 prev_wqe_size;
+};
+
+/* Second element (16 bytes) of bind wqe */
+struct rdma_sq_bind_wqe_2nd {
+ u8 bind_ctrl;
+#define RDMA_SQ_BIND_WQE_2ND_ZERO_BASED_MASK 0x1
+#define RDMA_SQ_BIND_WQE_2ND_ZERO_BASED_SHIFT 0
+#define RDMA_SQ_BIND_WQE_2ND_MW_TYPE_MASK 0x1
+#define RDMA_SQ_BIND_WQE_2ND_MW_TYPE_SHIFT 1
+#define RDMA_SQ_BIND_WQE_2ND_RESERVED1_MASK 0x3F
+#define RDMA_SQ_BIND_WQE_2ND_RESERVED1_SHIFT 2
+ u8 access_ctrl;
+#define RDMA_SQ_BIND_WQE_2ND_REMOTE_READ_MASK 0x1
+#define RDMA_SQ_BIND_WQE_2ND_REMOTE_READ_SHIFT 0
+#define RDMA_SQ_BIND_WQE_2ND_REMOTE_WRITE_MASK 0x1
+#define RDMA_SQ_BIND_WQE_2ND_REMOTE_WRITE_SHIFT 1
+#define RDMA_SQ_BIND_WQE_2ND_ENABLE_ATOMIC_MASK 0x1
+#define RDMA_SQ_BIND_WQE_2ND_ENABLE_ATOMIC_SHIFT 2
+#define RDMA_SQ_BIND_WQE_2ND_LOCAL_READ_MASK 0x1
+#define RDMA_SQ_BIND_WQE_2ND_LOCAL_READ_SHIFT 3
+#define RDMA_SQ_BIND_WQE_2ND_LOCAL_WRITE_MASK 0x1
+#define RDMA_SQ_BIND_WQE_2ND_LOCAL_WRITE_SHIFT 4
+#define RDMA_SQ_BIND_WQE_2ND_RESERVED2_MASK 0x7
+#define RDMA_SQ_BIND_WQE_2ND_RESERVED2_SHIFT 5
+ u8 reserved3;
+ u8 length_hi;
+ __le32 length_lo;
+ __le32 parent_l_key;
+ __le32 reserved4;
+};
+
+/* Structure with only the SQ WQE common
+ * fields. Size is of one SQ element (16B)
+ */
+struct rdma_sq_common_wqe {
+ __le32 reserved1[3];
+ u8 req_type;
+ u8 flags;
+#define RDMA_SQ_COMMON_WQE_COMP_FLG_MASK 0x1
+#define RDMA_SQ_COMMON_WQE_COMP_FLG_SHIFT 0
+#define RDMA_SQ_COMMON_WQE_RD_FENCE_FLG_MASK 0x1
+#define RDMA_SQ_COMMON_WQE_RD_FENCE_FLG_SHIFT 1
+#define RDMA_SQ_COMMON_WQE_INV_FENCE_FLG_MASK 0x1
+#define RDMA_SQ_COMMON_WQE_INV_FENCE_FLG_SHIFT 2
+#define RDMA_SQ_COMMON_WQE_SE_FLG_MASK 0x1
+#define RDMA_SQ_COMMON_WQE_SE_FLG_SHIFT 3
+#define RDMA_SQ_COMMON_WQE_INLINE_FLG_MASK 0x1
+#define RDMA_SQ_COMMON_WQE_INLINE_FLG_SHIFT 4
+#define RDMA_SQ_COMMON_WQE_RESERVED0_MASK 0x7
+#define RDMA_SQ_COMMON_WQE_RESERVED0_SHIFT 5
+ u8 wqe_size;
+ u8 prev_wqe_size;
+};
+
+struct rdma_sq_fmr_wqe {
+ struct regpair addr;
+ __le32 l_key;
+ u8 req_type;
+ u8 flags;
+#define RDMA_SQ_FMR_WQE_COMP_FLG_MASK 0x1
+#define RDMA_SQ_FMR_WQE_COMP_FLG_SHIFT 0
+#define RDMA_SQ_FMR_WQE_RD_FENCE_FLG_MASK 0x1
+#define RDMA_SQ_FMR_WQE_RD_FENCE_FLG_SHIFT 1
+#define RDMA_SQ_FMR_WQE_INV_FENCE_FLG_MASK 0x1
+#define RDMA_SQ_FMR_WQE_INV_FENCE_FLG_SHIFT 2
+#define RDMA_SQ_FMR_WQE_SE_FLG_MASK 0x1
+#define RDMA_SQ_FMR_WQE_SE_FLG_SHIFT 3
+#define RDMA_SQ_FMR_WQE_INLINE_FLG_MASK 0x1
+#define RDMA_SQ_FMR_WQE_INLINE_FLG_SHIFT 4
+#define RDMA_SQ_FMR_WQE_DIF_ON_HOST_FLG_MASK 0x1
+#define RDMA_SQ_FMR_WQE_DIF_ON_HOST_FLG_SHIFT 5
+#define RDMA_SQ_FMR_WQE_RESERVED0_MASK 0x3
+#define RDMA_SQ_FMR_WQE_RESERVED0_SHIFT 6
+ u8 wqe_size;
+ u8 prev_wqe_size;
+ u8 fmr_ctrl;
+#define RDMA_SQ_FMR_WQE_PAGE_SIZE_LOG_MASK 0x1F
+#define RDMA_SQ_FMR_WQE_PAGE_SIZE_LOG_SHIFT 0
+#define RDMA_SQ_FMR_WQE_ZERO_BASED_MASK 0x1
+#define RDMA_SQ_FMR_WQE_ZERO_BASED_SHIFT 5
+#define RDMA_SQ_FMR_WQE_BIND_EN_MASK 0x1
+#define RDMA_SQ_FMR_WQE_BIND_EN_SHIFT 6
+#define RDMA_SQ_FMR_WQE_RESERVED1_MASK 0x1
+#define RDMA_SQ_FMR_WQE_RESERVED1_SHIFT 7
+ u8 access_ctrl;
+#define RDMA_SQ_FMR_WQE_REMOTE_READ_MASK 0x1
+#define RDMA_SQ_FMR_WQE_REMOTE_READ_SHIFT 0
+#define RDMA_SQ_FMR_WQE_REMOTE_WRITE_MASK 0x1
+#define RDMA_SQ_FMR_WQE_REMOTE_WRITE_SHIFT 1
+#define RDMA_SQ_FMR_WQE_ENABLE_ATOMIC_MASK 0x1
+#define RDMA_SQ_FMR_WQE_ENABLE_ATOMIC_SHIFT 2
+#define RDMA_SQ_FMR_WQE_LOCAL_READ_MASK 0x1
+#define RDMA_SQ_FMR_WQE_LOCAL_READ_SHIFT 3
+#define RDMA_SQ_FMR_WQE_LOCAL_WRITE_MASK 0x1
+#define RDMA_SQ_FMR_WQE_LOCAL_WRITE_SHIFT 4
+#define RDMA_SQ_FMR_WQE_RESERVED2_MASK 0x7
+#define RDMA_SQ_FMR_WQE_RESERVED2_SHIFT 5
+ u8 reserved3;
+ u8 length_hi;
+ __le32 length_lo;
+ struct regpair pbl_addr;
+ __le32 dif_base_ref_tag;
+ __le16 dif_app_tag;
+ __le16 dif_app_tag_mask;
+ __le16 dif_runt_crc_value;
+ __le16 dif_flags;
+#define RDMA_SQ_FMR_WQE_DIF_IO_DIRECTION_FLG_MASK 0x1
+#define RDMA_SQ_FMR_WQE_DIF_IO_DIRECTION_FLG_SHIFT 0
+#define RDMA_SQ_FMR_WQE_DIF_BLOCK_SIZE_MASK 0x1
+#define RDMA_SQ_FMR_WQE_DIF_BLOCK_SIZE_SHIFT 1
+#define RDMA_SQ_FMR_WQE_DIF_RUNT_VALID_FLG_MASK 0x1
+#define RDMA_SQ_FMR_WQE_DIF_RUNT_VALID_FLG_SHIFT 2
+#define RDMA_SQ_FMR_WQE_DIF_VALIDATE_CRC_GUARD_MASK 0x1
+#define RDMA_SQ_FMR_WQE_DIF_VALIDATE_CRC_GUARD_SHIFT 3
+#define RDMA_SQ_FMR_WQE_DIF_VALIDATE_REF_TAG_MASK 0x1
+#define RDMA_SQ_FMR_WQE_DIF_VALIDATE_REF_TAG_SHIFT 4
+#define RDMA_SQ_FMR_WQE_DIF_VALIDATE_APP_TAG_MASK 0x1
+#define RDMA_SQ_FMR_WQE_DIF_VALIDATE_APP_TAG_SHIFT 5
+#define RDMA_SQ_FMR_WQE_DIF_CRC_SEED_MASK 0x1
+#define RDMA_SQ_FMR_WQE_DIF_CRC_SEED_SHIFT 6
+#define RDMA_SQ_FMR_WQE_RESERVED4_MASK 0x1FF
+#define RDMA_SQ_FMR_WQE_RESERVED4_SHIFT 7
+ __le32 Reserved5;
+};
+
+/* First element (16 bytes) of fmr wqe */
+struct rdma_sq_fmr_wqe_1st {
+ struct regpair addr;
+ __le32 l_key;
+ u8 req_type;
+ u8 flags;
+#define RDMA_SQ_FMR_WQE_1ST_COMP_FLG_MASK 0x1
+#define RDMA_SQ_FMR_WQE_1ST_COMP_FLG_SHIFT 0
+#define RDMA_SQ_FMR_WQE_1ST_RD_FENCE_FLG_MASK 0x1
+#define RDMA_SQ_FMR_WQE_1ST_RD_FENCE_FLG_SHIFT 1
+#define RDMA_SQ_FMR_WQE_1ST_INV_FENCE_FLG_MASK 0x1
+#define RDMA_SQ_FMR_WQE_1ST_INV_FENCE_FLG_SHIFT 2
+#define RDMA_SQ_FMR_WQE_1ST_SE_FLG_MASK 0x1
+#define RDMA_SQ_FMR_WQE_1ST_SE_FLG_SHIFT 3
+#define RDMA_SQ_FMR_WQE_1ST_INLINE_FLG_MASK 0x1
+#define RDMA_SQ_FMR_WQE_1ST_INLINE_FLG_SHIFT 4
+#define RDMA_SQ_FMR_WQE_1ST_DIF_ON_HOST_FLG_MASK 0x1
+#define RDMA_SQ_FMR_WQE_1ST_DIF_ON_HOST_FLG_SHIFT 5
+#define RDMA_SQ_FMR_WQE_1ST_RESERVED0_MASK 0x3
+#define RDMA_SQ_FMR_WQE_1ST_RESERVED0_SHIFT 6
+ u8 wqe_size;
+ u8 prev_wqe_size;
+};
+
+/* Second element (16 bytes) of fmr wqe */
+struct rdma_sq_fmr_wqe_2nd {
+ u8 fmr_ctrl;
+#define RDMA_SQ_FMR_WQE_2ND_PAGE_SIZE_LOG_MASK 0x1F
+#define RDMA_SQ_FMR_WQE_2ND_PAGE_SIZE_LOG_SHIFT 0
+#define RDMA_SQ_FMR_WQE_2ND_ZERO_BASED_MASK 0x1
+#define RDMA_SQ_FMR_WQE_2ND_ZERO_BASED_SHIFT 5
+#define RDMA_SQ_FMR_WQE_2ND_BIND_EN_MASK 0x1
+#define RDMA_SQ_FMR_WQE_2ND_BIND_EN_SHIFT 6
+#define RDMA_SQ_FMR_WQE_2ND_RESERVED1_MASK 0x1
+#define RDMA_SQ_FMR_WQE_2ND_RESERVED1_SHIFT 7
+ u8 access_ctrl;
+#define RDMA_SQ_FMR_WQE_2ND_REMOTE_READ_MASK 0x1
+#define RDMA_SQ_FMR_WQE_2ND_REMOTE_READ_SHIFT 0
+#define RDMA_SQ_FMR_WQE_2ND_REMOTE_WRITE_MASK 0x1
+#define RDMA_SQ_FMR_WQE_2ND_REMOTE_WRITE_SHIFT 1
+#define RDMA_SQ_FMR_WQE_2ND_ENABLE_ATOMIC_MASK 0x1
+#define RDMA_SQ_FMR_WQE_2ND_ENABLE_ATOMIC_SHIFT 2
+#define RDMA_SQ_FMR_WQE_2ND_LOCAL_READ_MASK 0x1
+#define RDMA_SQ_FMR_WQE_2ND_LOCAL_READ_SHIFT 3
+#define RDMA_SQ_FMR_WQE_2ND_LOCAL_WRITE_MASK 0x1
+#define RDMA_SQ_FMR_WQE_2ND_LOCAL_WRITE_SHIFT 4
+#define RDMA_SQ_FMR_WQE_2ND_RESERVED2_MASK 0x7
+#define RDMA_SQ_FMR_WQE_2ND_RESERVED2_SHIFT 5
+ u8 reserved3;
+ u8 length_hi;
+ __le32 length_lo;
+ struct regpair pbl_addr;
+};
+
+/* Third element (16 bytes) of fmr wqe */
+struct rdma_sq_fmr_wqe_3rd {
+ __le32 dif_base_ref_tag;
+ __le16 dif_app_tag;
+ __le16 dif_app_tag_mask;
+ __le16 dif_runt_crc_value;
+ __le16 dif_flags;
+#define RDMA_SQ_FMR_WQE_3RD_DIF_IO_DIRECTION_FLG_MASK 0x1
+#define RDMA_SQ_FMR_WQE_3RD_DIF_IO_DIRECTION_FLG_SHIFT 0
+#define RDMA_SQ_FMR_WQE_3RD_DIF_BLOCK_SIZE_MASK 0x1
+#define RDMA_SQ_FMR_WQE_3RD_DIF_BLOCK_SIZE_SHIFT 1
+#define RDMA_SQ_FMR_WQE_3RD_DIF_RUNT_VALID_FLG_MASK 0x1
+#define RDMA_SQ_FMR_WQE_3RD_DIF_RUNT_VALID_FLG_SHIFT 2
+#define RDMA_SQ_FMR_WQE_3RD_DIF_VALIDATE_CRC_GUARD_MASK 0x1
+#define RDMA_SQ_FMR_WQE_3RD_DIF_VALIDATE_CRC_GUARD_SHIFT 3
+#define RDMA_SQ_FMR_WQE_3RD_DIF_VALIDATE_REF_TAG_MASK 0x1
+#define RDMA_SQ_FMR_WQE_3RD_DIF_VALIDATE_REF_TAG_SHIFT 4
