RE: [RFC 11/11] scsi: storvsc: Support PAGE_SIZE larger than 4K

[Michael Kelley <mikelley@microsoft.com>]

From: boqun.feng@gmail.com <boqun.feng@gmail.com> Sent: Wednesday, July 22, 2020 6:52 PM
> 
> On Thu, Jul 23, 2020 at 12:13:07AM +0000, Michael Kelley wrote:
> > From: Boqun Feng <boqun.feng@gmail.com> Sent: Monday, July 20, 2020 6:42 PM
> > >
> > > Hyper-V always use 4k page size (HV_HYP_PAGE_SIZE), so when
> > > communicating with Hyper-V, a guest should always use HV_HYP_PAGE_SIZE
> > > as the unit for page related data. For storvsc, the data is
> > > vmbus_packet_mpb_array. And since in scsi_cmnd, sglist of pages (in unit
> > > of PAGE_SIZE) is used, we need convert pages in the sglist of scsi_cmnd
> > > into Hyper-V pages in vmbus_packet_mpb_array.
> > >
> > > This patch does the conversion by dividing pages in sglist into Hyper-V
> > > pages, offset and indexes in vmbus_packet_mpb_array are recalculated
> > > accordingly.
> > >
> > > Signed-off-by: Boqun Feng <boqun.feng@gmail.com>
> > > ---
> > >  drivers/scsi/storvsc_drv.c | 27 +++++++++++++++++++++------
> > >  1 file changed, 21 insertions(+), 6 deletions(-)
> > >
> > > diff --git a/drivers/scsi/storvsc_drv.c b/drivers/scsi/storvsc_drv.c
> > > index fb41636519ee..c54d25f279bc 100644
> > > --- a/drivers/scsi/storvsc_drv.c
> > > +++ b/drivers/scsi/storvsc_drv.c
> > > @@ -1561,7 +1561,7 @@ static int storvsc_queuecommand(struct Scsi_Host *host,
> struct
> > > scsi_cmnd *scmnd)
> > >  	struct hv_host_device *host_dev = shost_priv(host);
> > >  	struct hv_device *dev = host_dev->dev;
> > >  	struct storvsc_cmd_request *cmd_request = scsi_cmd_priv(scmnd);
> > > -	int i;
> > > +	int i, j, k;
> > >  	struct scatterlist *sgl;
> > >  	unsigned int sg_count = 0;
> > >  	struct vmscsi_request *vm_srb;
> > > @@ -1569,6 +1569,8 @@ static int storvsc_queuecommand(struct Scsi_Host *host,
> struct
> > > scsi_cmnd *scmnd)
> > >  	struct vmbus_packet_mpb_array  *payload;
> > >  	u32 payload_sz;
> > >  	u32 length;
> > > +	int subpage_idx = 0;
> > > +	unsigned int hvpg_count = 0;
> > >
> > >  	if (vmstor_proto_version <= VMSTOR_PROTO_VERSION_WIN8) {
> > >  		/*
> > > @@ -1643,23 +1645,36 @@ static int storvsc_queuecommand(struct Scsi_Host *host,
> struct
> > > scsi_cmnd *scmnd)
> > >  	payload_sz = sizeof(cmd_request->mpb);
> > >
> > >  	if (sg_count) {
> > > -		if (sg_count > MAX_PAGE_BUFFER_COUNT) {
> > > +		hvpg_count = sg_count * (PAGE_SIZE / HV_HYP_PAGE_SIZE);
> >
> > The above calculation doesn't take into account the offset in the
> > first sglist or the overall length of the transfer, so the value of hvpg_count
> > could be quite a bit bigger than it needs to be.  For example, with a 64K
> > page size and an 8 Kbyte transfer size that starts at offset 60K in the
> > first page, hvpg_count will be 32 when it really only needs to be 2.
> >
> > The nested loops below that populate the pfn_array take the
> > offset into account when starting, so that's good.  But it will potentially
> > leave allocated entries unused.  Furthermore, the nested loops could
> > terminate early when enough Hyper-V size pages are mapped to PFNs
> > based on the length of the transfer, even if all of the last guest size
> > page has not been mapped to PFNs.  Like the offset at the beginning of
> > first guest size page in the sglist, there's potentially an unused portion
> > at the end of the last guest size page in the sglist.
> >
> 
> Good point. I think we could calculate the exact hvpg_count as follow:
> 
> 	hvpg_count = 0;
> 	cur_sgl = sgl;
> 
> 	for (i = 0; i < sg_count; i++) {
> 		hvpg_count += HVPFN_UP(cur_sg->length)
> 		cur_sgl = sg_next(cur_sgl);
> 	}
> 

The downside would be going around that loop a lot of times when
the page size is 4K bytes and the I/O transfer size is something like
256K bytes.  I think this gives the right result in constant time:  the
starting offset within a Hyper-V page, plus the transfer length,
rounded up to a Hyper-V page size, and divided by the Hyper-V
page size.


