Re: SQ overflow seen running isert traffic with high block sizes

[Potnuri Bharat Teja <bharat@chelsio.com>]

Hi Nicholas, 
thanks for the suggestions. Comments below.

On Thursday, January 01/18/18, 2018 at 15:28:42 +0530, Nicholas A. Bellinger wrote:
> Hi Shiraz, Michal & Co,
> 
> Thanks for the feedback.  Comments below.
> 
> On Mon, 2018-01-15 at 09:22 -0600, Shiraz Saleem wrote:
> > On Mon, Jan 15, 2018 at 03:12:36AM -0700, Kalderon, Michal wrote:
> > > > From: linux-rdma-owner@vger.kernel.org [mailto:linux-rdma-
> > > > owner@vger.kernel.org] On Behalf Of Nicholas A. Bellinger
> > > > Sent: Monday, January 15, 2018 6:57 AM
> > > > To: Shiraz Saleem <shiraz.saleem@intel.com>
> > > > Cc: Amrani, Ram <Ram.Amrani@cavium.com>; Sagi Grimberg
> > > > <sagi@grimberg.me>; linux-rdma@vger.kernel.org; Elior, Ariel
> > > > <Ariel.Elior@cavium.com>; target-devel <target-devel@vger.kernel.org>;
> > > > Potnuri Bharat Teja <bharat@chelsio.com>
> > > > Subject: Re: SQ overflow seen running isert traffic with high block sizes
> > > > 
> > > > Hi Shiraz, Ram, Ariel, & Potnuri,
> > > > 
> > > > Following up on this old thread, as it relates to Potnuri's recent fix for a iser-
> > > > target queue-full memory leak:
> > > > 
> > > > https://www.spinics.net/lists/target-devel/msg16282.html
> > > > 
> > > > Just curious how frequent this happens in practice with sustained large block
> > > > workloads, as it appears to effect at least three different iwarp RNICS (i40iw,
> > > > qedr and iw_cxgb4)..?
> > > > 
> > > > Is there anything else from an iser-target consumer level that should be
> > > > changed for iwarp to avoid repeated ib_post_send() failures..?
> > > > 
> > > Would like to mention, that although we are an iWARP RNIC as well, we've hit this
> > > Issue when running RoCE. It's not iWARP related. 
> > > This is easily reproduced within seconds with IO size of 5121K
> > > Using 5 Targets with 2 Ram Disk each and 5 targets with FileIO Disks each.
> > > 
> > > IO Command used:
> > > maim -b512k -T32 -t2 -Q8 -M0 -o -u -n -m17 -ftargets.dat -d1
> > > 
> > > thanks,
> > > Michal
> > 
> > Its seen with block size >= 2M on a single target 1 RAM disk config. And similar to Michals report;
> > rather quickly, in a matter of seconds.
> > 
> > fio --rw=read --bs=2048k --numjobs=1 --iodepth=128 --runtime=30 --size=20g --loops=1 --ioengine=libaio 
> > --direct=1 --invalidate=1 --fsync_on_close=1 --norandommap --exitall --filename=/dev/sdb --name=sdb 
> > 
> 
> A couple of thoughts.
> 
> First, would it be helpful to limit maximum payload size per I/O for
> consumers based on number of iser-target sq hw sges..?
yes, I think HW num sge needs to be propagated to iscsi target.
> 
> That is, if rdma_rw_ctx_post() -> ib_post_send() failures are related to
> maximum payload size per I/O being too large there is an existing

Yes they are IO size specific, I observed SQ overflow with fio for IO sizes above
256k and for READ tests only with chelsio(iw_cxgb4) adapters.
> target_core_fabric_ops mechanism for limiting using SCSI residuals,
> originally utilized by qla2xxx here:
> 
> target/qla2xxx: Honor max_data_sg_nents I/O transfer limit
> https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=8f9b565482c537821588444e09ff732c7d65ed6e
> 
> Note this patch also will return a smaller Block Limits VPD (0x86)
> MAXIMUM TRANSFER LENGTH based on max_data_sg_nents * PAGE_SIZE, which
> means for modern SCSI initiators honoring MAXIMUM TRANSFER LENGTH will
> automatically limit maximum outgoing payload transfer length, and avoid
> SCSI residual logic.
> 
> As-is, iser-target doesn't a propagate max_data_sg_ents limit into
> iscsi-target, but you can try testing with a smaller value to see if
> it's useful.  Eg:
> 
> diff --git a/drivers/target/iscsi/iscsi_target_configfs.c b/drivers/target/iscsi/iscsi_target_configf
> index 0ebc481..d8a4cc5 100644
> --- a/drivers/target/iscsi/iscsi_target_configfs.c
> +++ b/drivers/target/iscsi/iscsi_target_configfs.c
> @@ -1553,6 +1553,7 @@ static void lio_release_cmd(struct se_cmd *se_cmd)
>         .module                         = THIS_MODULE,
>         .name                           = "iscsi",
>         .node_acl_size                  = sizeof(struct iscsi_node_acl),
> +       .max_data_sg_nents              = 32, /* 32 * PAGE_SIZE = MAXIMUM TRANSFER LENGTH */
>         .get_fabric_name                = iscsi_get_fabric_name,
>         .tpg_get_wwn                    = lio_tpg_get_endpoint_wwn,
>         .tpg_get_tag                    = lio_tpg_get_tag,
> 
With above change, SQ overflow isn't observed. I started of with max_data_sg_nents = 16.
> Second, if the failures are not SCSI transfer length specific, another
> option would be to limit the total command sequence number depth (CmdSN)
> per session.
> 
> This is controlled at runtime by default_cmdsn_depth TPG attribute:
> 
> /sys/kernel/config/target/iscsi/$TARGET_IQN/$TPG/attrib/default_cmdsn_depth
> 
> and on per initiator context with cmdsn_depth NodeACL attribute:
> 
> /sys/kernel/config/target/iscsi/$TARGET_IQN/$TPG/acls/$ACL_IQN/cmdsn_depth
> 
> Note these default to 64, and can be changed at build time via
> include/target/iscsi/iscsi_target_core.h:TA_DEFAULT_CMDSN_DEPTH.
> 
> That said, Sagi, any further comments as what else iser-target should be
> doing to avoid repeated queue-fulls with limited hw sges..?
>