+#define RDMA_SQ_FMR_WQE_3RD_DIF_VALIDATE_APP_TAG_MASK 0x1
+#define RDMA_SQ_FMR_WQE_3RD_DIF_VALIDATE_APP_TAG_SHIFT 5
+#define RDMA_SQ_FMR_WQE_3RD_DIF_CRC_SEED_MASK 0x1
+#define RDMA_SQ_FMR_WQE_3RD_DIF_CRC_SEED_SHIFT 6
+#define RDMA_SQ_FMR_WQE_3RD_RESERVED4_MASK 0x1FF
+#define RDMA_SQ_FMR_WQE_3RD_RESERVED4_SHIFT 7
+ __le32 Reserved5;
+};
+
+struct rdma_sq_local_inv_wqe {
+ struct regpair reserved;
+ __le32 inv_l_key;
+ u8 req_type;
+ u8 flags;
+#define RDMA_SQ_LOCAL_INV_WQE_COMP_FLG_MASK 0x1
+#define RDMA_SQ_LOCAL_INV_WQE_COMP_FLG_SHIFT 0
+#define RDMA_SQ_LOCAL_INV_WQE_RD_FENCE_FLG_MASK 0x1
+#define RDMA_SQ_LOCAL_INV_WQE_RD_FENCE_FLG_SHIFT 1
+#define RDMA_SQ_LOCAL_INV_WQE_INV_FENCE_FLG_MASK 0x1
+#define RDMA_SQ_LOCAL_INV_WQE_INV_FENCE_FLG_SHIFT 2
+#define RDMA_SQ_LOCAL_INV_WQE_SE_FLG_MASK 0x1
+#define RDMA_SQ_LOCAL_INV_WQE_SE_FLG_SHIFT 3
+#define RDMA_SQ_LOCAL_INV_WQE_INLINE_FLG_MASK 0x1
+#define RDMA_SQ_LOCAL_INV_WQE_INLINE_FLG_SHIFT 4
+#define RDMA_SQ_LOCAL_INV_WQE_DIF_ON_HOST_FLG_MASK 0x1
+#define RDMA_SQ_LOCAL_INV_WQE_DIF_ON_HOST_FLG_SHIFT 5
+#define RDMA_SQ_LOCAL_INV_WQE_RESERVED0_MASK 0x3
+#define RDMA_SQ_LOCAL_INV_WQE_RESERVED0_SHIFT 6
+ u8 wqe_size;
+ u8 prev_wqe_size;
+};
+
+struct rdma_sq_rdma_wqe {
+ __le32 imm_data;
+ __le32 length;
+ __le32 xrc_srq;
+ u8 req_type;
+ u8 flags;
+#define RDMA_SQ_RDMA_WQE_COMP_FLG_MASK 0x1
+#define RDMA_SQ_RDMA_WQE_COMP_FLG_SHIFT 0
+#define RDMA_SQ_RDMA_WQE_RD_FENCE_FLG_MASK 0x1
+#define RDMA_SQ_RDMA_WQE_RD_FENCE_FLG_SHIFT 1
+#define RDMA_SQ_RDMA_WQE_INV_FENCE_FLG_MASK 0x1
+#define RDMA_SQ_RDMA_WQE_INV_FENCE_FLG_SHIFT 2
+#define RDMA_SQ_RDMA_WQE_SE_FLG_MASK 0x1
+#define RDMA_SQ_RDMA_WQE_SE_FLG_SHIFT 3
+#define RDMA_SQ_RDMA_WQE_INLINE_FLG_MASK 0x1
+#define RDMA_SQ_RDMA_WQE_INLINE_FLG_SHIFT 4
+#define RDMA_SQ_RDMA_WQE_DIF_ON_HOST_FLG_MASK 0x1
+#define RDMA_SQ_RDMA_WQE_DIF_ON_HOST_FLG_SHIFT 5
+#define RDMA_SQ_RDMA_WQE_RESERVED0_MASK 0x3
+#define RDMA_SQ_RDMA_WQE_RESERVED0_SHIFT 6
+ u8 wqe_size;
+ u8 prev_wqe_size;
+ struct regpair remote_va;
+ __le32 r_key;
+ u8 dif_flags;
+#define RDMA_SQ_RDMA_WQE_DIF_BLOCK_SIZE_MASK 0x1
+#define RDMA_SQ_RDMA_WQE_DIF_BLOCK_SIZE_SHIFT 0
+#define RDMA_SQ_RDMA_WQE_DIF_FIRST_RDMA_IN_IO_FLG_MASK 0x1
+#define RDMA_SQ_RDMA_WQE_DIF_FIRST_RDMA_IN_IO_FLG_SHIFT 1
+#define RDMA_SQ_RDMA_WQE_DIF_LAST_RDMA_IN_IO_FLG_MASK 0x1
+#define RDMA_SQ_RDMA_WQE_DIF_LAST_RDMA_IN_IO_FLG_SHIFT 2
+#define RDMA_SQ_RDMA_WQE_RESERVED1_MASK 0x1F
+#define RDMA_SQ_RDMA_WQE_RESERVED1_SHIFT 3
+ u8 reserved2[3];
+};
+
+/* First element (16 bytes) of rdma wqe */
+struct rdma_sq_rdma_wqe_1st {
+ __le32 imm_data;
+ __le32 length;
+ __le32 xrc_srq;
+ u8 req_type;
+ u8 flags;
+#define RDMA_SQ_RDMA_WQE_1ST_COMP_FLG_MASK 0x1
+#define RDMA_SQ_RDMA_WQE_1ST_COMP_FLG_SHIFT 0
+#define RDMA_SQ_RDMA_WQE_1ST_RD_FENCE_FLG_MASK 0x1
+#define RDMA_SQ_RDMA_WQE_1ST_RD_FENCE_FLG_SHIFT 1
+#define RDMA_SQ_RDMA_WQE_1ST_INV_FENCE_FLG_MASK 0x1
+#define RDMA_SQ_RDMA_WQE_1ST_INV_FENCE_FLG_SHIFT 2
+#define RDMA_SQ_RDMA_WQE_1ST_SE_FLG_MASK 0x1
+#define RDMA_SQ_RDMA_WQE_1ST_SE_FLG_SHIFT 3
+#define RDMA_SQ_RDMA_WQE_1ST_INLINE_FLG_MASK 0x1
+#define RDMA_SQ_RDMA_WQE_1ST_INLINE_FLG_SHIFT 4
+#define RDMA_SQ_RDMA_WQE_1ST_DIF_ON_HOST_FLG_MASK 0x1
+#define RDMA_SQ_RDMA_WQE_1ST_DIF_ON_HOST_FLG_SHIFT 5
+#define RDMA_SQ_RDMA_WQE_1ST_RESERVED0_MASK 0x3
+#define RDMA_SQ_RDMA_WQE_1ST_RESERVED0_SHIFT 6
+ u8 wqe_size;
+ u8 prev_wqe_size;
+};
+
+/* Second element (16 bytes) of rdma wqe */
+struct rdma_sq_rdma_wqe_2nd {
+ struct regpair remote_va;
+ __le32 r_key;
+ u8 dif_flags;
+#define RDMA_SQ_RDMA_WQE_2ND_DIF_BLOCK_SIZE_MASK 0x1
+#define RDMA_SQ_RDMA_WQE_2ND_DIF_BLOCK_SIZE_SHIFT 0
+#define RDMA_SQ_RDMA_WQE_2ND_DIF_FIRST_SEGMENT_FLG_MASK 0x1
+#define RDMA_SQ_RDMA_WQE_2ND_DIF_FIRST_SEGMENT_FLG_SHIFT 1
+#define RDMA_SQ_RDMA_WQE_2ND_DIF_LAST_SEGMENT_FLG_MASK 0x1
+#define RDMA_SQ_RDMA_WQE_2ND_DIF_LAST_SEGMENT_FLG_SHIFT 2
+#define RDMA_SQ_RDMA_WQE_2ND_RESERVED1_MASK 0x1F
+#define RDMA_SQ_RDMA_WQE_2ND_RESERVED1_SHIFT 3
+ u8 reserved2[3];
+};
+
+/* SQ WQE req type enumeration */
+enum rdma_sq_req_type {
+ RDMA_SQ_REQ_TYPE_SEND,
+ RDMA_SQ_REQ_TYPE_SEND_WITH_IMM,
+ RDMA_SQ_REQ_TYPE_SEND_WITH_INVALIDATE,
+ RDMA_SQ_REQ_TYPE_RDMA_WR,
+ RDMA_SQ_REQ_TYPE_RDMA_WR_WITH_IMM,
+ RDMA_SQ_REQ_TYPE_RDMA_RD,
+ RDMA_SQ_REQ_TYPE_ATOMIC_CMP_AND_SWAP,
+ RDMA_SQ_REQ_TYPE_ATOMIC_ADD,
+ RDMA_SQ_REQ_TYPE_LOCAL_INVALIDATE,
+ RDMA_SQ_REQ_TYPE_FAST_MR,
+ RDMA_SQ_REQ_TYPE_BIND,
+ RDMA_SQ_REQ_TYPE_INVALID,
+ MAX_RDMA_SQ_REQ_TYPE
+};
+
+struct rdma_sq_send_wqe {
+ __le32 inv_key_or_imm_data;
+ __le32 length;
+ __le32 xrc_srq;
+ u8 req_type;
+ u8 flags;
+#define RDMA_SQ_SEND_WQE_COMP_FLG_MASK 0x1
+#define RDMA_SQ_SEND_WQE_COMP_FLG_SHIFT 0
+#define RDMA_SQ_SEND_WQE_RD_FENCE_FLG_MASK 0x1
+#define RDMA_SQ_SEND_WQE_RD_FENCE_FLG_SHIFT 1
+#define RDMA_SQ_SEND_WQE_INV_FENCE_FLG_MASK 0x1
+#define RDMA_SQ_SEND_WQE_INV_FENCE_FLG_SHIFT 2
+#define RDMA_SQ_SEND_WQE_SE_FLG_MASK 0x1
+#define RDMA_SQ_SEND_WQE_SE_FLG_SHIFT 3
+#define RDMA_SQ_SEND_WQE_INLINE_FLG_MASK 0x1
+#define RDMA_SQ_SEND_WQE_INLINE_FLG_SHIFT 4
+#define RDMA_SQ_SEND_WQE_DIF_ON_HOST_FLG_MASK 0x1
+#define RDMA_SQ_SEND_WQE_DIF_ON_HOST_FLG_SHIFT 5
+#define RDMA_SQ_SEND_WQE_RESERVED0_MASK 0x3
+#define RDMA_SQ_SEND_WQE_RESERVED0_SHIFT 6
+ u8 wqe_size;
+ u8 prev_wqe_size;
+ __le32 reserved1[4];
+};
+
+struct rdma_sq_send_wqe_1st {
+ __le32 inv_key_or_imm_data;
+ __le32 length;
+ __le32 xrc_srq;
+ u8 req_type;
+ u8 flags;
+#define RDMA_SQ_SEND_WQE_1ST_COMP_FLG_MASK 0x1
+#define RDMA_SQ_SEND_WQE_1ST_COMP_FLG_SHIFT 0
+#define RDMA_SQ_SEND_WQE_1ST_RD_FENCE_FLG_MASK 0x1
+#define RDMA_SQ_SEND_WQE_1ST_RD_FENCE_FLG_SHIFT 1
+#define RDMA_SQ_SEND_WQE_1ST_INV_FENCE_FLG_MASK 0x1
+#define RDMA_SQ_SEND_WQE_1ST_INV_FENCE_FLG_SHIFT 2
+#define RDMA_SQ_SEND_WQE_1ST_SE_FLG_MASK 0x1
+#define RDMA_SQ_SEND_WQE_1ST_SE_FLG_SHIFT 3
+#define RDMA_SQ_SEND_WQE_1ST_INLINE_FLG_MASK 0x1
+#define RDMA_SQ_SEND_WQE_1ST_INLINE_FLG_SHIFT 4
+#define RDMA_SQ_SEND_WQE_1ST_RESERVED0_MASK 0x7
+#define RDMA_SQ_SEND_WQE_1ST_RESERVED0_SHIFT 5
+ u8 wqe_size;
+ u8 prev_wqe_size;
+};
+
+struct rdma_sq_send_wqe_2st {
+ __le32 reserved1[4];
+};
+
#endif /* __QED_HSI_RDMA__ */
diff --git a/drivers/infiniband/hw/qedr/verbs.c b/drivers/infiniband/hw/qedr/verbs.c
index feff3d2..30b71e3 100644
--- a/drivers/infiniband/hw/qedr/verbs.c
+++ b/drivers/infiniband/hw/qedr/verbs.c
@@ -2416,3 +2416,1032 @@ err1:
kfree(mr);
return ERR_PTR(rc);
}
+
+static inline int qedr_wq_is_full(struct qedr_qp_hwq_info *wq)
+{
+ return (((wq->prod + 1) % wq->max_wr) == wq->cons);
+}
+
+static int sge_data_len(struct ib_sge *sg_list, int num_sge)
+{
+ int i, len = 0;
+
+ for (i = 0; i < num_sge; i++)
+ len += sg_list[i].length;
+
+ return len;
+}
+
+static void swap_wqe_data64(u64 *p)
+{
+ int i;
+
+ for (i = 0; i < QEDR_SQE_ELEMENT_SIZE / sizeof(u64); i++, p++)
+ *p = cpu_to_be64(cpu_to_le64(*p));
+}
+
+static u32 qedr_prepare_sq_inline_data(struct qedr_dev *dev,
+ struct qedr_qp *qp, u8 *wqe_size,
+ struct ib_send_wr *wr,
+ struct ib_send_wr **bad_wr, u8 *bits,
+ u8 bit)
+{
+ u32 data_size = sge_data_len(wr->sg_list, wr->num_sge);
+ char *seg_prt, *wqe;
+ int i, seg_siz;
+
+ if (data_size > ROCE_REQ_MAX_INLINE_DATA_SIZE) {
+ DP_ERR(dev, "Too much inline data in WR: %d\n", data_size);
+ *bad_wr = wr;
+ return 0;
+ }
+
+ if (!data_size)
+ return data_size;
+
+ *bits |= bit;
+
+ seg_prt = NULL;
+ wqe = NULL;
+ seg_siz = 0;
+
+ /* Copy data inline */
+ for (i = 0; i < wr->num_sge; i++) {
+ u32 len = wr->sg_list[i].length;