> > > +		if (hvpg_count > MAX_PAGE_BUFFER_COUNT) {
> > >
> > > -			payload_sz = (sg_count * sizeof(u64) +
> > > +			payload_sz = (hvpg_count * sizeof(u64) +
> > >  				      sizeof(struct vmbus_packet_mpb_array));
> > >  			payload = kzalloc(payload_sz, GFP_ATOMIC);
> > >  			if (!payload)
> > >  				return SCSI_MLQUEUE_DEVICE_BUSY;
> > >  		}
> > >
> > > +		/*
> > > +		 * sgl is a list of PAGEs, and payload->range.pfn_array
> > > +		 * expects the page number in the unit of HV_HYP_PAGE_SIZE (the
> > > +		 * page size that Hyper-V uses, so here we need to divide PAGEs
> > > +		 * into HV_HYP_PAGE in case that PAGE_SIZE > HV_HYP_PAGE_SIZE.
> > > +		 */
> > >  		payload->range.len = length;
> > > -		payload->range.offset = sgl[0].offset;
> > > +		payload->range.offset = sgl[0].offset & ~HV_HYP_PAGE_MASK;
> > > +		subpage_idx = sgl[0].offset >> HV_HYP_PAGE_SHIFT;
> > >
> > >  		cur_sgl = sgl;
> > > +		k = 0;
> > >  		for (i = 0; i < sg_count; i++) {
> > > -			payload->range.pfn_array[i] =
> > > -				page_to_pfn(sg_page((cur_sgl)));
> > > +			for (j = subpage_idx; j < (PAGE_SIZE / HV_HYP_PAGE_SIZE); j++) {
> >
> > In the case where PAGE_SIZE == HV_HYP_PAGE_SIZE, would it help the compiler
> > eliminate the loop if local variable j is declared as unsigned?  In that case the test in the
> > for statement will always be false.
> >
> 
> Good point! I did the following test:
> 
> test.c:
> 
> 	int func(unsigned int input, int *arr)
> 	{
> 		unsigned int i;
> 		int result = 0;
> 
> 		for (i = input; i < 1; i++)
> 			result += arr[i];
> 
> 		return result;
> 	}
> 
> if I define i as "int", I got:
> 
> 	0000000000000000 <func>:
> 	   0:	85 ff                	test   %edi,%edi
> 	   2:	7f 2c                	jg     30 <func+0x30>
> 	   4:	48 63 d7             	movslq %edi,%rdx
> 	   7:	f7 df                	neg    %edi
> 	   9:	45 31 c0             	xor    %r8d,%r8d
> 	   c:	89 ff                	mov    %edi,%edi
> 	   e:	48 8d 04 96          	lea    (%rsi,%rdx,4),%rax
> 	  12:	48 01 d7             	add    %rdx,%rdi
> 	  15:	48 8d 54 be 04       	lea    0x4(%rsi,%rdi,4),%rdx
> 	  1a:	66 0f 1f 44 00 00    	nopw   0x0(%rax,%rax,1)
> 	  20:	44 03 00             	add    (%rax),%r8d
> 	  23:	48 83 c0 04          	add    $0x4,%rax
> 	  27:	48 39 d0             	cmp    %rdx,%rax
> 	  2a:	75 f4                	jne    20 <func+0x20>
> 	  2c:	44 89 c0             	mov    %r8d,%eax
> 	  2f:	c3                   	retq
> 	  30:	45 31 c0             	xor    %r8d,%r8d
> 	  33:	44 89 c0             	mov    %r8d,%eax
> 	  36:	c3                   	retq
> 
> and when I define i as "unsigned int", I got:
> 
> 	0000000000000000 <func>:
> 	   0:	85 ff                	test   %edi,%edi
> 	   2:	75 03                	jne    7 <func+0x7>
> 	   4:	8b 06                	mov    (%rsi),%eax
> 	   6:	c3                   	retq
> 	   7:	31 c0                	xor    %eax,%eax
> 	   9:	c3                   	retq
> 
> So clearly it helps, I will change this in the next version.

Wow!  The compiler is good ....

> 
> Regards,
> Boqun
> 
> > > +				payload->range.pfn_array[k] =
> > > +					page_to_hvpfn(sg_page((cur_sgl))) + j;
> > > +				k++;
> > > +			}
> > >  			cur_sgl = sg_next(cur_sgl);
> > > +			subpage_idx = 0;
> > >  		}
> > >  	}
> > >
> > > --
> > > 2.27.0
> >