On Wed, Jul 10, 2024 at 1:19 AM Liao, Chang <liaochang1@xxxxxxxxxx> wrote:
>
>
>
> 在 2024/7/10 7:55, Andrii Nakryiko 写道:
> > On Mon, Jul 8, 2024 at 6:00 PM Liao Chang <liaochang1@xxxxxxxxxx> wrote:
> >>
> >> Reduce the runtime overhead for struct return_instance data managed by
> >> uretprobe. This patch replaces the dynamic allocation with statically
> >> allocated array, leverage two facts that are limited nesting depth of
> >> uretprobe (max 64) and the function call style of return_instance usage
> >> (create at entry, free at exit).
> >>
> >> This patch has been tested on Kunpeng916 (Hi1616), 4 NUMA nodes, 64
> >> cores @ 2.4GHz. Redis benchmarks show a throughput gain by 2% for Redis
> >> GET and SET commands:
> >>
> >> ------------------------------------------------------------------
> >> Test case | No uretprobes | uretprobes | uretprobes
> >> | | (current) | (optimized)
> >> ==================================================================
> >> Redis SET (RPS) | 47025 | 40619 (-13.6%) | 41529 (-11.6%)
> >> ------------------------------------------------------------------
> >> Redis GET (RPS) | 46715 | 41426 (-11.3%) | 42306 (-9.4%)
> >> ------------------------------------------------------------------
> >>
> >> Signed-off-by: Liao Chang <liaochang1@xxxxxxxxxx>
> >> ---
> >> include/linux/uprobes.h | 10 ++-
> >> kernel/events/uprobes.c | 162 ++++++++++++++++++++++++----------------
> >> 2 files changed, 105 insertions(+), 67 deletions(-)
> >>
> >
> > [...]
> >
> >> +static void cleanup_return_instances(struct uprobe_task *utask, bool chained,
> >> + struct pt_regs *regs)
> >> +{
> >> + struct return_frame *frame = &utask->frame;
> >> + struct return_instance *ri = frame->return_instance;
> >> + enum rp_check ctx = chained ? RP_CHECK_CHAIN_CALL : RP_CHECK_CALL;
> >> +
> >> + while (ri && !arch_uretprobe_is_alive(ri, ctx, regs)) {
> >> + ri = next_ret_instance(frame, ri);
> >> + utask->depth--;
> >> + }
> >> + frame->return_instance = ri;
> >> +}
> >> +
> >> +static struct return_instance *alloc_return_instance(struct uprobe_task *task)
> >> +{
> >> + struct return_frame *frame = &task->frame;
> >> +
> >> + if (!frame->vaddr) {
> >> + frame->vaddr = kcalloc(MAX_URETPROBE_DEPTH,
> >> + sizeof(struct return_instance), GFP_KERNEL);
> >
> > Are you just pre-allocating MAX_URETPROBE_DEPTH instances always?
> > I.e., even if we need just one (because there is no recursion), you'd
> > still waste memory for all 64 ones?
>
> This is the truth. On my testing machines, each struct return_instance data
> is 28 bytes, resulting in a total pre-allocated 1792 bytes when the first
> instrumented function is hit.
>
> >
> > That seems rather wasteful.
> >
> > Have you considered using objpool for fast reuse across multiple CPUs?
> > Check lib/objpool.c.
>
> After studying how kretprobe uses objpool, I'm convinced it is a right solution for
> managing return_instance in uretporbe. While I need some time to fully understand
> the objpool code itself and run some benchmark to verify its performance.
>
> Thanks for the suggestion.
Keep in mind that there are two patch sets under development/review,
both of which touch this code. [0] will make return_instance
variable-sized, so think how to accommodate that. And [1] in general
touches a bunch of this code. So I'd let those two settle and land
before optimizing return_instance allocations further.
[0] https://lore.kernel.org/linux-trace-kernel/20240701164115.723677-1-jolsa@xxxxxxxxxx/
[1] https://lore.kernel.org/linux-kernel/20240708091241.544262971@xxxxxxxxxxxxx/
>
> >
> >> + if (!frame->vaddr)
> >> + return NULL;
> >> + }
> >> +
> >> + if (!frame->return_instance) {
> >> + frame->return_instance = frame->vaddr;
> >> + return frame->return_instance;
> >> + }
> >> +
> >> + return ++frame->return_instance;
> >> +}
> >> +
> >> +static inline bool return_frame_empty(struct uprobe_task *task)
> >> +{
> >> + return !task->frame.return_instance;
> >> }
> >>
> >> /*
> >
> > [...]
>
> --
> BR
> Liao, Chang
