A write request with the REQ_RAHEAD flag enabled cause the nvme driver
to send a write command with access frequency value that is reserved
(at least on specification version 1.4c). NVME devices might fail this
write command with an unsupported error. So we now clear the request's
flag based on its direction.
Signed-off-by: Gal Hammer <gal.hammer@huawei.com>
Signed-off-by: Shai Fultheim <shai.fultheim@huawei.com>
Signed-off-by: Robert Baldyga <robert.baldyga@huawei.com>
When dir is ignored, and 0 is passed instead, each flush request will
appear as READ request, which is not supported by some block device
drivers.
Signed-off-by: Robert Baldyga <robert.baldyga@intel.com>
This is a suboptimal solution to CAS on top of MD RAID1 device. If using
only submit_bio API RAID1 would process all IOs in single thread.
Plugging bypasses this thread and processess IOs in blk_finish_plug
caller context improving performance drastically.
Testing showed no negative impact to other usecases and it's a thing
that Linux does in AIO, so it's vetted and proven to work.
Signed-off-by: Jan Musial <jan.musial@intel.com>
It's defined on every single supported kernel, so there is actually no need
for this define at all.
Signed-off-by: Robert Baldyga <robert.baldyga@intel.com>
This counter is not accurate (missing required memory barrier
to avoid unwanted behavior due to processor optimizations)
and performance gain is not clear - generally global
atomic variables are something we would like to avoid
going forward.
Signed-off-by: Adam Rutkowski <adam.j.rutkowski@intel.com>
Returning non-const char* doesn't have any benefits since
the only caller casts the output pointer to const char *.
This change fixes compilation on 5.3 kernel.
Signed-off-by: Adam Rutkowski <adam.j.rutkowski@intel.com>
While unloading cas_cache module, volume types were deinitialized, although core
pool still wasn't empty. Now this deinitialization can be done after removing
cores from core pool, and before context is freed.
Signed-off-by: Michal Mielewczyk <michal.mielewczyk@intel.com>
Functions and macros dependent on different kernel versions are now generated
before compilation basing on current kernel capabilities instead of hardcoding
them for specific kernels.
Signed-off-by: Michal Mielewczyk <michal.mielewczyk@intel.com>
This needs to be done in caller application context. As part of asynchronous
management API rework, volume_open callback was moved to kernel worker thread
context. To work around this, atomic parameters are now read before cache
attach/load/probe and provided to bottom volume via volume params.
Signed-off-by: Adam Rutkowski <adam.j.rutkowski@intel.com>