+ void *src = (void *)(uintptr_t)wr->sg_list[i].addr;
+
+ while (len > 0) {
+ u32 cur;
+
+ /* New segment required */
+ if (!seg_siz) {
+ wqe = (char *)qed_chain_produce(&qp->sq.pbl);
+ seg_prt = wqe;
+ seg_siz = sizeof(struct rdma_sq_common_wqe);
+ (*wqe_size)++;
+ }
+
+ /* Calculate currently allowed length */
+ cur = min_t(u32, len, seg_siz);
+ memcpy(seg_prt, src, cur);
+
+ /* Update segment variables */
+ seg_prt += cur;
+ seg_siz -= cur;
+
+ /* Update sge variables */
+ src += cur;
+ len -= cur;
+
+ /* Wwap fully-completed segments */
+ if (!seg_siz)
+ swap_wqe_data64((u64 *)wqe);
+ }
+ }
+
+ /* swap last not completed segment */
+ if (seg_siz)
+ swap_wqe_data64((u64 *)wqe);
+
+ return data_size;
+}
+
+#define RQ_SGE_SET(sge, vaddr, vlength, vflags) \
+ do { \
+ DMA_REGPAIR_LE(sge->addr, vaddr); \
+ (sge)->length = cpu_to_le32(vlength); \
+ (sge)->flags = cpu_to_le32(vflags); \
+ } while (0)
+
+#define SRQ_HDR_SET(hdr, vwr_id, num_sge) \
+ do { \
+ DMA_REGPAIR_LE(hdr->wr_id, vwr_id); \
+ (hdr)->num_sges = num_sge; \
+ } while (0)
+
+#define SRQ_SGE_SET(sge, vaddr, vlength, vlkey) \
+ do { \
+ DMA_REGPAIR_LE(sge->addr, vaddr); \
+ (sge)->length = cpu_to_le32(vlength); \
+ (sge)->l_key = cpu_to_le32(vlkey); \
+ } while (0)
+
+static u32 qedr_prepare_sq_sges(struct qedr_qp *qp, u8 *wqe_size,
+ struct ib_send_wr *wr)
+{
+ u32 data_size = 0;
+ int i;
+
+ for (i = 0; i < wr->num_sge; i++) {
+ struct rdma_sq_sge *sge = qed_chain_produce(&qp->sq.pbl);
+
+ DMA_REGPAIR_LE(sge->addr, wr->sg_list[i].addr);
+ sge->l_key = cpu_to_le32(wr->sg_list[i].lkey);
+ sge->length = cpu_to_le32(wr->sg_list[i].length);
+ data_size += wr->sg_list[i].length;
+ }
+
+ if (wqe_size)
+ *wqe_size += wr->num_sge;
+
+ return data_size;
+}
+
+static u32 qedr_prepare_sq_rdma_data(struct qedr_dev *dev,
+ struct qedr_qp *qp,
+ struct rdma_sq_rdma_wqe_1st *rwqe,
+ struct rdma_sq_rdma_wqe_2nd *rwqe2,
+ struct ib_send_wr *wr,
+ struct ib_send_wr **bad_wr)
+{
+ rwqe2->r_key = cpu_to_le32(rdma_wr(wr)->rkey);
+ DMA_REGPAIR_LE(rwqe2->remote_va, rdma_wr(wr)->remote_addr);
+
+ if (wr->send_flags & IB_SEND_INLINE) {
+ u8 flags = 0;
+
+ SET_FIELD2(flags, RDMA_SQ_RDMA_WQE_1ST_INLINE_FLG, 1);
+ return qedr_prepare_sq_inline_data(dev, qp, &rwqe->wqe_size, wr,
+ bad_wr, &rwqe->flags, flags);
+ }
+
+ return qedr_prepare_sq_sges(qp, &rwqe->wqe_size, wr);
+}
+
+static u32 qedr_prepare_sq_send_data(struct qedr_dev *dev,
+ struct qedr_qp *qp,
+ struct rdma_sq_send_wqe_1st *swqe,
+ struct rdma_sq_send_wqe_2st *swqe2,
+ struct ib_send_wr *wr,
+ struct ib_send_wr **bad_wr)
+{
+ memset(swqe2, 0, sizeof(*swqe2));
+ if (wr->send_flags & IB_SEND_INLINE) {
+ u8 flags = 0;
+
+ SET_FIELD2(flags, RDMA_SQ_SEND_WQE_INLINE_FLG, 1);
+ return qedr_prepare_sq_inline_data(dev, qp, &swqe->wqe_size, wr,
+ bad_wr, &swqe->flags, flags);
+ }
+
+ return qedr_prepare_sq_sges(qp, &swqe->wqe_size, wr);
+}
+
+static int qedr_prepare_reg(struct qedr_qp *qp,
+ struct rdma_sq_fmr_wqe_1st *fwqe1,
+ struct ib_reg_wr *wr)
+{
+ struct qedr_mr *mr = get_qedr_mr(wr->mr);
+ struct rdma_sq_fmr_wqe_2nd *fwqe2;
+
+ fwqe2 = (struct rdma_sq_fmr_wqe_2nd *)qed_chain_produce(&qp->sq.pbl);
+ fwqe1->addr.hi = upper_32_bits(mr->ibmr.iova);
+ fwqe1->addr.lo = lower_32_bits(mr->ibmr.iova);
+ fwqe1->l_key = wr->key;
+
+ SET_FIELD2(fwqe2->access_ctrl, RDMA_SQ_FMR_WQE_2ND_REMOTE_READ,
+ !!(wr->access & IB_ACCESS_REMOTE_READ));
+ SET_FIELD2(fwqe2->access_ctrl, RDMA_SQ_FMR_WQE_2ND_REMOTE_WRITE,
+ !!(wr->access & IB_ACCESS_REMOTE_WRITE));
+ SET_FIELD2(fwqe2->access_ctrl, RDMA_SQ_FMR_WQE_2ND_ENABLE_ATOMIC,
+ !!(wr->access & IB_ACCESS_REMOTE_ATOMIC));
+ SET_FIELD2(fwqe2->access_ctrl, RDMA_SQ_FMR_WQE_2ND_LOCAL_READ, 1);
+ SET_FIELD2(fwqe2->access_ctrl, RDMA_SQ_FMR_WQE_2ND_LOCAL_WRITE,
+ !!(wr->access & IB_ACCESS_LOCAL_WRITE));
+ fwqe2->fmr_ctrl = 0;
+
+ SET_FIELD2(fwqe2->fmr_ctrl, RDMA_SQ_FMR_WQE_2ND_PAGE_SIZE_LOG,
+ ilog2(mr->ibmr.page_size) - 12);
+
+ fwqe2->length_hi = 0;
+ fwqe2->length_lo = mr->ibmr.length;
+ fwqe2->pbl_addr.hi = upper_32_bits(mr->info.pbl_table->pa);
+ fwqe2->pbl_addr.lo = lower_32_bits(mr->info.pbl_table->pa);
+
+ qp->wqe_wr_id[qp->sq.prod].mr = mr;
+
+ return 0;
+}
+
+enum ib_wc_opcode qedr_ib_to_wc_opcode(enum ib_wr_opcode opcode)
+{
+ switch (opcode) {
+ case IB_WR_RDMA_WRITE:
+ case IB_WR_RDMA_WRITE_WITH_IMM:
+ return IB_WC_RDMA_WRITE;
+ case IB_WR_SEND_WITH_IMM:
+ case IB_WR_SEND:
+ case IB_WR_SEND_WITH_INV:
+ return IB_WC_SEND;
+ case IB_WR_RDMA_READ:
+ return IB_WC_RDMA_READ;
+ case IB_WR_ATOMIC_CMP_AND_SWP:
+ return IB_WC_COMP_SWAP;
+ case IB_WR_ATOMIC_FETCH_AND_ADD:
+ return IB_WC_FETCH_ADD;
+ case IB_WR_REG_MR:
+ return IB_WC_REG_MR;
+ case IB_WR_LOCAL_INV:
+ return IB_WC_LOCAL_INV;
+ default:
+ return IB_WC_SEND;
+ }
+}
+
+inline bool qedr_can_post_send(struct qedr_qp *qp, struct ib_send_wr *wr)
+{
+ int wq_is_full, err_wr, pbl_is_full;
+ struct qedr_dev *dev = qp->dev;
+
+ /* prevent SQ overflow and/or processing of a bad WR */
+ err_wr = wr->num_sge > qp->sq.max_sges;
+ wq_is_full = qedr_wq_is_full(&qp->sq);
+ pbl_is_full = qed_chain_get_elem_left_u32(&qp->sq.pbl) <
+ QEDR_MAX_SQE_ELEMENTS_PER_SQE;
+ if (wq_is_full || err_wr || pbl_is_full) {
+ if (wq_is_full && !(qp->err_bitmap & QEDR_QP_ERR_SQ_FULL)) {
+ DP_ERR(dev,
+ "error: WQ is full. Post send on QP %p failed (this error appears only once)\n",
+ qp);
+ qp->err_bitmap |= QEDR_QP_ERR_SQ_FULL;
+ }
+
+ if (err_wr && !(qp->err_bitmap & QEDR_QP_ERR_BAD_SR)) {
+ DP_ERR(dev,
+ "error: WR is bad. Post send on QP %p failed (this error appears only once)\n",
+ qp);
+ qp->err_bitmap |= QEDR_QP_ERR_BAD_SR;
+ }
+
+ if (pbl_is_full &&
+ !(qp->err_bitmap & QEDR_QP_ERR_SQ_PBL_FULL)) {
+ DP_ERR(dev,
+ "error: WQ PBL is full. Post send on QP %p failed (this error appears only once)\n",
+ qp);
+ qp->err_bitmap |= QEDR_QP_ERR_SQ_PBL_FULL;
+ }
+ return false;
+ }
+ return true;
+}
+
+int __qedr_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
+ struct ib_send_wr **bad_wr)
+{
+ struct qedr_dev *dev = get_qedr_dev(ibqp->device);
+ struct qedr_qp *qp = get_qedr_qp(ibqp);
+ struct rdma_sq_atomic_wqe_1st *awqe1;
+ struct rdma_sq_atomic_wqe_2nd *awqe2;
+ struct rdma_sq_atomic_wqe_3rd *awqe3;
+ struct rdma_sq_send_wqe_2st *swqe2;
+ struct rdma_sq_local_inv_wqe *iwqe;
+ struct rdma_sq_rdma_wqe_2nd *rwqe2;
+ struct rdma_sq_send_wqe_1st *swqe;
+ struct rdma_sq_rdma_wqe_1st *rwqe;
+ struct rdma_sq_fmr_wqe_1st *fwqe1;
+ struct rdma_sq_common_wqe *wqe;
+ u32 length;
+ int rc = 0;
+ bool comp;
+
+ if (!qedr_can_post_send(qp, wr)) {
+ *bad_wr = wr;
+ return -ENOMEM;
+ }
+
+ wqe = qed_chain_produce(&qp->sq.pbl);
+ qp->wqe_wr_id[qp->sq.prod].signaled =
+ !!(wr->send_flags & IB_SEND_SIGNALED) || qp->signaled;
+
+ wqe->flags = 0;
+ SET_FIELD2(wqe->flags, RDMA_SQ_SEND_WQE_SE_FLG,
+ !!(wr->send_flags & IB_SEND_SOLICITED));
+ comp = (!!(wr->send_flags & IB_SEND_SIGNALED)) || qp->signaled;
+ SET_FIELD2(wqe->flags, RDMA_SQ_SEND_WQE_COMP_FLG, comp);
+ SET_FIELD2(wqe->flags, RDMA_SQ_SEND_WQE_RD_FENCE_FLG,
+ !!(wr->send_flags & IB_SEND_FENCE));
+ wqe->prev_wqe_size = qp->prev_wqe_size;
+
+ qp->wqe_wr_id[qp->sq.prod].opcode = qedr_ib_to_wc_opcode(wr->opcode);
+
+ switch (wr->opcode) {
+ case IB_WR_SEND_WITH_IMM:
+ wqe->req_type = RDMA_SQ_REQ_TYPE_SEND_WITH_IMM;
+ swqe = (struct rdma_sq_send_wqe_1st *)wqe;
+ swqe->wqe_size = 2;
+ swqe2 = qed_chain_produce(&qp->sq.pbl);
+
+ swqe->inv_key_or_imm_data = cpu_to_le32(wr->ex.imm_data);
+ length = qedr_prepare_sq_send_data(dev, qp, swqe, swqe2,
