stack_depot_save() cannot accept __GFP_NOFAIL flag because
__stack_depot_save() drops gfp flags which are not in
GFP_KERNEL | GFP_ATOMIC | __GFP_NOWARN. Also, changing
__stack_depot_save() to accept __GFP_NOFAIL is not possible
because rmqueue() does not want __GFP_NOFAIL flag for
order == DEPOT_POOL_ORDER allocation request.
Therefore, assume that stack_depot_save(GFP_KERNEL | __GFP_NOFAIL) from
ref_tracker_alloc() can silently fail, and emit "unreliable refcount
tracker." message.
Signed-off-by: Tetsuo Handa <[email protected]>
---
lib/ref_tracker.c | 6 ++++++
1 file changed, 6 insertions(+)
diff --git a/lib/ref_tracker.c b/lib/ref_tracker.c
index dc7b14aa3431..ad48ff19adb2 100644
--- a/lib/ref_tracker.c
+++ b/lib/ref_tracker.c
@@ -84,12 +84,18 @@ int ref_tracker_alloc(struct ref_tracker_dir *dir,
gfp_mask |= __GFP_NOFAIL;
*trackerp = tracker = kzalloc(sizeof(*tracker), gfp_mask);
if (unlikely(!tracker)) {
+nomem:
pr_err_once("memory allocation failure, unreliable refcount tracker.\n");
refcount_inc(&dir->untracked);
return -ENOMEM;
}
nr_entries = stack_trace_save(entries, ARRAY_SIZE(entries), 1);
tracker->alloc_stack_handle = stack_depot_save(entries, nr_entries, gfp);
+ if (!tracker->alloc_stack_handle) {
+ *trackerp = NULL;
+ kfree(tracker);
+ goto nomem;
+ }
spin_lock_irqsave(&dir->lock, flags);
list_add(&tracker->head, &dir->list);
--
2.18.4
On Sat, 27 May 2023 20:04:11 +0900 Tetsuo Handa wrote:
> stack_depot_save() cannot accept __GFP_NOFAIL flag because
> __stack_depot_save() drops gfp flags which are not in
> GFP_KERNEL | GFP_ATOMIC | __GFP_NOWARN. Also, changing
> __stack_depot_save() to accept __GFP_NOFAIL is not possible
> because rmqueue() does not want __GFP_NOFAIL flag for
> order == DEPOT_POOL_ORDER allocation request.
>
> Therefore, assume that stack_depot_save(GFP_KERNEL | __GFP_NOFAIL) from
> ref_tracker_alloc() can silently fail, and emit "unreliable refcount
> tracker." message.
It's probably a good idea to CC netdev@vger. I'm not sure if anyone
will pick this up from LKML.
For the patch itself - I'm not sure it's needed, even if we don't
record the stack we'll have a tracker object and still detect the leak.
So printing the "unreliable refcount" message is not very precise.
At least to me; Eric's opinion matters most.
On Tue, May 30, 2023 at 4:05 AM Jakub Kicinski <[email protected]> wrote:
>
> On Sat, 27 May 2023 20:04:11 +0900 Tetsuo Handa wrote:
> > stack_depot_save() cannot accept __GFP_NOFAIL flag because
> > __stack_depot_save() drops gfp flags which are not in
> > GFP_KERNEL | GFP_ATOMIC | __GFP_NOWARN. Also, changing
> > __stack_depot_save() to accept __GFP_NOFAIL is not possible
> > because rmqueue() does not want __GFP_NOFAIL flag for
> > order == DEPOT_POOL_ORDER allocation request.
> >
> > Therefore, assume that stack_depot_save(GFP_KERNEL | __GFP_NOFAIL) from
> > ref_tracker_alloc() can silently fail, and emit "unreliable refcount
> > tracker." message.
>
> It's probably a good idea to CC netdev@vger. I'm not sure if anyone
> will pick this up from LKML.
>
> For the patch itself - I'm not sure it's needed, even if we don't
> record the stack we'll have a tracker object and still detect the leak.
> So printing the "unreliable refcount" message is not very precise.
> At least to me; Eric's opinion matters most.
Thanks Jakub (I was on a 3-days week end, computer turned off)
This patch looks wrong to me, or at very least not complete ?
If we really want this, why not remove all the code dealing with
tracker->alloc_stack_handle
being potentially NULL ?
