BLKZEROOUT ioctl and WriteZero opcode

BLKZEROOUT ioctl and WriteZero opcode

[אייל בן דוד]

Hi all,

There is an ioctl command BLKZEROOUT that accepts range (lba + number
of block to zero)

Suppose:
Block device is an nvme device (e.g. /dev/nvme0nX)
Device does support the WriteZeroes opcode.

If so, does the kernel make use of the WriteZeroes opcode in the
implementation of BLKZEROOUT ?
Same question for BLKDISCARD.

Thanks
Eyal.

BLKZEROOUT ioctl and WriteZero opcode

[Keith Busch]

On Thu, May 03, 2018@06:15:54PM +0300, ???? ?? ??? wrote:
> There is an ioctl command BLKZEROOUT that accepts range (lba + number
> of block to zero)
> 
> Suppose:
> Block device is an nvme device (e.g. /dev/nvme0nX)
> Device does support the WriteZeroes opcode.
> 
> If so, does the kernel make use of the WriteZeroes opcode in the
> implementation of BLKZEROOUT ?
> Same question for BLKDISCARD.

The NVMe driver supports BLKDISACRD if the device supports the DSM
command.

The NVMe driver at the moment doesn't support the NVME Writes Zeroes
command. We had it briefly, but apparently broke Linus' machine; never
did figure out why. Relevent link:

  https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=cdb98c2698b4af287925abcba4d77d92af82a0c3

For devices that claim discard determinisitcally returns 0's on reads,
the driver will map the user BLKZEROOUT request to the DSM command
rather than NVMe Write Zeroes.

For testing NVMe Write Zeroes, the nvme passthrough ioctl may still
be used.

BLKZEROOUT ioctl and WriteZero opcode

[Eyal BenDavid]

Thanks Keith,

So:
BLKZEROOUT is not implemented by NVME passthru + Write Zeroes (for now).
BLKZEROOUT Is implemented by DSM provided the device guarantees 0's
are returned after "Unmap"
In order to to zero out a range, NVMe passthru with Write Zeroes
opcode must be used.

Is it correct?

Thanks,
Eyal


On Fri, May 4, 2018 at 11:05 PM, Keith Busch
<keith.busch@linux.intel.com> wrote:
> On Thu, May 03, 2018@06:15:54PM +0300, ???? ?? ??? wrote:
>> There is an ioctl command BLKZEROOUT that accepts range (lba + number
>> of block to zero)
>>
>> Suppose:
>> Block device is an nvme device (e.g. /dev/nvme0nX)
>> Device does support the WriteZeroes opcode.
>>
>> If so, does the kernel make use of the WriteZeroes opcode in the
>> implementation of BLKZEROOUT ?
>> Same question for BLKDISCARD.
>
> The NVMe driver supports BLKDISACRD if the device supports the DSM
> command.
>
> The NVMe driver at the moment doesn't support the NVME Writes Zeroes
> command. We had it briefly, but apparently broke Linus' machine; never
> did figure out why. Relevent link:
>
>   https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=cdb98c2698b4af287925abcba4d77d92af82a0c3
>
> For devices that claim discard determinisitcally returns 0's on reads,
> the driver will map the user BLKZEROOUT request to the DSM command
> rather than NVMe Write Zeroes.
>
> For testing NVMe Write Zeroes, the nvme passthrough ioctl may still
> be used.

BLKZEROOUT ioctl and WriteZero opcode

[Keith Busch]

On Mon, May 07, 2018@03:17:56PM +0300, Eyal BenDavid wrote:
> Thanks Keith,
> 
> So:
> BLKZEROOUT is not implemented by NVME passthru + Write Zeroes (for now).
> BLKZEROOUT Is implemented by DSM provided the device guarantees 0's
> are returned after "Unmap"
> In order to to zero out a range, NVMe passthru with Write Zeroes
> opcode must be used.
> 
> Is it correct?

Yes, that is correct.

BLKZEROOUT ioctl and WriteZero opcode

[Eyal BenDavid]

Thanks!


On Mon, May 7, 2018@6:14 PM, Keith Busch <keith.busch@linux.intel.com> wrote:
> On Mon, May 07, 2018@03:17:56PM +0300, Eyal BenDavid wrote:
>> Thanks Keith,
>>
>> So:
>> BLKZEROOUT is not implemented by NVME passthru + Write Zeroes (for now).
>> BLKZEROOUT Is implemented by DSM provided the device guarantees 0's
>> are returned after "Unmap"
>> In order to to zero out a range, NVMe passthru with Write Zeroes
>> opcode must be used.
>>
>> Is it correct?
>
> Yes, that is correct.