Messages in this thread Patch in this message | | | From | Nadav Amit <> | Subject | Re: [PATCH v2 -tip] x86/percpu: Use C for arch_raw_cpu_ptr() | Date | Thu, 12 Oct 2023 15:19:03 +0000 |
| |
> On Oct 12, 2023, at 12:54 AM, Linus Torvalds <torvalds@linux-foundation.org> wrote: > > !! External Email > > On Wed, 11 Oct 2023 at 14:33, Uros Bizjak <ubizjak@gmail.com> wrote: >> >> Reading the above, it looks to me that we don't want to play games >> with "const aliased" versions of current_task [1], as proposed by >> Nadav in his patch series. > > Well, maybe I'd like it if I saw what the effect of it was, but that > patch mentions "sync_mm_rss()" which doesn't actually exist > (SPLIT_RSS_COUNTING is never defined, the split version is gone and > hasn't existed since commit f1a7941243c1 "mm: convert mm's rss stats > into percpu_counter")
So I added a new version of the current aliasing (well, actually pcpu_hot in the new version) on top of Uros’s patches, and the effect can be seen in many functions. I don’t want to bother with many examples so here is a common and simple one:
Currently syscall_exit_work() that starts with:
0xffffffff8111e120 <+0>: push %rbp 0xffffffff8111e121 <+1>: mov %rdi,%rbp 0xffffffff8111e124 <+4>: push %rbx 0xffffffff8111e125 <+5>: mov %rsi,%rbx 0xffffffff8111e128 <+8>: and $0x20,%esi 0xffffffff8111e12b <+11>: je 0xffffffff8111e143 <syscall_exit_work+35> 0xffffffff8111e12d <+13>: mov %gs:0x2ac80,%rax 0xffffffff8111e136 <+22>: cmpb $0x0,0x800(%rax) 0xffffffff8111e13d <+29>: jne 0xffffffff8111e22a <syscall_exit_work+266> 0xffffffff8111e143 <+35>: mov %gs:0x2ac80,%rax 0xffffffff8111e14c <+44>: cmpq $0x0,0x7c8(%rax)
Using the const-alias changes the beginning of syscall_exit_work to:
0xffffffff8111cb80 <+0>: push %r12 0xffffffff8111cb82 <+2>: mov %gs:0x7ef0e0f6(%rip),%r12 # 0x2ac80 <pcpu_hot> 0xffffffff8111cb8a <+10>: push %rbp 0xffffffff8111cb8b <+11>: mov %rdi,%rbp 0xffffffff8111cb8e <+14>: push %rbx 0xffffffff8111cb8f <+15>: mov %rsi,%rbx 0xffffffff8111cb92 <+18>: and $0x20,%esi 0xffffffff8111cb95 <+21>: je 0xffffffff8111cba6 <syscall_exit_work+38> 0xffffffff8111cb97 <+23>: cmpb $0x0,0x800(%r12) 0xffffffff8111cba0 <+32>: jne 0xffffffff8111cc7a <syscall_exit_work+250> 0xffffffff8111cba6 <+38>: cmpq $0x0,0x7c8(%r12)
So we both see RIP-relative addressing is being used (hence the instruction is one byte shorter) and the reload going away.
Now, I am not a compiler expert as for the rationale, but it googling around I can see Nick explaining the rationale [1] - if you use “p” your read memory. BTW: It is related to discussion you had [2], in which you encountered an issue I also encountered before [3]. My bad for pushing it in.
Anyhow, I created a similar code on godbolt ( https://godbolt.org/z/dPqKKzPs4 ) to show this behavior - how compiler barriers cause reload. It seems that this behavior happens on GCC and CLANG on various versions.
The idea behind the patch is that it communicates - in the compilation unit granularity - that current is fixed. There is an issue of whether it works with LTO, which I have never checked.
[1] https://reviews.llvm.org/D145416 [2] https://lore.kernel.org/lkml/20230306120106.GE1267364@hirez.programming.kicks-ass.net/ [3] https://lore.kernel.org/all/20190823224424.15296-5-namit@vmware.com/
--
Here’s the updated patch - but I didn’t really boot a machine with it so new issues might have come since my last patch-set:
-- >8 --
Date: Thu, 12 Oct 2023 06:02:03 -0700 Subject: [PATCH] Const current
--- arch/x86/include/asm/current.h | 17 ++++++++++++++++- arch/x86/kernel/cpu/common.c | 4 ++++ include/linux/compiler.h | 2 +- 3 files changed, 21 insertions(+), 2 deletions(-)
diff --git a/arch/x86/include/asm/current.h b/arch/x86/include/asm/current.h index a1168e7b69e5..d05fbb6a8bd7 100644 --- a/arch/x86/include/asm/current.h +++ b/arch/x86/include/asm/current.h @@ -36,9 +36,24 @@ static_assert(sizeof(struct pcpu_hot) == 64); DECLARE_PER_CPU_ALIGNED(struct pcpu_hot, pcpu_hot); +/* + * Hold a constant alias for current_task, which would allow to avoid caching of + * current task. + * + * We must mark const_current_task with the segment qualifiers, as otherwise gcc + * would do redundant reads of const_current_task. + */ +DECLARE_PER_CPU(struct pcpu_hot const __percpu_seg_override, const_pcpu_hot); + static __always_inline struct task_struct *get_current(void) { - return this_cpu_read_stable(pcpu_hot.current_task); + + /* + * GCC is missing functionality of removing segment qualifiers, which + * messes with per-cpu infrastructure that holds local copies. Use + * __raw_cpu_read to avoid holding any copy. + */ + return __raw_cpu_read(, const_pcpu_hot.current_task); } #define current get_current() diff --git a/arch/x86/kernel/cpu/common.c b/arch/x86/kernel/cpu/common.c index 382d4e6b848d..94590af11388 100644 --- a/arch/x86/kernel/cpu/common.c +++ b/arch/x86/kernel/cpu/common.c @@ -2052,6 +2052,10 @@ DEFINE_PER_CPU_ALIGNED(struct pcpu_hot, pcpu_hot) = { }; EXPORT_PER_CPU_SYMBOL(pcpu_hot); +DECLARE_PER_CPU_ALIGNED(struct pcpu_hot const __percpu_seg_override, const_pcpu_hot) + __attribute__((alias("pcpu_hot"))); +EXPORT_PER_CPU_SYMBOL(const_pcpu_hot); + #ifdef CONFIG_X86_64 DEFINE_PER_CPU_FIRST(struct fixed_percpu_data, fixed_percpu_data) __aligned(PAGE_SIZE) __visible; diff --git a/include/linux/compiler.h b/include/linux/compiler.h index d7779a18b24f..e7059292085e 100644 --- a/include/linux/compiler.h +++ b/include/linux/compiler.h @@ -212,7 +212,7 @@ void ftrace_likely_update(struct ftrace_likely_data *f, int val, */ #define ___ADDRESSABLE(sym, __attrs) \ static void * __used __attrs \ - __UNIQUE_ID(__PASTE(__addressable_,sym)) = (void *)&sym; + __UNIQUE_ID(__PASTE(__addressable_,sym)) = (void *)(uintptr_t)&sym; #define __ADDRESSABLE(sym) \ ___ADDRESSABLE(sym, __section(".discard.addressable")) -- 2.25.1
| |