Backport commit 38adf94e166e3cb4eb89683458ca578051e8218d and it's dependencies to linux-stable 5.4.y.
Dependent commits: 314d48dd224897e35ddcaf5a1d7d133b5adddeb7 e08f2ae850929d40e66268ee47e443e7ea56eeb7
When running test cases to stress NVMe device, a race condition / deadlocks is seen every couple of days or so where multiple threads are trying to acquire ctrl->subsystem->lock or ctrl->scan_lock.
The test cases send a lot nvme-cli requests to do Sanitize, Format, FW Download, FW Activate, Flush, Get Log, Identify, and reset requests to two controllers that share a namespace. Some of those commands target a namespace, some target a controller. The commands are sent in random order and random mix to the two controllers.
The test cases does not wait for nvme-cli requests to finish before sending more. So for example, there could be multiple reset requests, multiple format requests, outstanding at the same time as a sanitize, on both paths at the same time, etc. Many of these test cases include combos that don't really make sense in the context of NVMe, however it is used to create as much stress as possible.
This patchset fixes this issue.
Similar issue with a detailed call trace/log was discussed in the LKML Link: https://lore.kernel.org/linux-nvme/04580CD6-7652-459D-ABDD-732947B4A6DF@javi...
Revanth Rajashekar (3): nvme: Cleanup and rename nvme_block_nr() nvme: Introduce nvme_lba_to_sect() nvme: consolidate chunk_sectors settings
drivers/nvme/host/core.c | 40 +++++++++++++++++++--------------------- drivers/nvme/host/nvme.h | 16 +++++++++++++--- 2 files changed, 32 insertions(+), 24 deletions(-)
-- 2.17.1