+ wr, bad_wr);
+ swqe->length = cpu_to_le32(length);
+ qp->wqe_wr_id[qp->sq.prod].wqe_size = swqe->wqe_size;
+ qp->prev_wqe_size = swqe->wqe_size;
+ qp->wqe_wr_id[qp->sq.prod].bytes_len = swqe->length;
+ break;
+ case IB_WR_SEND:
+ wqe->req_type = RDMA_SQ_REQ_TYPE_SEND;
+ swqe = (struct rdma_sq_send_wqe_1st *)wqe;
+
+ swqe->wqe_size = 2;
+ swqe2 = qed_chain_produce(&qp->sq.pbl);
+ length = qedr_prepare_sq_send_data(dev, qp, swqe, swqe2,
+ wr, bad_wr);
+ swqe->length = cpu_to_le32(length);
+ qp->wqe_wr_id[qp->sq.prod].wqe_size = swqe->wqe_size;
+ qp->prev_wqe_size = swqe->wqe_size;
+ qp->wqe_wr_id[qp->sq.prod].bytes_len = swqe->length;
+ break;
+ case IB_WR_SEND_WITH_INV:
+ wqe->req_type = RDMA_SQ_REQ_TYPE_SEND_WITH_INVALIDATE;
+ swqe = (struct rdma_sq_send_wqe_1st *)wqe;
+ swqe2 = qed_chain_produce(&qp->sq.pbl);
+ swqe->wqe_size = 2;
+ swqe->inv_key_or_imm_data = cpu_to_le32(wr->ex.invalidate_rkey);
+ length = qedr_prepare_sq_send_data(dev, qp, swqe, swqe2,
+ wr, bad_wr);
+ swqe->length = cpu_to_le32(length);
+ qp->wqe_wr_id[qp->sq.prod].wqe_size = swqe->wqe_size;
+ qp->prev_wqe_size = swqe->wqe_size;
+ qp->wqe_wr_id[qp->sq.prod].bytes_len = swqe->length;
+ break;
+
+ case IB_WR_RDMA_WRITE_WITH_IMM:
+ wqe->req_type = RDMA_SQ_REQ_TYPE_RDMA_WR_WITH_IMM;
+ rwqe = (struct rdma_sq_rdma_wqe_1st *)wqe;
+
+ rwqe->wqe_size = 2;
+ rwqe->imm_data = htonl(cpu_to_le32(wr->ex.imm_data));
+ rwqe2 = qed_chain_produce(&qp->sq.pbl);
+ length = qedr_prepare_sq_rdma_data(dev, qp, rwqe, rwqe2,
+ wr, bad_wr);
+ rwqe->length = cpu_to_le32(length);
+ qp->wqe_wr_id[qp->sq.prod].wqe_size = rwqe->wqe_size;
+ qp->prev_wqe_size = rwqe->wqe_size;
+ qp->wqe_wr_id[qp->sq.prod].bytes_len = rwqe->length;
+ break;
+ case IB_WR_RDMA_WRITE:
+ wqe->req_type = RDMA_SQ_REQ_TYPE_RDMA_WR;
+ rwqe = (struct rdma_sq_rdma_wqe_1st *)wqe;
+
+ rwqe->wqe_size = 2;
+ rwqe2 = qed_chain_produce(&qp->sq.pbl);
+ length = qedr_prepare_sq_rdma_data(dev, qp, rwqe, rwqe2,
+ wr, bad_wr);
+ rwqe->length = cpu_to_le32(length);
+ qp->wqe_wr_id[qp->sq.prod].wqe_size = rwqe->wqe_size;
+ qp->prev_wqe_size = rwqe->wqe_size;
+ qp->wqe_wr_id[qp->sq.prod].bytes_len = rwqe->length;
+ break;
+ case IB_WR_RDMA_READ_WITH_INV:
+ DP_ERR(dev,
+ "RDMA READ WITH INVALIDATE not supported\n");
+ *bad_wr = wr;
+ rc = -EINVAL;
+ break;
+
+ case IB_WR_RDMA_READ:
+ wqe->req_type = RDMA_SQ_REQ_TYPE_RDMA_RD;
+ rwqe = (struct rdma_sq_rdma_wqe_1st *)wqe;
+
+ rwqe->wqe_size = 2;
+ rwqe2 = qed_chain_produce(&qp->sq.pbl);
+ length = qedr_prepare_sq_rdma_data(dev, qp, rwqe, rwqe2,
+ wr, bad_wr);
+ rwqe->length = cpu_to_le32(length);
+ qp->wqe_wr_id[qp->sq.prod].wqe_size = rwqe->wqe_size;
+ qp->prev_wqe_size = rwqe->wqe_size;
+ qp->wqe_wr_id[qp->sq.prod].bytes_len = rwqe->length;
+ break;
+
+ case IB_WR_ATOMIC_CMP_AND_SWP:
+ case IB_WR_ATOMIC_FETCH_AND_ADD:
+ awqe1 = (struct rdma_sq_atomic_wqe_1st *)wqe;
+ awqe1->wqe_size = 4;
+
+ awqe2 = qed_chain_produce(&qp->sq.pbl);
+ DMA_REGPAIR_LE(awqe2->remote_va, atomic_wr(wr)->remote_addr);
+ awqe2->r_key = cpu_to_le32(atomic_wr(wr)->rkey);
+
+ awqe3 = qed_chain_produce(&qp->sq.pbl);
+
+ if (wr->opcode == IB_WR_ATOMIC_FETCH_AND_ADD) {
+ wqe->req_type = RDMA_SQ_REQ_TYPE_ATOMIC_ADD;
+ DMA_REGPAIR_LE(awqe3->swap_data,
+ atomic_wr(wr)->compare_add);
+ } else {
+ wqe->req_type = RDMA_SQ_REQ_TYPE_ATOMIC_CMP_AND_SWAP;
+ DMA_REGPAIR_LE(awqe3->swap_data,
+ atomic_wr(wr)->swap);
+ DMA_REGPAIR_LE(awqe3->cmp_data,
+ atomic_wr(wr)->compare_add);
+ }
+
+ qedr_prepare_sq_sges(qp, NULL, wr);
+
+ qp->wqe_wr_id[qp->sq.prod].wqe_size = awqe1->wqe_size;
+ qp->prev_wqe_size = awqe1->wqe_size;
+ break;
+
+ case IB_WR_LOCAL_INV:
+ iwqe = (struct rdma_sq_local_inv_wqe *)wqe;
+ iwqe->wqe_size = 1;
+
+ iwqe->req_type = RDMA_SQ_REQ_TYPE_LOCAL_INVALIDATE;
+ iwqe->inv_l_key = wr->ex.invalidate_rkey;
+ qp->wqe_wr_id[qp->sq.prod].wqe_size = iwqe->wqe_size;
+ qp->prev_wqe_size = iwqe->wqe_size;
+ break;
+ case IB_WR_REG_MR:
+ DP_VERBOSE(dev, QEDR_MSG_CQ, "REG_MR\n");
+ wqe->req_type = RDMA_SQ_REQ_TYPE_FAST_MR;
+ fwqe1 = (struct rdma_sq_fmr_wqe_1st *)wqe;
+ fwqe1->wqe_size = 2;
+
+ rc = qedr_prepare_reg(qp, fwqe1, reg_wr(wr));
+ if (rc) {
+ DP_ERR(dev, "IB_REG_MR failed rc=%d\n", rc);
+ *bad_wr = wr;
+ break;
+ }
+
+ qp->wqe_wr_id[qp->sq.prod].wqe_size = fwqe1->wqe_size;
+ qp->prev_wqe_size = fwqe1->wqe_size;
+ break;
+ default:
+ DP_ERR(dev, "invalid opcode 0x%x!\n", wr->opcode);
+ rc = -EINVAL;
+ *bad_wr = wr;
+ break;
+ }
+
+ if (*bad_wr) {
+ u16 value;
+
+ /* Restore prod to its position before
+ * this WR was processed
+ */
+ value = le16_to_cpu(qp->sq.db_data.data.value);
+ qed_chain_set_prod(&qp->sq.pbl, value, wqe);
+
+ /* Restore prev_wqe_size */
+ qp->prev_wqe_size = wqe->prev_wqe_size;
+ rc = -EINVAL;
+ DP_ERR(dev, "POST SEND FAILED\n");
+ }
+
+ return rc;
+}
+
+int qedr_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
+ struct ib_send_wr **bad_wr)
+{
+ struct qedr_dev *dev = get_qedr_dev(ibqp->device);
+ struct qedr_qp *qp = get_qedr_qp(ibqp);
+ unsigned long flags;
+ int rc = 0;
+
+ *bad_wr = NULL;
+
+ spin_lock_irqsave(&qp->q_lock, flags);
+
+ if ((qp->state == QED_ROCE_QP_STATE_RESET) ||
+ (qp->state == QED_ROCE_QP_STATE_ERR)) {
+ spin_unlock_irqrestore(&qp->q_lock, flags);
+ *bad_wr = wr;
+ DP_VERBOSE(dev, QEDR_MSG_CQ,
+ "QP in wrong state! QP icid=0x%x state %d\n",
+ qp->icid, qp->state);
+ return -EINVAL;
+ }
+
+ if (!wr) {
+ DP_ERR(dev, "Got an empty post send.\n");
+ return -EINVAL;
+ }
+
+ while (wr) {
+ rc = __qedr_post_send(ibqp, wr, bad_wr);
+ if (rc)
+ break;
+
+ qp->wqe_wr_id[qp->sq.prod].wr_id = wr->wr_id;
+
+ qedr_inc_sw_prod(&qp->sq);
+
+ qp->sq.db_data.data.value++;
+
+ wr = wr->next;
+ }
+
+ /* Trigger doorbell
+ * If there was a failure in the first WR then it will be triggered in
+ * vane. However this is not harmful (as long as the producer value is
+ * unchanged). For performance reasons we avoid checking for this
+ * redundant doorbell.
+ */
+ wmb();
+ writel(qp->sq.db_data.raw, qp->sq.db);
+
+ /* Make sure write sticks */
+ mmiowb();
+
+ spin_unlock_irqrestore(&qp->q_lock, flags);
+
+ return rc;
+}
+
+int qedr_post_recv(struct ib_qp *ibqp, struct ib_recv_wr *wr,
+ struct ib_recv_wr **bad_wr)
+{
+ struct qedr_qp *qp = get_qedr_qp(ibqp);
+ struct qedr_dev *dev = qp->dev;
+ unsigned long flags;
+ int status = 0;
+
+ spin_lock_irqsave(&qp->q_lock, flags);
+
+ if ((qp->state == QED_ROCE_QP_STATE_RESET) ||
+ (qp->state == QED_ROCE_QP_STATE_ERR)) {
+ spin_unlock_irqrestore(&qp->q_lock, flags);
+ *bad_wr = wr;
+ return -EINVAL;
+ }
+
+ while (wr) {
+ int i;
+
+ if (qed_chain_get_elem_left_u32(&qp->rq.pbl) <
+ QEDR_MAX_RQE_ELEMENTS_PER_RQE ||
+ wr->num_sge > qp->rq.max_sges) {
+ DP_ERR(dev, "Can't post WR (%d < %d) || (%d > %d)\n",
+ qed_chain_get_elem_left_u32(&qp->rq.pbl),
+ QEDR_MAX_RQE_ELEMENTS_PER_RQE, wr->num_sge,
+ qp->rq.max_sges);
+ status = -ENOMEM;
+ *bad_wr = wr;
+ break;
+ }
+ for (i = 0; i < wr->num_sge; i++) {
+ u32 flags = 0;
+ struct rdma_rq_sge *rqe =
+ qed_chain_produce(&qp->rq.pbl);
+
+ /* First one must include the number
+ * of SGE in the list
+ */
+ if (!i)
+ SET_FIELD(flags, RDMA_RQ_SGE_NUM_SGES,
+ wr->num_sge);
+
+ SET_FIELD(flags, RDMA_RQ_SGE_L_KEY,
+ wr->sg_list[i].lkey);
+
+ RQ_SGE_SET(rqe, wr->sg_list[i].addr,
+ wr->sg_list[i].length, flags);
+ }
+
+ /* Special case of no sges. FW requires between 1-4 sges...
+ * in this case we need to post 1 sge with length zero. this is
+ * because rdma write with immediate consumes an RQ.
