On 5/8/25 6:15 PM, Damien Le Moal wrote:
On 5/9/25 9:42 AM, Bart Van Assche wrote:
If a bio is split, the bio fragments are added in reverse LBA order into
the plug list. This triggers write errors with zoned storage and
sequential writes. Fix this by preserving the LBA order when inserting in
the plug list.
Preserving the order of the fragment would be a good thing for all block
devices. But what I fail to see here is how this lack of ordering affects zoned
block device writes since zone write plugging will split large BIOs when a
write BIO goes through zone write plugging. That happens before we have a
request, so we should never endup needing to split a zone write request.
Hi Damien,
From a comment in source file block/blk-zoned.c:
"We always receive BIOs after they are split and ready to be issued."
Does this mean that splitting bios happens before these are passed to
the code in block/blk-zoned.c?
Note: as you probably know Christoph's patch series "don't reorder
requests passed to ->queue_rqs" affects bio splitting. The bio splitting
code pushes bio fragments onto the plug list in decreasing LBA order.
Before that patch series the requests pushed onto the plug list last
were processed first and hence bio fragments are submitted in LBA order.
Since that patch series the requests pushed onto the plug list first are
processed first and hence bio fragments are submitted in decreasing LBA
order.
See also
https://lore.kernel.org/linux-block/20241113152050.157179-1-hch@xxxxxx/.
Thanks,
Bart.
