Re: [PATCH RFC v5 00/29] io_uring getdents

[Dave Chinner <david@fromorbit.com>]

On Fri, Aug 25, 2023 at 09:54:02PM +0800, Hao Xu wrote:
> From: Hao Xu <howeyxu@tencent.com>
> 
> This series introduce getdents64 to io_uring, the code logic is similar
> with the snychronized version's. It first try nowait issue, and offload
> it to io-wq threads if the first try fails.
> 
> Patch1 and Patch2 are some preparation
> Patch3 supports nowait for xfs getdents code
> Patch4-11 are vfs change, include adding helpers and trylock for locks
> Patch12-29 supports nowait for involved xfs journal stuff
> note, Patch24 and 27 are actually two questions, might be removed later.
> an xfs test may come later.

You need to drop all the XFS journal stuff. It's fundamentally
broken as it stands, and we cannot support non-blocking
transactional changes without first putting a massive investment in
transaction and intent chain rollback to allow correctly undoing
partially complete modifications.

Regardless, non-blocking transactions are completely unnecessary for
a non-blocking readdir implementation. readdir should only be
touching atime, and with relatime it should only occur once every 24
hours per inode. If that's a problem, then we have noatime mount
options. Hence I just don't see any point in worrying about having a
timestamp update block occasionally...

I also don't really don't see why you need to fiddle with xfs buffer
cache semantics - it already has the functionality "nowait" buffer
reads require (i.e.  XBF_INCORE|XBF_TRYLOCK).

However, the readahead IO that the xfs readdir code issues cannot
use your defined NOWAIT semantics - it must be able to allocate
memory and issue IO. Readahead already avoids blocking on memory
allocation and blocking on IO via the XBF_READ_AHEAD flag. This sets
__GFP_NORETRY for buffer allocation and REQ_RAHEAD for IO. Hence
readahead only needs the existing XBF_TRYLOCK flag to be set to be
compatible with the required NOWAIT semantics....

As for the NOIO memory allocation restrictions io_uring requires,
that should be enforced at the io_uring layer before calling into
the VFS using memalloc_noio_save/restore.  At that point no memory
allocation will trigger IO and none of the code running under NOWAIT
conditions even needs to be aware that io_uring has a GFP_NOIO
restriction on memory allocation....

Please go back to the simple "do non-blocking buffer IO"
implementation we started with and don't try to solve every little
blocking problem that might exist in the VFS and filesystems...

-Dave
-- 
Dave Chinner
david@fromorbit.com