On Apr 8 15:56, Alan Adamson wrote:
> While testing Linux atomic writes with qemu-nvme v10.0.0-rc1, Linux was
> incorrectly displaying atomic_write_max_bytes
> # cat /sys/block/nvme0n1/queue/atomic_write_max_bytes
> 0
> # nvme id-ctrl /dev/nvme0n1 | grep awupf
> awupf : 15
> #
> Since AWUPF was set to 15, it was expected atomic_write_max_bytes would
> be set to 8192.
>
> The commit cd59f50ab017 ("hw/nvme: always initialize a subsystem")
> introduced this behavior. The commit hardcodes the subsystem cmic bit
> to ON which caused the Linux NVMe driver to treat the namespace as
> multi-pathed which uncovered a bug with how Atomic Write Queue Limits
> were being inherited. This Linux issue is being addressed, but the
> question was asked of why the subsystem cmic bit was hardcoded to ON.
> Most NVMe devices today don't set cmic to ON. Shouldn't the setting of
> this bit be a settable parameter?
>
> <subsystem>,cmic=BOOLEAN (default: off)
>
> Example:
> -device nvme-subsys,id=subsys0,cmic=on \
> -device
> nvme,serial=deadbeef,id=nvme0,subsys=subsys0,atomic.dn=off,atomic.awun=31,atomic.awupf=15
> \
> -drive id=ns1,file=/dev/nullb3,if=none \
> -device nvme-ns,drive=ns1,bus=nvme0,nsid=1,shared=false
>
> Alan Adamson (1):
> hw/nvme: create parameter to enable/disable cmic on subsystem
>
> hw/nvme/ctrl.c | 5 ++++-
> hw/nvme/nvme.h | 1 +
> hw/nvme/subsys.c | 1 +
> 3 files changed, 6 insertions(+), 1 deletion(-)
>
> --
> 2.43.5
>
> Hi Alan, I agree that it would be better for CMIC.MCTRS to remain zero for single-controller subsystems, but I'd rather not add the parameter without verifying it (it would be invalid to have a multi-controller subsystem with CMIC.MCTRS zeroed). An improvement would be to just set it if the subsystems contains multiple controllers. Prior to commit cd59f50ab017 we would also statically set CMIC.MCTRS to 1 if a subsystem was configured, regardless of the number of controllers.
signature.asc
Description: PGP signature
