Re: git: 02ea6033020e - main - LinuxKPI: Allow spin_lock_irqsave to be called within a critical section

On 19.01.2022 12:50, Hans Petter Selasky wrote:
> On 1/18/22 22:35, Vladimir Kondratyev wrote:
>>
>> Any ideas how to avoid it in a generic way?
> 
> Hi,
> 
> On a non-SMP system this will lead to deadlock.
> 
> Is it possible you can pre-lock this spin lock earlier, so that it is already 
> locked, so instead of
> 
> while(trylock());
> 
> You have:
> 
> assert (trylock() XXX)
> 

Unfortunately, vulnerable functions are called in too many code paths to patch 
them all.

> Or else,
> 
> convert this particular lock to a native FreeBSD spinlock mutex.
> 
It can be done for wake_up() but not for dma_fence_signal() which suffers from 
this problem too. Some code that uses that lock expect it to be spinlock_t

I think we can just drop critical section in seqlock-related part of dma-buf 
code and replace it with rwlock as kib@ and mjg@ suggested. Leave seqlock for 
actual locking to preserve semantics as much as possible and add rwlock to 
implement reader's blocking. Following snippets show code conversion required 
for this change:


Lock seqlock as reader:

retry:
	seq = read_seqcount_begin(&obj->seq);

... reader payload ...
	
	if (read_seqcount_retry(&obj->seq, seq)) {
#ifdef __FreeBSD__
		/* Wait for rwlock to be released by writer */
		rw_rlock(&obj->rwlock);
		rw_runlock(&obj->rwlock);
#endif
		goto retry;
	}


Lock seqlock as writer:

#ifdef __linux__
	preempt_disable();
#elif defined (__FreeBSD__)
	rw_wlock(&obj->rwlock);
#endif
	write_seqcount_begin(&obj->seq);

... writer payload ...

	write_seqcount_end(&obj->seq);
#ifdef __linux__
	preempt_enable();
#elif defined (__FreeBSD__)
	rw_wunlock(&obj->rwlock);
#endif

-- 
WBR
Vladimir Kondratyev