+ */
+ if (!wr->num_sge) {
+ u32 flags = 0;
+ struct rdma_rq_sge *rqe =
+ qed_chain_produce(&qp->rq.pbl);
+
+ /* First one must include the number
+ * of SGE in the list
+ */
+ SET_FIELD(flags, RDMA_RQ_SGE_L_KEY, 0);
+ SET_FIELD(flags, RDMA_RQ_SGE_NUM_SGES, 1);
+
+ RQ_SGE_SET(rqe, 0, 0, flags);
+ i = 1;
+ }
+
+ qp->rqe_wr_id[qp->rq.prod].wr_id = wr->wr_id;
+ qp->rqe_wr_id[qp->rq.prod].wqe_size = i;
+
+ qedr_inc_sw_prod(&qp->rq);
+
+ /* Flush all the writes before signalling doorbell */
+ wmb();
+
+ qp->rq.db_data.data.value++;
+
+ writel(qp->rq.db_data.raw, qp->rq.db);
+
+ /* Make sure write sticks */
+ mmiowb();
+
+ wr = wr->next;
+ }
+
+ spin_unlock_irqrestore(&qp->q_lock, flags);
+
+ return status;
+}
+
+static int is_valid_cqe(struct qedr_cq *cq, union rdma_cqe *cqe)
+{
+ struct rdma_cqe_requester *resp_cqe = &cqe->req;
+
+ return (resp_cqe->flags & RDMA_CQE_REQUESTER_TOGGLE_BIT_MASK) ==
+ cq->pbl_toggle;
+}
+
+static struct qedr_qp *cqe_get_qp(union rdma_cqe *cqe)
+{
+ struct rdma_cqe_requester *resp_cqe = &cqe->req;
+ struct qedr_qp *qp;
+
+ qp = (struct qedr_qp *)(uintptr_t)HILO_GEN(resp_cqe->qp_handle.hi,
+ resp_cqe->qp_handle.lo,
+ u64);
+ return qp;
+}
+
+static enum rdma_cqe_type cqe_get_type(union rdma_cqe *cqe)
+{
+ struct rdma_cqe_requester *resp_cqe = &cqe->req;
+
+ return GET_FIELD(resp_cqe->flags, RDMA_CQE_REQUESTER_TYPE);
+}
+
+/* Return latest CQE (needs processing) */
+static union rdma_cqe *get_cqe(struct qedr_cq *cq)
+{
+ return cq->latest_cqe;
+}
+
+/* In fmr we need to increase the number of fmr completed counter for the fmr
+ * algorithm determining whether we can free a pbl or not.
+ * we need to perform this whether the work request was signaled or not. for
+ * this purpose we call this function from the condition that checks if a wr
+ * should be skipped, to make sure we don't miss it ( possibly this fmr
+ * operation was not signalted)
+ */
+static inline void qedr_chk_if_fmr(struct qedr_qp *qp)
+{
+ if (qp->wqe_wr_id[qp->sq.cons].opcode == IB_WC_REG_MR)
+ qp->wqe_wr_id[qp->sq.cons].mr->info.completed++;
+}
+
+static int process_req(struct qedr_dev *dev, struct qedr_qp *qp,
+ struct qedr_cq *cq, int num_entries,
+ struct ib_wc *wc, u16 hw_cons, enum ib_wc_status status,
+ int force)
+{
+ u16 cnt = 0;
+
+ while (num_entries && qp->sq.wqe_cons != hw_cons) {
+ if (!qp->wqe_wr_id[qp->sq.cons].signaled && !force) {
+ qedr_chk_if_fmr(qp);
+ /* skip WC */
+ goto next_cqe;
+ }
+
+ /* fill WC */
+ wc->status = status;
+ wc->wc_flags = 0;
+ wc->src_qp = qp->id;
+ wc->qp = &qp->ibqp;
+
+ wc->wr_id = qp->wqe_wr_id[qp->sq.cons].wr_id;
+ wc->opcode = qp->wqe_wr_id[qp->sq.cons].opcode;
+
+ switch (wc->opcode) {
+ case IB_WC_RDMA_WRITE:
+ wc->byte_len = qp->wqe_wr_id[qp->sq.cons].bytes_len;
+ break;
+ case IB_WC_COMP_SWAP:
+ case IB_WC_FETCH_ADD:
+ wc->byte_len = 8;
+ break;
+ case IB_WC_REG_MR:
+ qp->wqe_wr_id[qp->sq.cons].mr->info.completed++;
+ break;
+ default:
+ break;
+ }
+
+ num_entries--;
+ wc++;
+ cnt++;
+next_cqe:
+ while (qp->wqe_wr_id[qp->sq.cons].wqe_size--)
+ qed_chain_consume(&qp->sq.pbl);
+ qedr_inc_sw_cons(&qp->sq);
+ }
+
+ return cnt;
+}
+
+static int qedr_poll_cq_req(struct qedr_dev *dev,
+ struct qedr_qp *qp, struct qedr_cq *cq,
+ int num_entries, struct ib_wc *wc,
+ struct rdma_cqe_requester *req)
+{
+ int cnt = 0;
+
+ switch (req->status) {
+ case RDMA_CQE_REQ_STS_OK:
+ cnt = process_req(dev, qp, cq, num_entries, wc, req->sq_cons,
+ IB_WC_SUCCESS, 0);
+ break;
+ case RDMA_CQE_REQ_STS_WORK_REQUEST_FLUSHED_ERR:
+ DP_ERR(dev,
+ "Error: POLL CQ with RDMA_CQE_REQ_STS_WORK_REQUEST_FLUSHED_ERR. CQ icid=0x%x, QP icid=0x%x\n",
+ cq->icid, qp->icid);
+ cnt = process_req(dev, qp, cq, num_entries, wc, req->sq_cons,
+ IB_WC_WR_FLUSH_ERR, 0);
+ break;
+ default:
+ /* process all WQE before the cosumer */
+ qp->state = QED_ROCE_QP_STATE_ERR;
+ cnt = process_req(dev, qp, cq, num_entries, wc,
+ req->sq_cons - 1, IB_WC_SUCCESS, 0);
+ wc += cnt;
+ /* if we have extra WC fill it with actual error info */
+ if (cnt < num_entries) {
+ enum ib_wc_status wc_status;
+
+ switch (req->status) {
+ case RDMA_CQE_REQ_STS_BAD_RESPONSE_ERR:
+ DP_ERR(dev,
+ "Error: POLL CQ with RDMA_CQE_REQ_STS_BAD_RESPONSE_ERR. CQ icid=0x%x, QP icid=0x%x\n",
+ cq->icid, qp->icid);
+ wc_status = IB_WC_BAD_RESP_ERR;
+ break;
+ case RDMA_CQE_REQ_STS_LOCAL_LENGTH_ERR:
+ DP_ERR(dev,
+ "Error: POLL CQ with RDMA_CQE_REQ_STS_LOCAL_LENGTH_ERR. CQ icid=0x%x, QP icid=0x%x\n",
+ cq->icid, qp->icid);
+ wc_status = IB_WC_LOC_LEN_ERR;
+ break;
+ case RDMA_CQE_REQ_STS_LOCAL_QP_OPERATION_ERR:
+ DP_ERR(dev,
+ "Error: POLL CQ with RDMA_CQE_REQ_STS_LOCAL_QP_OPERATION_ERR. CQ icid=0x%x, QP icid=0x%x\n",
+ cq->icid, qp->icid);
+ wc_status = IB_WC_LOC_QP_OP_ERR;
+ break;
+ case RDMA_CQE_REQ_STS_LOCAL_PROTECTION_ERR:
+ DP_ERR(dev,
+ "Error: POLL CQ with RDMA_CQE_REQ_STS_LOCAL_PROTECTION_ERR. CQ icid=0x%x, QP icid=0x%x\n",
+ cq->icid, qp->icid);
+ wc_status = IB_WC_LOC_PROT_ERR;
+ break;
+ case RDMA_CQE_REQ_STS_MEMORY_MGT_OPERATION_ERR:
+ DP_ERR(dev,
+ "Error: POLL CQ with RDMA_CQE_REQ_STS_MEMORY_MGT_OPERATION_ERR. CQ icid=0x%x, QP icid=0x%x\n",
+ cq->icid, qp->icid);
+ wc_status = IB_WC_MW_BIND_ERR;
+ break;
+ case RDMA_CQE_REQ_STS_REMOTE_INVALID_REQUEST_ERR:
+ DP_ERR(dev,
+ "Error: POLL CQ with RDMA_CQE_REQ_STS_REMOTE_INVALID_REQUEST_ERR. CQ icid=0x%x, QP icid=0x%x\n",
+ cq->icid, qp->icid);
+ wc_status = IB_WC_REM_INV_REQ_ERR;
+ break;
+ case RDMA_CQE_REQ_STS_REMOTE_ACCESS_ERR:
+ DP_ERR(dev,
+ "Error: POLL CQ with RDMA_CQE_REQ_STS_REMOTE_ACCESS_ERR. CQ icid=0x%x, QP icid=0x%x\n",
+ cq->icid, qp->icid);
+ wc_status = IB_WC_REM_ACCESS_ERR;
+ break;
+ case RDMA_CQE_REQ_STS_REMOTE_OPERATION_ERR:
+ DP_ERR(dev,
+ "Error: POLL CQ with RDMA_CQE_REQ_STS_REMOTE_OPERATION_ERR. CQ icid=0x%x, QP icid=0x%x\n",
+ cq->icid, qp->icid);
+ wc_status = IB_WC_REM_OP_ERR;
+ break;
+ case RDMA_CQE_REQ_STS_RNR_NAK_RETRY_CNT_ERR:
+ DP_ERR(dev,
+ "Error: POLL CQ with RDMA_CQE_REQ_STS_RNR_NAK_RETRY_CNT_ERR. CQ icid=0x%x, QP icid=0x%x\n",
+ cq->icid, qp->icid);
+ wc_status = IB_WC_RNR_RETRY_EXC_ERR;
+ break;
+ case RDMA_CQE_REQ_STS_TRANSPORT_RETRY_CNT_ERR:
+ DP_ERR(dev,
+ "Error: POLL CQ with ROCE_CQE_REQ_STS_TRANSPORT_RETRY_CNT_ERR. CQ icid=0x%x, QP icid=0x%x\n",
+ cq->icid, qp->icid);
+ wc_status = IB_WC_RETRY_EXC_ERR;
+ break;
+ default:
+ DP_ERR(dev,
+ "Error: POLL CQ with IB_WC_GENERAL_ERR. CQ icid=0x%x, QP icid=0x%x\n",
+ cq->icid, qp->icid);
+ wc_status = IB_WC_GENERAL_ERR;
+ }
+ cnt += process_req(dev, qp, cq, 1, wc, req->sq_cons,
+ wc_status, 1);
+ }
+ }
+
+ return cnt;
+}
+
+static void __process_resp_one(struct qedr_dev *dev, struct qedr_qp *qp,
+ struct qedr_cq *cq, struct ib_wc *wc,
+ struct rdma_cqe_responder *resp, u64 wr_id)
+{
+ enum ib_wc_status wc_status = IB_WC_SUCCESS;
+ u8 flags;
+
+ wc->opcode = IB_WC_RECV;
+ wc->wc_flags = 0;
+
+ switch (resp->status) {
+ case RDMA_CQE_RESP_STS_LOCAL_ACCESS_ERR:
+ wc_status = IB_WC_LOC_ACCESS_ERR;
+ break;
+ case RDMA_CQE_RESP_STS_LOCAL_LENGTH_ERR:
+ wc_status = IB_WC_LOC_LEN_ERR;
+ break;
+ case RDMA_CQE_RESP_STS_LOCAL_QP_OPERATION_ERR:
+ wc_status = IB_WC_LOC_QP_OP_ERR;
+ break;
+ case RDMA_CQE_RESP_STS_LOCAL_PROTECTION_ERR:
+ wc_status = IB_WC_LOC_PROT_ERR;
+ break;
+ case RDMA_CQE_RESP_STS_MEMORY_MGT_OPERATION_ERR:
+ wc_status = IB_WC_MW_BIND_ERR;
+ break;
+ case RDMA_CQE_RESP_STS_REMOTE_INVALID_REQUEST_ERR:
+ wc_status = IB_WC_REM_INV_RD_REQ_ERR;
+ break;
+ case RDMA_CQE_RESP_STS_OK:
+ wc_status = IB_WC_SUCCESS;
+ wc->byte_len = le32_to_cpu(resp->length);
+
+ flags = resp->flags & QEDR_RESP_RDMA_IMM;
+
+ switch (flags) {
+ case QEDR_RESP_RDMA_IMM:
+ /* update opcode */
+ wc->opcode = IB_WC_RECV_RDMA_WITH_IMM;
+ /* fall to set imm data */
+ case QEDR_RESP_IMM:
+ wc->ex.imm_data =
+ le32_to_cpu(resp->imm_data_or_inv_r_Key);
+ wc->wc_flags |= IB_WC_WITH_IMM;
+ break;
+ case QEDR_RESP_RDMA:
+ DP_ERR(dev, "Invalid flags detected\n");
+ break;
+ default:
+ DP_ERR(dev, "Invalid opcode\n");
+ }
+
+ break;
+ default:
+ wc->status = IB_WC_GENERAL_ERR;
+ DP_ERR(dev, "Invalid CQE status detected\n");
+ }
+
+ /* fill WC */
+ wc->status = wc_status;
+ wc->src_qp = qp->id;
+ wc->qp = &qp->ibqp;
+ wc->wr_id = wr_id;
+}
+
+static int process_resp_one(struct qedr_dev *dev, struct qedr_qp *qp,
+ struct qedr_cq *cq, struct ib_wc *wc,
+ struct rdma_cqe_responder *resp)
+{
+ u64 wr_id = qp->rqe_wr_id[qp->rq.cons].wr_id;
+
+ __process_resp_one(dev, qp, cq, wc, resp, wr_id);
+
+ while (qp->rqe_wr_id[qp->rq.cons].wqe_size--)
+ qed_chain_consume(&qp->rq.pbl);
+ qedr_inc_sw_cons(&qp->rq);
+
+ return 1;
+}
+
+static int process_resp_flush(struct qedr_qp *qp, struct qedr_cq *cq,
+ int num_entries, struct ib_wc *wc, u16 hw_cons)
+{
+ u16 cnt = 0;
+
+ while (num_entries && qp->rq.wqe_cons != hw_cons) {
+ /* fill WC */
+ wc->status = IB_WC_WR_FLUSH_ERR;
+ wc->wc_flags = 0;
