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/usb/bulk-streams.rst | 83 +++++++++++++++++++++++++++ 1 file changed, 83 insertions(+) create mode 100644 Documentation/driver-api/usb/bulk-streams.rst (limited to 'Documentation/driver-api/usb/bulk-streams.rst') diff --git a/Documentation/driver-api/usb/bulk-streams.rst b/Documentation/driver-api/usb/bulk-streams.rst new file mode 100644 index 000000000..eeefe582f --- /dev/null +++ b/Documentation/driver-api/usb/bulk-streams.rst @@ -0,0 +1,83 @@ +USB bulk streams +~~~~~~~~~~~~~~~~ + +Background +========== + +Bulk endpoint streams were added in the USB 3.0 specification. Streams allow a +device driver to overload a bulk endpoint so that multiple transfers can be +queued at once. + +Streams are defined in sections 4.4.6.4 and 8.12.1.4 of the Universal Serial Bus +3.0 specification at https://www.usb.org/developers/docs/ The USB Attached SCSI +Protocol, which uses streams to queue multiple SCSI commands, can be found on +the T10 website (https://t10.org/). + + +Device-side implications +======================== + +Once a buffer has been queued to a stream ring, the device is notified (through +an out-of-band mechanism on another endpoint) that data is ready for that stream +ID. The device then tells the host which "stream" it wants to start. The host +can also initiate a transfer on a stream without the device asking, but the +device can refuse that transfer. Devices can switch between streams at any +time. + + +Driver implications +=================== + +:: + + int usb_alloc_streams(struct usb_interface *interface, + struct usb_host_endpoint **eps, unsigned int num_eps, + unsigned int num_streams, gfp_t mem_flags); + +Device drivers will call this API to request that the host controller driver +allocate memory so the driver can use up to num_streams stream IDs. They must +pass an array of usb_host_endpoints that need to be setup with similar stream +IDs. This is to ensure that a UASP driver will be able to use the same stream +ID for the bulk IN and OUT endpoints used in a Bi-directional command sequence. + +The return value is an error condition (if one of the endpoints doesn't support +streams, or the xHCI driver ran out of memory), or the number of streams the +host controller allocated for this endpoint. The xHCI host controller hardware +declares how many stream IDs it can support, and each bulk endpoint on a +SuperSpeed device will say how many stream IDs it can handle. Therefore, +drivers should be able to deal with being allocated less stream IDs than they +requested. + +Do NOT call this function if you have URBs enqueued for any of the endpoints +passed in as arguments. Do not call this function to request less than two +streams. + +Drivers will only be allowed to call this API once for the same endpoint +without calling usb_free_streams(). This is a simplification for the xHCI host +controller driver, and may change in the future. + + +Picking new Stream IDs to use +============================= + +Stream ID 0 is reserved, and should not be used to communicate with devices. If +usb_alloc_streams() returns with a value of N, you may use streams 1 though N. +To queue an URB for a specific stream, set the urb->stream_id value. If the +endpoint does not support streams, an error will be returned. + +Note that new API to choose the next stream ID will have to be added if the xHCI +driver supports secondary stream IDs. + + +Clean up +======== + +If a driver wishes to stop using streams to communicate with the device, it +should call:: + + void usb_free_streams(struct usb_interface *interface, + struct usb_host_endpoint **eps, unsigned int num_eps, + gfp_t mem_flags); + +All stream IDs will be deallocated when the driver releases the interface, to +ensure that drivers that don't support streams will be able to use the endpoint. -- cgit v1.3.1