[PATCH v3] tracing: fgraph: fix the quadratic shadow stack walk

[PATCH v3] tracing: fgraph: fix the quadratic shadow stack walk

[Vineet Gupta]

alloc_retstack_tasklist() assigns a shadow stack to every thread when
fgraph is enabled. It did so FTRACE_RETSTACK_ALLOC_SIZE (32) tasks at a
time, and since for_each_process_thread() has no cursor, every sweep
restarted from init_task and re-walked the tasks already served, making
the ftrace_graph_active 0 -> 1 transition quadratic in the thread count.
On a 60-core Sapphire Rapids machine with 400000 idle threads it took
227 s, inside a single bpf() syscall for the kprobe_multi case.

Allocating inline with GFP_NOWAIT removes the batching: a single sweep
serves every task, 227 s -> 0.092 s.

v1: https://lore.kernel.org/all/20260922225526.1554758-1-vineet.gupta@linux.dev/
v2: https://lore.kernel.org/all/20260929005411.4105448-1-vineet.gupta@linux.dev/

Changes since v2:
 - update the stale alloc_retstack_tasklist() comment, which still
   described the old batch-of-32 behaviour (bot+bpf-ci)

Sashiko flagged the mix of goto-based cleanup and scoped_guard() in
this function. Peter and Andrii both considered it fine as is, so it
is unchanged:
https://lore.kernel.org/all/20260929080346.GR4120091@noisy.programming.kicks-ass.net/

Changes since v1:
 - replace both v1 patches with Peter's inline GFP_NOWAIT approach,
   which removes the quadratic behaviour rather than dividing it by a
   constant. With a single sweep there is no retry loop left, so the
   v1 cond_resched() patch has nothing to attach to and is dropped.
 - comment why returning from inside scoped_guard() skips the free
   loop safely

Vineet Gupta (1):
  tracing: fgraph: allocate shadow stacks inline with GFP_NOWAIT

 kernel/trace/fgraph.c | 44 ++++++++++++++++++++++++++++---------------
 1 file changed, 29 insertions(+), 15 deletions(-)

-- 
2.53.0-Meta

[PATCH v3 1/1] tracing: fgraph: allocate shadow stacks inline with GFP_NOWAIT

[Vineet Gupta]

When ftrace graphing is turned on, all tasks in the system missing
return stack page are assigned one. This is done in a simplistic
multi-sweep loop of FTRACE_RETSTACK_ALLOC_SIZE (currently 32) tasks
at a time as follows:

   start_graph_tracing()
	do {
		alloc_retstack_tasklist
	} while (-EAGAIN);

   alloc_retstack_tasklist()
	alloc x32           # GFP_KERNEL, may sleep
	rcu_read_lock()     # preempt off
		for_each_process_thread    walk N_total, no cond_resched
                     t->ret_stack = new_page
	rcu_read_unlock()   # preempt enable but no explicit yield

Each successive iteration of loop invokes for_each_process_thread()
which doesn't support cursor based resume and always restarts from the
init_task. Thus each successive loop needs to skip the tasks assigned
ret_stack in prior sweeps and thus take longer and longer to find the
candidate 32 tasks. And while the loop end calls rcu unlock,
and re-enables preemption briefly, there is no explicit yield.

The total number of iterations turn out to be

	N_total * N_null / (2 * FTRACE_RETSTACK_ALLOC_SIZE)

where N_total is the number of threads on the system and N_null the number
of those whose ret_stack is still NULL. After boot with no fgraph user that
is every thread.

This is not a tracing-only path. Since commit 4346ba160409 ("fprobe:
Rewrite fprobe on function-graph tracer") fprobe is built on fgraph,
so an ordinary kprobe_multi BPF attach that flips ftrace_graph_active
from 0 to 1 pays the whole cost inside one bpf() syscall. Commit
2c67dc457bc6 ("tracing: fprobe: optimization for entry only case")
narrows that to return and session probes but does not remove it.

A host running hundreds of thousands of threads, with the current batch
size of 32, turns the ftrace_graph_active 0 -> 1 transition into a
multi-second, and eventually multi-minute, operation which showed up as
RCU stall and softlockup_panic on heavily loaded Meta Fleet machines
with preemption disabled.

    rcu: INFO: rcu_sched self-detected stall on CPU
    rcu: 0-....: (20718 ticks this GP) (t=21000 jiffies g=913 q=8 ncpus=8)
    CPU: 0 UID: 0 PID: 160141 Comm: bpftrace
     ___slab_alloc+0x549/0xa20
     kmem_cache_alloc_noprof+0x16c/0x340
     register_ftrace_graph+0x3d7/0x6c0
     register_fprobe_ips+0x2d5/0x310
     bpf_kprobe_multi_link_attach+0x218/0x7e0
     __sys_bpf+0x267a/0x27a0

Fix this by allocating inline with GFP_NOWAIT instead of pre-allocating
a fixed batch. GFP_NOWAIT does not sleep, so it is legal inside the RCU
read-side critical section.

