For block host devices, I/O can happen through either the kernel file
descriptor I/O system calls (preadv/pwritev, io_submit, io_uring)
or the SCSI passthrough ioctl SG_IO.

In the latter case, the size of each transfer can be limited by the
HBA, while for file descriptor I/O the kernel is able to split and
merge I/O in smaller pieces as needed.  Applying the HBA limits to
file descriptor I/O results in more system calls and suboptimal
performance, so this patch splits the max_transfer limit in two:
max_transfer remains valid and is used in general, while max_hw_transfer
is limited to the maximum hardware size.  max_hw_transfer can then be
included by the scsi-generic driver in the block limits page, to ensure
that the stricter hardware limit is used.

Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
---
 block/block-backend.c          | 12 ++++++++++++
 block/file-posix.c             |  2 +-
 block/io.c                     |  1 +
 hw/scsi/scsi-generic.c         |  2 +-
 include/block/block_int.h      |  7 +++++++
 include/sysemu/block-backend.h |  1 +
 6 files changed, 23 insertions(+), 2 deletions(-)

diff --git a/block/block-backend.c b/block/block-backend.c
index 15f1ea4288..2ea1412a54 100644
--- a/block/block-backend.c
+++ b/block/block-backend.c
@@ -1953,6 +1953,18 @@ uint32_t blk_get_request_alignment(BlockBackend *blk)
     return bs ? bs->bl.request_alignment : BDRV_SECTOR_SIZE;
 }
 
+/* Returns the maximum hardware transfer length, in bytes; guaranteed nonzero */
+uint64_t blk_get_max_hw_transfer(BlockBackend *blk)
+{
+    BlockDriverState *bs = blk_bs(blk);
+    uint64_t max = INT_MAX;
+
+    if (bs) {
+        max = MIN_NON_ZERO(bs->bl.max_hw_transfer, bs->bl.max_transfer);
+    }
+    return max;
+}
+
 /* Returns the maximum transfer length, in bytes; guaranteed nonzero */
 uint32_t blk_get_max_transfer(BlockBackend *blk)
 {
diff --git a/block/file-posix.c b/block/file-posix.c
index 670c577bfe..c9746d3eb6 100644
--- a/block/file-posix.c
+++ b/block/file-posix.c
@@ -1234,7 +1234,7 @@ static void raw_refresh_limits(BlockDriverState *bs, Error **errp)
         int ret = sg_get_max_transfer_length(s->fd);
 
         if (ret > 0 && ret <= BDRV_REQUEST_MAX_BYTES) {
-            bs->bl.max_transfer = pow2floor(ret);
+            bs->bl.max_hw_transfer = pow2floor(ret);
         }
 
         ret = sg_get_max_segments(s->fd);
diff --git a/block/io.c b/block/io.c
index 323854d063..089b99bb0c 100644
--- a/block/io.c
+++ b/block/io.c
@@ -127,6 +127,7 @@ static void bdrv_merge_limits(BlockLimits *dst, const BlockLimits *src)
 {
     dst->opt_transfer = MAX(dst->opt_transfer, src->opt_transfer);
     dst->max_transfer = MIN_NON_ZERO(dst->max_transfer, src->max_transfer);
+    dst->max_hw_transfer = MIN_NON_ZERO(dst->max_hw_transfer, src->max_hw_transfer);
     dst->opt_mem_alignment = MAX(dst->opt_mem_alignment,
                                  src->opt_mem_alignment);
     dst->min_mem_alignment = MAX(dst->min_mem_alignment,
diff --git a/hw/scsi/scsi-generic.c b/hw/scsi/scsi-generic.c
index 82e1e2ee79..3762dce749 100644
--- a/hw/scsi/scsi-generic.c
+++ b/hw/scsi/scsi-generic.c
@@ -179,7 +179,7 @@ static void scsi_handle_inquiry_reply(SCSIGenericReq *r, SCSIDevice *s)
         (r->req.cmd.buf[1] & 0x01)) {
         page = r->req.cmd.buf[2];
         if (page == 0xb0) {
-            uint32_t max_transfer = blk_get_max_transfer(s->conf.blk);
+            uint64_t max_transfer = blk_get_max_hw_transfer(s->conf.blk);
             uint32_t max_iov = blk_get_max_iov(s->conf.blk);
 
