linux-kernel - [PATCH] seqlock: annotate spinning as unlikely() in __read_seqcount

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>] [day] [month] [year] [list]

Message-Id: <20230727180355.813995-1-mjguzik@gmail.com>
Date:   Thu, 27 Jul 2023 20:03:55 +0200
From:   Mateusz Guzik <mjguzik@...il.com>
To:     peterz@...radead.org
Cc:     linux-kernel@...r.kernel.org, Mateusz Guzik <mjguzik@...il.com>
Subject: [PATCH] seqlock: annotate spinning as unlikely() in __read_seqcount_begin

Annotation already used to be there, but got lost in 52ac39e5db5148f7
("seqlock: seqcount_t: Implement all read APIs as statement expressions").
Does not look like it was intentional.

Without it gcc 12 decides to compile the following in path_init:
        nd->m_seq = __read_seqcount_begin(&mount_lock.seqcount);
        nd->r_seq = __read_seqcount_begin(&rename_lock.seqcount);

into 2 cases of conditional jumps forward if the value is even, aka
branch prediction miss by default in the common case on x86-64.

With the patch jumps are only for odd values.

before:
[snip]
    mov    0x104fe96(%rip),%eax        # 0xffffffff82409680 <mount_lock>
    test   $0x1,%al
    je     0xffffffff813b97fa <path_init+122>
    pause
    mov    0x104fe8a(%rip),%eax        # 0xffffffff82409680 <mount_lock>
    test   $0x1,%al
    jne    0xffffffff813b97ee <path_init+110>
    mov    %eax,0x48(%rbx)
    mov    0x104fdfd(%rip),%eax        # 0xffffffff82409600 <rename_lock>
    test   $0x1,%al
    je     0xffffffff813b9813 <path_init+147>
    pause
    mov    0x104fdf1(%rip),%eax        # 0xffffffff82409600 <rename_lock>
    test   $0x1,%al
    jne    0xffffffff813b9807 <path_init+135>
[/snip]

after:
[snip]
   mov    0x104fec6(%rip),%eax        # 0xffffffff82409680 <mount_lock>
   test   $0x1,%al
   jne    0xffffffff813b99af <path_init+607>
   mov    %eax,0x48(%rbx)
   mov    0x104fe35(%rip),%eax        # 0xffffffff82409600 <rename_lock>
   test   $0x1,%al
   jne    0xffffffff813b999d <path_init+589>
[/snip]

Interestingly .text gets slightly smaller (as reported by size(1)):
before:	20702563
after:	20702429

Signed-off-by: Mateusz Guzik <mjguzik@...il.com>
---
 include/linux/seqlock.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/include/linux/seqlock.h b/include/linux/seqlock.h
index 987a59d977c5..225eb1db35b8 100644
--- a/include/linux/seqlock.h
+++ b/include/linux/seqlock.h
@@ -326,7 +326,7 @@ SEQCOUNT_LOCKNAME(mutex,        struct mutex,    true,     s->lock,        mutex
 ({									\
 	unsigned __seq;							\
 									\
-	while ((__seq = seqprop_sequence(s)) & 1)			\
+	while (unlikely((__seq = seqprop_sequence(s)) & 1))		\
 		cpu_relax();						\
 									\
 	kcsan_atomic_next(KCSAN_SEQLOCK_REGION_MAX);			\
-- 
2.39.2