RE: [PATCH] scsi: storvsc: Correct sysfs parameters as per Hyper-V storvsc requirement

[Michael Kelley (LINUX)]

From: Saurabh Singh Sengar <ssengar@xxxxxxxxxxxxxxxxxxx> Sent: Friday, June 10, 2022 9:37 AM
> 
> CC : linux-scsi@xxxxxxxxxxxxxxx, jejb@xxxxxxxxxxxxx, martin.petersen@xxxxxxxxxx
> 
> On Fri, Jun 10, 2022 at 09:33:44AM -0700, Saurabh Sengar wrote:
> > This patch corrects 3 parameters:
> > 1. Correct the sysfs entry for maximum hardware transfer limit of single
> >    transfer (max_hw_sectors_kb) by setting max_sectors, this was set to
> >    default value 512kb before.
> > 2. Correct SGL memory offset alignment as per Hyper-V page size.
> > 3. Correct sg_tablesize which accounts for max SGL segments entries in a
> >    single SGL.

I think a richer explanation in the commit message is warranted.
Something like:

  Current code is based on the idea that the max number of SGL entries also
  determines the max size of an I/O request.  While this idea was true in older
  versions of the storvsc driver when SGL entry length was limited to 4 Kbytes,
  commit 3d9c3dcc58e9 ("scsi: storvsc: Enable scatter list entry lengths > 4Kbytes")
  removed that limitation.  It's now theoretically possible for the block
  layer to send requests that exceed the maximum size supported by Hyper-V.
  This problem doesn't currently happen in practice because the block layer
  defaults to a 512 Kbyte maximum, while Hyper-V in Azure supports 2 Mbyte
  I/O sizes.  But some future configuration of Hyper-V could have a smaller
  max I/O size, and the block layer could exceed that max.

  Fix this by correctly setting max_sectors as well as sg_tablesize to reflect
  the maximum I/O size that Hyper-V reports.  While allowing larger I/O sizes
  larger than the block layer default of 512 Kbytes doesn't provide any
  noticeable performance benefit in the tests we ran, it's still appropriate
  to report the correct underlying Hyper-V capabilities to the Linux block layer.

  Also tweak the virt_boundary_mask to reflect that the needed alignment
  derives from Hyper-V communication using a 4 Kbyte page size, and not
  on the guest page size, which might be bigger (on ARM64, for example).

I don't think the title of the commit should focus on sysfs.  This
commit is about correctly reporting Hyper-V I/O size limits; the sysfs
entries just provide visibility into the values.

And given that the problem was introduced by the above mentioned
commit, it would be appropriate to add a "Fixes:" tag.

> >
> > Signed-off-by: Saurabh Sengar <ssengar@xxxxxxxxxxxxxxxxxxx>
> > ---
> >  drivers/scsi/storvsc_drv.c | 28 ++++++++++++++++++++++++----
> >  1 file changed, 24 insertions(+), 4 deletions(-)
> >
> > diff --git a/drivers/scsi/storvsc_drv.c b/drivers/scsi/storvsc_drv.c
> > index ca3530982e52..3e032660ae36 100644
> > --- a/drivers/scsi/storvsc_drv.c
> > +++ b/drivers/scsi/storvsc_drv.c
> > @@ -1844,7 +1844,7 @@ static struct scsi_host_template scsi_driver = {
> >  	.cmd_per_lun =		2048,
> >  	.this_id =		-1,
> >  	/* Ensure there are no gaps in presented sgls */
> > -	.virt_boundary_mask =	PAGE_SIZE-1,
> > +	.virt_boundary_mask =	HV_HYP_PAGE_SIZE - 1,
> >  	.no_write_same =	1,
> >  	.track_queue_depth =	1,
> >  	.change_queue_depth =	storvsc_change_queue_depth,
> > @@ -1969,11 +1969,31 @@ static int storvsc_probe(struct hv_device *device,
> >  	/* max cmd length */
> >  	host->max_cmd_len = STORVSC_MAX_CMD_LEN;
> >
> > +	/* max_hw_sectors_kb */
> > +	host->max_sectors = (stor_device->max_transfer_bytes) >> 9;
> >  	/*
> > -	 * set the table size based on the info we got
> > -	 * from the host.
> > +	 * There are 2 requirements for Hyper-V storvsc sgl segments,
> > +	 * based on which the below calculation for max segments is
> > +	 * done:
> > +	 *
> > +	 * 1. Except for the first and last sgl segment, all sgl segments
> > +	 *    should be align to HV_HYP_PAGE_SIZE, that also means the
> > +	 *    maximum number of segments in a sgl can be calculated by
> > +	 *    dividing the total max transfer length by HV_HYP_PAGE_SIZE.
> > +	 *
> > +	 * 2. Except for the first and last, each entry in the SGL must
> > +	 *    have an offset that is a multiple of HV_HYP_PAGE_SIZE,
> > +	 *    whereas the complete length of transfer may not be aligned
> > +	 *    to HV_HYP_PAGE_SIZE always. This can result in 2 cases:
> > +	 *    Example for unaligned case: Let's say the total transfer
> > +	 *    length is 6 KB, the max segments will be 3 (1,4,1).
> > +	 *    Example for aligned case: Let's say the total transfer length
> > +	 *    is 8KB, then max segments will still be 3(2,4,2) and not 4.
> > +	 *    4 (read next higher value) segments will only be required
> > +	 *    once the length is at least 2 bytes more then 8KB (read any
> > +	 *    HV_HYP_PAGE_SIZE aligned length).
> >  	 */
> > -	host->sg_tablesize = (stor_device->max_transfer_bytes >> PAGE_SHIFT);
> > +	host->sg_tablesize = ((stor_device->max_transfer_bytes - 2) >> HV_HYP_PAGE_SHIFT) + 2;

This calculation covers all possible I/O request sizes up to and including
the value of max_transfer_bytes, even if max_transfer_bytes is some
weird number that's not a multiple of 512.   So I think it works as
intended.

But setting host->max_sectors means that storvsc won't see an I/O request
with a weird size, and some of the cases handled by the calculation don't
actually occur.  You could use a simpler calculation that's a bit easier to
understand:

host->sg_tablesize = (stor_device->max_transfer_bytes >> HV_HYP_PAGE_SIZE) + 1;

The "+1" handles the unaligned case you mention above.

Michael

> >  	/*
> >  	 * For non-IDE disks, the host supports multiple channels.
> >  	 * Set the number of HW queues we are supporting.
> > --
> > 2.25.1