Re: [PATCH] ext4_sb_breadahead_unmovable: Change to be no-blocking

[Matthew Wilcox]

On Fri, Aug 11, 2023 at 07:15:19AM +0000, Hui Zhu wrote:
> From: Hui Zhu <teawater@xxxxxxxxxxxx>
> 
> This version fix the gfp flags in the callers instead of working this
> new "bool" flag through the buffer head layers according to the comments
> from Matthew Wilcox.

FYI, this paragraph should have been below the --- so it gets excluded
from the commit log.

> Meanwhile, it was observed that the task holding the ext4 journal lock
> was blocked for an extended period of time on "shrink_page_list" due to
> "ext4_sb_breadahead_unmovable".
> 0 [] __schedule at xxxxxxxxxxxxxxx
> 1 [] _cond_resched at xxxxxxxxxxxxxxx
> 2 [] shrink_page_list at xxxxxxxxxxxxxxx
> 3 [] shrink_inactive_list at xxxxxxxxxxxxxxx
> 4 [] shrink_lruvec at xxxxxxxxxxxxxxx
> 5 [] shrink_node_memcgs at xxxxxxxxxxxxxxx
> 6 [] shrink_node at xxxxxxxxxxxxxxx
> 7 [] shrink_zones at xxxxxxxxxxxxxxx
> 8 [] do_try_to_free_pages at xxxxxxxxxxxxxxx
> 9 [] try_to_free_mem_cgroup_pages at xxxxxxxxxxxxxxx
> 10 [] try_charge at xxxxxxxxxxxxxxx
> 11 [] mem_cgroup_charge at xxxxxxxxxxxxxxx
> 12 [] __add_to_page_cache_locked at xxxxxxxxxxxxxxx
> 13 [] add_to_page_cache_lru at xxxxxxxxxxxxxxx
> 14 [] pagecache_get_page at xxxxxxxxxxxxxxx
> 15 [] grow_dev_page at xxxxxxxxxxxxxxx

After applying your patch, we'd still get into trouble with
folio_alloc_buffers() also specifying __GFP_NOWAIT.  So I decided
to pass the GFP flags into folio_alloc_buffers() -- see the patch
series I just sent out.

> @@ -1050,18 +1051,27 @@ grow_dev_page(struct block_device *bdev, sector_t block,
>  	int ret = 0;
>  	gfp_t gfp_mask;
>  
> -	gfp_mask = mapping_gfp_constraint(inode->i_mapping, ~__GFP_FS) | gfp;
> +	gfp_mask = mapping_gfp_constraint(inode->i_mapping, ~__GFP_FS);
> +	if (gfp == ~__GFP_DIRECT_RECLAIM)
> +		gfp_mask &= ~__GFP_DIRECT_RECLAIM;

This isn't how we normally use gfp_mask.  OTOH, how buffer.c uses GFP
masks is also a bit weird.  The bdev_getblk() I just added is more
normal.

Please try the patchset I cc'd you on (with the __GFP_ACCOUNT added);
I'm currently running it through xfstests and it's holding up fine.
I suppose I should play around with memcgs to try to make it happen a
bit more often.