[PATCH 0/5] SRP kernel patches for kernel v5.14

[PATCH 0/5] SRP kernel patches for kernel v5.14

[Bart Van Assche]

Hi Jason,

Please consider the five patches in this series for kernel v5.14.

Thanks,

Bart.

Bart Van Assche (5):
  RDMA/ib_hdrs.h: Remove a superfluous cast
  RDMA/srp: Add more structure size checks
  RDMA/srp: Apply the __packed attribute to members instead of
    structures
  RDMA/srp: Fix a recently introduced memory leak
  RDMA/srp: Make struct scsi_cmnd and struct srp_request adjacent

 drivers/infiniband/ulp/srp/ib_srp.c | 154 ++++++++++++----------------
 drivers/infiniband/ulp/srp/ib_srp.h |   2 -
 include/rdma/ib_hdrs.h              |   2 +-
 include/scsi/srp.h                  |  26 ++---
 4 files changed, 76 insertions(+), 108 deletions(-)

[PATCH 1/5] RDMA/ib_hdrs.h: Remove a superfluous cast

[Bart Van Assche]

be16_to_cpu() returns a u16 value. Remove the superfluous u16 cast. That
cast was introduced by commit 7dafbab3753f ("IB/hfi1: Add functions to
parse BTH/IB headers").

Cc: Don Hiatt <don.hiatt@intel.com>
Cc: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Bart Van Assche <bvanassche@acm.org>
---
 include/rdma/ib_hdrs.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/include/rdma/ib_hdrs.h b/include/rdma/ib_hdrs.h
index 57c1ac881d08..82483120539f 100644
--- a/include/rdma/ib_hdrs.h
+++ b/include/rdma/ib_hdrs.h
@@ -208,7 +208,7 @@ static inline u8 ib_get_lver(struct ib_header *hdr)
 
 static inline u16 ib_get_len(struct ib_header *hdr)
 {
-	return (u16)(be16_to_cpu(hdr->lrh[2]));
+	return be16_to_cpu(hdr->lrh[2]);
 }
 
 static inline u32 ib_get_qkey(struct ib_other_headers *ohdr)

Re: [PATCH 1/5] RDMA/ib_hdrs.h: Remove a superfluous cast

[Leon Romanovsky]

On Tue, May 11, 2021 at 08:27:48PM -0700, Bart Van Assche wrote:
> be16_to_cpu() returns a u16 value. Remove the superfluous u16 cast. That
> cast was introduced by commit 7dafbab3753f ("IB/hfi1: Add functions to
> parse BTH/IB headers").
> 
> Cc: Don Hiatt <don.hiatt@intel.com>
> Cc: Dennis Dalessandro <dennis.dalessandro@intel.com>
> Signed-off-by: Bart Van Assche <bvanassche@acm.org>
> ---
>  include/rdma/ib_hdrs.h | 2 +-
>  1 file changed, 1 insertion(+), 1 deletion(-)
> 
> diff --git a/include/rdma/ib_hdrs.h b/include/rdma/ib_hdrs.h
> index 57c1ac881d08..82483120539f 100644
> --- a/include/rdma/ib_hdrs.h
> +++ b/include/rdma/ib_hdrs.h
> @@ -208,7 +208,7 @@ static inline u8 ib_get_lver(struct ib_header *hdr)
>  
>  static inline u16 ib_get_len(struct ib_header *hdr)
>  {
> -	return (u16)(be16_to_cpu(hdr->lrh[2]));
> +	return be16_to_cpu(hdr->lrh[2]);
>  }
>  
>  static inline u32 ib_get_qkey(struct ib_other_headers *ohdr)

It is unclear why this function in the header. It is called only once.

Thanks,
Reviewed-by: Leon Romanovsky <leonro@nvidia.com>

Re: [PATCH 1/5] RDMA/ib_hdrs.h: Remove a superfluous cast

[Bart Van Assche]

On 5/19/21 2:00 AM, Leon Romanovsky wrote:
> On Tue, May 11, 2021 at 08:27:48PM -0700, Bart Van Assche wrote:
>> diff --git a/include/rdma/ib_hdrs.h b/include/rdma/ib_hdrs.h
>> index 57c1ac881d08..82483120539f 100644
>> --- a/include/rdma/ib_hdrs.h
>> +++ b/include/rdma/ib_hdrs.h
>> @@ -208,7 +208,7 @@ static inline u8 ib_get_lver(struct ib_header *hdr)
>>  
>>  static inline u16 ib_get_len(struct ib_header *hdr)
>>  {
>> -	return (u16)(be16_to_cpu(hdr->lrh[2]));
>> +	return be16_to_cpu(hdr->lrh[2]);
>>  }
>>  
>>  static inline u32 ib_get_qkey(struct ib_other_headers *ohdr)
> 
> It is unclear why this function in the header. It is called only once.

That's a good point. I will move it into the .c file that uses this
function.

Thanks,

Bart.

[PATCH 2/5] RDMA/srp: Add more structure size checks

[Bart Van Assche]

Before modifying how the __packed attribute is used, add compile time
size checks for the structures that will be modified.

