arm64: cmpxchg_double*: hazard against entire exchange variable
authorMark Rutland <mark.rutland@arm.com>
Wed, 4 Jan 2023 15:16:26 +0000 (15:16 +0000)
committerWill Deacon <will@kernel.org>
Thu, 5 Jan 2023 15:28:15 +0000 (15:28 +0000)
commit031af50045ea97ed4386eb3751ca2c134d0fc911
tree0e215470f74a3910eb70084521c18365f062336e
parentef08c0fadd8a17ebe429b85e23952dac3263ad34
arm64: cmpxchg_double*: hazard against entire exchange variable

The inline assembly for arm64's cmpxchg_double*() implementations use a
+Q constraint to hazard against other accesses to the memory location
being exchanged. However, the pointer passed to the constraint is a
pointer to unsigned long, and thus the hazard only applies to the first
8 bytes of the location.

GCC can take advantage of this, assuming that other portions of the
location are unchanged, leading to a number of potential problems.

This is similar to what we fixed back in commit:

  fee960bed5e857eb ("arm64: xchg: hazard against entire exchange variable")

... but we forgot to adjust cmpxchg_double*() similarly at the same
time.

The same problem applies, as demonstrated with the following test:

| struct big {
|         u64 lo, hi;
| } __aligned(128);
|
| unsigned long foo(struct big *b)
| {
|         u64 hi_old, hi_new;
|
|         hi_old = b->hi;
|         cmpxchg_double_local(&b->lo, &b->hi, 0x12, 0x34, 0x56, 0x78);
|         hi_new = b->hi;
|
|         return hi_old ^ hi_new;
| }

... which GCC 12.1.0 compiles as:

0000000000000000 <foo>:
|    0:   d503233f        paciasp
|    4:   aa0003e4        mov     x4, x0
|    8:   1400000e        b       40 <foo+0x40>
|    c:   d2800240        mov     x0, #0x12                       // #18
|   10:   d2800681        mov     x1, #0x34                       // #52
|   14:   aa0003e5        mov     x5, x0
|   18:   aa0103e6        mov     x6, x1
|   1c:   d2800ac2        mov     x2, #0x56                       // #86
|   20:   d2800f03        mov     x3, #0x78                       // #120
|   24:   48207c82        casp    x0, x1, x2, x3, [x4]
|   28:   ca050000        eor     x0, x0, x5
|   2c:   ca060021        eor     x1, x1, x6
|   30:   aa010000        orr     x0, x0, x1
|   34:   d2800000        mov     x0, #0x0                        // #0    <--- BANG
|   38:   d50323bf        autiasp
|   3c:   d65f03c0        ret
|   40:   d2800240        mov     x0, #0x12                       // #18
|   44:   d2800681        mov     x1, #0x34                       // #52
|   48:   d2800ac2        mov     x2, #0x56                       // #86
|   4c:   d2800f03        mov     x3, #0x78                       // #120
|   50:   f9800091        prfm    pstl1strm, [x4]
|   54:   c87f1885        ldxp    x5, x6, [x4]
|   58:   ca0000a5        eor     x5, x5, x0
|   5c:   ca0100c6        eor     x6, x6, x1
|   60:   aa0600a6        orr     x6, x5, x6
|   64:   b5000066        cbnz    x6, 70 <foo+0x70>
|   68:   c8250c82        stxp    w5, x2, x3, [x4]
|   6c:   35ffff45        cbnz    w5, 54 <foo+0x54>
|   70:   d2800000        mov     x0, #0x0                        // #0     <--- BANG
|   74:   d50323bf        autiasp
|   78:   d65f03c0        ret

Notice that at the lines with "BANG" comments, GCC has assumed that the
higher 8 bytes are unchanged by the cmpxchg_double() call, and that
`hi_old ^ hi_new` can be reduced to a constant zero, for both LSE and
LL/SC versions of cmpxchg_double().