             assert(max_transfer);
diff --git a/include/block/block_int.h b/include/block/block_int.h
index 057d88b1fc..f1a54db0f8 100644
--- a/include/block/block_int.h
+++ b/include/block/block_int.h
@@ -695,6 +695,13 @@ typedef struct BlockLimits {
      * clamped down. */
     uint32_t max_transfer;
 
+    /* Maximal hardware transfer length in bytes.  Applies whenever
+     * transfers to the device bypass the kernel I/O scheduler, for
+     * example with SG_IO.  If larger than max_transfer or if zero,
+     * blk_get_max_hw_transfer will fall back to max_transfer.
+     */
+    uint64_t max_hw_transfer;
+
     /* memory alignment, in bytes so that no bounce buffer is needed */
     size_t min_mem_alignment;
 
diff --git a/include/sysemu/block-backend.h b/include/sysemu/block-backend.h
index 5423e3d9c6..9ac5f7bbd3 100644
--- a/include/sysemu/block-backend.h
+++ b/include/sysemu/block-backend.h
@@ -208,6 +208,7 @@ void blk_eject(BlockBackend *blk, bool eject_flag);
 int blk_get_flags(BlockBackend *blk);
 uint32_t blk_get_request_alignment(BlockBackend *blk);
 uint32_t blk_get_max_transfer(BlockBackend *blk);
+uint64_t blk_get_max_hw_transfer(BlockBackend *blk);
 int blk_get_max_iov(BlockBackend *blk);
 void blk_set_guest_block_size(BlockBackend *blk, int align);
 void *blk_try_blockalign(BlockBackend *blk, size_t size);
-- 
2.31.1

On Tue, Jun 08, 2021 at 03:16:30PM +0200, Paolo Bonzini wrote:
> For block host devices, I/O can happen through either the kernel file
> descriptor I/O system calls (preadv/pwritev, io_submit, io_uring)
> or the SCSI passthrough ioctl SG_IO.
> 
> In the latter case, the size of each transfer can be limited by the
> HBA, while for file descriptor I/O the kernel is able to split and
> merge I/O in smaller pieces as needed.  Applying the HBA limits to
> file descriptor I/O results in more system calls and suboptimal
> performance, so this patch splits the max_transfer limit in two:
> max_transfer remains valid and is used in general, while max_hw_transfer
> is limited to the maximum hardware size.  max_hw_transfer can then be
> included by the scsi-generic driver in the block limits page, to ensure
> that the stricter hardware limit is used.
> 
> Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
> ---
>  block/block-backend.c          | 12 ++++++++++++
>  block/file-posix.c             |  2 +-
>  block/io.c                     |  1 +
>  hw/scsi/scsi-generic.c         |  2 +-
>  include/block/block_int.h      |  7 +++++++
>  include/sysemu/block-backend.h |  1 +

[you can use git's orderfile option to put .h changes first]

>  6 files changed, 23 insertions(+), 2 deletions(-)
> 
> diff --git a/block/block-backend.c b/block/block-backend.c
> index 15f1ea4288..2ea1412a54 100644
> --- a/block/block-backend.c
> +++ b/block/block-backend.c
> @@ -1953,6 +1953,18 @@ uint32_t blk_get_request_alignment(BlockBackend *blk)
>      return bs ? bs->bl.request_alignment : BDRV_SECTOR_SIZE;
>  }
>  
> +/* Returns the maximum hardware transfer length, in bytes; guaranteed nonzero */
> +uint64_t blk_get_max_hw_transfer(BlockBackend *blk)

Since we have declared (elsewhere) that the maximum block device is
signed, would this be better as int64_t?  (Our reasoning is that off_t
is also signed, and we are unlikely to need to handle anything bigger
than what off_t can handle; plus it leaves room for returning errors,
although this particular function is not giving errors; see also
Vladimir's work on making the block layer 64-bit clean).  I'm not
opposed to unsigned here to represent lack of errors, but consistency
with the rest of the block layer may argue for signed.