+ wc->src_qp = qp->id;
+ wc->byte_len = 0;
+ wc->wr_id = qp->rqe_wr_id[qp->rq.cons].wr_id;
+ wc->qp = &qp->ibqp;
+ num_entries--;
+ wc++;
+ cnt++;
+ while (qp->rqe_wr_id[qp->rq.cons].wqe_size--)
+ qed_chain_consume(&qp->rq.pbl);
+ qedr_inc_sw_cons(&qp->rq);
+ }
+
+ return cnt;
+}
+
+static void try_consume_resp_cqe(struct qedr_cq *cq, struct qedr_qp *qp,
+ struct rdma_cqe_responder *resp, int *update)
+{
+ if (le16_to_cpu(resp->rq_cons) == qp->rq.wqe_cons) {
+ consume_cqe(cq);
+ *update |= 1;
+ }
+}
+
+static int qedr_poll_cq_resp(struct qedr_dev *dev, struct qedr_qp *qp,
+ struct qedr_cq *cq, int num_entries,
+ struct ib_wc *wc, struct rdma_cqe_responder *resp,
+ int *update)
+{
+ int cnt;
+
+ if (resp->status == RDMA_CQE_RESP_STS_WORK_REQUEST_FLUSHED_ERR) {
+ cnt = process_resp_flush(qp, cq, num_entries, wc,
+ resp->rq_cons);
+ try_consume_resp_cqe(cq, qp, resp, update);
+ } else {
+ cnt = process_resp_one(dev, qp, cq, wc, resp);
+ consume_cqe(cq);
+ *update |= 1;
+ }
+
+ return cnt;
+}
+
+static void try_consume_req_cqe(struct qedr_cq *cq, struct qedr_qp *qp,
+ struct rdma_cqe_requester *req, int *update)
+{
+ if (le16_to_cpu(req->sq_cons) == qp->sq.wqe_cons) {
+ consume_cqe(cq);
+ *update |= 1;
+ }
+}
+
+int qedr_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc)
+{
+ struct qedr_dev *dev = get_qedr_dev(ibcq->device);
+ struct qedr_cq *cq = get_qedr_cq(ibcq);
+ union rdma_cqe *cqe = cq->latest_cqe;
+ u32 old_cons, new_cons;
+ unsigned long flags;
+ int update = 0;
+ int done = 0;
+
+ spin_lock_irqsave(&cq->cq_lock, flags);
+ old_cons = qed_chain_get_cons_idx_u32(&cq->pbl);
+ while (num_entries && is_valid_cqe(cq, cqe)) {
+ struct qedr_qp *qp;
+ int cnt = 0;
+
+ /* prevent speculative reads of any field of CQE */
+ rmb();
+
+ qp = cqe_get_qp(cqe);
+ if (!qp) {
+ WARN(1, "Error: CQE QP pointer is NULL. CQE=%p\n", cqe);
+ break;
+ }
+
+ wc->qp = &qp->ibqp;
+
+ switch (cqe_get_type(cqe)) {
+ case RDMA_CQE_TYPE_REQUESTER:
+ cnt = qedr_poll_cq_req(dev, qp, cq, num_entries, wc,
+ &cqe->req);
+ try_consume_req_cqe(cq, qp, &cqe->req, &update);
+ break;
+ case RDMA_CQE_TYPE_RESPONDER_RQ:
+ cnt = qedr_poll_cq_resp(dev, qp, cq, num_entries, wc,
+ &cqe->resp, &update);
+ break;
+ case RDMA_CQE_TYPE_INVALID:
+ default:
+ DP_ERR(dev, "Error: invalid CQE type = %d\n",
+ cqe_get_type(cqe));
+ }
+ num_entries -= cnt;
+ wc += cnt;
+ done += cnt;
+
+ cqe = get_cqe(cq);
+ }
+ new_cons = qed_chain_get_cons_idx_u32(&cq->pbl);
+
+ cq->cq_cons += new_cons - old_cons;
+
+ if (update)
+ /* doorbell notifies abount latest VALID entry,
+ * but chain already point to the next INVALID one
+ */
+ doorbell_cq(cq, cq->cq_cons - 1, cq->arm_flags);
+
+ spin_unlock_irqrestore(&cq->cq_lock, flags);
+ return done;
+}
diff --git a/drivers/infiniband/hw/qedr/verbs.h b/drivers/infiniband/hw/qedr/verbs.h
index e7cb3fd..9eb2722b 100644
--- a/drivers/infiniband/hw/qedr/verbs.h
+++ b/drivers/infiniband/hw/qedr/verbs.h
@@ -58,4 +58,9 @@ int qedr_map_mr_sg(struct ib_mr *ibmr, struct scatterlist *sg,
struct ib_mr *qedr_alloc_mr(struct ib_pd *pd, enum ib_mr_type mr_type,
u32 max_num_sg);
+int qedr_poll_cq(struct ib_cq *, int num_entries, struct ib_wc *wc);
+int qedr_post_send(struct ib_qp *, struct ib_send_wr *,
+ struct ib_send_wr **bad_wr);
+int qedr_post_recv(struct ib_qp *, struct ib_recv_wr *,
+ struct ib_recv_wr **bad_wr);
#endif
--
1.8.3.1
^ permalink raw reply related
* [RFC v2 07/12] qedr: Add support for memory registeration verbs
From: Ram Amrani @ 2016-09-20 10:35 UTC (permalink / raw)
To: davem-fT/PcQaiUtIeIZ0/mPfg9Q, dledford-H+wXaHxf7aLQT0dZR+AlfA
Cc: Ariel.Elior-YGCgFSpz5w/QT0dZR+AlfA,
Michal.Kalderon-YGCgFSpz5w/QT0dZR+AlfA,
Yuval.Mintz-YGCgFSpz5w/QT0dZR+AlfA,
rajesh.borundia-YGCgFSpz5w/QT0dZR+AlfA,
linux-rdma-u79uwXL29TY76Z2rM5mHXA, netdev-u79uwXL29TY76Z2rM5mHXA,
Ram amrani
In-Reply-To: <1474367764-9555-1-git-send-email-Ram.Amrani-YGCgFSpz5w/QT0dZR+AlfA@public.gmane.org>
From: Ram amrani <Ram.Amrani-YGCgFSpz5w/QT0dZR+AlfA@public.gmane.org>
Add support for user, dma and memory regions registration.
Signed-off-by: Rajesh Borundia <rajesh.borundia-YGCgFSpz5w/QT0dZR+AlfA@public.gmane.org>
Signed-off-by: Ram Amrani <Ram.Amrani-YGCgFSpz5w/QT0dZR+AlfA@public.gmane.org>
---
drivers/infiniband/hw/qedr/main.c | 11 +-
drivers/infiniband/hw/qedr/qedr.h | 40 ++++
drivers/infiniband/hw/qedr/verbs.c | 365 +++++++++++++++++++++++++++++
drivers/infiniband/hw/qedr/verbs.h | 14 ++
drivers/net/ethernet/qlogic/qed/qed_roce.c | 240 +++++++++++++++++++
include/linux/qed/qed_roce_if.h | 6 +
6 files changed, 674 insertions(+), 2 deletions(-)
diff --git a/drivers/infiniband/hw/qedr/main.c b/drivers/infiniband/hw/qedr/main.c
index 10ad9ed..66d33fa 100644
--- a/drivers/infiniband/hw/qedr/main.c
+++ b/drivers/infiniband/hw/qedr/main.c
@@ -103,7 +103,9 @@ static int qedr_register_device(struct qedr_dev *dev)
QEDR_UVERBS(CREATE_QP) |
QEDR_UVERBS(MODIFY_QP) |
QEDR_UVERBS(QUERY_QP) |
- QEDR_UVERBS(DESTROY_QP);
+ QEDR_UVERBS(DESTROY_QP) |
+ QEDR_UVERBS(REG_MR) |
+ QEDR_UVERBS(DEREG_MR);
dev->ibdev.phys_port_cnt = 1;
dev->ibdev.num_comp_vectors = dev->num_cnq;
@@ -135,8 +137,13 @@ static int qedr_register_device(struct qedr_dev *dev)
dev->ibdev.query_pkey = qedr_query_pkey;
+ dev->ibdev.get_dma_mr = qedr_get_dma_mr;
+ dev->ibdev.dereg_mr = qedr_dereg_mr;
+ dev->ibdev.reg_user_mr = qedr_reg_user_mr;
+ dev->ibdev.alloc_mr = qedr_alloc_mr;
+ dev->ibdev.map_mr_sg = qedr_map_mr_sg;
+
dev->ibdev.dma_device = &dev->pdev->dev;
-
dev->ibdev.get_link_layer = qedr_link_layer;
dev->ibdev.get_dev_fw_str = qedr_get_dev_fw_str;
diff --git a/drivers/infiniband/hw/qedr/qedr.h b/drivers/infiniband/hw/qedr/qedr.h
index 05017be..c16844b 100644
--- a/drivers/infiniband/hw/qedr/qedr.h
+++ b/drivers/infiniband/hw/qedr/qedr.h
@@ -373,6 +373,41 @@ struct qedr_qp {
struct qedr_userq urq;
};
+struct qedr_ah {
+ struct ib_ah ibah;
+ struct ib_ah_attr attr;
+};
+
+enum qedr_mr_type {
+ QEDR_MR_USER,
+ QEDR_MR_KERNEL,
+ QEDR_MR_DMA,
+ QEDR_MR_FRMR
+};
+
+struct mr_info {
+ struct qedr_pbl *pbl_table;
+ struct qedr_pbl_info pbl_info;
+ struct list_head free_pbl_list;
+ struct list_head inuse_pbl_list;
+ u32 completed;
+ u32 completed_handled;
+};
+
+struct qedr_mr {
+ struct ib_mr ibmr;
+ struct ib_umem *umem;
+
+ struct qed_rdma_register_tid_in_params hw_mr;
+ enum qedr_mr_type type;
+
+ struct qedr_dev *dev;
+ struct mr_info info;
+
+ u64 *pages;
+ u32 npages;
+};
+
static inline int qedr_get_dmac(struct qedr_dev *dev,
struct ib_ah_attr *ah_attr, u8 *mac_addr)
{
@@ -417,4 +452,9 @@ static inline struct qedr_qp *get_qedr_qp(struct ib_qp *ibqp)
{
return container_of(ibqp, struct qedr_qp, ibqp);
}
+
+static inline struct qedr_mr *get_qedr_mr(struct ib_mr *ibmr)
+{
+ return container_of(ibmr, struct qedr_mr, ibmr);
+}
#endif
diff --git a/drivers/infiniband/hw/qedr/verbs.c b/drivers/infiniband/hw/qedr/verbs.c
index 616d656..feff3d2 100644
--- a/drivers/infiniband/hw/qedr/verbs.c
+++ b/drivers/infiniband/hw/qedr/verbs.c
@@ -2051,3 +2051,368 @@ int qedr_destroy_qp(struct ib_qp *ibqp)
return rc;
}
+
+static void free_mr_info(struct qedr_dev *dev, struct mr_info *info)
+{
+ struct qedr_pbl *pbl, *tmp;
+
+ if (info->pbl_table)
+ list_add_tail(&info->pbl_table->list_entry,
+ &info->free_pbl_list);
+
+ if (!list_empty(&info->inuse_pbl_list))
+ list_splice(&info->inuse_pbl_list, &info->free_pbl_list);
+
+ list_for_each_entry_safe(pbl, tmp, &info->free_pbl_list, list_entry) {
+ list_del(&pbl->list_entry);
+ qedr_free_pbl(dev, &info->pbl_info, pbl);
+ }
+}
+
+static int init_mr_info(struct qedr_dev *dev, struct mr_info *info,
+ size_t page_list_len, bool two_layered)
+{
+ struct qedr_pbl *tmp;
+ int rc;
+
+ INIT_LIST_HEAD(&info->free_pbl_list);
+ INIT_LIST_HEAD(&info->inuse_pbl_list);
+
+ rc = qedr_prepare_pbl_tbl(dev, &info->pbl_info,
+ page_list_len, two_layered);
+ if (rc)
+ goto done;
+
+ info->pbl_table = qedr_alloc_pbl_tbl(dev, &info->pbl_info, GFP_KERNEL);
+ if (!info->pbl_table) {
+ rc = -ENOMEM;
+ goto done;
+ }
+
+ DP_VERBOSE(dev, QEDR_MSG_MR, "pbl_table_pa = %pa\n",
+ &info->pbl_table->pa);
+
+ /* in usual case we use 2 PBLs, so we add one to free
+ * list and allocating another one
+ */
+ tmp = qedr_alloc_pbl_tbl(dev, &info->pbl_info, GFP_KERNEL);
+ if (!tmp) {
+ DP_VERBOSE(dev, QEDR_MSG_MR, "Extra PBL is not allocated\n");
+ goto done;
+ }
+
+ list_add_tail(&tmp->list_entry, &info->free_pbl_list);
+
+ DP_VERBOSE(dev, QEDR_MSG_MR, "extra pbl_table_pa = %pa\n", &tmp->pa);
+
+done:
+ if (rc)
+ free_mr_info(dev, info);
+
+ return rc;
+}
+
+struct ib_mr *qedr_reg_user_mr(struct ib_pd *ibpd, u64 start, u64 len,
+ u64 usr_addr, int acc, struct ib_udata *udata)
+{
+ struct qedr_dev *dev = get_qedr_dev(ibpd->device);
+ struct qedr_mr *mr;
+ struct qedr_pd *pd;
+ int rc = -ENOMEM;
+
+ pd = get_qedr_pd(ibpd);
+ DP_VERBOSE(dev, QEDR_MSG_MR,
