Re: [f2fs-dev] [PATCH v3] f2fs: fix race of pending_pages in decompression

[Eric Biggers <ebiggers@kernel.org>]

On Wed, Dec 09, 2020 at 09:34:06AM +0800, Chao Yu wrote:
> On 2020/12/9 7:55, Jaegeuk Kim wrote:
> > On 12/07, Eric Biggers wrote:
> > > On Tue, Dec 08, 2020 at 08:51:45AM +0900, Daeho Jeong wrote:
> > > > > I am trying to review this but it is very hard, as the f2fs compression code is
> > > > > very hard to understand.
> > > > > 
> > > > > It looks like a 'struct decompress_io_ctx' represents the work to decompress a
> > > > > particular cluster.  Since the compressed data of the cluster can be read using
> > > > > multiple bios, there is a reference count of how many pages are remaining to be
> > > > > read before all the cluster's pages have been read and decompression can start.
> > > > > 
> > > > > What I don't understand is why that reference counting needs to work differently
> > > > > depending on whether verity is enabled or not.  Shouldn't it be exactly the
> > > > > same?
> > > > > 
> > > > > There also seems to be some confusion about the scope of STEP_VERITY.  Before
> > > > > f2fs compression was added, it was a per-bio thing.  But now in a compressed
> > > > > file, it's really a per-cluster thing, since all decompressed pages in a
> > > > > compressed cluster are verified (or not verified) at once.
> > > > > 
> > > > > Wouldn't it make a lot more sense to, when a cluster needs both compression and
> > > > > verity, *not* set STEP_VERITY on the bios, but rather set a similar flag in the
> > > > > decompress_io_ctx?
> > > > > 
> > > > 
> > > > Eric,
> > > > 
> > > > Decompression and verity can be executed in different thread contexts
> > > > in different timing, so we need separate counts for each.
> > > > 
> > > > We already use STEP_VERITY for non-compression case, so I think using
> > > > this flag in here looks more making sense.
> > > > 
> > > > Thanks,
> > > 
> > > That didn't really answer my questions.
> > > 
> > > I gave up trying to review this patch as the compression post-read handling is
> > > just way too weird and hard to understand.  I wrote a patch to clean it all up
> > > instead, please take a look:
> > > https://lkml.kernel.org/r/20201208060328.2237091-1-ebiggers@kernel.org
> > 
> > Eric,
> > I also tried to review your patch, but it's quite hard to follow quickly and
> 
> Me too, it needs more time to check whether the cleanup doesn't miss any cases.
> 
> Thanks,
> 
> > requires stress tests for a while. Given upcoming merge window and urgency of
> > the bug, let me apply Daeho's fix first. By any chance, may I ask revisiting
> > your clean-up on top of the fix in the next cycle?
> > 
> > Thanks,

I'm not in a hurry, please just take a look when you have time.

- Eric