Re: [PATCH] memcg: calling reclaim_high(GFP_KERNEL) in GFP_NOFS context deadlocks

[Dave Chinner <david-FqsqvQoI3Ljby3iVrkZq2A@public.gmane.org>]

[oops, cc should have been linux-mm-Bw31MaZKKs3YtjvyW6yDsg@public.gmane.org]

On Fri, Sep 30, 2022 at 07:54:40AM +1000, Dave Chinner wrote:
> From: Dave Chinner <dchinner-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>
> 
> This should be more obvious, but gfpflags_allow_blocking() is not
> the same thing as a GFP_KERNEL reclaim contexts. The former checks
> GFP_DIRECT_RECLAIM which tells us if direct reclaim is allowed. The
> latter (GFP_KERNEL) allows blocking on anything, including
> filesystem and IO structures during reclaim.
> 
> However, we do lots of memory allocation from various filesystems we
> are under GFP_NOFS contexts, including page cache folios. Hence if
> direct reclaim in GFP_NOFS context waits on filesystem progress
> (e.g. waits on folio writeback) then memory reclaim can deadlock.
> 
> e.g. page cache allocation (which is GFP_NOFS context) gets stuck
> waiting on page writeback like so:
> 
> [   75.943494] task:test_write      state:D stack:12560 pid: 3728 ppid:  3613 flags:0x00004002
> [   75.944788] Call Trace:
> [   75.945183]  <TASK>
> [   75.945543]  __schedule+0x2f9/0xa30
> [   75.946118]  ? __mod_memcg_lruvec_state+0x41/0x90
> [   75.946895]  schedule+0x5a/0xc0
> [   75.947397]  io_schedule+0x42/0x70
> [   75.947992]  folio_wait_bit_common+0x159/0x3d0
> [   75.948732]  ? dio_warn_stale_pagecache.part.0+0x50/0x50
> [   75.949505]  folio_wait_writeback+0x28/0x80
> [   75.950163]  shrink_page_list+0x96e/0xc30
> [   75.950843]  shrink_lruvec+0x558/0xb80
> [   75.951440]  shrink_node+0x2c6/0x700
> [   75.952059]  do_try_to_free_pages+0xd5/0x570
> [   75.952771]  try_to_free_mem_cgroup_pages+0x105/0x220
> [   75.953548]  reclaim_high.constprop.0+0xa3/0xf0
> [   75.954209]  mem_cgroup_handle_over_high+0x8f/0x280
> [   75.955025]  ? kmem_cache_alloc_lru+0x1c6/0x3f0
> [   75.955781]  try_charge_memcg+0x6c3/0x820
> [   75.956436]  ? __mem_cgroup_threshold+0x16/0x150
> [   75.957204]  charge_memcg+0x76/0xf0
> [   75.957810]  __mem_cgroup_charge+0x29/0x80
> [   75.958464]  __filemap_add_folio+0x225/0x590
> [   75.959112]  ? scan_shadow_nodes+0x30/0x30
> [   75.959794]  filemap_add_folio+0x37/0xa0
> [   75.960432]  __filemap_get_folio+0x1fd/0x340
> [   75.961141]  ? xas_load+0x5/0xa0
> [   75.961712]  iomap_write_begin+0x103/0x6a0
> [   75.962390]  ? filemap_dirty_folio+0x5c/0x80
> [   75.963106]  ? iomap_write_end+0xa2/0x2b0
> [   75.963744]  iomap_file_buffered_write+0x17c/0x380
> [   75.964546]  xfs_file_buffered_write+0xb1/0x2e0
> [   75.965286]  ? xfs_file_buffered_write+0x2b2/0x2e0
> [   75.966097]  vfs_write+0x2ca/0x3d0
> [   75.966702]  __x64_sys_pwrite64+0x8c/0xc0
> [   75.967349]  do_syscall_64+0x35/0x80
> 
> At this point, the system has 58 pending XFS IO completions that are
> stuck waiting for workqueue progress:
> 
> [ 1664.460579] workqueue xfs-conv/dm-0: flags=0x4c
> [ 1664.461332]   pwq 48: cpus=24 node=3 flags=0x0 nice=0 active=58/256 refcnt=59
> [ 1664.461335]     pending: xfs_end_io, xfs_end_io, xfs_end_io, xfs_end_io, xfs_end_io, xfs_end_io, xfs_end_io, xfs_end_io, xfs_end_io, xfs_end_io, xfs_end_io, xfs_end_io, xfs_end_io, xfs_end_io, xfs_end_io, xfs_end_io, xfs_end_io, xfs_end_io, xfs_end_io, xfs_end_io, xfs_end_io, xfs_end_io, xfs_end_io, xfs_end_io, xfs_end_io, xfs_end_io, xfs_end_io, xfs_end_io, xfs_end_io, xfs_end_io, xfs_end_io, xfs_end_io, xfs_end_io, xfs_end_io, xfs_end_io, xfs_end_io, xfs_end_io, xfs_end_io, xfs_end_io, xfs_end_io, xfs_end_io, xfs_end_io, xfs_end_io, xfs_end_io, xfs_end_io, xfs_end_io, xfs_end_io, xfs_end_io, xfs_end_io, xfs_end_io, xfs_end_io, xfs_end_io, xfs_end_io, xfs_end_io, xfs_end_io, xfs_end_io, xfs_end_io, xfs_end_io
