Re: [PATCH AUTOSEL 5.19 17/40] arm64: atomics: remove LL/SC trampolines

From: Sasha Levin
Date: Thu Oct 13 2022 - 14:07:07 EST


On Wed, Oct 12, 2022 at 09:53:53AM +0100, Catalin Marinas wrote:
Hi Sasha,

On Tue, Oct 11, 2022 at 10:51:06AM -0400, Sasha Levin wrote:
From: Mark Rutland <mark.rutland@xxxxxxx>

[ Upstream commit b2c3ccbd0011bb3b51d0fec24cb3a5812b1ec8ea ]

When CONFIG_ARM64_LSE_ATOMICS=y, each use of an LL/SC atomic results in
a fragment of code being generated in a subsection without a clear
association with its caller. A trampoline in the caller branches to the
LL/SC atomic with with a direct branch, and the atomic directly branches
back into its trampoline.

This breaks backtracing, as any PC within the out-of-line fragment will
be symbolized as an offset from the nearest prior symbol (which may not
be the function using the atomic), and since the atomic returns with a
direct branch, the caller's PC may be missing from the backtrace.

For example, with secondary_start_kernel() hacked to contain
atomic_inc(NULL), the resulting exception can be reported as being taken
from cpus_are_stuck_in_kernel():

| Unable to handle kernel NULL pointer dereference at virtual address 0000000000000000
| Mem abort info:
| ESR = 0x0000000096000004
| EC = 0x25: DABT (current EL), IL = 32 bits
| SET = 0, FnV = 0
| EA = 0, S1PTW = 0
| FSC = 0x04: level 0 translation fault
| Data abort info:
| ISV = 0, ISS = 0x00000004
| CM = 0, WnR = 0
| [0000000000000000] user address but active_mm is swapper
| Internal error: Oops: 96000004 [#1] PREEMPT SMP
| Modules linked in:
| CPU: 1 PID: 0 Comm: swapper/1 Not tainted 5.19.0-11219-geb555cb5b794-dirty #3
| Hardware name: linux,dummy-virt (DT)
| pstate: 60000005 (nZCv daif -PAN -UAO -TCO -DIT -SSBS BTYPE=--)
| pc : cpus_are_stuck_in_kernel+0xa4/0x120
| lr : secondary_start_kernel+0x164/0x170
| sp : ffff80000a4cbe90
| x29: ffff80000a4cbe90 x28: 0000000000000000 x27: 0000000000000000
| x26: 0000000000000000 x25: 0000000000000000 x24: 0000000000000000
| x23: 0000000000000000 x22: 0000000000000000 x21: 0000000000000000
| x20: 0000000000000001 x19: 0000000000000001 x18: 0000000000000008
| x17: 3030383832343030 x16: 3030303030307830 x15: ffff80000a4cbab0
| x14: 0000000000000001 x13: 5d31666130663133 x12: 3478305b20313030
| x11: 3030303030303078 x10: 3020726f73736563 x9 : 726f737365636f72
| x8 : ffff800009ff2ef0 x7 : 0000000000000003 x6 : 0000000000000000
| x5 : 0000000000000000 x4 : 0000000000000000 x3 : 0000000000000100
| x2 : 0000000000000000 x1 : ffff0000029bd880 x0 : 0000000000000000
| Call trace:
| cpus_are_stuck_in_kernel+0xa4/0x120
| __secondary_switched+0xb0/0xb4
| Code: 35ffffa3 17fffc6c d53cd040 f9800011 (885f7c01)
| ---[ end trace 0000000000000000 ]---

This is confusing and hinders debugging, and will be problematic for
CONFIG_LIVEPATCH as these cases cannot be unwound reliably.

This is very similar to recent issues with out-of-line exception fixups,
which were removed in commits:

35d67794b8828333 ("arm64: lib: __arch_clear_user(): fold fixups into body")
4012e0e22739eef9 ("arm64: lib: __arch_copy_from_user(): fold fixups into body")
139f9ab73d60cf76 ("arm64: lib: __arch_copy_to_user(): fold fixups into body")

When the trampolines were introduced in commit:

addfc38672c73efd ("arm64: atomics: avoid out-of-line ll/sc atomics")

The rationale was to improve icache performance by grouping the LL/SC
atomics together. This has never been measured, and this theoretical
benefit is outweighed by other factors:

* As the subsections are collapsed into sections at object file
granularity, these are spread out throughout the kernel and can share
cachelines with unrelated code regardless.

* GCC 12.1.0 has been observed to place the trampoline out-of-line in
specialised __ll_sc_*() functions, introducing more branching than was
intended.

* Removing the trampolines has been observed to shrink a defconfig
kernel Image by 64KiB when building with GCC 12.1.0.

This patch removes the LL/SC trampolines, meaning that the LL/SC atomics
will be inlined into their callers (or placed in out-of line functions
using regular BL/RET pairs). When CONFIG_ARM64_LSE_ATOMICS=y, the LL/SC
atomics are always called in an unlikely branch, and will be placed in a
cold portion of the function, so this should have minimal impact to the
hot paths.

Other than the improved backtracing, there should be no functional
change as a result of this patch.

Signed-off-by: Mark Rutland <mark.rutland@xxxxxxx>
Cc: Will Deacon <will@xxxxxxxxxx>
Link: https://lore.kernel.org/r/20220817155914.3975112-2-mark.rutland@xxxxxxx
Signed-off-by: Catalin Marinas <catalin.marinas@xxxxxxx>
Signed-off-by: Sasha Levin <sashal@xxxxxxxxxx>

I don't think we should back-port this. There is no functional change,
more of a clean-up in preparation for RELIABLE_STACKTRACE. The oops
message in the log is to show how reporting works rather than a real
bug.

I went by the "This breaks backtracing" line when backporting :)

I'll drop it, thanks!

--
Thanks,
Sasha