+ "qedr_register user mr pd = %d start = %lld, len = %lld, usr_addr = %lld, acc = %d\n",
+ pd->pd_id, start, len, usr_addr, acc);
+
+ if (acc & IB_ACCESS_REMOTE_WRITE && !(acc & IB_ACCESS_LOCAL_WRITE))
+ return ERR_PTR(-EINVAL);
+
+ mr = kzalloc(sizeof(*mr), GFP_KERNEL);
+ if (!mr)
+ return ERR_PTR(rc);
+
+ mr->type = QEDR_MR_USER;
+
+ mr->umem = ib_umem_get(ibpd->uobject->context, start, len, acc, 0);
+ if (IS_ERR(mr->umem)) {
+ rc = -EFAULT;
+ goto err0;
+ }
+
+ rc = init_mr_info(dev, &mr->info, ib_umem_page_count(mr->umem), 1);
+ if (rc)
+ goto err1;
+
+ qedr_populate_pbls(dev, mr->umem, mr->info.pbl_table,
+ &mr->info.pbl_info);
+
+ rc = dev->ops->rdma_alloc_tid(dev->rdma_ctx, &mr->hw_mr.itid);
+ if (rc) {
+ DP_ERR(dev, "roce alloc tid returned an error %d\n", rc);
+ goto err1;
+ }
+
+ /* Index only, 18 bit long, lkey = itid << 8 | key */
+ mr->hw_mr.tid_type = QED_RDMA_TID_REGISTERED_MR;
+ mr->hw_mr.key = 0;
+ mr->hw_mr.pd = pd->pd_id;
+ mr->hw_mr.local_read = 1;
+ mr->hw_mr.local_write = (acc & IB_ACCESS_LOCAL_WRITE) ? 1 : 0;
+ mr->hw_mr.remote_read = (acc & IB_ACCESS_REMOTE_READ) ? 1 : 0;
+ mr->hw_mr.remote_write = (acc & IB_ACCESS_REMOTE_WRITE) ? 1 : 0;
+ mr->hw_mr.remote_atomic = (acc & IB_ACCESS_REMOTE_ATOMIC) ? 1 : 0;
+ mr->hw_mr.mw_bind = false;
+ mr->hw_mr.pbl_ptr = mr->info.pbl_table[0].pa;
+ mr->hw_mr.pbl_two_level = mr->info.pbl_info.two_layered;
+ mr->hw_mr.pbl_page_size_log = ilog2(mr->info.pbl_info.pbl_size);
+ mr->hw_mr.page_size_log = ilog2(mr->umem->page_size);
+ mr->hw_mr.fbo = ib_umem_offset(mr->umem);
+ mr->hw_mr.length = len;
+ mr->hw_mr.vaddr = usr_addr;
+ mr->hw_mr.zbva = false;
+ mr->hw_mr.phy_mr = false;
+ mr->hw_mr.dma_mr = false;
+
+ rc = dev->ops->rdma_register_tid(dev->rdma_ctx, &mr->hw_mr);
+ if (rc) {
+ DP_ERR(dev, "roce register tid returned an error %d\n", rc);
+ goto err2;
+ }
+
+ mr->ibmr.lkey = mr->hw_mr.itid << 8 | mr->hw_mr.key;
+ if (mr->hw_mr.remote_write || mr->hw_mr.remote_read ||
+ mr->hw_mr.remote_atomic)
+ mr->ibmr.rkey = mr->hw_mr.itid << 8 | mr->hw_mr.key;
+
+ DP_VERBOSE(dev, QEDR_MSG_MR, "register user mr lkey: %x\n",
+ mr->ibmr.lkey);
+ return &mr->ibmr;
+
+err2:
+ dev->ops->rdma_free_tid(dev->rdma_ctx, mr->hw_mr.itid);
+err1:
+ qedr_free_pbl(dev, &mr->info.pbl_info, mr->info.pbl_table);
+err0:
+ kfree(mr);
+ return ERR_PTR(rc);
+}
+
+int qedr_dereg_mr(struct ib_mr *ib_mr)
+{
+ struct qedr_mr *mr = get_qedr_mr(ib_mr);
+ struct qedr_dev *dev = get_qedr_dev(ib_mr->device);
+ int rc = 0;
+
+ rc = dev->ops->rdma_deregister_tid(dev->rdma_ctx, mr->hw_mr.itid);
+ if (rc)
+ return rc;
+
+ dev->ops->rdma_free_tid(dev->rdma_ctx, mr->hw_mr.itid);
+
+ if ((mr->type != QEDR_MR_DMA) && (mr->type != QEDR_MR_FRMR))
+ qedr_free_pbl(dev, &mr->info.pbl_info, mr->info.pbl_table);
+
+ /* it could be user registered memory. */
+ if (mr->umem)
+ ib_umem_release(mr->umem);
+
+ kfree(mr);
+
+ return rc;
+}
+
+struct qedr_mr *__qedr_alloc_mr(struct ib_pd *ibpd, int max_page_list_len)
+{
+ struct qedr_pd *pd = get_qedr_pd(ibpd);
+ struct qedr_dev *dev = get_qedr_dev(ibpd->device);
+ struct qedr_mr *mr;
+ int rc = -ENOMEM;
+
+ DP_VERBOSE(dev, QEDR_MSG_MR,
+ "qedr_alloc_frmr pd = %d max_page_list_len= %d\n", pd->pd_id,
+ max_page_list_len);
+
+ mr = kzalloc(sizeof(*mr), GFP_KERNEL);
+ if (!mr)
+ return ERR_PTR(rc);
+
+ mr->dev = dev;
+ mr->type = QEDR_MR_FRMR;
+
+ rc = init_mr_info(dev, &mr->info, max_page_list_len, 1);
+ if (rc)
+ goto err0;
+
+ rc = dev->ops->rdma_alloc_tid(dev->rdma_ctx, &mr->hw_mr.itid);
+ if (rc) {
+ DP_ERR(dev, "roce alloc tid returned an error %d\n", rc);
+ goto err0;
+ }
+
+ /* Index only, 18 bit long, lkey = itid << 8 | key */
+ mr->hw_mr.tid_type = QED_RDMA_TID_FMR;
+ mr->hw_mr.key = 0;
+ mr->hw_mr.pd = pd->pd_id;
+ mr->hw_mr.local_read = 1;
+ mr->hw_mr.local_write = 0;
+ mr->hw_mr.remote_read = 0;
+ mr->hw_mr.remote_write = 0;
+ mr->hw_mr.remote_atomic = 0;
+ mr->hw_mr.mw_bind = false;
+ mr->hw_mr.pbl_ptr = 0;
+ mr->hw_mr.pbl_two_level = mr->info.pbl_info.two_layered;
+ mr->hw_mr.pbl_page_size_log = ilog2(mr->info.pbl_info.pbl_size);
+ mr->hw_mr.fbo = 0;
+ mr->hw_mr.length = 0;
+ mr->hw_mr.vaddr = 0;
+ mr->hw_mr.zbva = false;
+ mr->hw_mr.phy_mr = true;
+ mr->hw_mr.dma_mr = false;
+
+ rc = dev->ops->rdma_register_tid(dev->rdma_ctx, &mr->hw_mr);
+ if (rc) {
+ DP_ERR(dev, "roce register tid returned an error %d\n", rc);
+ goto err1;
+ }
+
+ mr->ibmr.lkey = mr->hw_mr.itid << 8 | mr->hw_mr.key;
+ mr->ibmr.rkey = mr->ibmr.lkey;
+
+ DP_VERBOSE(dev, QEDR_MSG_MR, "alloc frmr: %x\n", mr->ibmr.lkey);
+ return mr;
+
+err1:
+ dev->ops->rdma_free_tid(dev->rdma_ctx, mr->hw_mr.itid);
+err0:
+ kfree(mr);
+ return ERR_PTR(rc);
+}
+
+struct ib_mr *qedr_alloc_mr(struct ib_pd *ibpd,
+ enum ib_mr_type mr_type, u32 max_num_sg)
+{
+ struct qedr_dev *dev;
+ struct qedr_mr *mr;
+
+ if (mr_type != IB_MR_TYPE_MEM_REG)
+ return ERR_PTR(-EINVAL);
+
+ mr = __qedr_alloc_mr(ibpd, max_num_sg);
+
+ if (IS_ERR(mr))
+ return ERR_PTR(-EINVAL);
+
+ dev = mr->dev;
+
+ return &mr->ibmr;
+}
+
+static int qedr_set_page(struct ib_mr *ibmr, u64 addr)
+{
+ struct qedr_mr *mr = get_qedr_mr(ibmr);
+ struct qedr_pbl *pbl_table;
+ struct regpair *pbe;
+ u32 pbes_in_page;
+
+ if (unlikely(mr->npages == mr->info.pbl_info.num_pbes)) {
+ DP_ERR(mr->dev, "qedr_set_page failes when %d\n", mr->npages);
+ return -ENOMEM;
+ }
+
+ DP_VERBOSE(mr->dev, QEDR_MSG_MR, "qedr_set_page pages[%d] = 0x%llx\n",
+ mr->npages, addr);
+
+ pbes_in_page = mr->info.pbl_info.pbl_size / sizeof(u64);
+ pbl_table = mr->info.pbl_table + (mr->npages / pbes_in_page);
+ pbe = (struct regpair *)pbl_table->va;
+ pbe += mr->npages % pbes_in_page;
+ pbe->lo = cpu_to_le32((u32)addr);
+ pbe->hi = cpu_to_le32((u32)upper_32_bits(addr));
+
+ mr->npages++;
+
+ return 0;
+}
+
+static void handle_completed_mrs(struct qedr_dev *dev, struct mr_info *info)
+{
+ int work = info->completed - info->completed_handled - 1;
+
+ DP_VERBOSE(dev, QEDR_MSG_MR, "Special FMR work = %d\n", work);
+ while (work-- > 0 && !list_empty(&info->inuse_pbl_list)) {
+ struct qedr_pbl *pbl;
+
+ /* Free all the page list that are possible to be freed
+ * (all the ones that were invalidated), under the assumption
+ * that if an FMR was completed successfully that means that
+ * if there was an invalidate operation before it also ended
+ */
+ pbl = list_first_entry(&info->inuse_pbl_list,
+ struct qedr_pbl, list_entry);
+ list_del(&pbl->list_entry);
+ list_add_tail(&pbl->list_entry, &info->free_pbl_list);
+ info->completed_handled++;
+ }
+}
+
+int qedr_map_mr_sg(struct ib_mr *ibmr, struct scatterlist *sg,
+ int sg_nents, unsigned int *sg_offset)
+{
+ struct qedr_mr *mr = get_qedr_mr(ibmr);
+
+ mr->npages = 0;
+
+ handle_completed_mrs(mr->dev, &mr->info);
+ return ib_sg_to_pages(ibmr, sg, sg_nents, NULL, qedr_set_page);
+}
+
+struct ib_mr *qedr_get_dma_mr(struct ib_pd *ibpd, int acc)
+{
+ struct qedr_dev *dev = get_qedr_dev(ibpd->device);
+ struct qedr_pd *pd = get_qedr_pd(ibpd);
+ struct qedr_mr *mr;
+ int rc;
+
+ mr = kzalloc(sizeof(*mr), GFP_KERNEL);
+ if (!mr)
+ return ERR_PTR(-ENOMEM);
+
+ mr->type = QEDR_MR_DMA;
+
+ rc = dev->ops->rdma_alloc_tid(dev->rdma_ctx, &mr->hw_mr.itid);
+ if (rc) {
+ DP_ERR(dev, "roce alloc tid returned an error %d\n", rc);
+ goto err1;
+ }
+
+ /* index only, 18 bit long, lkey = itid << 8 | key */
+ mr->hw_mr.tid_type = QED_RDMA_TID_REGISTERED_MR;
+ mr->hw_mr.pd = pd->pd_id;
+ mr->hw_mr.local_read = 1;
+ mr->hw_mr.local_write = (acc & IB_ACCESS_LOCAL_WRITE) ? 1 : 0;
+ mr->hw_mr.remote_read = (acc & IB_ACCESS_REMOTE_READ) ? 1 : 0;
+ mr->hw_mr.remote_write = (acc & IB_ACCESS_REMOTE_WRITE) ? 1 : 0;
+ mr->hw_mr.remote_atomic = (acc & IB_ACCESS_REMOTE_ATOMIC) ? 1 : 0;
+ mr->hw_mr.dma_mr = true;
+
+ rc = dev->ops->rdma_register_tid(dev->rdma_ctx, &mr->hw_mr);
+ if (rc) {
+ DP_ERR(dev, "roce register tid returned an error %d\n", rc);
+ goto err2;
+ }
+
+ mr->ibmr.lkey = mr->hw_mr.itid << 8 | mr->hw_mr.key;
+ if (mr->hw_mr.remote_write || mr->hw_mr.remote_read ||
+ mr->hw_mr.remote_atomic)
+ mr->ibmr.rkey = mr->hw_mr.itid << 8 | mr->hw_mr.key;
+
+ DP_VERBOSE(dev, QEDR_MSG_MR, "get dma mr: lkey = %x\n", mr->ibmr.lkey);
+ return &mr->ibmr;
+
+err2:
+ dev->ops->rdma_free_tid(dev->rdma_ctx, mr->hw_mr.itid);
+err1:
+ kfree(mr);
+ return ERR_PTR(rc);
+}
diff --git a/drivers/infiniband/hw/qedr/verbs.h b/drivers/infiniband/hw/qedr/verbs.h
index 80dbac1..e7cb3fd 100644
--- a/drivers/infiniband/hw/qedr/verbs.h
+++ b/drivers/infiniband/hw/qedr/verbs.h
@@ -44,4 +44,18 @@ int qedr_query_qp(struct ib_qp *, struct ib_qp_attr *qp_attr,
int qp_attr_mask, struct ib_qp_init_attr *);
int qedr_destroy_qp(struct ib_qp *ibqp);
+struct ib_ah *qedr_create_ah(struct ib_pd *ibpd, struct ib_ah_attr *attr);
+int qedr_destroy_ah(struct ib_ah *ibah);
+
+int qedr_dereg_mr(struct ib_mr *);
+struct ib_mr *qedr_get_dma_mr(struct ib_pd *, int acc);