Cc: Nicolas Morey-Chaisemartin <nmoreychaisemartin@suse.com>
Signed-off-by: Bart Van Assche <bvanassche@acm.org>
---
 drivers/infiniband/ulp/srp/ib_srp.c | 5 ++++-
 1 file changed, 4 insertions(+), 1 deletion(-)

diff --git a/drivers/infiniband/ulp/srp/ib_srp.c b/drivers/infiniband/ulp/srp/ib_srp.c
index 31f8aa2c40ed..0f66bf015db6 100644
--- a/drivers/infiniband/ulp/srp/ib_srp.c
+++ b/drivers/infiniband/ulp/srp/ib_srp.c
@@ -4078,10 +4078,13 @@ static int __init srp_init_module(void)
 {
 	int ret;
 
+	BUILD_BUG_ON(sizeof(struct srp_aer_req) != 36);
+	BUILD_BUG_ON(sizeof(struct srp_cmd) != 48);
 	BUILD_BUG_ON(sizeof(struct srp_imm_buf) != 4);
+	BUILD_BUG_ON(sizeof(struct srp_indirect_buf) != 20);
 	BUILD_BUG_ON(sizeof(struct srp_login_req) != 64);
 	BUILD_BUG_ON(sizeof(struct srp_login_req_rdma) != 56);
-	BUILD_BUG_ON(sizeof(struct srp_cmd) != 48);
+	BUILD_BUG_ON(sizeof(struct srp_rsp) != 36);
 
 	if (srp_sg_tablesize) {
 		pr_warn("srp_sg_tablesize is deprecated, please use cmd_sg_entries\n");

Re: [PATCH 2/5] RDMA/srp: Add more structure size checks

[Leon Romanovsky]

On Tue, May 11, 2021 at 08:27:49PM -0700, Bart Van Assche wrote:
> Before modifying how the __packed attribute is used, add compile time
> size checks for the structures that will be modified.
> 
> Cc: Nicolas Morey-Chaisemartin <nmoreychaisemartin@suse.com>
> Signed-off-by: Bart Van Assche <bvanassche@acm.org>
> ---
>  drivers/infiniband/ulp/srp/ib_srp.c | 5 ++++-
>  1 file changed, 4 insertions(+), 1 deletion(-)
> 

Thanks,
Reviewed-by: Leon Romanovsky <leonro@nvidia.com>

[PATCH 3/5] RDMA/srp: Apply the __packed attribute to members instead of structures

[Bart Van Assche]

Applying the __packed attribute to an entire data structure results in
suboptimal code on architectures that do not support unaligned accesses.
Hence apply the __packed attribute only to those data members that are
not naturally aligned.

Cc: Nicolas Morey-Chaisemartin <nmoreychaisemartin@suse.com>
Signed-off-by: Bart Van Assche <bvanassche@acm.org>
---
 include/scsi/srp.h | 26 +++++++++-----------------
 1 file changed, 9 insertions(+), 17 deletions(-)

diff --git a/include/scsi/srp.h b/include/scsi/srp.h
index 177d8026e96f..84d2b5fc1409 100644
--- a/include/scsi/srp.h
+++ b/include/scsi/srp.h
@@ -107,10 +107,10 @@ struct srp_direct_buf {
  * having the 20-byte structure padded to 24 bytes on 64-bit architectures.
  */
 struct srp_indirect_buf {
-	struct srp_direct_buf	table_desc;
+	struct srp_direct_buf	table_desc __packed;
 	__be32			len;
-	struct srp_direct_buf	desc_list[];
-} __attribute__((packed));
+	struct srp_direct_buf	desc_list[] __packed;
+};
 
 /* Immediate data buffer descriptor as defined in SRP2. */
 struct srp_imm_buf {
@@ -175,13 +175,13 @@ struct srp_login_rsp {
 	u8	opcode;
 	u8	reserved1[3];
 	__be32	req_lim_delta;
-	u64	tag;
+	u64	tag __packed;
 	__be32	max_it_iu_len;
 	__be32	max_ti_iu_len;
 	__be16	buf_fmt;
 	u8	rsp_flags;
 	u8	reserved2[25];
-} __attribute__((packed));
+};
 
 struct srp_login_rej {
 	u8	opcode;
@@ -207,10 +207,6 @@ struct srp_t_logout {
 	u64	tag;
 };
 
-/*
- * We need the packed attribute because the SRP spec only aligns the
- * 8-byte LUN field to 4 bytes.
- */
 struct srp_tsk_mgmt {
 	u8	opcode;
 	u8	sol_not;
@@ -225,10 +221,6 @@ struct srp_tsk_mgmt {
 	u8	reserved5[8];
 };
 
-/*
- * We need the packed attribute because the SRP spec only aligns the
- * 8-byte LUN field to 4 bytes.
- */
 struct srp_cmd {
 	u8	opcode;
 	u8	sol_not;
@@ -266,7 +258,7 @@ struct srp_rsp {
 	u8	sol_not;
 	u8	reserved1[2];
 	__be32	req_lim_delta;
-	u64	tag;
+	u64	tag __packed;
 	u8	reserved2[2];
 	u8	flags;
 	u8	status;
@@ -275,7 +267,7 @@ struct srp_rsp {
 	__be32	sense_data_len;
 	__be32	resp_data_len;
 	u8	data[];
-} __attribute__((packed));
+};
 
 struct srp_cred_req {
 	u8	opcode;
@@ -301,13 +293,13 @@ struct srp_aer_req {
 	u8	sol_not;
 	u8	reserved[2];
 	__be32	req_lim_delta;
-	u64	tag;
+	u64	tag __packed;
 	u32	reserved2;
 	struct scsi_lun	lun;
 	__be32	sense_data_len;
 	u32	reserved3;
 	u8	sense_data[];
-} __attribute__((packed));
+};
 
 struct srp_aer_rsp {
 	u8	opcode;

Re: [PATCH 3/5] RDMA/srp: Apply the __packed attribute to members instead of structures

[Jason Gunthorpe]