> +++ b/include/block/block_int.h
> @@ -695,6 +695,13 @@ typedef struct BlockLimits {
>       * clamped down. */
>      uint32_t max_transfer;
>  
> +    /* Maximal hardware transfer length in bytes.  Applies whenever
> +     * transfers to the device bypass the kernel I/O scheduler, for
> +     * example with SG_IO.  If larger than max_transfer or if zero,
> +     * blk_get_max_hw_transfer will fall back to max_transfer.
> +     */
> +    uint64_t max_hw_transfer;
> +

Reviewed-by: Eric Blake <eblake@redhat.com>

-- 
Eric Blake, Principal Software Engineer
Red Hat, Inc.           +1-919-301-3266
Virtualization:  qemu.org | libvirt.org

On Tue, 2021-06-08 at 15:16 +0200, Paolo Bonzini wrote:
> For block host devices, I/O can happen through either the kernel file
> descriptor I/O system calls (preadv/pwritev, io_submit, io_uring)
> or the SCSI passthrough ioctl SG_IO.
> 
> In the latter case, the size of each transfer can be limited by the
> HBA, while for file descriptor I/O the kernel is able to split and
> merge I/O in smaller pieces as needed.  Applying the HBA limits to
> file descriptor I/O results in more system calls and suboptimal
> performance, so this patch splits the max_transfer limit in two:
> max_transfer remains valid and is used in general, while max_hw_transfer
> is limited to the maximum hardware size.  max_hw_transfer can then be
> included by the scsi-generic driver in the block limits page, to ensure
> that the stricter hardware limit is used.
> 
> Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

This is mostly the same as my patch 

https://www.mail-archive.com/qemu-devel@nongnu.org/msg768264.html

I called this max_ioctl_transfer, since this limit is only relevant
to the .ioctl, but max_hw_transfer is fine as well.

So this patch looks OK, but I might have missed something
as I haven't touched this area for a long time.