+
+struct ib_mr *qedr_reg_user_mr(struct ib_pd *, u64 start, u64 length,
+ u64 virt, int acc, struct ib_udata *);
+
+int qedr_map_mr_sg(struct ib_mr *ibmr, struct scatterlist *sg,
+ int sg_nents, unsigned int *sg_offset);
+
+struct ib_mr *qedr_alloc_mr(struct ib_pd *pd, enum ib_mr_type mr_type,
+ u32 max_num_sg);
#endif
diff --git a/drivers/net/ethernet/qlogic/qed/qed_roce.c b/drivers/net/ethernet/qlogic/qed/qed_roce.c
index cf92ff7..04e33ee 100644
--- a/drivers/net/ethernet/qlogic/qed/qed_roce.c
+++ b/drivers/net/ethernet/qlogic/qed/qed_roce.c
@@ -690,6 +690,17 @@ struct qed_rdma_device *qed_rdma_query_device(void *rdma_cxt)
return p_hwfn->p_rdma_info->dev;
}
+void qed_rdma_free_tid(void *rdma_cxt, u32 itid)
+{
+ struct qed_hwfn *p_hwfn = (struct qed_hwfn *)rdma_cxt;
+
+ DP_VERBOSE(p_hwfn, QED_MSG_RDMA, "itid = %08x\n", itid);
+
+ spin_lock_bh(&p_hwfn->p_rdma_info->lock);
+ qed_bmap_release_id(p_hwfn, &p_hwfn->p_rdma_info->tid_map, itid);
+ spin_unlock_bh(&p_hwfn->p_rdma_info->lock);
+}
+
int qed_rdma_alloc_tid(void *rdma_cxt, u32 *itid)
{
struct qed_hwfn *p_hwfn = (struct qed_hwfn *)rdma_cxt;
@@ -2321,6 +2332,231 @@ int qed_rdma_modify_qp(void *rdma_cxt,
return rc;
}
+int qed_rdma_register_tid(void *rdma_cxt,
+ struct qed_rdma_register_tid_in_params *params)
+{
+ struct qed_hwfn *p_hwfn = (struct qed_hwfn *)rdma_cxt;
+ struct rdma_register_tid_ramrod_data *p_ramrod;
+ struct qed_sp_init_data init_data;
+ struct qed_spq_entry *p_ent;
+ enum rdma_tid_type tid_type;
+ u8 fw_return_code;
+ int rc;
+
+ DP_VERBOSE(p_hwfn, QED_MSG_RDMA, "itid = %08x\n", params->itid);
+
+ /* Get SPQ entry */
+ memset(&init_data, 0, sizeof(init_data));
+ init_data.opaque_fid = p_hwfn->hw_info.opaque_fid;
+ init_data.comp_mode = QED_SPQ_MODE_EBLOCK;
+
+ rc = qed_sp_init_request(p_hwfn, &p_ent, RDMA_RAMROD_REGISTER_MR,
+ p_hwfn->p_rdma_info->proto, &init_data);
+ if (rc) {
+ DP_VERBOSE(p_hwfn, QED_MSG_RDMA, "rc = %d\n", rc);
+ return rc;
+ }
+
+ if (p_hwfn->p_rdma_info->last_tid < params->itid)
+ p_hwfn->p_rdma_info->last_tid = params->itid;
+
+ p_ramrod = &p_ent->ramrod.rdma_register_tid;
+
+ p_ramrod->flags = 0;
+ SET_FIELD(p_ramrod->flags,
+ RDMA_REGISTER_TID_RAMROD_DATA_TWO_LEVEL_PBL,
+ params->pbl_two_level);
+
+ SET_FIELD(p_ramrod->flags,
+ RDMA_REGISTER_TID_RAMROD_DATA_ZERO_BASED, params->zbva);
+
+ SET_FIELD(p_ramrod->flags,
+ RDMA_REGISTER_TID_RAMROD_DATA_PHY_MR, params->phy_mr);
+
+ /* Don't initialize D/C field, as it may override other bits. */
+ if (!(params->tid_type == QED_RDMA_TID_FMR) && !(params->dma_mr))
+ SET_FIELD(p_ramrod->flags,
+ RDMA_REGISTER_TID_RAMROD_DATA_PAGE_SIZE_LOG,
+ params->page_size_log - 12);
+
+ SET_FIELD(p_ramrod->flags,
+ RDMA_REGISTER_TID_RAMROD_DATA_MAX_ID,
+ p_hwfn->p_rdma_info->last_tid);
+
+ SET_FIELD(p_ramrod->flags,
+ RDMA_REGISTER_TID_RAMROD_DATA_REMOTE_READ,
+ params->remote_read);
+
+ SET_FIELD(p_ramrod->flags,
+ RDMA_REGISTER_TID_RAMROD_DATA_REMOTE_WRITE,
+ params->remote_write);
+
+ SET_FIELD(p_ramrod->flags,
+ RDMA_REGISTER_TID_RAMROD_DATA_REMOTE_ATOMIC,
+ params->remote_atomic);
+
+ SET_FIELD(p_ramrod->flags,
+ RDMA_REGISTER_TID_RAMROD_DATA_LOCAL_WRITE,
+ params->local_write);
+
+ SET_FIELD(p_ramrod->flags,
+ RDMA_REGISTER_TID_RAMROD_DATA_LOCAL_READ, params->local_read);
+
+ SET_FIELD(p_ramrod->flags,
+ RDMA_REGISTER_TID_RAMROD_DATA_ENABLE_MW_BIND,
+ params->mw_bind);
+
+ SET_FIELD(p_ramrod->flags1,
+ RDMA_REGISTER_TID_RAMROD_DATA_PBL_PAGE_SIZE_LOG,
+ params->pbl_page_size_log - 12);
+
+ SET_FIELD(p_ramrod->flags2,
+ RDMA_REGISTER_TID_RAMROD_DATA_DMA_MR, params->dma_mr);
+
+ switch (params->tid_type) {
+ case QED_RDMA_TID_REGISTERED_MR:
+ tid_type = RDMA_TID_REGISTERED_MR;
+ break;
+ case QED_RDMA_TID_FMR:
+ tid_type = RDMA_TID_FMR;
+ break;
+ case QED_RDMA_TID_MW_TYPE1:
+ tid_type = RDMA_TID_MW_TYPE1;
+ break;
+ case QED_RDMA_TID_MW_TYPE2A:
+ tid_type = RDMA_TID_MW_TYPE2A;
+ break;
+ default:
+ rc = -EINVAL;
+ DP_VERBOSE(p_hwfn, QED_MSG_RDMA, "rc = %d\n", rc);
+ return rc;
+ }
+ SET_FIELD(p_ramrod->flags1,
+ RDMA_REGISTER_TID_RAMROD_DATA_TID_TYPE, tid_type);
+
+ p_ramrod->itid = cpu_to_le32(params->itid);
+ p_ramrod->key = params->key;
+ p_ramrod->pd = cpu_to_le16(params->pd);
+ p_ramrod->length_hi = (u8)(params->length >> 32);
+ p_ramrod->length_lo = DMA_LO_LE(params->length);
+ if (params->zbva) {
+ /* Lower 32 bits of the registered MR address.
+ * In case of zero based MR, will hold FBO
+ */
+ p_ramrod->va.hi = 0;
+ p_ramrod->va.lo = cpu_to_le32(params->fbo);
+ } else {
+ DMA_REGPAIR_LE(p_ramrod->va, params->vaddr);
+ }
+ DMA_REGPAIR_LE(p_ramrod->pbl_base, params->pbl_ptr);
+
+ /* DIF */
+ if (params->dif_enabled) {
+ SET_FIELD(p_ramrod->flags2,
+ RDMA_REGISTER_TID_RAMROD_DATA_DIF_ON_HOST_FLG, 1);
+ DMA_REGPAIR_LE(p_ramrod->dif_error_addr,
+ params->dif_error_addr);
+ DMA_REGPAIR_LE(p_ramrod->dif_runt_addr, params->dif_runt_addr);
+ }
+
+ rc = qed_spq_post(p_hwfn, p_ent, &fw_return_code);
+
+ if (fw_return_code != RDMA_RETURN_OK) {
+ DP_NOTICE(p_hwfn, "fw_return_code = %d\n", fw_return_code);
+ return -EINVAL;
+ }
+
+ DP_VERBOSE(p_hwfn, QED_MSG_RDMA, "Register TID, rc = %d\n", rc);
+ return rc;
+}
+
+int qed_rdma_deregister_tid(void *rdma_cxt, u32 itid)
+{
+ struct qed_hwfn *p_hwfn = (struct qed_hwfn *)rdma_cxt;
+ struct rdma_deregister_tid_ramrod_data *p_ramrod;
+ struct qed_sp_init_data init_data;
+ struct qed_spq_entry *p_ent;
+ struct qed_ptt *p_ptt;
+ u8 fw_return_code;
+ int rc;
+
+ DP_VERBOSE(p_hwfn, QED_MSG_RDMA, "itid = %08x\n", itid);
+
+ /* Get SPQ entry */
+ memset(&init_data, 0, sizeof(init_data));
+ init_data.opaque_fid = p_hwfn->hw_info.opaque_fid;
+ init_data.comp_mode = QED_SPQ_MODE_EBLOCK;
+
+ rc = qed_sp_init_request(p_hwfn, &p_ent, RDMA_RAMROD_DEREGISTER_MR,
+ p_hwfn->p_rdma_info->proto, &init_data);
+ if (rc) {
+ DP_VERBOSE(p_hwfn, QED_MSG_RDMA, "rc = %d\n", rc);
+ return rc;
+ }
+
+ p_ramrod = &p_ent->ramrod.rdma_deregister_tid;
+ p_ramrod->itid = cpu_to_le32(itid);
+
+ rc = qed_spq_post(p_hwfn, p_ent, &fw_return_code);
+ if (rc) {
+ DP_VERBOSE(p_hwfn, QED_MSG_RDMA, "rc = %d\n", rc);
+ return rc;
+ }
+
+ if (fw_return_code == RDMA_RETURN_DEREGISTER_MR_BAD_STATE_ERR) {
+ DP_NOTICE(p_hwfn, "fw_return_code = %d\n", fw_return_code);
+ return -EINVAL;
+ } else if (fw_return_code == RDMA_RETURN_NIG_DRAIN_REQ) {
+ /* Bit indicating that the TID is in use and a nig drain is
+ * required before sending the ramrod again
+ */
+ p_ptt = qed_ptt_acquire(p_hwfn);
+ if (!p_ptt) {
+ rc = -EBUSY;
+ DP_VERBOSE(p_hwfn, QED_MSG_RDMA,
+ "Failed to acquire PTT\n");
+ return rc;
+ }
+
+ rc = qed_mcp_drain(p_hwfn, p_ptt);
+ if (rc) {
+ qed_ptt_release(p_hwfn, p_ptt);
+ DP_VERBOSE(p_hwfn, QED_MSG_RDMA,
+ "Drain failed\n");
+ return rc;
+ }
+
+ qed_ptt_release(p_hwfn, p_ptt);
+
+ /* Resend the ramrod */
+ rc = qed_sp_init_request(p_hwfn, &p_ent,
+ RDMA_RAMROD_DEREGISTER_MR,
+ p_hwfn->p_rdma_info->proto,
+ &init_data);
+ if (rc) {
+ DP_VERBOSE(p_hwfn, QED_MSG_RDMA,
+ "Failed to init sp-element\n");
+ return rc;
+ }
+
+ rc = qed_spq_post(p_hwfn, p_ent, &fw_return_code);
+ if (rc) {
+ DP_VERBOSE(p_hwfn, QED_MSG_RDMA,
+ "Ramrod failed\n");
+ return rc;
+ }
+
+ if (fw_return_code != RDMA_RETURN_OK) {
+ DP_NOTICE(p_hwfn, "fw_return_code = %d\n",
+ fw_return_code);
+ return rc;
+ }
+ }
+
+ DP_VERBOSE(p_hwfn, QED_MSG_RDMA, "De-registered TID, rc = %d\n", rc);
+ return rc;
+}
+
static void *qed_rdma_get_rdma_ctx(struct qed_dev *cdev)
{
return QED_LEADING_HWFN(cdev);
@@ -2419,6 +2655,10 @@ static const struct qed_rdma_ops qed_rdma_ops_pass = {
.rdma_modify_qp = &qed_rdma_modify_qp,
.rdma_query_qp = &qed_rdma_query_qp,
.rdma_destroy_qp = &qed_rdma_destroy_qp,
+ .rdma_alloc_tid = &qed_rdma_alloc_tid,
+ .rdma_free_tid = &qed_rdma_free_tid,
+ .rdma_register_tid = &qed_rdma_register_tid,
+ .rdma_deregister_tid = &qed_rdma_deregister_tid,
};
const struct qed_rdma_ops *qed_get_rdma_ops()
diff --git a/include/linux/qed/qed_roce_if.h b/include/linux/qed/qed_roce_if.h
index 02321e3..0b6df6e 100644
--- a/include/linux/qed/qed_roce_if.h
+++ b/include/linux/qed/qed_roce_if.h
@@ -512,6 +512,12 @@ struct qed_rdma_ops {
int (*rdma_query_qp)(void *rdma_cxt, struct qed_rdma_qp *qp,
struct qed_rdma_query_qp_out_params *oparams);
int (*rdma_destroy_qp)(void *rdma_cxt, struct qed_rdma_qp *qp);
+ int
+ (*rdma_register_tid)(void *rdma_cxt,
+ struct qed_rdma_register_tid_in_params *iparams);
+ int (*rdma_deregister_tid)(void *rdma_cxt, u32 itid);
+ int (*rdma_alloc_tid)(void *rdma_cxt, u32 *itid);
+ void (*rdma_free_tid)(void *rdma_cxt, u32 itid);
};
const struct qed_rdma_ops *qed_get_rdma_ops(void);
--
1.8.3.1
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
^ permalink raw reply related
page: next (older) | prev (newer) | latest
- recent:[subjects (threaded)|topics (new)|topics (active)]
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox