mailbox: mtk-cmdq: fix runtime PM usage counter leak in cmdq_mbox_flush()
Subsystems (2)
PATCH v4 1/1 mailbox: mtk-cmdq: fix runtime PM usage counter leak in cmdq_mbox_flush()
cmdq_mbox_flush() calls pm_runtime_get_sync() and returns its error code directly on failure. Per Documentation/power/runtime_pm.rst, pm_runtime_get_sync() does not drop the usage counter on error, so the caller is expected to release the reference itself. The early return here skipped that, leaking a runtime PM usage count on cmdq->mbox.dev on every failed resume. v2 fixed this by switching to pm_runtime_resume_and_get(), which performs the get-and-put internally on failure.
Markus Elfring additionally pointed out that the "out:" and "wait:" labels duplicate the same pm_runtime_mark_last_busy() + pm_runtime_put_autosuspend() + return sequence. That duplication hides a second, separate leak: the "wait:" path's timeout branch (readl_poll_timeout_atomic() failing) returns -EFAULT directly, without ever calling pm_runtime_mark_last_busy() / pm_runtime_put_autosuspend(), so the pm_runtime reference taken at function entry is leaked on every polling timeout too.
Fix both by merging the two exit sequences into one common "out_pm:" label reached by both "out:" and "wait:", using "ret" to carry the return value. "ret" is explicitly reset to 0 on both success paths before reaching "out_pm:", since pm_runtime_resume_and_get() can return a positive value (e.g. 1) on success, not just 0, and that stale value must not leak into the function's return value.
Verified the merged exit path with a standalone model of the four possible outcomes (initial get failure, "out:" success, "wait:" success, "wait:" timeout): the runtime PM refcount returns to its starting value and the return code is correct on every path.
---
v4:
- Add parentheses to the function name in the summary phrase (cmdq_mbox_flush -> cmdq_mbox_flush()), per Markus Elfring's review.
- Reformat this changelog so each version identifier stands on its own line and blocks are separated by blank lines, per Markus Elfring's review.
- No functional change from v3. Still no Fixes tag: I don't have a git history for this file available in my current environment to identify the commit that introduced this exit path with confidence, and would rather leave the tag off than guess a SHA. Still happy to add it in a v5 if Markus or anyone else on Cc can point at the right commit.
v3:
- Fix the pm_runtime leak on the "wait:" timeout path found while addressing Markus Elfring's duplicate-code comment, by merging the "out:" and "wait:" exit sequences into a single "out_pm:" label.
- Add Matthias Brugger and AngeloGioacchino Del Regno plus the linux-mediatek and linux-arm-kernel lists to Cc, per MAINTAINERS' "ARM/Mediatek SoC support" entry (matches any drivers/mailbox/mtk-*.c filename), per Markus Elfring's review.
v2:
- Use pm_runtime_resume_and_get() instead of pm_runtime_get_sync() plus a manual pm_runtime_put_noidle() on the error path, per Markus Elfring's review.
drivers/mailbox/mtk-cmdq-mailbox.c | 15 +++++++-------- 1 file changed, 7 insertions(+), 8 deletions(-)
drivers/mailbox/mtk-cmdq-mailbox.c +7 -8
@@ -565,7 +565,7 @@ static int cmdq_mbox_flush(struct mbox_chan *chan, unsigned long timeout)int ret; - ret =pm_runtime_get_sync (cmdq->mbox.dev); + ret =pm_runtime_resume_and_get (cmdq->mbox.dev);if (ret <0 )return ret;
@@ -590,22 +590,21 @@ static int cmdq_mbox_flush(struct mbox_chan *chan, unsigned long timeout)cmdq_thread_disable (cmdq, thread); out:spin_unlock_irqrestore (&thread->chan->lock, flags); -pm_runtime_mark_last_busy (cmdq->mbox.dev); -pm_runtime_put_autosuspend (cmdq->mbox.dev); - -return 0 ; + ret =0 ; + goto out_pm; wait:cmdq_thread_resume (thread);spin_unlock_irqrestore (&thread->chan->lock, flags); + ret =0 ;if (readl_poll_timeout_atomic (thread->base +CMDQ_THR_ENABLE_TASK , enable, enable ==0 ,1 , timeout)) {dev_err (cmdq->mbox.dev,"Fail to wait GCE thread 0x%x done\n" , (u32)(thread->base - cmdq->base)); - -return -EFAULT ; + ret = -EFAULT ; } + +out_pm:pm_runtime_mark_last_busy (cmdq->mbox.dev);pm_runtime_put_autosuspend (cmdq->mbox.dev); -return 0 ;
+ return ret;
}
--
2.46.2.windows.1
…
Quoted 2 lines
---
v4:
…
Quoted 2 lines
- No functional change from v3. Still no Fixes tag: I don't have a git history for this file available in my current environment to
How can this happen?
Are you still looking for more convenient user interfaces?
See also:
* https://git-scm.com/docs/git-blame
* https://www.geeksforgeeks.org/git/how-to-use-git-blame/#an-alternate-option-to-the-command-line-gui-based
* https://stackoverflow.com/questions/31203001/what-does-git-blame-do
Quoted 1 line
identify the commit that introduced this exit path with confidence,
I assume that corresponding patching experiences can grow.
Quoted 3 lines
and would rather leave the tag off than guess a SHA. Still happy to add it in a v5 if Markus or anyone else on Cc can point at the right commit.
It would be nice if further advice would be added accordingly.
…
Quoted 1 line
drivers/mailbox/mtk-cmdq-mailbox.c| 15 +++++++--------…
Regards,
Markus