Messages in this thread | | | Date | Tue, 7 Mar 2023 12:35:45 +0100 | From | Peter Zijlstra <> | Subject | Re: [PATCH] x86/resctrl: avoid compiler optimization in __resctrl_sched_in |
| |
On Mon, Mar 06, 2023 at 04:16:52PM -0800, Nick Desaulniers wrote: > Start of Lore thread: > https://lore.kernel.org/lkml/20230303231133.1486085-1-eranian@google.com/ > > On Mon, Mar 6, 2023 at 4:01 AM Peter Zijlstra <peterz@infradead.org> wrote: > > > > On Fri, Mar 03, 2023 at 03:11:33PM -0800, Stephane Eranian wrote: > > > > > The problem is located in the __resctrl_sched_in() routine which rewrites > > > the active closid via the PQR_ASSOC register. Because this is an expensive > > > operation, the kernel only does it when the context switch involves tasks > > > with different CLOSID. And to check that, it needs to access the current > > > task's closid field using current->closid. current is actually a macro > > > that reads the per-cpu variable pcpu_hot.current_task. > > > > > > After an investigation by compiler experts, the problem has been tracked down > > > to the usage of the get_current() macro in the __resctrl_sched_in() code and > > > in particular the per-cpu macro: > > > > > > static __always_inline struct task_struct *get_current(void) > > > { > > > return this_cpu_read_stable(pcpu_hot.current_task); > > > } > > > > > > And as per percpu.h: > > > > > > /* > > > * this_cpu_read() makes gcc load the percpu variable every time it is > > > * accessed while this_cpu_read_stable() allows the value to be cached. > > > * this_cpu_read_stable() is more efficient and can be used if its value > > > * is guaranteed to be valid across cpus. The current users include > > > * get_current() and get_thread_info() both of which are actually > > > * per-thread variables implemented as per-cpu variables and thus > > > * stable for the duration of the respective task. > > > */ > > > > > > The _stable version of the macro allows the value to be cached, meaning it > > > does not force a reload. > > > > Right, so afaict the difference between this_cpu_read() and > > this_cpu_read_stable() is the volatile qualifier. > > > > this_cpu_read() is asm volatile(), while this_cpu_read_stable() and > > raw_cpu_read() are both an unqualified asm(). > > > > Now, afaiu we're inlining all of this into __switch_to(), which has > > raw_cpu_write(pcpu_hot.current_task, next_p). > > > > And I suppose what the compiler is doing is lifting the 'current' load > > over that store, but how is it allowed that? I thought C was supposed to > > have PO consistency, That raw_cpu_write() should be seen as a store to > > to pcpu_hot.current_task, why can it lift a load over the store? > > > > Specifically, percpu_to_op() has a "+m" output constaint while > > percpu_stable_op() has a "p" input constraint on the same address. > > I definitely think the issue is specific to "p" constraints. > https://godbolt.org/z/34YeG6WbY is the test case I reduced which I > think demonstrates the issue. > > https://reviews.llvm.org/D145416 > -> click "Show Older Changes" for the ongoing discussion.
So per that summary, I'm going to nit-pick and state we very much want CSE. CSE good. What we don't want it violating store-load ordering.
> I don't have a satisfactory answer yet, but am looking into this.
Oh, geez, what a twisty tale that... So Linus knew back in '09 that "p" was icky, but it sorta was the only thing and it 'worked' -- until now :/
Is there a way to explicitly order these things? barrier() obviously isn't going to help here.
So ideally we'd get something that respects the whole store-load ordering but still allows agressive CSE. And works for both toolchains. Small ask, I know :-)
| |