Note: I never assumed stack_depot_save() would enforce/use NOFAIL.
diff --git a/lib/ref_tracker.c b/lib/ref_tracker.c
index dc7b14aa3431e2bf7a97a7e78220f04da144563d..530c51ab31f227a64e1210d108e9780f0bad72f7
100644
--- a/lib/ref_tracker.c
+++ b/lib/ref_tracker.c
@@ -29,8 +29,7 @@ void ref_tracker_dir_exit(struct ref_tracker_dir *dir)
}
list_for_each_entry_safe(tracker, n, &dir->list, head) {
pr_err("leaked reference.\n");
- if (tracker->alloc_stack_handle)
- stack_depot_print(tracker->alloc_stack_handle);
+ stack_depot_print(tracker->alloc_stack_handle);
leak = true;
list_del(&tracker->head);
kfree(tracker);
@@ -53,8 +52,7 @@ void ref_tracker_dir_print(struct ref_tracker_dir *dir,
list_for_each_entry(tracker, &dir->list, head) {
if (i < display_limit) {
pr_err("leaked reference.\n");
- if (tracker->alloc_stack_handle)
- stack_depot_print(tracker->alloc_stack_handle);
+ stack_depot_print(tracker->alloc_stack_handle);
i++;
} else {
break;
@@ -124,10 +122,8 @@ int ref_tracker_free(struct ref_tracker_dir *dir,
spin_lock_irqsave(&dir->lock, flags);
if (tracker->dead) {
pr_err("reference already released.\n");
- if (tracker->alloc_stack_handle) {
- pr_err("allocated in:\n");
- stack_depot_print(tracker->alloc_stack_handle);
- }
+ pr_err("allocated in:\n");
+ stack_depot_print(tracker->alloc_stack_handle);
if (tracker->free_stack_handle) {
pr_err("freed in:\n");
stack_depot_print(tracker->free_stack_handle);
On 2023/05/30 16:22, Eric Dumazet wrote:
>>> Therefore, assume that stack_depot_save(GFP_KERNEL | __GFP_NOFAIL) from
>>> ref_tracker_alloc() can silently fail, and emit "unreliable refcount
>>> tracker." message.
>
> Note: I never assumed stack_depot_save() would enforce/use NOFAIL.
Hmm, I misread this function.
if (gfp & __GFP_DIRECT_RECLAIM)
gfp_mask |= __GFP_NOFAIL; // Or'ing to "gfp_mask" than "gfp".
*trackerp = tracker = kzalloc(sizeof(*tracker), gfp_mask); // <= This is "gfp_mask".
tracker->alloc_stack_handle = stack_depot_save(entries, nr_entries, gfp); // <= This is "gfp".
So, stack_depot_save(GFP_KERNEL | __GFP_NOFAIL) is not happening.
Then, question becomes whether we want tracker->alloc_stack_handle != NULL or not.
If tracker->alloc_stack_handle == NULL is still useful, this patch will be useless...
On Tue, May 30, 2023 at 11:52 AM Tetsuo Handa
<[email protected]> wrote:
>
> On 2023/05/30 16:22, Eric Dumazet wrote:
> >>> Therefore, assume that stack_depot_save(GFP_KERNEL | __GFP_NOFAIL) from
> >>> ref_tracker_alloc() can silently fail, and emit "unreliable refcount
> >>> tracker." message.
> >
> > Note: I never assumed stack_depot_save() would enforce/use NOFAIL.
>
> Hmm, I misread this function.
>
> if (gfp & __GFP_DIRECT_RECLAIM)
> gfp_mask |= __GFP_NOFAIL; // Or'ing to "gfp_mask" than "gfp".
> *trackerp = tracker = kzalloc(sizeof(*tracker), gfp_mask); // <= This is "gfp_mask".
> tracker->alloc_stack_handle = stack_depot_save(entries, nr_entries, gfp); // <= This is "gfp".
>
> So, stack_depot_save(GFP_KERNEL | __GFP_NOFAIL) is not happening.
Yes.1
>
> Then, question becomes whether we want tracker->alloc_stack_handle != NULL or not.
> If tracker->alloc_stack_handle == NULL is still useful, this patch will be useless...
>
I think it is useful to have the tracker (as Jakub hinted).
It is better than nothing.
We even might be able to allocate memory later for the
free_stack_handle which could give us
developers enough clues for bug hunting.
Thanks.