On Tue, May 11, 2021 at 08:27:50PM -0700, Bart Van Assche wrote:
> @@ -107,10 +107,10 @@ struct srp_direct_buf {
>   * having the 20-byte structure padded to 24 bytes on 64-bit architectures.
>   */
>  struct srp_indirect_buf {
> -	struct srp_direct_buf	table_desc;
> +	struct srp_direct_buf	table_desc __packed;
>  	__be32			len;
> -	struct srp_direct_buf	desc_list[];
> -} __attribute__((packed));
> +	struct srp_direct_buf	desc_list[] __packed;
> +};
>  
>  /* Immediate data buffer descriptor as defined in SRP2. */
>  struct srp_imm_buf {
> @@ -175,13 +175,13 @@ struct srp_login_rsp {
>  	u8	opcode;
>  	u8	reserved1[3];
>  	__be32	req_lim_delta;
> -	u64	tag;
> +	u64	tag __packed;

What you really want is just something like this:

typedef u64 __attribute__((aligned(4))) compat_u64;

And then use that every place you have the u64 and forget about packed
entirely.

Except for a couple exceptions IBA mads are always aligned to 4 bytes,
only the 64 bit quantities are unaligned.

But really this whole thing should be replaced with the IBA_FIELD
macros like include/rdma/ibta_vol1_c12.h demos.

Then it would be sparse safe and obviously endian correct as well.

I suppose you are respinning this due to the other comments?

Jason

Re: [PATCH 3/5] RDMA/srp: Apply the __packed attribute to members instead of structures

[Bart Van Assche]

On 5/20/21 7:48 AM, Jason Gunthorpe wrote:
> On Tue, May 11, 2021 at 08:27:50PM -0700, Bart Van Assche wrote:
>> @@ -107,10 +107,10 @@ struct srp_direct_buf {
>>   * having the 20-byte structure padded to 24 bytes on 64-bit architectures.
>>   */
>>  struct srp_indirect_buf {
>> -	struct srp_direct_buf	table_desc;
>> +	struct srp_direct_buf	table_desc __packed;
>>  	__be32			len;
>> -	struct srp_direct_buf	desc_list[];
>> -} __attribute__((packed));
>> +	struct srp_direct_buf	desc_list[] __packed;
>> +};
>>  
>>  /* Immediate data buffer descriptor as defined in SRP2. */
>>  struct srp_imm_buf {
>> @@ -175,13 +175,13 @@ struct srp_login_rsp {
>>  	u8	opcode;
>>  	u8	reserved1[3];
>>  	__be32	req_lim_delta;
>> -	u64	tag;
>> +	u64	tag __packed;
> 
> What you really want is just something like this:
> 
> typedef u64 __attribute__((aligned(4))) compat_u64;
> 
> And then use that every place you have the u64 and forget about packed
> entirely.

Really? My understanding is that the aligned attribute can be used to
increase alignment of a structure member but not to decrease it. From
https://gcc.gnu.org/onlinedocs/gcc/Common-Variable-Attributes.html#Common-Variable-Attributes:
"When used on a struct, or struct member, the aligned attribute can only
increase the alignment; in order to decrease it, the packed attribute
must be specified as well."

Additionally, there is a recommendation in
Documentation/process/coding-style.rst not to introduce new typedefs.

> Except for a couple exceptions IBA mads are always aligned to 4 bytes,
> only the 64 bit quantities are unaligned.
> 
> But really this whole thing should be replaced with the IBA_FIELD
> macros like include/rdma/ibta_vol1_c12.h demos.
> 
> Then it would be sparse safe and obviously endian correct as well.

I prefer a structure over the IBA_FIELD macros so I will change __packed
into __packed __aligned(4).

Thanks,

Bart.

Re: [PATCH 3/5] RDMA/srp: Apply the __packed attribute to members instead of structures

[Jason Gunthorpe]

On Sun, May 23, 2021 at 08:41:25PM -0700, Bart Van Assche wrote:
> On 5/20/21 7:48 AM, Jason Gunthorpe wrote:
> > On Tue, May 11, 2021 at 08:27:50PM -0700, Bart Van Assche wrote:
> >> @@ -107,10 +107,10 @@ struct srp_direct_buf {
> >>   * having the 20-byte structure padded to 24 bytes on 64-bit architectures.
> >>   */
> >>  struct srp_indirect_buf {
> >> -	struct srp_direct_buf	table_desc;
> >> +	struct srp_direct_buf	table_desc __packed;
> >>  	__be32			len;
> >> -	struct srp_direct_buf	desc_list[];
> >> -} __attribute__((packed));
> >> +	struct srp_direct_buf	desc_list[] __packed;
> >> +};
> >>  
> >>  /* Immediate data buffer descriptor as defined in SRP2. */
> >>  struct srp_imm_buf {
> >> @@ -175,13 +175,13 @@ struct srp_login_rsp {
> >>  	u8	opcode;
> >>  	u8	reserved1[3];
> >>  	__be32	req_lim_delta;
> >> -	u64	tag;
> >> +	u64	tag __packed;
> > 
> > What you really want is just something like this:
> > 
> > typedef u64 __attribute__((aligned(4))) compat_u64;
> > 
> > And then use that every place you have the u64 and forget about packed
> > entirely.
> 
> Really? My understanding is that the aligned attribute can be used to
> increase alignment of a structure member but not to decrease it. 

I didn't test it, but that is pre-existing code in Linux.. Maybe it
doesn't work and/or needs packed too

> Additionally, there is a recommendation in
> Documentation/process/coding-style.rst not to introduce new typedefs.

That we tend to selective ignore <shrug>

> > Except for a couple exceptions IBA mads are always aligned to 4 bytes,
> > only the 64 bit quantities are unaligned.
> > 
> > But really this whole thing should be replaced with the IBA_FIELD
> > macros like include/rdma/ibta_vol1_c12.h demos.
> > 
> > Then it would be sparse safe and obviously endian correct as well.
> 
> I prefer a structure over the IBA_FIELD macros so I will change __packed
> into __packed __aligned(4).

IMHO the struct is alot worse due to lack of endian safety and
complexity in establishing the layout.

Jason

[PATCH 4/5] RDMA/srp: Fix a recently introduced memory leak

[Bart Van Assche]

Only allocate an mr_list if it will be used and if it will be freed.

Cc: Max Gurtovoy <maxg@mellanox.com>
Cc: Nicolas Morey-Chaisemartin <nmoreychaisemartin@suse.com>
Fixes: f273ad4f8d90 ("RDMA/srp: Remove support for FMR memory registration")
Signed-off-by: Bart Van Assche <bvanassche@acm.org>
---
 drivers/infiniband/ulp/srp/ib_srp.c | 11 ++++++-----
 1 file changed, 6 insertions(+), 5 deletions(-)

diff --git a/drivers/infiniband/ulp/srp/ib_srp.c b/drivers/infiniband/ulp/srp/ib_srp.c
index 0f66bf015db6..52db42af421b 100644
--- a/drivers/infiniband/ulp/srp/ib_srp.c
+++ b/drivers/infiniband/ulp/srp/ib_srp.c
@@ -1009,12 +1009,13 @@ static int srp_alloc_req_data(struct srp_rdma_ch *ch)
 
 	for (i = 0; i < target->req_ring_size; ++i) {
 		req = &ch->req_ring[i];
-		mr_list = kmalloc_array(target->mr_per_cmd, sizeof(void *),
-					GFP_KERNEL);
-		if (!mr_list)
-			goto out;
-		if (srp_dev->use_fast_reg)
+		if (srp_dev->use_fast_reg) {
+			mr_list = kmalloc_array(target->mr_per_cmd,
+						sizeof(void *), GFP_KERNEL);
+			if (!mr_list)
+				goto out;
 			req->fr_list = mr_list;
+		}
 		req->indirect_desc = kmalloc(target->indirect_size, GFP_KERNEL);
 		if (!req->indirect_desc)
 			goto out;

Re: [PATCH 4/5] RDMA/srp: Fix a recently introduced memory leak

[Max Gurtovoy]

On 5/12/2021 6:27 AM, Bart Van Assche wrote:
> Only allocate an mr_list if it will be used and if it will be freed.
>
> Cc: Max Gurtovoy <maxg@mellanox.com>
> Cc: Nicolas Morey-Chaisemartin <nmoreychaisemartin@suse.com>
> Fixes: f273ad4f8d90 ("RDMA/srp: Remove support for FMR memory registration")
> Signed-off-by: Bart Van Assche <bvanassche@acm.org>
> ---
>   drivers/infiniband/ulp/srp/ib_srp.c | 11 ++++++-----
>   1 file changed, 6 insertions(+), 5 deletions(-)
>
> diff --git a/drivers/infiniband/ulp/srp/ib_srp.c b/drivers/infiniband/ulp/srp/ib_srp.c
> index 0f66bf015db6..52db42af421b 100644
> --- a/drivers/infiniband/ulp/srp/ib_srp.c
> +++ b/drivers/infiniband/ulp/srp/ib_srp.c
> @@ -1009,12 +1009,13 @@ static int srp_alloc_req_data(struct srp_rdma_ch *ch)
>   
>   	for (i = 0; i < target->req_ring_size; ++i) {
>   		req = &ch->req_ring[i];
> -		mr_list = kmalloc_array(target->mr_per_cmd, sizeof(void *),
> -					GFP_KERNEL);
> -		if (!mr_list)
> -			goto out;
> -		if (srp_dev->use_fast_reg)
> +		if (srp_dev->use_fast_reg) {
> +			mr_list = kmalloc_array(target->mr_per_cmd,
> +						sizeof(void *), GFP_KERNEL);
> +			if (!mr_list)
> +				goto out;
>   			req->fr_list = mr_list;
> +		}
>   		req->indirect_desc = kmalloc(target->indirect_size, GFP_KERNEL);
>   		if (!req->indirect_desc)
>   			goto out;

Looks good,

maybe we can remove the local mr_list variable while we're here ?

otherwise,

Reviewed-by: Max Gurtovoy <mgurtovoy@nvidia.com>

Re: [PATCH 4/5] RDMA/srp: Fix a recently introduced memory leak

[Bart Van Assche]

On 5/12/21 1:15 AM, Max Gurtovoy wrote:
> maybe we can remove the local mr_list variable while we're here ?
> 
> otherwise,
> 
> Reviewed-by: Max Gurtovoy <mgurtovoy@nvidia.com>

Hi Max,

I will remove the mr_list variable if I have to repost this patch series.

Thanks,

Bart.

[PATCH 5/5] RDMA/srp: Make struct scsi_cmnd and struct srp_request adjacent

[Bart Van Assche]

Define .init_cmd_priv and .exit_cmd_priv callback functions in struct
scsi_host_template. Set .cmd_size such that the SCSI core allocates
per-command private data. Use scsi_cmd_priv() to access that private
data. Remove the req_ring pointer from struct srp_rdma_ch since it is
no longer necessary. Convert srp_alloc_req_data() and srp_free_req_data()
into functions that initialize one instance of the SRP-private command
data. This is a micro-optimization since this patch removes several
pointer dereferences from the hot path.

