GitLab

Use async function calls to make namespace scanning happen in parallel.

Without the patch, NVME namespaces are scanned serially, so it can take
a long time for all of a controller's namespaces to become available,
especially with a slower (TCP) interface with large number of
namespaces.

It is not uncommon to have large numbers (hundreds or thousands) of
namespaces on nvme-of with storage servers.

The time it took for all namespaces to show up after connecting (via
TCP) to a controller with 1002 namespaces was measured on one system:

network latency   without patch   with patch
     0                 6s            1s
    50ms             210s           10s
   100ms             417s           18s

Measurements taken on another system show the effect of the patch on the
time nvme_scan_work() took to complete, when connecting to a linux
nvme-of target with varying numbers of namespaces, on a network of
400us.

namespaces    without patch   with patch
     1            16ms           14ms
     2            24ms           16ms
     4            49ms           22ms
     8           101ms           33ms
    16           207ms           56ms
   100           1.4s           0.6s
  1000          12.9s           2.0s

On the same system, connecting to a local PCIe NVMe drive (a Samsung
PM1733) instead of a network target:

namespaces    without patch   with patch
     1            13ms           12ms
     2            41ms           13ms
Signed-off-by: Stuart Hayes <stuart.w.hayes@gmail.com>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>