Re: [RFC PATCH] sbitmap: fix io hung due to race on sbitmap_word::cleared

[Ming Lei]

On Mon, May 27, 2024 at 12:26:50PM +0800, Yang Yang wrote:
> Configuration for sbq:
>   depth=64, wake_batch=6, shift=6, map_nr=1
> 
> 1. There are 64 requests in progress:
>   map->word = 0xFFFFFFFFFFFFFFFF
> 2. After all the 64 requests complete, and no more requests come:
>   map->word = 0xFFFFFFFFFFFFFFFF, map->cleared = 0xFFFFFFFFFFFFFFFF
> 3. Now two tasks try to allocate requests:
>   T1:                                       T2:
>   __blk_mq_get_tag                          .
>   __sbitmap_queue_get                       .
>   sbitmap_get                               .
>   sbitmap_find_bit                          .
>   sbitmap_find_bit_in_word                  .
>   __sbitmap_get_word  -> nr=-1              __blk_mq_get_tag
>   sbitmap_deferred_clear                    __sbitmap_queue_get
>   /* map->cleared=0xFFFFFFFFFFFFFFFF */     sbitmap_find_bit
>     if (!READ_ONCE(map->cleared))           sbitmap_find_bit_in_word
>       return false;                         __sbitmap_get_word -> nr=-1
>     mask = xchg(&map->cleared, 0)           sbitmap_deferred_clear
>     atomic_long_andnot()                    /* map->cleared=0 */
>                                               if (!(map->cleared))
>                                                 return false;
>                                      /*
>                                       * map->cleared is cleared by T1
>                                       * T2 fail to acquire the tag
>                                       */
> 
> 4. T2 is the sole tag waiter. When T1 puts the tag, T2 cannot be woken
> up due to the wake_batch being set at 6. If no more requests come, T1
> will wait here indefinitely.
> 
> Fix this issue by adding a new flag swap_inprogress to indicate whether
> the swap is ongoing.
> 
> Fixes: 661d4f55a794 ("sbitmap: remove swap_lock")

Indeed, commit 661d4f55a794 ("sbitmap: remove swap_lock") causes this
issue, and ->cleared and ->word should have been checked & updated
atomically.

> Signed-off-by: Yang Yang <yang.yang@xxxxxxxx>
> ---
>  include/linux/sbitmap.h |  5 +++++
>  lib/sbitmap.c           | 22 ++++++++++++++++++++--
>  2 files changed, 25 insertions(+), 2 deletions(-)
> 
> diff --git a/include/linux/sbitmap.h b/include/linux/sbitmap.h
> index d662cf136021..b88a9e4997ab 100644
> --- a/include/linux/sbitmap.h
> +++ b/include/linux/sbitmap.h
> @@ -36,6 +36,11 @@ struct sbitmap_word {
>  	 * @cleared: word holding cleared bits
>  	 */
>  	unsigned long cleared ____cacheline_aligned_in_smp;
> +
> +	/**
> +	 * @swap_inprogress: set to 1 when swapping word <-> cleared
> +	 */
> +	atomic_t swap_inprogress;
>  } ____cacheline_aligned_in_smp;
>  
>  /**
> diff --git a/lib/sbitmap.c b/lib/sbitmap.c
> index 1e453f825c05..d4bb258fe8b0 100644
> --- a/lib/sbitmap.c
> +++ b/lib/sbitmap.c
> @@ -62,10 +62,19 @@ static inline void update_alloc_hint_after_get(struct sbitmap *sb,
>   */
>  static inline bool sbitmap_deferred_clear(struct sbitmap_word *map)
>  {
> -	unsigned long mask;
> +	unsigned long mask, flags;
> +	int zero = 0;
>  
> -	if (!READ_ONCE(map->cleared))
> +	if (!READ_ONCE(map->cleared)) {
> +		if (atomic_read(&map->swap_inprogress))
> +			goto out_wait;
>  		return false;
> +	}
> +
> +	if (!atomic_try_cmpxchg(&map->swap_inprogress, &zero, 1))
> +		goto out_wait;
> +
> +	local_irq_save(flags);
>  
>  	/*
>  	 * First get a stable cleared mask, setting the old mask to 0.
> @@ -77,6 +86,15 @@ static inline bool sbitmap_deferred_clear(struct sbitmap_word *map)
>  	 */
>  	atomic_long_andnot(mask, (atomic_long_t *)&map->word);
>  	BUILD_BUG_ON(sizeof(atomic_long_t) != sizeof(map->word));
> +
> +	atomic_set(&map->swap_inprogress, 0);
> +	smp_mb__after_atomic();
> +	local_irq_restore(flags);
> +	return true;
> +
> +out_wait:
> +	while (atomic_read(&map->swap_inprogress))
> +		;
>  	return true;
>  }

IMO, the above fix looks a bit complicated, and I'd suggest to revert
661d4f55a794 ("sbitmap: remove swap_lock") if no clean & simple solution
can be figured out.

Thanks,
Ming