Note: due to commit e73a5e8e8003 ("scsi: core: Only return started requests
from scsi_host_find_tag()"), it is no longer necessary to protect the
completion path against duplicate responses.

Cc: Nicolas Morey-Chaisemartin <nmoreychaisemartin@suse.com>
Signed-off-by: Bart Van Assche <bvanassche@acm.org>
---
 drivers/infiniband/ulp/srp/ib_srp.c | 156 ++++++++++++----------------
 drivers/infiniband/ulp/srp/ib_srp.h |   2 -
 2 files changed, 65 insertions(+), 93 deletions(-)

diff --git a/drivers/infiniband/ulp/srp/ib_srp.c b/drivers/infiniband/ulp/srp/ib_srp.c
index 52db42af421b..773ac5929082 100644
--- a/drivers/infiniband/ulp/srp/ib_srp.c
+++ b/drivers/infiniband/ulp/srp/ib_srp.c
@@ -965,69 +965,55 @@ static void srp_disconnect_target(struct srp_target_port *target)
 	}
 }
 
-static void srp_free_req_data(struct srp_target_port *target,
-			      struct srp_rdma_ch *ch)
+static int srp_exit_cmd_priv(struct Scsi_Host *shost, struct scsi_cmnd *cmd)
 {
+	struct srp_target_port *target = host_to_target(shost);
 	struct srp_device *dev = target->srp_host->srp_dev;
 	struct ib_device *ibdev = dev->dev;
-	struct srp_request *req;
-	int i;
+	struct srp_request *req = scsi_cmd_priv(cmd);
 
-	if (!ch->req_ring)
-		return;
-
-	for (i = 0; i < target->req_ring_size; ++i) {
-		req = &ch->req_ring[i];
-		if (dev->use_fast_reg)
-			kfree(req->fr_list);
-		if (req->indirect_dma_addr) {
-			ib_dma_unmap_single(ibdev, req->indirect_dma_addr,
-					    target->indirect_size,
-					    DMA_TO_DEVICE);
-		}
-		kfree(req->indirect_desc);
+	if (dev->use_fast_reg)
+		kfree(req->fr_list);
+	if (req->indirect_dma_addr) {
+		ib_dma_unmap_single(ibdev, req->indirect_dma_addr,
+				    target->indirect_size,
+				    DMA_TO_DEVICE);
 	}
+	kfree(req->indirect_desc);
 
-	kfree(ch->req_ring);
-	ch->req_ring = NULL;
+	return 0;
 }
 
-static int srp_alloc_req_data(struct srp_rdma_ch *ch)
+static int srp_init_cmd_priv(struct Scsi_Host *shost, struct scsi_cmnd *cmd)
 {
-	struct srp_target_port *target = ch->target;
+	struct srp_target_port *target = host_to_target(shost);
 	struct srp_device *srp_dev = target->srp_host->srp_dev;
 	struct ib_device *ibdev = srp_dev->dev;
-	struct srp_request *req;
+	struct srp_request *req = scsi_cmd_priv(cmd);
 	void *mr_list;
 	dma_addr_t dma_addr;
-	int i, ret = -ENOMEM;
-
-	ch->req_ring = kcalloc(target->req_ring_size, sizeof(*ch->req_ring),
-			       GFP_KERNEL);
-	if (!ch->req_ring)
-		goto out;
+	int ret = -ENOMEM;
 
-	for (i = 0; i < target->req_ring_size; ++i) {
-		req = &ch->req_ring[i];
-		if (srp_dev->use_fast_reg) {
-			mr_list = kmalloc_array(target->mr_per_cmd,
-						sizeof(void *), GFP_KERNEL);
-			if (!mr_list)
-				goto out;
-			req->fr_list = mr_list;
-		}
-		req->indirect_desc = kmalloc(target->indirect_size, GFP_KERNEL);
-		if (!req->indirect_desc)
-			goto out;
-
-		dma_addr = ib_dma_map_single(ibdev, req->indirect_desc,
-					     target->indirect_size,
-					     DMA_TO_DEVICE);
-		if (ib_dma_mapping_error(ibdev, dma_addr))
+	if (srp_dev->use_fast_reg) {
+		mr_list = kmalloc_array(target->mr_per_cmd, sizeof(void *),
+					GFP_KERNEL);
+		if (!mr_list)
 			goto out;
+		req->fr_list = mr_list;
+	}
+	req->indirect_desc = kmalloc(target->indirect_size, GFP_KERNEL);
+	if (!req->indirect_desc)
+		goto out;
 
-		req->indirect_dma_addr = dma_addr;
+	dma_addr = ib_dma_map_single(ibdev, req->indirect_desc,
+				     target->indirect_size,
+				     DMA_TO_DEVICE);
+	if (ib_dma_mapping_error(ibdev, dma_addr)) {
+		srp_exit_cmd_priv(shost, cmd);
+		goto out;
 	}
+
+	req->indirect_dma_addr = dma_addr;
 	ret = 0;
 
 out:
@@ -1069,10 +1055,6 @@ static void srp_remove_target(struct srp_target_port *target)
 	}
 	cancel_work_sync(&target->tl_err_work);
 	srp_rport_put(target->rport);
-	for (i = 0; i < target->ch_count; i++) {
-		ch = &target->ch[i];
-		srp_free_req_data(target, ch);
-	}
 	kfree(target->ch);
 	target->ch = NULL;
 
@@ -1291,22 +1273,32 @@ static void srp_finish_req(struct srp_rdma_ch *ch, struct srp_request *req,
 	}
 }
 
