Re: [PATCH AUTOSEL 6.0 17/46] arm64: atomics: remove LL/SC trampolines

From: Catalin Marinas
Date: Wed Oct 12 2022 - 04:56:57 EST


On Tue, Oct 11, 2022 at 10:49:45AM -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>

Please also drop this here. Thanks.

--
Catalin