Re: [PATCH] fs/buffer.c: Revoke LRU when trying to drop buffers

[Matthew Wilcox <willy@infradead.org>]

On Mon, Nov 23, 2020 at 10:49:38PM -0800, Chris Goldsworthy wrote:
> +static void __evict_bh_lru(void *arg)
> +{
> +	struct bh_lru *b = &get_cpu_var(bh_lrus);
> +	struct buffer_head *bh = arg;
> +	int i;
> +
> +	for (i = 0; i < BH_LRU_SIZE; i++) {
> +		if (b->bhs[i] == bh) {
> +			brelse(b->bhs[i]);
> +			b->bhs[i] = NULL;
> +			goto out;

That's an odd way to spell 'break' ...

> +		}
> +	}
> +out:
> +	put_cpu_var(bh_lrus);
> +}

...

> @@ -3245,8 +3281,15 @@ drop_buffers(struct page *page, struct buffer_head **buffers_to_free)
>  
>  	bh = head;
>  	do {
> -		if (buffer_busy(bh))
> -			goto failed;
> +		if (buffer_busy(bh)) {
> +			/*
> +			 * Check if the busy failure was due to an
> +			 * outstanding LRU reference
> +			 */
> +			evict_bh_lrus(bh);
> +			if (buffer_busy(bh))
> +				goto failed;

Do you see any performance problems with this?  I'm concerned that we
need to call all CPUs for each buffer on a page, so with a 4kB page
and 512-byte block, we'd call each CPU eight times (with a 64kB page
size and 4kB page, we'd call each CPU 16 times!).  We might be better
off just calling invalidate_bh_lrus() -- we'd flush the entire LRU,
but we'd only need to do it once, not once per buffer head.

We could have a more complex 'evict' that iterates each busy buffer on a
page so transforming:

for_each_buffer
	for_each_cpu
		for_each_lru_entry

to:

for_each_cpu
	for_each_buffer
		for_each_lru_entry

(and i suggest that way because it's more expensive to iterate the buffers
than it is to iterate the lru entries)