-static void srp_terminate_io(struct srp_rport *rport)
+struct srp_terminate_context {
+	struct srp_target_port *srp_target;
+	int scsi_result;
+};
+
+static bool srp_terminate_cmd(struct scsi_cmnd *scmnd, void *context_ptr,
+			      bool reserved)
 {
-	struct srp_target_port *target = rport->lld_data;
-	struct srp_rdma_ch *ch;
-	int i, j;
+	struct srp_terminate_context *context = context_ptr;
+	struct srp_target_port *target = context->srp_target;
+	u32 tag = blk_mq_unique_tag(scmnd->request);
+	struct srp_rdma_ch *ch = &target->ch[blk_mq_unique_tag_to_hwq(tag)];
+	struct srp_request *req = scsi_cmd_priv(scmnd);
 
-	for (i = 0; i < target->ch_count; i++) {
-		ch = &target->ch[i];
+	srp_finish_req(ch, req, NULL, context->scsi_result);
 
-		for (j = 0; j < target->req_ring_size; ++j) {
-			struct srp_request *req = &ch->req_ring[j];
+	return true;
+}
 
-			srp_finish_req(ch, req, NULL,
-				       DID_TRANSPORT_FAILFAST << 16);
-		}
-	}
+static void srp_terminate_io(struct srp_rport *rport)
+{
+	struct srp_target_port *target = rport->lld_data;
+	struct srp_terminate_context context = { .srp_target = target,
+		.scsi_result = DID_TRANSPORT_FAILFAST << 16 };
+
+	scsi_host_busy_iter(target->scsi_host, srp_terminate_cmd, &context);
 }
 
 /* Calculate maximum initiator to target information unit length. */
@@ -1362,13 +1354,12 @@ static int srp_rport_reconnect(struct srp_rport *rport)
 		ch = &target->ch[i];
 		ret += srp_new_cm_id(ch);
 	}
-	for (i = 0; i < target->ch_count; i++) {
-		ch = &target->ch[i];
-		for (j = 0; j < target->req_ring_size; ++j) {
-			struct srp_request *req = &ch->req_ring[j];
+	{
+		struct srp_terminate_context context = {
+			.srp_target = target, .scsi_result = DID_RESET << 16};
 
-			srp_finish_req(ch, req, NULL, DID_RESET << 16);
-		}
+		scsi_host_busy_iter(target->scsi_host, srp_terminate_cmd,
+				    &context);
 	}
 	for (i = 0; i < target->ch_count; i++) {
 		ch = &target->ch[i];
@@ -1964,13 +1955,10 @@ static void srp_process_rsp(struct srp_rdma_ch *ch, struct srp_rsp *rsp)
 		spin_unlock_irqrestore(&ch->lock, flags);
 	} else {
 		scmnd = scsi_host_find_tag(target->scsi_host, rsp->tag);
-		if (scmnd && scmnd->host_scribble) {
-			req = (void *)scmnd->host_scribble;
+		if (scmnd) {
+			req = scsi_cmd_priv(scmnd);
 			scmnd = srp_claim_req(ch, req, NULL, scmnd);
 		} else {
-			scmnd = NULL;
-		}
-		if (!scmnd) {
 			shost_printk(KERN_ERR, target->scsi_host,
 				     "Null scmnd for RSP w/tag %#016llx received on ch %td / QP %#x\n",
 				     rsp->tag, ch - target->ch, ch->qp->qp_num);
@@ -2002,7 +1990,6 @@ static void srp_process_rsp(struct srp_rdma_ch *ch, struct srp_rsp *rsp)
 		srp_free_req(ch, req, scmnd,
 			     be32_to_cpu(rsp->req_lim_delta));
 
-		scmnd->host_scribble = NULL;
 		scmnd->scsi_done(scmnd);
 	}
 }
@@ -2170,13 +2157,12 @@ static int srp_queuecommand(struct Scsi_Host *shost, struct scsi_cmnd *scmnd)
 {
 	struct srp_target_port *target = host_to_target(shost);
 	struct srp_rdma_ch *ch;
-	struct srp_request *req;
+	struct srp_request *req = scsi_cmd_priv(scmnd);
 	struct srp_iu *iu;
 	struct srp_cmd *cmd;
 	struct ib_device *dev;
 	unsigned long flags;
 	u32 tag;
-	u16 idx;
 	int len, ret;
 
 	scmnd->result = srp_chkready(target->rport);
@@ -2186,10 +2172,6 @@ static int srp_queuecommand(struct Scsi_Host *shost, struct scsi_cmnd *scmnd)
 	WARN_ON_ONCE(scmnd->request->tag < 0);
 	tag = blk_mq_unique_tag(scmnd->request);
 	ch = &target->ch[blk_mq_unique_tag_to_hwq(tag)];
-	idx = blk_mq_unique_tag_to_tag(tag);
-	WARN_ONCE(idx >= target->req_ring_size, "%s: tag %#x: idx %d >= %d\n",
-		  dev_name(&shost->shost_gendev), tag, idx,
-		  target->req_ring_size);
 
 	spin_lock_irqsave(&ch->lock, flags);
 	iu = __srp_get_tx_iu(ch, SRP_IU_CMD);
@@ -2198,13 +2180,10 @@ static int srp_queuecommand(struct Scsi_Host *shost, struct scsi_cmnd *scmnd)
 	if (!iu)
 		goto err;
 
-	req = &ch->req_ring[idx];
 	dev = target->srp_host->srp_dev->dev;
 	ib_dma_sync_single_for_cpu(dev, iu->dma, ch->max_it_iu_len,
 				   DMA_TO_DEVICE);
 
