From 0d6526f82c3cdefcca47f73f5fc08dc6f335eac6 Mon Sep 17 00:00:00 2001 From: Tim Chen Date: Mon, 21 Sep 2026 17:37:22 -0700 Subject: sched/cache: Keep nr_pref_llc_running in the runnable domain, to fix LLC mis-scheduling bug alb_break_llc() decides whether to break LLC preference during active load balance. It does so by testing that every runnable fair task on the source rq prefers its LLC: env->src_rq->nr_pref_llc_running == env->src_rq->cfs.h_nr_runnable But the two counters cover different sets. nr_pref_llc_running is updated in account_llc_enqueue()/account_llc_dequeue(), next to cfs_rq->nr_queued, so it follows queued tasks. h_nr_runnable is updated in set_delayed()/ clear_delayed() and drops delay-dequeued tasks. So under DELAY_DEQUEUE, a preferring task that goes to sleep stays counted in nr_pref_llc_running while h_nr_runnable falls. The equality then breaks, alb_break_llc() returns false, and active balance is free to pull a task off its preferred LLC. Active balance only moves runnable tasks, and this is the only LLC check it consults: once the stopper runs, LBF_ACTIVE_LB skips the per-task test in can_migrate_task(). The runnable set is the one we want. Fix it on the counter side. A task should be counted in nr_pref_llc_running exactly while it is both queued on its preferred LLC (pref_llc_queued) and runnable (!sched_delayed). Define that membership once in task_pref_llc_runnable(), and adjust the counter only through pref_llc_running_inc()/pref_llc_running_dec() from the four sites that change either input: account_llc_enqueue(), account_llc_dequeue(), set_delayed() and clear_delayed(). Gating every update on the same predicate keeps the delay, wake and dequeue paths from double-counting or underflowing; see the comments at those sites for the ordering. nr_llc_running and sd->llc_counts are not touched and stay on queued semantics. Fixes: 714059f79ff0 ("sched/cache: Handle moving single tasks to/from their preferred LLC") Closes: https://lore.kernel.org/lkml/20260827135000.735138-1-zhanxusheng@xiaomi.com/ Reported-by: Zhan Xusheng Suggested-by: Chen Yu Signed-off-by: Tim Chen Signed-off-by: Peter Zijlstra (Intel) Signed-off-by: Ingo Molnar Reviewed-by: Kayra Cizmeci Cc: # v7.2.x Link: https://patch.msgid.link/06af61afedac32e6477f57feb4d658f6c411c3af.1790035273.git.tim.c.chen@linux.intel.com --- Documentation/driver-api/sync_file.rst | 86 ++++++++++++++++++++++++++++++++++ 1 file changed, 86 insertions(+) create mode 100644 Documentation/driver-api/sync_file.rst (limited to 'Documentation/driver-api/sync_file.rst') diff --git a/Documentation/driver-api/sync_file.rst b/Documentation/driver-api/sync_file.rst new file mode 100644 index 000000000..496fb2c3b --- /dev/null +++ b/Documentation/driver-api/sync_file.rst @@ -0,0 +1,86 @@ +=================== +Sync File API Guide +=================== + +:Author: Gustavo Padovan + +This document serves as a guide for device drivers writers on what the +sync_file API is, and how drivers can support it. Sync file is the carrier of +the fences(struct dma_fence) that are needed to synchronize between drivers or +across process boundaries. + +The sync_file API is meant to be used to send and receive fence information +to/from userspace. It enables userspace to do explicit fencing, where instead +of attaching a fence to the buffer a producer driver (such as a GPU or V4L +driver) sends the fence related to the buffer to userspace via a sync_file. + +The sync_file then can be sent to the consumer (DRM driver for example), that +will not use the buffer for anything before the fence(s) signals, i.e., the +driver that issued the fence is not using/processing the buffer anymore, so it +signals that the buffer is ready to use. And vice-versa for the consumer -> +producer part of the cycle. + +Sync files allows userspace awareness on buffer sharing synchronization between +drivers. + +Sync file was originally added in the Android kernel but current Linux Desktop +can benefit a lot from it. + +in-fences and out-fences +------------------------ + +Sync files can go either to or from userspace. When a sync_file is sent from +the driver to userspace we call the fences it contains 'out-fences'. They are +related to a buffer that the driver is processing or is going to process, so +the driver creates an out-fence to be able to notify, through +dma_fence_signal(), when it has finished using (or processing) that buffer. +Out-fences are fences that the driver creates. + +On the other hand if the driver receives fence(s) through a sync_file from +userspace we call these fence(s) 'in-fences'. Receiving in-fences means that +we need to wait for the fence(s) to signal before using any buffer related to +the in-fences. + +Creating Sync Files +------------------- + +When a driver needs to send an out-fence userspace it creates a sync_file. + +Interface:: + + struct sync_file *sync_file_create(struct dma_fence *fence); + +The caller pass the out-fence and gets back the sync_file. That is just the +first step, next it needs to install an fd on sync_file->file. So it gets an +fd:: + + fd = get_unused_fd_flags(O_CLOEXEC); + +and installs it on sync_file->file:: + + fd_install(fd, sync_file->file); + +The sync_file fd now can be sent to userspace. + +If the creation process fail, or the sync_file needs to be released by any +other reason fput(sync_file->file) should be used. + +Receiving Sync Files from Userspace +----------------------------------- + +When userspace needs to send an in-fence to the driver it passes file descriptor +of the Sync File to the kernel. The kernel can then retrieve the fences +from it. + +Interface:: + + struct dma_fence *sync_file_get_fence(int fd); + + +The returned reference is owned by the caller and must be disposed of +afterwards using dma_fence_put(). In case of error, a NULL is returned instead. + +References: + +1. struct sync_file in include/linux/sync_file.h +2. All interfaces mentioned above defined in include/linux/sync_file.h -- cgit v1.3.1