On Mon, Feb 13, 2017 at 09:19:54PM -0500, Laurence Oberman wrote: > > > ----- Original Message ----- > > From: "Bart Van Assche" <Bart.VanAssche@xxxxxxxxxxx> > > To: leon@xxxxxxxxxx, loberman@xxxxxxxxxx > > Cc: hch@xxxxxx, maxg@xxxxxxxxxxxx, israelr@xxxxxxxxxxxx, linux-rdma@xxxxxxxxxxxxxxx, dledford@xxxxxxxxxx > > Sent: Monday, February 13, 2017 4:52:28 PM > > Subject: Re: v4.10-rc SRP + mlx5 regression > > > > On Mon, 2017-02-13 at 16:46 -0500, Laurence Oberman wrote: > > > I will have to run through this again and see where the bisect went wrong. > > > > Hello Laurence, > > > > If you would be considering to repeat the bisect, did you know that a bisect > > can be sped up by specifying the names of the files and/or directories that > > are suspected? An example: > > > > git bisect start */infiniband */net > > > > Bart.-- > > To unsubscribe from this list: send the line "unsubscribe linux-rdma" in > > the body of a message to majordomo@xxxxxxxxxxxxxxx > > More majordomo info at http://vger.kernel.org/majordomo-info.html > > > > Hello Bart, > > Much better news this time :), worked late on this but got it figured out. > > OK, so we got to this one, which makes a lot more sense and is right in the area where we are having issues. > I must have answered wrong to one of the steps the first time I did the bisect. > > Reverted this in the master tree of rc8 and rebuilt the kernel > Now all tests pass on Linus's tree - 4.10.0_rc8+ > > The interesting point here is that this commit is in rc5 but rc5 was not failing so we have an interoperability issue with this commit > > > [loberman@ibclient linux]$ git bisect good > Bisecting: 0 revisions left to test after this (roughly 1 step) > [ad8e66b4a80182174f73487ed25fd2140cf43361] IB/srp: fix mr allocation when the device supports sg gaps > > [loberman@ibclient linux]$ git show ad8e66b4a80182174f73487ed25fd2140cf43361 > commit ad8e66b4a80182174f73487ed25fd2140cf43361 > Author: Israel Rukshin <israelr@xxxxxxxxxxxx> > Date: Wed Dec 28 12:48:28 2016 +0200 > > IB/srp: fix mr allocation when the device supports sg gaps > > If the device support arbitrary sg list mapping (device cap > IB_DEVICE_SG_GAPS_REG set) we allocate the memory regions with > IB_MR_TYPE_SG_GAPS. > > Fixes: 509c5f33f4f6 ("IB/srp: Prevent mapping failures") > Cc: <stable@xxxxxxxxxxxxxxx> # 4.7+ > Signed-off-by: Israel Rukshin <israelr@xxxxxxxxxxxx> > Signed-off-by: Max Gurtovoy <maxg@xxxxxxxxxxxx> > Reviewed-by: Leon Romanovsky <leonro@xxxxxxxxxxxx> > Reviewed-by: Mark Bloch <markb@xxxxxxxxxxxx> > Reviewed-by: Yuval Shaia <yuval.shaia@xxxxxxxxxx> > Reviewed-by: Bart Van Assche <bart.vanassche@xxxxxxxxxxx> > Signed-off-by: Doug Ledford <dledford@xxxxxxxxxx> > > diff --git a/drivers/infiniband/ulp/srp/ib_srp.c b/drivers/infiniband/ulp/srp/ib_srp.c > index 8ddc071..0f67cf9 100644 > --- a/drivers/infiniband/ulp/srp/ib_srp.c > +++ b/drivers/infiniband/ulp/srp/ib_srp.c > @@ -371,6 +371,7 @@ static struct srp_fr_pool *srp_create_fr_pool(struct ib_device *device, > struct srp_fr_desc *d; > struct ib_mr *mr; > int i, ret = -EINVAL; > + enum ib_mr_type mr_type; > > if (pool_size <= 0) > goto err; > @@ -384,9 +385,13 @@ static struct srp_fr_pool *srp_create_fr_pool(struct ib_device *device, > spin_lock_init(&pool->lock); > INIT_LIST_HEAD(&pool->free_list); > > + if (device->attrs.device_cap_flags & IB_DEVICE_SG_GAPS_REG) > + mr_type = IB_MR_TYPE_SG_GAPS; > + else > + mr_type = IB_MR_TYPE_MEM_REG; > + > for (i = 0, d = &pool->desc[0]; i < pool->size; i++, d++) { > - mr = ib_alloc_mr(pd, IB_MR_TYPE_MEM_REG, > - max_page_list_len); > + mr = ib_alloc_mr(pd, mr_type, max_page_list_len); First, ib_alloc_mr receives u32 as a third parameter, but int was supplied. Second (I can be wrong here), shouldn't max_page_list_len be replaced with max_fast_reg_page_list_len? Thanks > if (IS_ERR(mr)) { > ret = PTR_ERR(mr); > if (ret == -ENOMEM) > (END) > > > So here is the revert patch, but you need to decide how you want to deal with this. > > Revert "IB/srp: fix mr allocation when the device supports sg gaps" > Laurence Oberman > Traced after bisection to a cause for this failure > > Tested-by: Laurence Oberman <loberman@xxxxxxxxxx> > Signed-off-by: Laurence Oberman <loberman@xxxxxxxxxx> > > commit 90d169d312a173d5350c1bb36d6daab04c592127 > Author: Laurence Oberman <loberman@xxxxxxxxxx> > Date: Mon Feb 13 20:33:32 2017 -0500 > > Revert "IB/srp: fix mr allocation when the device supports sg gaps" > Laurence Oberman > Traced after bisection to a cause for this failure > > [ 130.437603] mlx5_0:dump_cqe:262:(pid 3812): dump error cqe > [ 130.437682] scsi host1: ib_srp: failed RECV status WR flushed (5) for CQE ffff8817f0edbfb0 > [ 130.510899] 00000000 00000000 00000000 00000000 > [ 130.536455] 00000000 00000000 00000000 00000000 > [ 130.561878] 00000000 00000000 00000000 00000000 > [ 130.585904] 00000000 0f007806 2500002a db0ec4d0 > [ 145.842925] fast_io_fail_tmo expired for SRP port-1:1 / host1. > [ 146.530439] scsi host1: ib_srp: reconnect succeeded > [ 146.566629] mlx5_0:dump_cqe:262:(pid 3293): dump error cqe > [ 146.597635] 00000000 00000000 00000000 00000000 > [ 146.623545] 00000000 00000000 00000000 00000000 > [ 146.649599] 00000000 00000000 00000000 00000000 > [ 146.673938] 00000000 0f007806 25000032 000c46d0 > [ 146.697969] scsi host1: ib_srp: failed FAST REG status memory management operation error (6) for CQE ffff88 > [ 162.225247] fast_io_fail_tmo expired for SRP port-1:1 / host1. > [ 162.256337] scsi host1: ib_srp: reconnect succeeded > [ 162.293396] scsi host1: ib_srp: failed RECV status WR flushed (5) for CQE ffff8817f0412ef0` > > This reverts commit ad8e66b4a80182174f73487ed25fd2140cf43361. > > diff --git a/drivers/infiniband/ulp/srp/ib_srp.c b/drivers/infiniband/ulp/srp/ib_srp.c > index 79bf484..01338c8 100644 > --- a/drivers/infiniband/ulp/srp/ib_srp.c > +++ b/drivers/infiniband/ulp/srp/ib_srp.c > @@ -371,7 +371,6 @@ static struct srp_fr_pool *srp_create_fr_pool(struct ib_device *device, > struct srp_fr_desc *d; > struct ib_mr *mr; > int i, ret = -EINVAL; > - enum ib_mr_type mr_type; > > if (pool_size <= 0) > goto err; > @@ -385,13 +384,9 @@ static struct srp_fr_pool *srp_create_fr_pool(struct ib_device *device, > spin_lock_init(&pool->lock); > INIT_LIST_HEAD(&pool->free_list); > > - if (device->attrs.device_cap_flags & IB_DEVICE_SG_GAPS_REG) > - mr_type = IB_MR_TYPE_SG_GAPS; > - else > - mr_type = IB_MR_TYPE_MEM_REG; > - > for (i = 0, d = &pool->desc[0]; i < pool->size; i++, d++) { > - mr = ib_alloc_mr(pd, mr_type, max_page_list_len); > + mr = ib_alloc_mr(pd, IB_MR_TYPE_MEM_REG, > + max_page_list_len); > if (IS_ERR(mr)) { > ret = PTR_ERR(mr); > if (ret == -ENOMEM) > > > > Now moving on to what got me here in the first place. > Bart, let me know if the 7 of the 8 patches in your most recent series are all still valid after this revert > Otherwise let me know which ones you want me to apply. > > patch 6 - I am thinking i sno longer valid. > " > If a HCA supports the SG_GAPS_REG feature then a single memory > region of type IB_MR_TYPE_SG_GAPS is sufficient. This patch > reduces the number of memory regions that is allocated per SRP > session. > " > > Thanks > Laurence
Attachment:
signature.asc
Description: PGP signature
