| Message ID | 1613682087-102535-1-git-send-email-mikelley@microsoft.com |
|---|---|
| State | New |
| Headers | show |
| Series | [1/1] scsi: storvsc: Enable scatterlist entry lengths > 4Kbytes | expand |
Michael Kelley <mikelley@microsoft.com> writes: > storvsc currently sets .dma_boundary to limit scatterlist entries > to 4 Kbytes, which is less efficient with huge pages that offer > large chunks of contiguous physical memory. Improve the algorithm > for creating the Hyper-V guest physical address PFN array so > that scatterlist entries with lengths > 4Kbytes are handled. > As a result, remove the .dma_boundary setting. > > The improved algorithm also adds support for scatterlist > entries with offsets >= 4Kbytes, which is supported by many > other SCSI low-level drivers. And it retains support for > architectures where possibly PAGE_SIZE != HV_HYP_PAGE_SIZE > (such as ARM64). > > Signed-off-by: Michael Kelley <mikelley@microsoft.com> > --- > drivers/scsi/storvsc_drv.c | 63 ++++++++++++++++------------------------------ > 1 file changed, 22 insertions(+), 41 deletions(-) > > diff --git a/drivers/scsi/storvsc_drv.c b/drivers/scsi/storvsc_drv.c > index 2e4fa77..5d06061 100644 > --- a/drivers/scsi/storvsc_drv.c > +++ b/drivers/scsi/storvsc_drv.c > @@ -1678,9 +1678,8 @@ static int storvsc_queuecommand(struct Scsi_Host *host, struct scsi_cmnd *scmnd) > struct storvsc_cmd_request *cmd_request = scsi_cmd_priv(scmnd); > int i; > struct scatterlist *sgl; > - unsigned int sg_count = 0; > + unsigned int sg_count; > struct vmscsi_request *vm_srb; > - struct scatterlist *cur_sgl; > struct vmbus_packet_mpb_array *payload; > u32 payload_sz; > u32 length; > @@ -1759,7 +1758,7 @@ static int storvsc_queuecommand(struct Scsi_Host *host, struct scsi_cmnd *scmnd) > payload_sz = sizeof(cmd_request->mpb); > > if (sg_count) { > - unsigned int hvpgoff = 0; > + unsigned int hvpgoff, sgl_size; > unsigned long offset_in_hvpg = sgl->offset & ~HV_HYP_PAGE_MASK; > unsigned int hvpg_count = HVPFN_UP(offset_in_hvpg + length); > u64 hvpfn; > @@ -1773,51 +1772,35 @@ static int storvsc_queuecommand(struct Scsi_Host *host, struct scsi_cmnd *scmnd) > 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. > - * Besides, payload->range.offset should be the offset in one > - * HV_HYP_PAGE. > - */ > payload->range.len = length; > payload->range.offset = offset_in_hvpg; > - hvpgoff = sgl->offset >> HV_HYP_PAGE_SHIFT; > > - cur_sgl = sgl; > - for (i = 0; i < hvpg_count; i++) { > + > + for (i = 0; sgl != NULL; sgl = sg_next(sgl)) { > /* > - * 'i' is the index of hv pages in the payload and > - * 'hvpgoff' is the offset (in hv pages) of the first > - * hv page in the the first page. The relationship > - * between the sum of 'i' and 'hvpgoff' and the offset > - * (in hv pages) in a payload page ('hvpgoff_in_page') > - * is as follow: > - * > - * |------------------ PAGE -------------------| > - * | NR_HV_HYP_PAGES_IN_PAGE hvpgs in total | > - * |hvpg|hvpg| ... |hvpg|... |hvpg| > - * ^ ^ ^ ^ > - * +-hvpgoff-+ +-hvpgoff_in_page-+ > - * ^ | > - * +--------------------- i ---------------------------+ > + * Init values for the current sgl entry. sgl_size > + * and hvpgoff are in units of Hyper-V size pages. > + * Handling the PAGE_SIZE != HV_HYP_PAGE_SIZE case > + * also handles values of sgl->offset that are > + * larger than PAGE_SIZE. Such offsets are handled > + * even