Re: [RFC PATCH 0/3] Remove mrlock

["Darrick J. Wong" <djwong@kernel.org>]

On Wed, Jan 13, 2021 at 02:17:29PM +0200, Nikolay Borisov wrote:
> 
> 
> On 13.01.21 г. 14:09 ч., Christoph Hellwig wrote:
> > On Wed, Jan 13, 2021 at 01:41:09PM +0200, Nikolay Borisov wrote:
> >>
> >>
> >> On 13.01.21 ??. 13:27 ??., Christoph Hellwig wrote:
> >>> Pavel has looked into this before and got stuck on the allocator
> >>> workqueue offloads:
> >>>
> >>> [PATCH v13 0/4] xfs: Remove wrappers for some semaphores
> >>
> >> I haven't looked into his series but I fail to see how lifting
> >> rwsemaphore out of the nested structure can change the behavior ? It
> >> just removes a level of indirection. My patches are semantically
> >> identical to the original code.
> > 
> > mrlocks have the mr_writer field that annotate that the is a writer
> > locking the lock.  The XFS asserts use it to assert that the lock that
> > the current thread holds it for exclusive protection, which isn't
> > actually what the field says, and this breaks when XFS uses synchronous
> > execution of work_struct as basically an extension of the kernel stack.
> 
> I'm still failing to see what's the problem of checking the last bit of
> the rwsem ->count field. It is set when the sem is held for writing
> (identical to what mr_write does). As I mention in the cover letter this
> might be considered a bit hacky because it exposes an internal detail of
> the rwsem i.e that the bit of interest is bit 0.

I don't want to tear into the implementation details of rwsems if I can
avoid it.  Just because we all have one big happy address space doesn't
mean shortcuts won't hose everyone.

> But I believe the same
> can be achieved using lockdep_is_held_type(xx, 0/1) and making XFS's
> debug routines depend on lockdep being on.

Pavel Reichl tried that two months ago in:
https://lore.kernel.org/linux-xfs/20201102194135.174806-2-preichl@redhat.com/

which resulted in fstests regressions:
https://lore.kernel.org/linux-xfs/20201104005345.GC7115@magnolia/

TLDR: the ILOCK semaphore is data-centric, but lockdep's debugging
chains are task-centric, which causes incorrect lock validation reports.

The solutions as I see them are: (a) figure out if we really still need
to defer bmbt splits to workers to avoid overflowing the kernel stack;
or (b) making it possible to transfer rwsem ownership to shut up
lockdep; or (c) fix the is_held predicate to ignore ownership.

--D

> 
> >