Re: [Qemu-devel] [PATCH for-2.9 v3] file-posix: Consider max_segments for BlockLimits.max_transfer

[Kevin Wolf <kwolf@redhat.com>]

Am 08.03.2017 um 13:08 hat Fam Zheng geschrieben:
> BlockLimits.max_transfer can be too high without this fix, guest will
> encounter I/O error or even get paused with werror=stop or rerror=stop. The
> cause is explained below.
> 
> Linux has a separate limit, /sys/block/.../queue/max_segments, which in
> the worst case can be more restrictive than the BLKSECTGET which we
> already consider (note that they are two different things). So, the
> failure scenario before this patch is:
> 
> 1) host device has max_sectors_kb = 4096 and max_segments = 64;
> 2) guest learns max_sectors_kb limit from QEMU, but doesn't know
>    max_segments;
> 3) guest issues e.g. a 512KB request thinking it's okay, but actually
>    it's not, because it will be passed through to host device as an
>    SG_IO req that has niov > 64;
> 4) host kernel doesn't like the segmenting of the request, and returns
>    -EINVAL;
> 
> This patch checks the max_segments sysfs entry for the host device and
> calculates a "conservative" bytes limit using the page size, which is
> then merged into the existing max_transfer limit. Guest will discover
> this from the usual virtual block device interfaces. (In the case of
> scsi-generic, it will be done in the INQUIRY reply interception in
> device model.)
> 
> The other possibility is to actually propagate it as a separate limit,
> but it's not better. On the one hand, there is a big complication: the
> limit is per-LUN in QEMU PoV (because we can attach LUNs from different
> host HBAs to the same virtio-scsi bus), but the channel to communicate
> it in a per-LUN manner is missing down the stack; on the other hand,
> two limits versus one doesn't change much about the valid size of I/O
> (because guest has no control over host segmenting).
> 
> Also, the idea to fall back to bounce buffering in QEMU, upon -EINVAL,
> was explored. Unfortunately there is no neat way to ensure the bounce
> buffer is less segmented (in terms of DMA addr) than the guest buffer.
> 
> Practically, this bug is not very common. It is only reported on a
> Emulex (lpfc), so it's okay to get it fixed in the easier way.
> 
> Reviewed-by: Paolo Bonzini <pbonzini@redhat.com>
> Signed-off-by: Fam Zheng <famz@redhat.com>

Thanks, applied to the block branch.

Kevin