Reviewed-by: Maxim Levitsky <mlevitsk@redhat.com>

Best regards,
	Maxim Levitsky


> ---
>  block/block-backend.c          | 12 ++++++++++++
>  block/file-posix.c             |  2 +-
>  block/io.c                     |  1 +
>  hw/scsi/scsi-generic.c         |  2 +-
>  include/block/block_int.h      |  7 +++++++
>  include/sysemu/block-backend.h |  1 +
>  6 files changed, 23 insertions(+), 2 deletions(-)
> 
> diff --git a/block/block-backend.c b/block/block-backend.c
> index 15f1ea4288..2ea1412a54 100644
> --- a/block/block-backend.c
> +++ b/block/block-backend.c
> @@ -1953,6 +1953,18 @@ uint32_t blk_get_request_alignment(BlockBackend *blk)
>      return bs ? bs->bl.request_alignment : BDRV_SECTOR_SIZE;
>  }
>  
> +/* Returns the maximum hardware transfer length, in bytes; guaranteed nonzero */
> +uint64_t blk_get_max_hw_transfer(BlockBackend *blk)
> +{
> +    BlockDriverState *bs = blk_bs(blk);
> +    uint64_t max = INT_MAX;
> +
> +    if (bs) {
> +        max = MIN_NON_ZERO(bs->bl.max_hw_transfer, bs->bl.max_transfer);
> +    }
> +    return max;
> +}
> +
>  /* Returns the maximum transfer length, in bytes; guaranteed nonzero */
>  uint32_t blk_get_max_transfer(BlockBackend *blk)
>  {
> diff --git a/block/file-posix.c b/block/file-posix.c
> index 670c577bfe..c9746d3eb6 100644
> --- a/block/file-posix.c
> +++ b/block/file-posix.c
> @@ -1234,7 +1234,7 @@ static void raw_refresh_limits(BlockDriverState *bs, Error **errp)
>          int ret = sg_get_max_transfer_length(s->fd);
>  
>          if (ret > 0 && ret <= BDRV_REQUEST_MAX_BYTES) {
> -            bs->bl.max_transfer = pow2floor(ret);
> +            bs->bl.max_hw_transfer = pow2floor(ret);
>          }
>  
>          ret = sg_get_max_segments(s->fd);
> diff --git a/block/io.c b/block/io.c
> index 323854d063..089b99bb0c 100644
> --- a/block/io.c
> +++ b/block/io.c
> @@ -127,6 +127,7 @@ static void bdrv_merge_limits(BlockLimits *dst, const BlockLimits *src)
>  {
>      dst->opt_transfer = MAX(dst->opt_transfer, src->opt_transfer);
>      dst->max_transfer = MIN_NON_ZERO(dst->max_transfer, src->max_transfer);
> +    dst->max_hw_transfer = MIN_NON_ZERO(dst->max_hw_transfer, src->max_hw_transfer);
>      dst->opt_mem_alignment = MAX(dst->opt_mem_alignment,
>                                   src->opt_mem_alignment);
>      dst->min_mem_alignment = MAX(dst->min_mem_alignment,
> diff --git a/hw/scsi/scsi-generic.c b/hw/scsi/scsi-generic.c
> index 82e1e2ee79..3762dce749 100644
> --- a/hw/scsi/scsi-generic.c
> +++ b/hw/scsi/scsi-generic.c
> @@ -179,7 +179,7 @@ static void scsi_handle_inquiry_reply(SCSIGenericReq *r, SCSIDevice *s)
>          (r->req.cmd.buf[1] & 0x01)) {
>          page = r->req.cmd.buf[2];
>          if (page == 0xb0) {
> -            uint32_t max_transfer = blk_get_max_transfer(s->conf.blk);
> +            uint64_t max_transfer = blk_get_max_hw_transfer(s->conf.blk);
>              uint32_t max_iov = blk_get_max_iov(s->conf.blk);
>  
>              assert(max_transfer);
> diff --git a/include/block/block_int.h b/include/block/block_int.h
> index 057d88b1fc..f1a54db0f8 100644
> --- a/include/block/block_int.h
> +++ b/include/block/block_int.h
> @@ -695,6 +695,13 @@ typedef struct BlockLimits {
>       * clamped down. */
>      uint32_t max_transfer;
>  
> +    /* Maximal hardware transfer length in bytes.  Applies whenever
> +     * transfers to the device bypass the kernel I/O scheduler, for
> +     * example with SG_IO.  If larger than max_transfer or if zero,
> +     * blk_get_max_hw_transfer will fall back to max_transfer.
> +     */
> +    uint64_t max_hw_transfer;
> +
>      /* memory alignment, in bytes so that no bounce buffer is needed */
>      size_t min_mem_alignment;
>  
> diff --git a/include/sysemu/block-backend.h b/include/sysemu/block-backend.h
> index 5423e3d9c6..9ac5f7bbd3 100644
> --- a/include/sysemu/block-backend.h
> +++ b/include/sysemu/block-backend.h
> @@ -208,6 +208,7 @@ void blk_eject(BlockBackend *blk, bool eject_flag);
>  int blk_get_flags(BlockBackend *blk);
>  uint32_t blk_get_request_alignment(BlockBackend *blk);
>  uint32_t blk_get_max_transfer(BlockBackend *blk);
> +uint64_t blk_get_max_hw_transfer(BlockBackend *blk);
>  int blk_get_max_iov(BlockBackend *blk);
>  void blk_set_guest_block_size(BlockBackend *blk, int align);
>  void *blk_try_blockalign(BlockBackend *blk, size_t size);