-	scmnd->host_scribble = (void *) req;
-
 	cmd = iu->buf;
 	memset(cmd, 0, sizeof *cmd);
 
@@ -3085,6 +3064,8 @@ static struct scsi_host_template srp_template = {
 	.target_alloc			= srp_target_alloc,
 	.slave_configure		= srp_slave_configure,
 	.info				= srp_target_info,
+	.init_cmd_priv			= srp_init_cmd_priv,
+	.exit_cmd_priv			= srp_exit_cmd_priv,
 	.queuecommand			= srp_queuecommand,
 	.change_queue_depth             = srp_change_queue_depth,
 	.eh_timed_out			= srp_timed_out,
@@ -3098,6 +3079,7 @@ static struct scsi_host_template srp_template = {
 	.cmd_per_lun			= SRP_DEFAULT_CMD_SQ_SIZE,
 	.shost_attrs			= srp_host_attrs,
 	.track_queue_depth		= 1,
+	.cmd_size			= sizeof(struct srp_request),
 };
 
 static int srp_sdev_count(struct Scsi_Host *host)
@@ -3677,8 +3659,6 @@ static ssize_t srp_create_target(struct device *dev,
 	if (ret)
 		goto out;
 
-	target->req_ring_size = target->queue_size - SRP_TSK_MGMT_SQ_SIZE;
-
 	if (!srp_conn_unique(target->srp_host, target)) {
 		if (target->using_rdma_cm) {
 			shost_printk(KERN_INFO, target->scsi_host,
@@ -3781,10 +3761,6 @@ static ssize_t srp_create_target(struct device *dev,
 		if (ret)
 			goto err_disconnect;
 
-		ret = srp_alloc_req_data(ch);
-		if (ret)
-			goto err_disconnect;
-
 		ret = srp_connect_ch(ch, max_iu_len, multich);
 		if (ret) {
 			char dst[64];
@@ -3803,7 +3779,6 @@ static ssize_t srp_create_target(struct device *dev,
 				goto free_ch;
 			} else {
 				srp_free_ch_ib(target, ch);
-				srp_free_req_data(target, ch);
 				target->ch_count = ch - target->ch;
 				goto connected;
 			}
@@ -3864,7 +3839,6 @@ static ssize_t srp_create_target(struct device *dev,
 	for (i = 0; i < target->ch_count; i++) {
 		ch = &target->ch[i];
 		srp_free_ch_ib(target, ch);
-		srp_free_req_data(target, ch);
 	}
 
 	kfree(target->ch);
diff --git a/drivers/infiniband/ulp/srp/ib_srp.h b/drivers/infiniband/ulp/srp/ib_srp.h
index 6818cac0a3b7..abccddeea1e3 100644
--- a/drivers/infiniband/ulp/srp/ib_srp.h
+++ b/drivers/infiniband/ulp/srp/ib_srp.h
@@ -174,7 +174,6 @@ struct srp_rdma_ch {
 
 	struct srp_iu	      **tx_ring;
 	struct srp_iu	      **rx_ring;
-	struct srp_request     *req_ring;
 	int			comp_vector;
 
 	u64			tsk_mgmt_tag;
@@ -220,7 +219,6 @@ struct srp_target_port {
 	int			mr_pool_size;
 	int			mr_per_cmd;
 	int			queue_size;
-	int			req_ring_size;
 	int			comp_vector;
 	int			tl_retry_count;
 

Re: [PATCH 5/5] RDMA/srp: Make struct scsi_cmnd and struct srp_request adjacent

[Leon Romanovsky]

On Tue, May 11, 2021 at 08:27:52PM -0700, Bart Van Assche wrote:
> Define .init_cmd_priv and .exit_cmd_priv callback functions in struct
> scsi_host_template. Set .cmd_size such that the SCSI core allocates
> per-command private data. Use scsi_cmd_priv() to access that private
> data. Remove the req_ring pointer from struct srp_rdma_ch since it is
> no longer necessary. Convert srp_alloc_req_data() and srp_free_req_data()
> into functions that initialize one instance of the SRP-private command
> data. This is a micro-optimization since this patch removes several
> pointer dereferences from the hot path.
> 
> Note: due to commit e73a5e8e8003 ("scsi: core: Only return started requests
> from scsi_host_find_tag()"), it is no longer necessary to protect the
> completion path against duplicate responses.
> 
> Cc: Nicolas Morey-Chaisemartin <nmoreychaisemartin@suse.com>
> Signed-off-by: Bart Van Assche <bvanassche@acm.org>
> ---
>  drivers/infiniband/ulp/srp/ib_srp.c | 156 ++++++++++++----------------
>  drivers/infiniband/ulp/srp/ib_srp.h |   2 -
>  2 files changed, 65 insertions(+), 93 deletions(-)
> 
> diff --git a/drivers/infiniband/ulp/srp/ib_srp.c b/drivers/infiniband/ulp/srp/ib_srp.c
> index 52db42af421b..773ac5929082 100644
> --- a/drivers/infiniband/ulp/srp/ib_srp.c
> +++ b/drivers/infiniband/ulp/srp/ib_srp.c
> @@ -965,69 +965,55 @@ static void srp_disconnect_target(struct srp_target_port *target)
>  	}
>  }
>  
> -static void srp_free_req_data(struct srp_target_port *target,
> -			      struct srp_rdma_ch *ch)
> +static int srp_exit_cmd_priv(struct Scsi_Host *shost, struct scsi_cmnd *cmd)
>  {
> +	struct srp_target_port *target = host_to_target(shost);
>  	struct srp_device *dev = target->srp_host->srp_dev;
>  	struct ib_device *ibdev = dev->dev;
> -	struct srp_request *req;
> -	int i;
> +	struct srp_request *req = scsi_cmd_priv(cmd);
>  
> -	if (!ch->req_ring)
> -		return;
> -
> -	for (i = 0; i < target->req_ring_size; ++i) {
> -		req = &ch->req_ring[i];
> -		if (dev->use_fast_reg)
> -			kfree(req->fr_list);
> -		if (req->indirect_dma_addr) {
> -			ib_dma_unmap_single(ibdev, req->indirect_dma_addr,
> -					    target->indirect_size,
> -					    DMA_TO_DEVICE);
> -		}
> -		kfree(req->indirect_desc);
> +	if (dev->use_fast_reg)
> +		kfree(req->fr_list);

