On Saturday, November 14, 2026 4:01 AM Patrick Fischer via B4 Relay wrote: > From: Patrick Fischer <[email protected]> > > host1x_syncpt_wait() creates its fence without the internal 30 second > timeout and cancels it only when dma_fence_wait_timeout() returns 0. > If the wait is interrupted, dma_fence_wait_timeout() returns > -ERESTARTSYS and the fence is left on the syncpoint's fence list, > holding the reference that host1x_syncpt_fence_enable_signaling() took > for the interrupt path. It is freed only once the syncpoint reaches the > threshold, which may never happen. > > Callers that retry on -ERESTARTSYS allocate a new fence each time. > tegra20_channel_capture_frame() retries in a loop without delay, and > while the syncpoint does not move, for example with a camera that > delivers no frames, every round leaks one fence until the system runs > out of memory. Svyatoslav Ryhel reported this as an out-of-memory crash > in the VI capture thread; the allocations in his report come from > host1x_fence_create(), called by host1x_syncpt_wait() from > tegra20_chan_capture_kthread_start(). The DRM syncpoint wait IOCTLs end > up in host1x_syncpt_wait() as well, so a process that is signalled while > waiting leaks a fence the same way. > > Cancel the fence whenever the wait did not succeed. host1x_fence_cancel() > is safe if the interrupt path has signalled the fence in the meantime, > as do_fence_timeout() then returns early. > > The leak was found with the help of an LLM assistant reading the host1x > and dma-fence code against the call chain in the report. > > Fixes: d5179020f5ce ("gpu: host1x: External timeout/cancellation for fences") > Cc: [email protected] > Reported-by: Svyatoslav Ryhel <[email protected]> > Signed-off-by: Patrick Fischer <[email protected]> > ---
Nice find, thank you! One thing I'd like to add to this patch is moving the call to INIT_LIST_HEAD from host1x_intr_add_fence_locked to host1x_fence_create. Otherwise, theoretically, a dma_fence_wait_timeout error path that returns before calling enable_signaling would with this patch result in host1x_fence_cancel being called, that would go into host1x_intr_remove_fence which assumes the list head has been initialized. Mikko > Svyatoslav reported the out-of-memory crash, and he confirmed off-list > that this patch stops the leak on his LG P895 (Tegra30). I left out a > Tested-by so that he can give it here. > > I could not test this myself: the Tegra Note 7 I work on has no > working camera path in mainline yet. Build-tested for arm and arm64. > The DRM IOCTL case in the changelog comes from reading the code; I have > not tried it. > > The retry loop in tegra20_channel_capture_frame() still spins without > delay as long as the wait keeps returning -ERESTARTSYS. With this patch > that costs CPU time instead of memory; how the VI driver should handle > an interrupted wait is a separate question. > > Tool use (see Documentation/process/generated-content.rst): an LLM > assistant (Claude, in a chat) found the leak by reading > host1x_syncpt_wait() and the fence code against the call chain in the > OOM report. Claude Code then checked the finding against the code and > wrote the change, the changelog and this note from my notes and > instructions. I reviewed the change. > --- > drivers/gpu/host1x/syncpt.c | 2 +- > 1 file changed, 1 insertion(+), 1 deletion(-) > > diff --git a/drivers/gpu/host1x/syncpt.c b/drivers/gpu/host1x/syncpt.c > index fe74ff6fc6f2..f9c871f218f3 100644 > --- a/drivers/gpu/host1x/syncpt.c > +++ b/drivers/gpu/host1x/syncpt.c > @@ -247,7 +247,7 @@ int host1x_syncpt_wait(struct host1x_syncpt *sp, u32 > thresh, long timeout, > return PTR_ERR(fence); > > wait_err = dma_fence_wait_timeout(fence, true, timeout); > - if (wait_err == 0) > + if (wait_err <= 0) > host1x_fence_cancel(fence); > dma_fence_put(fence); > > > --- > base-commit: 9d66f26fefacebfeed15de31c1123451a17ff119 > change-id: 20261113-upstream-host1x-fence-cancel-0b32753150d6 > > Best regards, > -- > Patrick Fischer <[email protected]> > > >
