[PATCH 4/5] nvmet: pci-epf: report transport errors in command completions

Wed Oct 14 05:46:46 PDT 2026

On Tue, Oct 13, 2026 at 05:12:16PM +0200, Damien Le Moal wrote:
> On 2026/10/13 16:27, Christoph Hellwig wrote:
> > I think this looks good.  Damien/Niklas, can you test this on hardware?
> 
> I do not have the hardware available right now. Niklas ? Can you test this?

I understand that these changes are not supposed to affect performance,
but posting it for completeness:

Before patches:

# ./fio-test.sh /dev/nvme0n1
Running on /dev/nvme0n1...
  Rnd read,    4KB,  QD=1, 1 job :  IOPS=3088, BW=12.1MiB/s (12.7MB/s)
  Rnd read,    4KB, QD=32, 1 job :  IOPS=45.4k, BW=177MiB/s (186MB/s)
  Rnd read,    4KB, QD=32, 4 jobs:  IOPS=47.8k, BW=187MiB/s (196MB/s)
  Rnd read,  128KB,  QD=1, 1 job :  IOPS=1803, BW=225MiB/s (236MB/s)
  Rnd read,  128KB, QD=32, 1 job :  IOPS=9078, BW=1135MiB/s (1190MB/s)
  Rnd read,  128KB, QD=32, 4 jobs:  IOPS=10.0k, BW=1253MiB/s (1314MB/s)
  Rnd read,  512KB,  QD=1, 1 job :  IOPS=1447, BW=724MiB/s (759MB/s)
  Rnd read,  512KB, QD=32, 1 job :  IOPS=2415, BW=1208MiB/s (1267MB/s)
  Rnd read,  512KB, QD=32, 4 jobs:  IOPS=2686, BW=1343MiB/s (1409MB/s)
  Rnd write,   4KB,  QD=1, 1 job :  IOPS=3046, BW=11.9MiB/s (12.5MB/s)
  Rnd write,   4KB, QD=32, 1 job :  IOPS=22.9k, BW=89.3MiB/s (93.6MB/s)
  Rnd write,   4KB, QD=32, 4 jobs:  IOPS=22.9k, BW=89.6MiB/s (94.0MB/s)
  Rnd write, 128KB,  QD=1, 1 job :  IOPS=1555, BW=194MiB/s (204MB/s)
  Rnd write, 128KB, QD=32, 1 job :  IOPS=6183, BW=773MiB/s (810MB/s)
  Rnd write, 128KB, QD=32, 4 jobs:  IOPS=6903, BW=863MiB/s (905MB/s)
  Seq read,  128KB,  QD=1, 1 job :  IOPS=1672, BW=209MiB/s (219MB/s)
  Seq read,  128KB, QD=32, 1 job :  IOPS=9033, BW=1129MiB/s (1184MB/s)
  Seq read,  512KB,  QD=1, 1 job :  IOPS=1580, BW=790MiB/s (829MB/s)
  Seq read,  512KB, QD=32, 1 job :  IOPS=2536, BW=1268MiB/s (1330MB/s)
  Seq read,    1MB, QD=32, 1 job :  IOPS=1297, BW=1297MiB/s (1360MB/s)
  Seq write, 128KB,  QD=1, 1 job :  IOPS=1547, BW=193MiB/s (203MB/s)
  Seq write, 128KB, QD=32, 1 job :  IOPS=6563, BW=820MiB/s (860MB/s)
  Seq write, 512KB,  QD=1, 1 job :  IOPS=893, BW=447MiB/s (469MB/s)
  Seq write, 512KB, QD=32, 1 job :  IOPS=1667, BW=834MiB/s (874MB/s)
  Seq write,   1MB, QD=32, 1 job :  IOPS=762, BW=763MiB/s (800MB/s)
  Rnd rdwr, 4K..1MB, QD=8, 4 jobs:  IOPS=788, BW=396MiB/s (415MB/s)
 IOPS=811, BW=408MiB/s (427MB/s)

After patches:

