diff options
author | Kent Overstreet <kent.overstreet@linux.dev> | 2025-03-10 18:44:12 -0400 |
---|---|---|
committer | Kent Overstreet <kent.overstreet@linux.dev> | 2025-07-01 00:20:01 -0400 |
commit | f8441fc0ffaac0d52dc6fa5454db82db2541d4a3 (patch) | |
tree | de720c31481281f045197b2478f9bbc72d6c20d7 /drivers/gpu/drm/amd/amdgpu/amdgpu_vm_cpu.c | |
parent | 7303d2b8b67e9f7dcb960aa3d397b135374aaa21 (diff) |
block: Allow REQ_FUA|REQ_READ
FUA is also allowed with reads, not just writes.
The specified behaviour is:
- If the location being read from in the drive cache is dirty, it's
flushed
- Read is serviced from media, not cache
It's documented in the NVME specification, and the nvme driver already
passes through REQ_FUA for reads, not just writes, so there's no reason
for the block layer to be disallowing it.
To validate behaviour, a simple test was run on a variety of hardware
that checks latency of repeated reads to the same location (cached
reads), random reads (uncached), and FUA reads to the same location.
Data:
- Samsung consumer SSDs
Reads appear to not be cached
- Seagate SCSI hard drives (ST20000NM002D)
Reads are cached, and FUA reads appear to work correctly
Link: https://lore.kernel.org/linux-block/20250311133517.3095878-1-kent.overstreet@linux.dev/
Link: https://lore.kernel.org/linux-bcachefs/26585.34711.506258.318405@quad.stoffel.home/T/#m5fffbc0e1c68cf0479c94b9f4ac1bef297333782
Cc: Jens Axboe <axboe@kernel.dk>
Cc: linux-block@vger.kernel.org
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
Diffstat (limited to 'drivers/gpu/drm/amd/amdgpu/amdgpu_vm_cpu.c')
0 files changed, 0 insertions, 0 deletions