Re: [PATCH v2] block-backend: Add new bds_io_in_flight counter

[Kevin Wolf <kwolf@redhat.com>]

Am 31.03.2023 um 18:23 hat Hanna Czenczek geschrieben:
> IDE TRIM is a BB user that wants to elevate its BB's in-flight counter
> for a "macro" operation that consists of several actual I/O operations.
> Each of those operations is individually started and awaited.  It does
> this so that blk_drain() will drain the whole TRIM, and not just a
> single one of the many discard operations it may encompass.
> 
> When request queuing is enabled, this leads to a deadlock: The currently
> ongoing discard is drained, and the next one is queued, waiting for the
> drain to stop.  Meanwhile, TRIM still keeps the in-flight counter
> elevated, waiting for all discards to stop -- which will never happen,
> because with the in-flight counter elevated, the BB is never considered
> drained, so the drained section does not begin and cannot end.

Alright, let's have another look at this now that another similar
deadlock was reported:
https://lists.gnu.org/archive/html/qemu-block/2023-09/msg00536.html

> There are two separate cases to look at here, namely bdrv_drain*() and
> blk_drain*().  As said above, we do want blk_drain*() to settle the
> whole operation: The only way to do so is to disable request queuing,
> then.  So, we do that: Have blk_drain() and blk_drain_all() temporarily
> disable request queuing, which prevents the deadlock.

Two separate cases with two separate fixes suggests that it could be
two separate patches. I feel the blk_*() case is uncontroversial and it
would fix John's case, so splitting wouldn't only make this easier to
understand, but could mean that we can fix a useful subset earlier.

> (The devil's in the details, though: blk_drain_all() runs
> bdrv_drain_all_begin() first, so when we get to the individual BB, there
> may already be queued requests.  Therefore, we have to not only disable
> request queuing then, but resume all already-queued requests, too.)

Why can't we just disable request queuing before calling bdrv_drain_*()?

Is it possible that the same problem occurs because someone else already
called bdrv_drain_*()? That is, blk_drain_*() would be called from a
callback in the nested event loop in bdrv_drain_*()? If so, we can't
avoid that there are already queued requests.

Restarting them seems correct anyway.

> For bdrv_drain*(), we want request queuing -- and macro requests such as
> IDE's TRIM request do not matter.  bdrv_drain*() wants to keep I/O
> requests from BDS nodes, and the TRIM does not issue such requests; it
> instead does so through blk_*() functions, which themselves elevate the
> BB's in-flight counter.  So the idea is to drain (and potentially queue)
> those blk_*() requests, but completely ignore the TRIM.
> 
> We can do that by splitting a new counter off of the existing BB
> counter: The new bds_io_in_flight counter counts all those blk_*()
> requests that can issue I/O to a BDS (so must be drained by
> bdrv_drain*()), but will never block waiting on another request on the
> BB.

You end up changing all of the existing blk_inc_in_flight() callers
except those in IDE and virtio-blk.

That makes me wonder if it shouldn't be approached the other way around:
BlockBackend users that want to be included in drain should use a
special function blk_inc_in_flight_external() or something that wouldn't
increase blk->in_flight, but only a new separate counter. And then only
blk_drain*() wait for it (by extending the AIO_WAIT_WHILE() condition),
but not the child_root callbacks.

This would give us more directly the semantics that we actually need:
The root BDS doesn't care if the operation on the device level has
completed as long as nothing new arrives, only external callers which
use blk_drain*() do. I believe internal/external is easier to reason
about than "requests that can issue I/O to a BDS [directly]", it keeps
the external callers the special ones that need extra care while the
normal I/O path is unaffected, and it would make the patch much smaller.

> In blk_drain*(), we disable request queuing and settle all requests (the
> full in_flight count).  In bdrv_drain*() (i.e. blk_root_drained_poll()),
> we only settle bds_io_in_flight_count, ignoring all requests that will
> not directly issue I/O requests to BDS nodes.
> 
> Reported-by: Fiona Ebner <f.ebner@proxmox.com>
> Fixes: 7e5cdb345f77d76cb4877fe6230c4e17a7d0d0ca
>        ("ide: Increment BB in-flight counter for TRIM BH")
> Signed-off-by: Hanna Czenczek <hreitz@redhat.com>

Kevin