# ./fio-test.sh /dev/nvme0n1
Running on /dev/nvme0n1...
  Rnd read,    4KB,  QD=1, 1 job :  IOPS=2958, BW=11.6MiB/s (12.1MB/s)
  Rnd read,    4KB, QD=32, 1 job :  IOPS=44.5k, BW=174MiB/s (182MB/s)
  Rnd read,    4KB, QD=32, 4 jobs:  IOPS=47.4k, BW=185MiB/s (194MB/s)
  Rnd read,  128KB,  QD=1, 1 job :  IOPS=1741, BW=218MiB/s (228MB/s)
  Rnd read,  128KB, QD=32, 1 job :  IOPS=9327, BW=1166MiB/s (1223MB/s)
  Rnd read,  128KB, QD=32, 4 jobs:  IOPS=10.2k, BW=1280MiB/s (1342MB/s)
  Rnd read,  512KB,  QD=1, 1 job :  IOPS=1444, BW=722MiB/s (758MB/s)
  Rnd read,  512KB, QD=32, 1 job :  IOPS=2505, BW=1253MiB/s (1313MB/s)
  Rnd read,  512KB, QD=32, 4 jobs:  IOPS=2516, BW=1258MiB/s (1319MB/s)
  Rnd write,   4KB,  QD=1, 1 job :  IOPS=3322, BW=13.0MiB/s (13.6MB/s)
  Rnd write,   4KB, QD=32, 1 job :  IOPS=22.8k, BW=89.1MiB/s (93.5MB/s)
  Rnd write,   4KB, QD=32, 4 jobs:  IOPS=22.9k, BW=89.4MiB/s (93.7MB/s)
  Rnd write, 128KB,  QD=1, 1 job :  IOPS=1529, BW=191MiB/s (201MB/s)
  Rnd write, 128KB, QD=32, 1 job :  IOPS=6369, BW=796MiB/s (835MB/s)
  Rnd write, 128KB, QD=32, 4 jobs:  IOPS=6449, BW=806MiB/s (845MB/s)
  Seq read,  128KB,  QD=1, 1 job :  IOPS=1690, BW=211MiB/s (222MB/s)
  Seq read,  128KB, QD=32, 1 job :  IOPS=9262, BW=1158MiB/s (1214MB/s)
  Seq read,  512KB,  QD=1, 1 job :  IOPS=1428, BW=714MiB/s (749MB/s)
  Seq read,  512KB, QD=32, 1 job :  IOPS=2652, BW=1326MiB/s (1391MB/s)
  Seq read,    1MB, QD=32, 1 job :  IOPS=1301, BW=1302MiB/s (1365MB/s)
  Seq write, 128KB,  QD=1, 1 job :  IOPS=1531, BW=191MiB/s (201MB/s)
  Seq write, 128KB, QD=32, 1 job :  IOPS=6144, BW=768MiB/s (805MB/s)
  Seq write, 512KB,  QD=1, 1 job :  IOPS=924, BW=462MiB/s (485MB/s)
  Seq write, 512KB, QD=32, 1 job :  IOPS=1468, BW=734MiB/s (770MB/s)
  Seq write,   1MB, QD=32, 1 job :  IOPS=749, BW=749MiB/s (786MB/s)
  Rnd rdwr, 4K..1MB, QD=8, 4 jobs:  IOPS=786, BW=394MiB/s (414MB/s)
 IOPS=809, BW=407MiB/s (426MB/s)

Also used ftrace on the EP side to see that SGLs were actually being used:
   kworker/u32:2-75      [005] .....   300.152187: nvmet_pci_epf_iod_parse_sgls <-nvmet_pci_epf_exec_iod_work
   kworker/u32:2-75      [005] .....   300.152560: nvmet_pci_epf_iod_parse_sgls <-nvmet_pci_epf_exec_iod_work
   kworker/u32:2-75      [005] .....   300.153126: nvmet_pci_epf_iod_parse_sgls <-nvmet_pci_epf_exec_iod_work
   kworker/u32:2-75      [005] .....   300.153702: nvmet_pci_epf_iod_parse_sgls <-nvmet_pci_epf_exec_iod_work
   kworker/u32:2-75      [005] .....   300.154273: nvmet_pci_epf_iod_parse_sgls <-nvmet_pci_epf_exec_iod_work
   kworker/u32:2-75      [005] .....   300.154807: nvmet_pci_epf_iod_parse_sgls <-nvmet_pci_epf_exec_iod_work
   kworker/u32:2-75      [005] .....   300.155445: nvmet_pci_epf_iod_parse_sgls <-nvmet_pci_epf_exec_iod_work
   kworker/u32:2-75      [005] .....   300.156263: nvmet_pci_epf_iod_parse_sgls <-nvmet_pci_epf_exec_iod_work
   kworker/u32:2-75      [005] .....   300.156676: nvmet_pci_epf_iod_parse_sgls <-nvmet_pci_epf_exec_iod_work
   kworker/u32:2-75      [005] .....   300.157340: nvmet_pci_epf_iod_parse_sgls <-nvmet_pci_epf_exec_iod_work
   kworker/u32:2-75      [005] .....   300.158096: nvmet_pci_epf_iod_parse_sgls <-nvmet_pci_epf_exec_iod_work

I can not see anything wrong, thus:

Tested-by: Niklas Cassel <cassel at kernel.org>