Isn't cleaner will be to ensure that fr_list is NULL for !dev->use_fast_reg path?
In patch #4 https://lore.kernel.org/linux-rdma/20210512032752.16611-5-bvanassche@acm.org

Thanks

Re: [PATCH 5/5] RDMA/srp: Make struct scsi_cmnd and struct srp_request adjacent

[Bart Van Assche]

On 5/19/21 2:09 AM, Leon Romanovsky wrote:
> On Tue, May 11, 2021 at 08:27:52PM -0700, Bart Van Assche wrote:
>> -static void srp_free_req_data(struct srp_target_port *target,
>> -			      struct srp_rdma_ch *ch)
>> +static int srp_exit_cmd_priv(struct Scsi_Host *shost, struct scsi_cmnd *cmd)
>>  {
>> +	struct srp_target_port *target = host_to_target(shost);
>>  	struct srp_device *dev = target->srp_host->srp_dev;
>>  	struct ib_device *ibdev = dev->dev;
>> -	struct srp_request *req;
>> -	int i;
>> +	struct srp_request *req = scsi_cmd_priv(cmd);
>>  
>> -	if (!ch->req_ring)
>> -		return;
>> -
>> -	for (i = 0; i < target->req_ring_size; ++i) {
>> -		req = &ch->req_ring[i];
>> -		if (dev->use_fast_reg)
>> -			kfree(req->fr_list);
>> -		if (req->indirect_dma_addr) {
>> -			ib_dma_unmap_single(ibdev, req->indirect_dma_addr,
>> -					    target->indirect_size,
>> -					    DMA_TO_DEVICE);
>> -		}
>> -		kfree(req->indirect_desc);
>> +	if (dev->use_fast_reg)
>> +		kfree(req->fr_list);
> 
> Isn't cleaner will be to ensure that fr_list is NULL for !dev->use_fast_reg path?
> In patch #4 https://lore.kernel.org/linux-rdma/20210512032752.16611-5-bvanassche@acm.org

Hi Leon,

I think that per-request private data is zero-initialized and hence that
it is not necessary to clear req->fr_list explicitly. blk_mq_alloc_rqs()
 passes __GFP_ZERO to alloc_pages_node(). blk_mq_alloc_rqs() does not
only allocate block layer requests (struct request) but also per-request
private data (set->cmd_size).

Thanks,

Bart.

Re: [PATCH 5/5] RDMA/srp: Make struct scsi_cmnd and struct srp_request adjacent

[Leon Romanovsky]

On Wed, May 19, 2021 at 08:13:09AM -0700, Bart Van Assche wrote:
> On 5/19/21 2:09 AM, Leon Romanovsky wrote:
> > On Tue, May 11, 2021 at 08:27:52PM -0700, Bart Van Assche wrote:
> >> -static void srp_free_req_data(struct srp_target_port *target,
> >> -			      struct srp_rdma_ch *ch)
> >> +static int srp_exit_cmd_priv(struct Scsi_Host *shost, struct scsi_cmnd *cmd)
> >>  {
> >> +	struct srp_target_port *target = host_to_target(shost);
> >>  	struct srp_device *dev = target->srp_host->srp_dev;
> >>  	struct ib_device *ibdev = dev->dev;
> >> -	struct srp_request *req;
> >> -	int i;
> >> +	struct srp_request *req = scsi_cmd_priv(cmd);
> >>  
> >> -	if (!ch->req_ring)
> >> -		return;
> >> -
> >> -	for (i = 0; i < target->req_ring_size; ++i) {
> >> -		req = &ch->req_ring[i];
> >> -		if (dev->use_fast_reg)
> >> -			kfree(req->fr_list);
> >> -		if (req->indirect_dma_addr) {
> >> -			ib_dma_unmap_single(ibdev, req->indirect_dma_addr,
> >> -					    target->indirect_size,
> >> -					    DMA_TO_DEVICE);
> >> -		}
> >> -		kfree(req->indirect_desc);
> >> +	if (dev->use_fast_reg)
> >> +		kfree(req->fr_list);
> > 
> > Isn't cleaner will be to ensure that fr_list is NULL for !dev->use_fast_reg path?
> > In patch #4 https://lore.kernel.org/linux-rdma/20210512032752.16611-5-bvanassche@acm.org
> 
> Hi Leon,
> 
> I think that per-request private data is zero-initialized and hence that
> it is not necessary to clear req->fr_list explicitly. blk_mq_alloc_rqs()
>  passes __GFP_ZERO to alloc_pages_node(). blk_mq_alloc_rqs() does not
> only allocate block layer requests (struct request) but also per-request
> private data (set->cmd_size).

So you don't need this "if (dev->use_fast_reg)" check.

Thanks

> 
> Thanks,
> 
> Bart.
> 
>