We have been chasing a memory corruption bug, which turned out to be
caused by very old gcc (4.3.4), which happily turned conditional load
into a non-conditional one, and that broke correctness (the condition
was met only if lock was held) and corrupted memory.
This particular problem with that particular code did not happen when
never gccs were used. I've brought this up with our gcc folks, as I
wanted to make sure that this can't really happen again, and it turns
out it actually can.
Quoting Martin Jambor <mjambor@suse.cz>:
"More current GCCs are more careful when it comes to replacing a
conditional load with a non-conditional one, most notably they check
that a store happens in each iteration of _a_ loop but they assume
loops are executed. They also perform a simple check whether the
store cannot trap which currently passes only for non-const
variables. A simple testcase demonstrating it on an x86_64 is for
example the following:
$ cat cond_store.c
int g_1 = 1;
int g_2[1024] __attribute__((section ("safe_section"), aligned (4096)));
int c = 4;
int __attribute__ ((noinline))
foo (void)
{
int l;
for (l = 0; (l != 4); l++) {
if (g_1)
return l;
for (g_2[0] = 0; (g_2[0] >= 26); ++g_2[0])
;
}
return 2;
}
int main (int argc, char* argv[])
{
if (mprotect (g_2, sizeof(g_2), PROT_READ) == -1)
{
int e = errno;
error (e, e, "mprotect error %i", e);
}
foo ();
__builtin_printf("OK\n");
return 0;
}
/* EOF */
$ ~/gcc/trunk/inst/bin/gcc cond_store.c -O2 --param allow-store-data-races=0
$ ./a.out
OK
$ ~/gcc/trunk/inst/bin/gcc cond_store.c -O2 --param allow-store-data-races=1
$ ./a.out
Segmentation fault
The testcase fails the same at least with 4.9, 4.8 and 4.7. Therefore
I would suggest building kernels with this parameter set to zero. I
also agree with Jikos that the default should be changed for -O2. I
have run most of the SPEC 2k6 CPU benchmarks (gamess and dealII
failed, at -O2, not sure why) compiled with and without this option
and did not see any real difference between respective run-times"
Hopefully the default will be changed in newer gccs, but let's force it
for kernel builds so that we are on a safe side even when older gcc are
used.
The code in question was out-of-tree printk-in-NMI (yeah, surprise
suprise, once again) patch written by Petr Mladek, let me quote his
comment from our internal bugzilla:
"I have spent few days investigating inconsistent state of kernel ring buffer.
It went out that it was caused by speculative store generated by
gcc-4.3.4.
The problem is in assembly generated for make_free_space(). The functions is
called the following way:
+ vprintk_emit();
+ log = MAIN_LOG; // with logbuf_lock
or
log = NMI_LOG; // with nmi_logbuf_lock
cont_add(log, ...);
+ cont_flush(log, ...);
+ log_store(log, ...);
+ log_make_free_space(log, ...);
If called with log = NMI_LOG then only nmi_log_* global variables are safe to
modify but the generated code does store also into (main_)log_* global
variables:
<log_make_free_space>:
55 push %rbp
89 f6 mov %esi,%esi
48 8b 05 03 99 51 01 mov 0x1519903(%rip),%rax # ffffffff82620868 <nmi_log_next_id>
44 8b 1d ec 98 51 01 mov 0x15198ec(%rip),%r11d # ffffffff82620858 <log_next_idx>
8b 35 36 60 14 01 mov 0x1146036(%rip),%esi # ffffffff8224cfa8 <log_buf_len>
44 8b 35 33 60 14 01 mov 0x1146033(%rip),%r14d # ffffffff8224cfac <nmi_log_buf_len>
4c 8b 2d d0 98 51 01 mov 0x15198d0(%rip),%r13 # ffffffff82620850 <log_next_seq>
4c 8b 25 11 61 14 01 mov 0x1146111(%rip),%r12 # ffffffff8224d098 <log_buf>
49 89 c2 mov %rax,%r10
48 21 c2 and %rax,%rdx
48 8b 1d 0c 99 55 01 mov 0x155990c(%rip),%rbx # ffffffff826608a0 <nmi_log_buf>
49 c1 ea 20 shr $0x20,%r10
48 89 55 d0 mov %rdx,-0x30(%rbp)
44 29 de sub %r11d,%esi
45 29 d6 sub %r10d,%r14d
4c 8b 0d 97 98 51 01 mov 0x1519897(%rip),%r9 # ffffffff82620840 <log_first_seq>
eb 7e jmp ffffffff81107029 <log_make_free_space+0xe9>
[...]
85 ff test %edi,%edi # edi = 1 for NMI_LOG
4c 89 e8 mov %r13,%rax
4c 89 ca mov %r9,%rdx
74 0a je ffffffff8110703d <log_make_free_space+0xfd>
8b 15 27 98 51 01 mov 0x1519827(%rip),%edx # ffffffff82620860 <nmi_log_first_id>
48 8b 45 d0 mov -0x30(%rbp),%rax
48 39 c2 cmp %rax,%rdx # end of loop
0f 84 da 00 00 00 je ffffffff81107120 <log_make_free_space+0x1e0>
[...]
85 ff test %edi,%edi # edi = 1 for NMI_LOG
4c 89 0d 17 97 51 01 mov %r9,0x1519717(%rip) # ffffffff82620840 <log_first_seq>
^^^^^^^^^^^^^^^^^^^^^^^^^^
KABOOOM
74 35 je ffffffff81107160 <log_make_free_space+0x220>
It stores log_first_seq when edi == NMI_LOG. This instructions are used also
when edi == MAIN_LOG but the store is done speculatively before the condition
is decided. It is unsafe because we do not have "logbuf_lock" in NMI context
and some other process migh modify "log_first_seq" in parallel"
I believe that the best course of action is both
- building kernel (and anything multi-threaded, I guess) with that
optimization turned off
- persuade gcc folks to change the default for future releases
Signed-off-by: Jiri Kosina <jkosina@suse.cz>
Cc: Martin Jambor <mjambor@suse.cz>
Cc: Petr Mladek <pmladek@suse.cz>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Marek Polacek <polacek@redhat.com>
Cc: Jakub Jelinek <jakub@redhat.com>
Cc: Steven Noonan <steven@uplinklabs.net>
Cc: Richard Biener <richard.guenther@gmail.com>
Cc: Dan Carpenter <dan.carpenter@oracle.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>