Thread (45 messages) 45 messages, 5 authors, 2024-07-27

Re: [PATCH v2 08/11] perf/uprobe: Convert (some) uprobe->refcount to SRCU

From: Andrii Nakryiko <hidden>
Date: 2024-07-12 21:21:47
Also in: bpf, lkml

+ bpf

On Thu, Jul 11, 2024 at 4:07 AM Peter Zijlstra [off-list ref] wrote:
quoted hunk ↗ jump to hunk
With handle_swbp() hitting concurrently on (all) CPUs, potentially on
the same uprobe, the uprobe->refcount can get *very* hot. Move the
struct uprobe lifetime into uprobes_srcu such that it covers both the
uprobe and the uprobe->consumers list.

With this, handle_swbp() can use a single large SRCU critical section
to avoid taking a refcount on the uprobe for it's duration.

Notably, the single-step and uretprobe paths need a reference that
leaves handle_swbp() and will, for now, still use ->refcount.

Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
---
 kernel/events/uprobes.c |   68 ++++++++++++++++++++++++++++--------------------
 1 file changed, 41 insertions(+), 27 deletions(-)
--- a/kernel/events/uprobes.c
+++ b/kernel/events/uprobes.c
@@ -51,7 +51,7 @@ static struct mutex uprobes_mmap_mutex[U
 DEFINE_STATIC_PERCPU_RWSEM(dup_mmap_sem);
[...]
quoted hunk ↗ jump to hunk
@@ -1982,22 +1990,31 @@ pre_ssout(struct uprobe *uprobe, struct
        if (!utask)
                return -ENOMEM;

+       utask->active_uprobe = try_get_uprobe(uprobe);
+       if (!utask->active_uprobe)
+               return -ESRCH;
+
        xol_vaddr = xol_get_insn_slot(uprobe);
-       if (!xol_vaddr)
-               return -ENOMEM;
+       if (!xol_vaddr) {
+               err = -ENOMEM;
+               goto err_uprobe;
+       }

        utask->xol_vaddr = xol_vaddr;
        utask->vaddr = bp_vaddr;

        err = arch_uprobe_pre_xol(&uprobe->arch, regs);
-       if (unlikely(err)) {
-               xol_free_insn_slot(current);
let's keep this here, because you later remove err_uprobe part and
err_xol is only jumped to from here; it's better to just drop err_xol
and err_uprobe altogether and keep the original xol_free_insn_slot()
here.
-               return err;
-       }
+       if (unlikely(err))
+               goto err_xol;

-       utask->active_uprobe = uprobe;
        utask->state = UTASK_SSTEP;
        return 0;
+
+err_xol:
+       xol_free_insn_slot(current);
+err_uprobe:
+       put_uprobe(utask->active_uprobe);
utask->active_uprobe = NULL;

let's not leave garbage in utask (even if you remove this later anyways)
+       return err;
 }

 /*
[...]
Keyboard shortcuts
hback out one level
jnext message in thread
kprevious message in thread
ldrill in
Escclose help / fold thread tree
?toggle this help