Date:   Sun, 23 Apr 2023 15:59:14 +0800
From:   "Huang, Ying" <ying.huang@...el.com>
To:     Douglas Anderson <dianders@...omium.org>
Cc:     Andrew Morton <akpm@...ux-foundation.org>,
        Mel Gorman <mgorman@...hsingularity.net>,
        Vlastimil Babka <vbabka@...e.cz>,
        Alexander Viro <viro@...iv.linux.org.uk>,
        Christian Brauner <brauner@...nel.org>,
        linux-kernel@...r.kernel.org, linux-mm@...ck.org,
        Yu Zhao <yuzhao@...gle.com>, linux-fsdevel@...r.kernel.org,
        Matthew Wilcox <willy@...radead.org>
Subject: Re: [PATCH v2 3/4] migrate_pages: Don't wait forever locking pages
 in MIGRATE_SYNC_LIGHT

Douglas Anderson <dianders@...omium.org> writes:

> The MIGRATE_SYNC_LIGHT mode is intended to block for things that will
> finish quickly but not for things that will take a long time. Exactly
> how long is too long is not well defined, but waits of tens of
> milliseconds is likely non-ideal.
>
> Waiting on the folio lock in isolate_movable_page() is something that
> usually is pretty quick, but is not officially bounded. Nothing stops
> another process from holding a folio lock while doing an expensive
> operation. Having an unbounded wait like this is not within the design
> goals of MIGRATE_SYNC_LIGHT.
>
> When putting a Chromebook under memory pressure (opening over 90 tabs
> on a 4GB machine) it was fairly easy to see delays waiting for the
> lock of > 100 ms. While the laptop wasn't amazingly usable in this
> state, it was still limping along and this state isn't something
> artificial. Sometimes we simply end up with a lot of memory pressure.
>
> Putting the same Chromebook under memory pressure while it was running
> Android apps (though not stressing them) showed a much worse result
> (NOTE: this was on a older kernel but the codepaths here are
> similar). Android apps on ChromeOS currently run from a 128K-block,
> zlib-compressed, loopback-mounted squashfs disk. If we get a page
> fault from something backed by the squashfs filesystem we could end up
> holding a folio lock while reading enough from disk to decompress 128K
> (and then decompressing it using the somewhat slow zlib algorithms).
> That reading goes through the ext4 subsystem (because it's a loopback
> mount) before eventually ending up in the block subsystem. This extra
> jaunt adds extra overhead. Without much work I could see cases where
> we ended up blocked on a folio lock for over a second. With more
> more extreme memory pressure I could see up to 25 seconds.
>
> Let's bound the amount of time we can wait for the folio lock. The
> SYNC_LIGHT migration mode can already handle failure for things that
> are slow, so adding this timeout in is fairly straightforward.
>
> With this timeout, it can be seen that kcompactd can move on to more
> productive tasks if it's taking a long time to acquire a lock.

How long is the max wait time of folio_lock_timeout()?

> NOTE: The reason I stated digging into this isn't because some
> benchmark had gone awry, but because we've received in-the-field crash
> reports where we have a hung task waiting on the page lock (which is
> the equivalent code path on old kernels). While the root cause of
> those crashes is likely unrelated and won't be fixed by this patch,
> analyzing those crash reports did point out this unbounded wait and it
> seemed like something good to fix.
>
> ALSO NOTE: the timeout mechanism used here uses "jiffies" and we also
> will retry up to 7 times. That doesn't give us much accuracy in
> specifying the timeout. On 1000 Hz machines we'll end up timing out in
> 7-14 ms. On 100 Hz machines we'll end up in 70-140 ms. Given that we
> don't have a strong definition of how long "too long" is, this is
> probably OK.

You can use HZ to work with different configuration.  It doesn't help
much if your target is 1ms.  But I think that it's possible to set it to
longer than that in the future.  So, some general definition looks
better.

Best Regards,
Huang, Ying

> Suggested-by: Mel Gorman <mgorman@...hsingularity.net>
> Signed-off-by: Douglas Anderson <dianders@...omium.org>
> ---
>
> Changes in v2:
> - Keep unbounded delay in "SYNC", delay with a timeout in "SYNC_LIGHT"
>
>  mm/migrate.c | 20 +++++++++++++++++++-
>  1 file changed, 19 insertions(+), 1 deletion(-)
>
> diff --git a/mm/migrate.c b/mm/migrate.c
> index db3f154446af..60982df71a93 100644
> --- a/mm/migrate.c
> +++ b/mm/migrate.c
> @@ -58,6 +58,23 @@
>  
>  #include "internal.h"
>  
> +/* Returns the schedule timeout for a non-async mode */
> +static long timeout_for_mode(enum migrate_mode mode)
> +{
> +	/*
> +	 * We'll always return 1 jiffy as the timeout. Since all places using
> +	 * this timeout are in a retry loop this means that the maximum time
> +	 * we might block is actually NR_MAX_MIGRATE_SYNC_RETRY jiffies.
> +	 * If a jiffy is 1 ms that's 7 ms, though with the accuracy of the
> +	 * timeouts it often ends up more like 14 ms; if a jiffy is 10 ms
> +	 * that's 70-140 ms.
> +	 */
> +	if (mode == MIGRATE_SYNC_LIGHT)
> +		return 1;
> +
> +	return MAX_SCHEDULE_TIMEOUT;
> +}
> +
>  bool isolate_movable_page(struct page *page, isolate_mode_t mode)
>  {
>  	struct folio *folio = folio_get_nontail_page(page);
> @@ -1162,7 +1179,8 @@ static int migrate_folio_unmap(new_page_t get_new_page, free_page_t put_new_page
>  		if (current->flags & PF_MEMALLOC)
>  			goto out;
>  
> -		folio_lock(src);
> +		if (folio_lock_timeout(src, timeout_for_mode(mode)))
> +			goto out;
>  	}
>  	locked = true;

Powered by blists - more mailing lists