This patch fixes the issue by passing a pointer to __uint128_t into the
+Q constraint, ensuring that the compiler hazards against the entire 16
bytes being modified.

With this change, GCC 12.1.0 compiles the above test as:

0000000000000000 <foo>:
|    0:   f9400407        ldr     x7, [x0, #8]
|    4:   d503233f        paciasp
|    8:   aa0003e4        mov     x4, x0
|    c:   1400000f        b       48 <foo+0x48>
|   10:   d2800240        mov     x0, #0x12                       // #18
|   14:   d2800681        mov     x1, #0x34                       // #52
|   18:   aa0003e5        mov     x5, x0
|   1c:   aa0103e6        mov     x6, x1
|   20:   d2800ac2        mov     x2, #0x56                       // #86
|   24:   d2800f03        mov     x3, #0x78                       // #120
|   28:   48207c82        casp    x0, x1, x2, x3, [x4]
|   2c:   ca050000        eor     x0, x0, x5
|   30:   ca060021        eor     x1, x1, x6
|   34:   aa010000        orr     x0, x0, x1
|   38:   f9400480        ldr     x0, [x4, #8]
|   3c:   d50323bf        autiasp
|   40:   ca0000e0        eor     x0, x7, x0
|   44:   d65f03c0        ret
|   48:   d2800240        mov     x0, #0x12                       // #18
|   4c:   d2800681        mov     x1, #0x34                       // #52
|   50:   d2800ac2        mov     x2, #0x56                       // #86
|   54:   d2800f03        mov     x3, #0x78                       // #120
|   58:   f9800091        prfm    pstl1strm, [x4]
|   5c:   c87f1885        ldxp    x5, x6, [x4]
|   60:   ca0000a5        eor     x5, x5, x0
|   64:   ca0100c6        eor     x6, x6, x1
|   68:   aa0600a6        orr     x6, x5, x6
|   6c:   b5000066        cbnz    x6, 78 <foo+0x78>
|   70:   c8250c82        stxp    w5, x2, x3, [x4]
|   74:   35ffff45        cbnz    w5, 5c <foo+0x5c>
|   78:   f9400480        ldr     x0, [x4, #8]
|   7c:   d50323bf        autiasp
|   80:   ca0000e0        eor     x0, x7, x0
|   84:   d65f03c0        ret

... sampling the high 8 bytes before and after the cmpxchg, and
performing an EOR, as we'd expect.

For backporting, I've tested this atop linux-4.9.y with GCC 5.5.0. Note
that linux-4.9.y is oldest currently supported stable release, and
mandates GCC 5.1+. Unfortunately I couldn't get a GCC 5.1 binary to run
on my machines due to library incompatibilities.

I've also used a standalone test to check that we can use a __uint128_t
pointer in a +Q constraint at least as far back as GCC 4.8.5 and LLVM
3.9.1.

Fixes: 5284e1b4bc8a ("arm64: xchg: Implement cmpxchg_double")
Fixes: e9a4b795652f ("arm64: cmpxchg_dbl: patch in lse instructions when supported by the CPU")
Reported-by: Boqun Feng <boqun.feng@gmail.com>
Link: https://lore.kernel.org/lkml/Y6DEfQXymYVgL3oJ@boqun-archlinux/
Reported-by: Peter Zijlstra <peterz@infradead.org>
Link: https://lore.kernel.org/lkml/Y6GXoO4qmH9OIZ5Q@hirez.programming.kicks-ass.net/
Signed-off-by: Mark Rutland <mark.rutland@arm.com>
Cc: stable@vger.kernel.org
Cc: Arnd Bergmann <arnd@arndb.de>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Steve Capper <steve.capper@arm.com>
Cc: Will Deacon <will@kernel.org>
Link: https://lore.kernel.org/r/20230104151626.3262137-1-mark.rutland@arm.com
Signed-off-by: Will Deacon <will@kernel.org>
arch/arm64/include/asm/atomic_ll_sc.h
arch/arm64/include/asm/atomic_lse.h