on other than the first sgl entry, provided > + * they are a multiple of PAGE_SIZE. > */ > - unsigned int hvpgoff_in_page = > - (i + hvpgoff) % NR_HV_HYP_PAGES_IN_PAGE; > + sgl_size = HVPFN_UP(sgl->offset + sgl->length); > + hvpgoff = sgl->offset >> HV_HYP_PAGE_SHIFT; > + hvpfn = page_to_hvpfn(sg_page(sgl)); > > /* > - * Two cases that we need to fetch a page: > - * 1) i == 0, the first step or > - * 2) hvpgoff_in_page == 0, when we reach the boundary > - * of a page. > + * Fill the next portion of the PFN array with > + * sequential Hyper-V PFNs for the continguous physical > + * memory described by the sgl entry. The end of the > + * last sgl should be reached at the same time that > + * the PFN array is filled. > */ > - if (hvpgoff_in_page == 0 || i == 0) { > - hvpfn = page_to_hvpfn(sg_page(cur_sgl)); > - cur_sgl = sg_next(cur_sgl); > + while (hvpgoff != sgl_size) { > + payload->range.pfn_array[i++] = > + hvpfn + hvpgoff++; > } Minor nitpicking: while this seems to be correct I, personally, find it a bit hard to read: 'hvpgoff' stands for "'sgl->offset' measured in Hyper-V pages' but we immediately re-use it as a cycle counter. If I'm not mistaken, we can count right away how many entries we're going to add. Also, we could've introduced HVPFN_DOWN() to complement HVPFN_UP(): ... #define HVPFN_DOWN(x) ((x) >> HV_HYP_PAGE_SHIFT) ... hvpgoff = HVPFN_DOWN(sgl->offset); hvpfn = page_to_hvpfn(sg_page(sgl)) + hvpgoff; hvpfns_to_add = HVPFN_UP(sgl->offset + sgl->length) - hvpgoff; and the cycle can look like: while (hvpfns_to_add) { payload->range.pfn_array[i++] = hvpfn++; hvpfns_to_add--; } > - > - payload->range.pfn_array[i] = hvpfn + hvpgoff_in_page; > } and then we can also make an explicit BUG_ON(i != hvpg_count) after the cycle to prove our math is correct :-) > } > > @@ -1851,8 +1834,6 @@ static int storvsc_queuecommand(struct Scsi_Host *host, struct scsi_cmnd *scmnd) > .slave_configure = storvsc_device_configure, > .cmd_per_lun = 2048, > .this_id = -1, > - /* Make sure we dont get a sg segment crosses a page boundary */ > - .dma_boundary = PAGE_SIZE-1, > /* Ensure there are no gaps in presented sgls */ > .virt_boundary_mask = PAGE_SIZE-1, > .no_write_same = 1, -- Vitaly
From: Vitaly Kuznetsov <vkuznets@redhat.com> Sent: Tuesday, February 23, 2021 6:30 AM > > Michael Kelley <mikelley@microsoft.com> writes: > > > storvsc currently sets .dma_boundary to limit scatterlist entries > > to 4 Kbytes, which is less efficient with huge pages that offer > > large chunks of contiguous physical memory. Improve the algorithm > > for creating the Hyper-V guest physical address PFN array so > > that scatterlist entries with lengths > 4Kbytes are handled. > > As a result, remove the .dma_boundary setting. > > > > The improved algorithm also adds support for scatterlist > > entries with offsets >= 4Kbytes, which is supported by many > > other SCSI low-level drivers. And it retains support for > > architectures where possibly PAGE_SIZE != HV_HYP_PAGE_SIZE > > (such as ARM64). > > > > Signed-off-by: Michael Kelley <mikelley@microsoft.com> > > --- > > drivers/scsi/storvsc_drv.c | 63 ++++++++++++++++------------------------------ > > 1 file changed, 22 insertions(+), 41 deletions(-) > > > > diff --git a/drivers/scsi/storvsc_drv.c b/drivers/scsi/storvsc_drv.c > > index 2e4fa77..5d06061 100644 > > --- a/drivers/scsi/storvsc_drv.c > > +++ b/drivers/scsi/storvsc_drv.c > > @@ -1678,9 +1678,8 @@ static int storvsc_queuecommand(struct Scsi_Host *host, struct > scsi_cmnd *scmnd) > > struct storvsc_cmd_request *cmd_request = scsi_cmd_priv(scmnd); > > int i; > > struct scatterlist *sgl; > > - unsigned int sg_count = 0; > > + unsigned int sg_count; > > struct vmscsi_request *vm_srb; > > - struct scatterlist *cur_sgl; > > struct vmbus_packet_mpb_array *payload; > > u32 payload_sz; > > u32 length; > > @@ -1759,7 +1758,7 @@ static int storvsc_queuecommand(struct Scsi_Host *host, struct > scsi_cmnd *scmnd) > > payload_sz = sizeof(cmd_request->mpb); > > > > if (sg_count) { > > - unsigned int hvpgoff = 0; > > + unsigned int hvpgoff, sgl_size; > > unsigned long offset_in_hvpg = sgl->offset & ~HV_HYP_PAGE_MASK; > > unsigned int hvpg_count = HVPFN_UP(offset_in_hvpg + length); > > u64 hvpfn; > > @@ -1773,51 +1772,35 @@ static int storvsc_queuecommand(struct Scsi_Host *host, > struct scsi_cmnd *scmnd) > > 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. > > - * Besides, payload->range.offset should be the offset in one > > - * HV_HYP_PAGE. > > - */ > > payload->range.len = length; > > payload->range.offset = offset_in_hvpg; > > - hvpgoff = sgl->offset >> HV_HYP_PAGE_SHIFT; > > > > - cur_sgl = sgl; > > - for (i = 0; i < hvpg_count; i++) { > > + > > + for (i = 0; sgl != NULL; sgl = sg_next(sgl)) { > > /* > > - * 'i' is the index of hv pages in the payload and > > - * 'hvpgoff' is the offset (in hv pages) of the first > > - * hv page in the the first page. The relationship > > - * between the sum of 'i' and 'hvpgoff' and the offset > > - * (in hv pages) in a payload page ('hvpgoff_in_page') > > - * is as follow: > > - * > > - * |------------------ PAGE -------------------| > > - * | NR_HV_HYP_PAGES_IN_PAGE hvpgs in total | > > - * |hvpg|hvpg| ... |hvpg|... |hvpg| > > - * ^ ^ ^ ^ > > - * +-hvpgoff-+ +-hvpgoff_in_page-+ > > - * ^ | > > - * +--------------------- i ---------------------------+ > > + * Init values for the current sgl entry. sgl_size > > + * and hvpgoff are in units of Hyper-V size pages. > > + * Handling the PAGE_SIZE != HV_HYP_PAGE_SIZE case > > + * also handles values of sgl->offset that are > > + * larger than PAGE_SIZE. Such offsets are handled > > + * even on other than the first sgl entry, provided > > + * they are a multiple of PAGE_SIZE. > > */ > > - unsigned int hvpgoff_in_page = > > - (i + hvpgoff) % NR_HV_HYP_PAGES_IN_PAGE; > > + sgl_size = HVPFN_UP(sgl->offset + sgl->length); > > + hvpgoff = sgl->offset >> HV_HYP_PAGE_SHIFT; > > + hvpfn = page_to_hvpfn(sg_page(sgl)); > > > > /* > > - * Two cases that we need to fetch a page: > > - * 1) i == 0, the first step or > > - * 2) hvpgoff_in_page == 0, when we reach the boundary > > - * of a page. > > + * Fill the next portion of the PFN array with > > + * sequential Hyper-V PFNs for the continguous physical > > + * memory described by the sgl entry. The end of the > > + * last sgl should be reached at the same time that > > + * the PFN array is filled. > > */ > > - if (hvpgoff_in_page == 0 || i == 0) { > > - hvpfn = page_to_hvpfn(sg_page(cur_sgl)); > > - cur_sgl = sg_next(cur_sgl); > > + while (hvpgoff != sgl_size) { > > + payload->range.pfn_array[i++] = > > + hvpfn + hvpgoff++; > > } > > Minor nitpicking: while this seems to be correct I, personally, find it > a bit hard to read: 'hvpgoff' stands for "'sgl->offset' measured in > Hyper-V pages' but we immediately re-use it as a cycle counter. > > If I'm not mistaken, we can count right away how many entries we're > going to add. Also, we could've introduced HVPFN_DOWN() to complement > HVPFN_UP(): > ... > #define HVPFN_DOWN(x) ((x) >> HV_HYP_PAGE_SHIFT) > ... > > hvpgoff = HVPFN_DOWN(sgl->offset); > hvpfn = page_to_hvpfn(sg_page(sgl)) + hvpgoff; > hvpfns_to_add = HVPFN_UP(sgl->offset + sgl->length) - hvpgoff; > > and the cycle can look like: > > while (hvpfns_to_add) { > payload->range.pfn_array[i++] = hvpfn++; > hvpfns_to_add--; > } > > > - > > - payload->range.pfn_array[i] = hvpfn + hvpgoff_in_page; > > } > > and then we can also make an explicit > > BUG_ON(i != hvpg_count) after the cycle to prove our math is correct :-) > Your proposal works for me, and it doesn't actually change the number of arithmetic operations that need to be done. But I don't think I'll actually add the BUG_ON(). The math is either right or it isn't, and we have it right. I'll spin a v2. Michael > > } > > > > @@ -1851,8 +1834,6 @@ static int storvsc_queuecommand(struct Scsi_Host *host, struct > scsi_cmnd *scmnd) > > .slave_configure = storvsc_device_configure, > > .cmd_per_lun = 2048, > > .this_id = -1, > > - /* Make sure we dont get a sg segment crosses a page boundary */ > > - .dma_boundary = PAGE_SIZE-1, > > /* Ensure there are no gaps in presented sgls */ > > .virt_boundary_mask = PAGE_SIZE-1, > > .no_write_same = 1, > > -- > Vitaly
Shouldn't storvsc just use blk_queue_virt_boundary instead of all this mess?
From: Christoph Hellwig <hch@infradead.org> Sent: Wednesday, February 24, 2021 7:53 AM > > Shouldn't storvsc just use blk_queue_virt_boundary instead of all this > mess? The storvsc driver does set the virt boundary to PAGE_SIZE - 1. But the driver still has to translate the scatterlist into a list of guest physical frame numbers (each representing 4K bytes) that the Hyper-V host understands so it can do the I/O. This patch improves that translation so it can handle a single scatterlist entry that represents more than PAGE_SIZE bytes of data. Then the SCSI dma_boundary (which turns into the blk level segment_boundary) no longer needs to be set to restrict scatterlist entries to just PAGE_SIZE bytes. We also have to preserve the ability to run guests on ARM64 with PAGE_SIZE of 16K or 64K, while Hyper-V still expects each PFN to represent only 4K bytes. Michael
diff --git a/drivers/scsi/storvsc_drv.c b/drivers/scsi/storvsc_drv.c index 2e4fa77..5d06061 100644 --- a/drivers/scsi/storvsc_drv.c +++ b/drivers/scsi/storvsc_drv.c @@ -1678,9 +1678,8 @@ static int storvsc_queuecommand(struct Scsi_Host *host, struct scsi_cmnd *scmnd) struct storvsc_cmd_request *cmd_request = scsi_cmd_priv(scmnd); int i; struct scatterlist *sgl; - unsigned int sg_count = 0; + unsigned int sg_count; struct vmscsi_request *vm_srb; - struct scatterlist *cur_sgl; struct vmbus_packet_mpb_array *payload; u32 payload_sz; u32 length; @@ -1759,7 +1758,7 @@ static int storvsc_queuecommand(struct Scsi_Host *host, struct scsi_cmnd *scmnd) payload_sz = sizeof(cmd_request->mpb); if (sg_count) { - unsigned int hvpgoff = 0; + unsigned int hvpgoff, sgl_size; unsigned long offset_in_hvpg = sgl->offset & ~HV_HYP_PAGE_MASK; unsigned int hvpg_count = HVPFN_UP(offset_in_hvpg + length); u64 hvpfn; @@ -1773,51 +1772,35 @@ static int