linux-kernel - Re: [PATCH v14 062/138] mm/migrate: Add folio_migrate

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Date:   Thu, 22 Jul 2021 13:29:23 +0100
From:   Matthew Wilcox <willy@...radead.org>
To:     Dmitry Osipenko <digetx@...il.com>
Cc:     linux-kernel@...r.kernel.org, linux-mm@...ck.org,
        linux-fsdevel@...r.kernel.org, "Theodore Y. Ts'o" <tytso@....edu>
Subject: Re: [PATCH v14 062/138] mm/migrate: Add folio_migrate_copy()

On Thu, Jul 22, 2021 at 02:52:28PM +0300, Dmitry Osipenko wrote:
> I'm getting warnings that might be related to this patch.

Thank you!  This is a good report.  I've trimmed away some of the
unnecessary bits from below:

> BUG: sleeping function called from invalid context at mm/util.c:761
> in_atomic(): 1, irqs_disabled(): 0, non_block: 0, pid: 29, name: kcompactd0

This is absolutely a result of this patch:

        for (i = 0; i < nr; i++) {
                cond_resched();
                copy_highpage(folio_page(dst, i), folio_page(src, i));
        }

cond_resched() can sleep, of course.  This is new; previously only
copying huge pages would call cond_resched().  Now every page copy
calls cond_resched().

> (___might_sleep) from (folio_copy+0x3f/0x84)
> (folio_copy) from (folio_migrate_copy+0x11/0x1c)
> (folio_migrate_copy) from (__buffer_migrate_page.part.0+0x215/0x238)
> (__buffer_migrate_page.part.0) from (buffer_migrate_page_norefs+0x19/0x28)

__buffer_migrate_page() is where we become atomic:

        if (check_refs)
                spin_lock(&mapping->private_lock);
...
                migrate_page_copy(newpage, page);
...
        if (check_refs)
                spin_unlock(&mapping->private_lock);

> (buffer_migrate_page_norefs) from (move_to_new_page+0x4d/0x200)
> (move_to_new_page) from (migrate_pages+0x521/0x72c)
> (migrate_pages) from (compact_zone+0x589/0xb60)

The obvious solution is just to change folio_copy():

 {
-       unsigned i, nr = folio_nr_pages(src);
+       unsigned i = 0;
+       unsigned nr = folio_nr_pages(src);

-       for (i = 0; i < nr; i++) {
-               cond_resched();
+       for (;;) {
                copy_highpage(folio_page(dst, i), folio_page(src, i));
+               if (i++ == nr)
+                       break;
+               cond_resched();
        }
 }

now it only calls cond_resched() for multi-page folios.

But that leaves us with a bit of an ... impediment to using multi-page
folios for buffer-head based filesystems (and block devices).  I must
admit to not knowing the buffer_head locking scheme quite as well as
I would like to.  Is it possible to drop this spinlock earlier?