> 
> and nothing is making progress. The reason progress is not being
> made is not clear from what I can gather from the steaming corpse,
> but it is clear that the memcg reclaim code should not be blocking
> on filesystem related objects in GFP_NOFS allocation contexts.
> 
> We have the reclaim context parameters right there when we call
> mem_cgroup_handle_over_high(), so pass them down the stack so memcg
> reclaim doesn't cause deadlocks. This makes the reclaim deadlocks in
> the test I've been running go away.
> 
> Signed-off-by: Dave Chinner <dchinner-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>
> ---
>  include/linux/memcontrol.h       | 4 ++--
>  include/linux/resume_user_mode.h | 2 +-
>  mm/memcontrol.c                  | 6 +++---
>  3 files changed, 6 insertions(+), 6 deletions(-)
> 
> diff --git a/include/linux/memcontrol.h b/include/linux/memcontrol.h
> index 6257867fbf95..575bb8cfc810 100644
> --- a/include/linux/memcontrol.h
> +++ b/include/linux/memcontrol.h
> @@ -919,7 +919,7 @@ unsigned long mem_cgroup_get_zone_lru_size(struct lruvec *lruvec,
>  	return READ_ONCE(mz->lru_zone_size[zone_idx][lru]);
>  }
>  
> -void mem_cgroup_handle_over_high(void);
> +void mem_cgroup_handle_over_high(gfp_t gfp_mask);
>  
>  unsigned long mem_cgroup_get_max(struct mem_cgroup *memcg);
>  
> @@ -1433,7 +1433,7 @@ static inline void folio_memcg_unlock(struct folio *folio)
>  {
>  }
>  
> -static inline void mem_cgroup_handle_over_high(void)
> +static inline void mem_cgroup_handle_over_high(gfp_t gfp_mask)
>  {
>  }
>  
> diff --git a/include/linux/resume_user_mode.h b/include/linux/resume_user_mode.h
> index 285189454449..f8f3e958e9cf 100644
> --- a/include/linux/resume_user_mode.h
> +++ b/include/linux/resume_user_mode.h
> @@ -55,7 +55,7 @@ static inline void resume_user_mode_work(struct pt_regs *regs)
>  	}
>  #endif
>  
> -	mem_cgroup_handle_over_high();
> +	mem_cgroup_handle_over_high(GFP_KERNEL);
>  	blkcg_maybe_throttle_current();
>  
>  	rseq_handle_notify_resume(NULL, regs);
> diff --git a/mm/memcontrol.c b/mm/memcontrol.c
> index b69979c9ced5..09fbebff9796 100644
> --- a/mm/memcontrol.c
> +++ b/mm/memcontrol.c
> @@ -2491,7 +2491,7 @@ static unsigned long calculate_high_delay(struct mem_cgroup *memcg,
>   * Scheduled by try_charge() to be executed from the userland return path
>   * and reclaims memory over the high limit.
>   */
> -void mem_cgroup_handle_over_high(void)
> +void mem_cgroup_handle_over_high(gfp_t gfp_mask)
>  {
>  	unsigned long penalty_jiffies;
>  	unsigned long pflags;
> @@ -2519,7 +2519,7 @@ void mem_cgroup_handle_over_high(void)
>  	 */
>  	nr_reclaimed = reclaim_high(memcg,
>  				    in_retry ? SWAP_CLUSTER_MAX : nr_pages,
> -				    GFP_KERNEL);
> +				    gfp_mask);
>  
>  	/*
>  	 * memory.high is breached and reclaim is unable to keep up. Throttle
> @@ -2755,7 +2755,7 @@ static int try_charge_memcg(struct mem_cgroup *memcg, gfp_t gfp_mask,
>  	if (current->memcg_nr_pages_over_high > MEMCG_CHARGE_BATCH &&
>  	    !(current->flags & PF_MEMALLOC) &&
>  	    gfpflags_allow_blocking(gfp_mask)) {
> -		mem_cgroup_handle_over_high();
> +		mem_cgroup_handle_over_high(gfp_mask);
>  	}
>  	return 0;
>  }
> -- 
> 2.37.2
> 
> 

-- 
Dave Chinner
david-FqsqvQoI3Ljby3iVrkZq2A@public.gmane.org