storvsc_queuecommand(struct Scsi_Host *host, struct scsi_cmnd *scmnd) 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. - * Besides, payload->range.offset should be the offset in one - * HV_HYP_PAGE. - */ payload->range.len = length; payload->range.offset = offset_in_hvpg; - hvpgoff = sgl->offset >> HV_HYP_PAGE_SHIFT; - cur_sgl = sgl; - for (i = 0; i < hvpg_count; i++) { + + for (i = 0; sgl != NULL; sgl = sg_next(sgl)) { /* - * 'i' is the index of hv pages in the payload and - * 'hvpgoff' is the offset (in hv pages) of the first - * hv page in the the first page. The relationship - * between the sum of 'i' and 'hvpgoff' and the offset - * (in hv pages) in a payload page ('hvpgoff_in_page') - * is as follow: - * - * |------------------ PAGE -------------------| - * | NR_HV_HYP_PAGES_IN_PAGE hvpgs in total | - * |hvpg|hvpg| ... |hvpg|... |hvpg| - * ^ ^ ^ ^ - * +-hvpgoff-+ +-hvpgoff_in_page-+ - * ^ | - * +--------------------- i ---------------------------+ + * Init values for the current sgl entry. sgl_size + * and hvpgoff are in units of Hyper-V size pages. + * Handling the PAGE_SIZE != HV_HYP_PAGE_SIZE case + * also handles values of sgl->offset that are + * larger than PAGE_SIZE. Such offsets are handled + * even on other than the first sgl entry, provided + * they are a multiple of PAGE_SIZE. */ - unsigned int hvpgoff_in_page = - (i + hvpgoff) % NR_HV_HYP_PAGES_IN_PAGE; + sgl_size = HVPFN_UP(sgl->offset + sgl->length); + hvpgoff = sgl->offset >> HV_HYP_PAGE_SHIFT; + hvpfn = page_to_hvpfn(sg_page(sgl)); /* - * Two cases that we need to fetch a page: - * 1) i == 0, the first step or - * 2) hvpgoff_in_page == 0, when we reach the boundary - * of a page. + * Fill the next portion of the PFN array with + * sequential Hyper-V PFNs for the continguous physical + * memory described by the sgl entry. The end of the + * last sgl should be reached at the same time that + * the PFN array is filled. */ - if (hvpgoff_in_page == 0 || i == 0) { - hvpfn = page_to_hvpfn(sg_page(cur_sgl)); - cur_sgl = sg_next(cur_sgl); + while (hvpgoff != sgl_size) { + payload->range.pfn_array[i++] = + hvpfn + hvpgoff++; } - - payload->range.pfn_array[i] = hvpfn + hvpgoff_in_page; } } @@ -1851,8 +1834,6 @@ static int storvsc_queuecommand(struct Scsi_Host *host, struct scsi_cmnd *scmnd) .slave_configure = storvsc_device_configure, .cmd_per_lun = 2048, .this_id = -1, - /* Make sure we dont get a sg segment crosses a page boundary */ - .dma_boundary = PAGE_SIZE-1, /* Ensure there are no gaps in presented sgls */ .virt_boundary_mask = PAGE_SIZE-1, .no_write_same = 1,
storvsc currently sets .dma_boundary to limit scatterlist entries to 4 Kbytes, which is less efficient with huge pages that offer large chunks of contiguous physical memory. Improve the algorithm for creating the Hyper-V guest physical address PFN array so that scatterlist entries with lengths > 4Kbytes are handled. As a result, remove the .dma_boundary setting. The improved algorithm also adds support for scatterlist entries with offsets >= 4Kbytes, which is supported by many other SCSI low-level drivers. And it retains support for architectures where possibly PAGE_SIZE != HV_HYP_PAGE_SIZE (such as ARM64). Signed-off-by: Michael Kelley <mikelley@microsoft.com> --- drivers/scsi/storvsc_drv.c | 63 ++++++++++++++++------------------------------ 1 file changed, 22 insertions(+), 41 deletions(-)