A single sweep now serves every task, so the walk becomes O(N) rather
than O(N_total * N_null), and the retry loop is no longer the normal
path.

scoped_guard(rcu) replaces the open-coded rcu_read_lock() and the
goto unlock, which is what makes returning directly from inside the
critical section legal.

The pre-allocated array remains as a fallback, although the testing
below never needed it.

smp_store_release() is a drop-in replacement for the smp_wmb() and
plain store it replaces.

Measured on a 60-core Sapphire Rapids machine running a PREEMPT_LAZY
kernel (CONFIG_PREEMPT_DYNAMIC=y, mode lazy), timing the write that
drives the ftrace_graph_active 0 -> 1 transition. Threads are spawned
while fgraph is inactive so every one of them has ret_stack == NULL,
and the loop was instrumented to count sweeps:

	threads   before              after
	          sweeps   time       sweeps   time
	100000     3126     15.3 s      1       0.019 s
	200000     6253     59.1 s      1       0.047 s
	400000    12503    227.8 s      1       0.092 s

The sweep count is 1 throughout. Time now scales linearly with the
thread count at roughly 230 ns per task, which is the allocation and
initialisation of one shadow stack.

Fixes: f201ae2356c7 ("tracing/function-return-tracer: store return stack into task_struct and allocate it dynamically")
Suggested-by: Peter Zijlstra <peterz@infradead.org>
Signed-off-by: Vineet Gupta <vineet.gupta@linux.dev>
Tested-by: Vineet Gupta <vineet.gupta@linux.dev>
Cc: stable@vger.kernel.org
---
 kernel/trace/fgraph.c | 44 ++++++++++++++++++++++++++++---------------
 1 file changed, 29 insertions(+), 15 deletions(-)

diff --git a/kernel/trace/fgraph.c b/kernel/trace/fgraph.c
index ed455b53513b..cdc5ac834d06 100644
--- a/kernel/trace/fgraph.c
+++ b/kernel/trace/fgraph.c
@@ -1033,13 +1033,16 @@ void ftrace_stub_graph(struct ftrace_graph_ret *trace, struct fgraph_ops *gops,
 trace_func_graph_ret_t ftrace_graph_return = ftrace_stub_graph;
 trace_func_graph_ent_t ftrace_graph_entry = ftrace_graph_entry_stub;
 
-/* Try to assign a return stack array on FTRACE_RETSTACK_ALLOC_SIZE tasks. */
+/*
+ * Loop thru all the tasks in the system and assign ret_stack.
+ * The @ret_stack_list is a small pre-allocated fallback, in case
+ * the GFP_NOWAIT allocations here fail.
+ */
 static int alloc_retstack_tasklist(unsigned long **ret_stack_list)
 {
-	int i;
-	int ret = 0;
 	int start = 0, end = FTRACE_RETSTACK_ALLOC_SIZE;
 	struct task_struct *g, *t;
+	int i, ret = 0;
 
 	if (WARN_ON_ONCE(!fgraph_stack_cachep))
 		return -ENOMEM;
@@ -1054,26 +1057,37 @@ static int alloc_retstack_tasklist(unsigned long **ret_stack_list)
 		}
 	}
 
-	rcu_read_lock();
-	for_each_process_thread(g, t) {
-		if (start == end) {
-			ret = -EAGAIN;
-			goto unlock;
-		}
+	scoped_guard (rcu) {
+		for_each_process_thread(g, t) {
+			unsigned long *rs;
+
+			if (t->ret_stack)
+				continue;
+
+			rs = kmem_cache_alloc(fgraph_stack_cachep, GFP_NOWAIT);
+			if (!rs) {
+				/*
+				 * Returning from inside scoped_guard() drops
+				 * the RCU read lock, but skips the free loop
+				 * below. That is only safe because start ==
+				 * end here, which leaves that loop nothing to
+				 * free. Keep the two in step if this
+				 * exhaustion check ever changes.
+				 */
+				if (start == end)
+					return -EAGAIN;
+				rs = ret_stack_list[start++];
+			}
 
-		if (t->ret_stack == NULL) {
 			atomic_set(&t->trace_overrun, 0);
-			ret_stack_init_task_vars(ret_stack_list[start]);
+			ret_stack_init_task_vars(rs);
 			t->curr_ret_stack = 0;
 			t->curr_ret_depth = -1;
 			/* Make sure the tasks see the 0 first: */
-			smp_wmb();
-			t->ret_stack = ret_stack_list[start++];
+			smp_store_release(&t->ret_stack, rs);
 		}
 	}
 
-unlock:
-	rcu_read_unlock();
 free:
 	for (i = start; i < end; i++)
 		kmem_cache_free(fgraph_stack_cachep, ret_stack_list[i]);
-- 
2.53.0-Meta

Re: [PATCH v3 1/1] tracing: fgraph: allocate shadow stacks inline with GFP_NOWAIT

[Andrii Nakryiko]

