Re: [PATCH 1/5] locking: Add rwsem_is_write_locked()

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On Tue, Sep 12, 2023 at 04:23:00PM +0200, Peter Zijlstra wrote:
> On Tue, Sep 12, 2023 at 02:58:32PM +0100, Matthew Wilcox wrote:
> > On Tue, Sep 12, 2023 at 03:52:13PM +0200, Peter Zijlstra wrote:
> > > On Tue, Sep 12, 2023 at 01:28:13PM +0100, Matthew Wilcox wrote:
> > > > On Tue, Sep 12, 2023 at 11:03:42AM +0200, Peter Zijlstra wrote:
> > > > > If not, then sure we can do this; it's not like I managed to get rid of
> > > > > muteX_is_locked() -- and I actually tried at some point :/
> > > > > 
> > > > > And just now I grepped for it, and look what I find:
> > > > > 
> > > > > drivers/hid/hid-nintendo.c:     if (unlikely(mutex_is_locked(&ctlr->output_mutex))) {
> > > > > drivers/nvdimm/btt.c:           if (mutex_is_locked(&arena->err_lock)
> > > > > 
> > > > > And there's more :-(
> > > > 
> > > > Are these actually abuse?  I looked at these two, and they both seem to
> > > > be asking "Does somebody else currently have this mutex?" rather than
> > > > "Do I have this mutex?".
> > > 
> > > It's effectively a random number generator in that capacity. Someone
> > > might have it or might have had it when you looked and no longer have
> > > it, or might have it now but not when you asked.
> > 
> > Well, no.
> > 
> >                 if (mutex_is_locked(&arena->err_lock)
> >                                 || arena->freelist[lane].has_err) {
> >                         nd_region_release_lane(btt->nd_region, lane);
> > 
> >                         ret = arena_clear_freelist_error(arena, lane);
> > 
> > So that's "Is somebody currently processing an error, or have they
> > already finished setting an error".  Sure, it's somewhat racy, but
> > it looks like a performance optimisation, not something that needs
> > 100% accuracy.
> 
> We're arguing past one another I think. Yes mutex_is_locked() is a
> random number generator when asked for something you don't own. But it
> might not be a bug because the code is ok with races.
> 
> It is still fully dodgy IMO, such usage is pretty close to UB.

My 2 cents here:

I could live with Longman's suggestion of an rwsem_assert_is_locked that
only exists if DEBUG_RWSEMS is enabled.  Something like:

#ifdef CONFIG_DEBUG_RWSEMS
static inline bool __rwsem_assert_is_locked(struct rw_semaphore *rwsem,
		const char *file, int line)
{
	bool ret = rwsem_is_locked(rwsem);
	if (!ret)
		WARN(1, "!rwsem_is_locked(rwsem) at %s line %d", file, line);
	return ret;
}
#define rwsem_assert_is_locked(r) \
	__rwsem_assert_is_locked((r), __FILE__, __LINE__)
#endif

and then XFS could do:

	ASSERT(rwsem_assert_is_locked(&VFS_I(ip)->i_rwsem));

Wherein ASSERT is only #defined if CONFIG_XFS_DEBUG, and XFS_DEBUG
selects DEBUG_RWSEMS, per Longman's suggestion.  That's work for what we
want it for (simple cheap lock checking) without becoming a general
lockabuse predicate.

--D



[Index of Archives]     [XFS Filesystem Development (older mail)]     [Linux Filesystem Development]     [Linux Audio Users]     [Yosemite Trails]     [Linux Kernel]     [Linux RAID]     [Linux SCSI]


  Powered by Linux