On Thu, Oct 1, 2026 at 11:29 AM Vineet Gupta <vineet.gupta@linux.dev> wrote:
>
> When ftrace graphing is turned on, all tasks in the system missing
> return stack page are assigned one. This is done in a simplistic
> multi-sweep loop of FTRACE_RETSTACK_ALLOC_SIZE (currently 32) tasks
> at a time as follows:
>
>    start_graph_tracing()
>         do {
>                 alloc_retstack_tasklist
>         } while (-EAGAIN);
>
>    alloc_retstack_tasklist()
>         alloc x32           # GFP_KERNEL, may sleep
>         rcu_read_lock()     # preempt off
>                 for_each_process_thread    walk N_total, no cond_resched
>                      t->ret_stack = new_page
>         rcu_read_unlock()   # preempt enable but no explicit yield
>
> Each successive iteration of loop invokes for_each_process_thread()
> which doesn't support cursor based resume and always restarts from the
> init_task. Thus each successive loop needs to skip the tasks assigned
> ret_stack in prior sweeps and thus take longer and longer to find the
> candidate 32 tasks. And while the loop end calls rcu unlock,
> and re-enables preemption briefly, there is no explicit yield.
>
> The total number of iterations turn out to be
>
>         N_total * N_null / (2 * FTRACE_RETSTACK_ALLOC_SIZE)
>
> where N_total is the number of threads on the system and N_null the number
> of those whose ret_stack is still NULL. After boot with no fgraph user that
> is every thread.
>
> This is not a tracing-only path. Since commit 4346ba160409 ("fprobe:
> Rewrite fprobe on function-graph tracer") fprobe is built on fgraph,
> so an ordinary kprobe_multi BPF attach that flips ftrace_graph_active
> from 0 to 1 pays the whole cost inside one bpf() syscall. Commit
> 2c67dc457bc6 ("tracing: fprobe: optimization for entry only case")
> narrows that to return and session probes but does not remove it.
>
> A host running hundreds of thousands of threads, with the current batch
> size of 32, turns the ftrace_graph_active 0 -> 1 transition into a
> multi-second, and eventually multi-minute, operation which showed up as
> RCU stall and softlockup_panic on heavily loaded Meta Fleet machines
> with preemption disabled.
>
>     rcu: INFO: rcu_sched self-detected stall on CPU
>     rcu: 0-....: (20718 ticks this GP) (t=21000 jiffies g=913 q=8 ncpus=8)
>     CPU: 0 UID: 0 PID: 160141 Comm: bpftrace
>      ___slab_alloc+0x549/0xa20
>      kmem_cache_alloc_noprof+0x16c/0x340
>      register_ftrace_graph+0x3d7/0x6c0
>      register_fprobe_ips+0x2d5/0x310
>      bpf_kprobe_multi_link_attach+0x218/0x7e0
>      __sys_bpf+0x267a/0x27a0
>
> Fix this by allocating inline with GFP_NOWAIT instead of pre-allocating
> a fixed batch. GFP_NOWAIT does not sleep, so it is legal inside the RCU
> read-side critical section.
>
> A single sweep now serves every task, so the walk becomes O(N) rather
> than O(N_total * N_null), and the retry loop is no longer the normal
> path.
>
> scoped_guard(rcu) replaces the open-coded rcu_read_lock() and the
> goto unlock, which is what makes returning directly from inside the
> critical section legal.
>
> The pre-allocated array remains as a fallback, although the testing
> below never needed it.
>
> smp_store_release() is a drop-in replacement for the smp_wmb() and
> plain store it replaces.
>
> Measured on a 60-core Sapphire Rapids machine running a PREEMPT_LAZY
> kernel (CONFIG_PREEMPT_DYNAMIC=y, mode lazy), timing the write that
> drives the ftrace_graph_active 0 -> 1 transition. Threads are spawned
> while fgraph is inactive so every one of them has ret_stack == NULL,
> and the loop was instrumented to count sweeps:
>
>         threads   before              after
>                   sweeps   time       sweeps   time
>         100000     3126     15.3 s      1       0.019 s
>         200000     6253     59.1 s      1       0.047 s
>         400000    12503    227.8 s      1       0.092 s
>
> The sweep count is 1 throughout. Time now scales linearly with the
> thread count at roughly 230 ns per task, which is the allocation and
> initialisation of one shadow stack.
>
> Fixes: f201ae2356c7 ("tracing/function-return-tracer: store return stack into task_struct and allocate it dynamically")
> Suggested-by: Peter Zijlstra <peterz@infradead.org>
> Signed-off-by: Vineet Gupta <vineet.gupta@linux.dev>
> Tested-by: Vineet Gupta <vineet.gupta@linux.dev>
> Cc: stable@vger.kernel.org
> ---
>  kernel/trace/fgraph.c | 44 ++++++++++++++++++++++++++++---------------
>  1 file changed, 29 insertions(+), 15 deletions(-)
>

Acked-by: Andrii Nakryiko <andrii@kernel.org>

[...]