** Description changed:
+ [ Impact ]
+
On xHCI controllers that support Contiguous Frame ID (CFC), the resolute 7.0
kernel puts gaps into High-Speed isochronous streams whose URBs are queued
less than about two frames ahead. The gaps are not caused by software
submitting late. The driver itself moves the stream forward by a frame.
The old xhci_get_isoc_frame_id() checks the first TD of every URB against
the valid Frame ID window. When the TD's frame falls on or before the window
start (MFINDEX + IST + 1, rounded up to a frame), the code moves it to
window start + 1 frame and continues the stream from there, even mid-stream.
The microframes in between get no transaction.
This is easy to hit with a USB Audio Class 2 device using asynchronous
explicit feedback where the feedback endpoint has bInterval=1. In that case
snd-usb-audio uses 1-packet (125 us) OUT URBs, so 12 URBs cover only
12 microframes.
Observed on Intel Meteor Lake-P xHCI [8086:7e7d] (HCSPARAMS2 IST = 4
microframes, HCCPARAMS1 CFC = 1), with a High-Speed UAC2 device
(Zephyr UAC2 explicit feedback sample, 2fe3:000e) on a root port:
- - A bus analyzer (OpenVizsla) shows a repeating pattern: about 11-12 SOFs
- with OUT data, then 3-4 SOFs with no IN and no OUT, plus occasional
- single SOFs without OUT. OUT always resumes on the first microframe of a
- new frame number.
- - The xhci trace events show that every isoc TD carries an explicit
- Frame ID, but only 4-7 TDs are assigned per Frame ID instead of 8. The
- OUT histogram of TDs per Frame ID is {4: 352, 5: 816, 6: 651, 7: 183}.
- URBs are resubmitted within a few microseconds of completion, and no
- Missed Service / Ring Underrun events occur.
- - The device receives no OUT data in roughly 25% of microframes. A device
- that expects data in every microframe cannot play the stream correctly.
- - The same hardware works correctly with Windows and macOS. Full-Speed
- mode of the same device is not affected.
+ - A bus analyzer (OpenVizsla) shows a repeating pattern: about 11-12 SOFs
+ with OUT data, then 3-4 SOFs with no IN and no OUT, plus occasional
+ single SOFs without OUT. OUT always resumes on the first microframe of a
+ new frame number.
+ - The xhci trace events show that every isoc TD carries an explicit
+ Frame ID, but only 4-7 TDs are assigned per Frame ID instead of 8. The
+ OUT histogram of TDs per Frame ID is {4: 352, 5: 816, 6: 651, 7: 183}.
+ URBs are resubmitted within a few microseconds of completion, and no
+ Missed Service / Ring Underrun events occur.
+ - The device receives no OUT data in roughly 25% of microframes. A device
+ that expects data in every microframe cannot play the stream correctly.
+ - The same hardware works correctly with Windows and macOS. Full-Speed
+ mode of the same device is not affected.
[ Fix ]
Mainline rewrote the Frame ID calculation in v7.3. Mid-stream URBs are no
longer moved forward, and a follow-up fixed a regression introduced by that
rewrite. All three are clean cherry-picks except for one context line in
xhci-mem.c in the first patch:
- add8469b3e00 ("xhci: fix frame id calculation and checks for isoc URBs")
- b08d153c95f1 ("xhci: Set frame ID field of isoc TRB when starting an isoch
stream")
- 05506a76f13a ("usb: xhci: Fix isochronous scheduling regression")
+ add8469b3e00 ("xhci: fix frame id calculation and checks for isoc URBs")
+ b08d153c95f1 ("xhci: Set frame ID field of isoc TRB when starting an isoch
stream")
+ 05506a76f13a ("usb: xhci: Fix isochronous scheduling regression")
05506a76f13a carries "Fixes: add8469b3e00". No other mainline commit up to
v7.4-rc3 has a Fixes: tag for any of the three. None of them was tagged for
stable. Upstream 7.0.y is no longer maintained, so this can only reach
resolute as an SRU.
[ Test Plan ]
1. Connect a High-Speed UAC2 device with asynchronous explicit feedback
- (feedback endpoint bInterval=1) to an xHCI port on a CFC-capable
- controller. Check that bit 11 of HCCPARAMS1 is set in
- /sys/kernel/debug/usb/xhci/<pci-dev>/reg-cap.
+ (feedback endpoint bInterval=1) to an xHCI port on a CFC-capable
+ controller. Check that bit 11 of HCCPARAMS1 is set in
+ /sys/kernel/debug/usb/xhci/<pci-dev>/reg-cap.
2. Start playback, e.g.
- aplay -D hw:<card>,0 --buffer-time=100000 --period-time=10000 -f dat
/dev/zero
+ aplay -D hw:<card>,0 --buffer-time=100000 --period-time=10000 -f dat
/dev/zero
3. While it plays, record the xhci trace:
- trace-cmd record -b 16384 -e xhci-hcd:xhci_queue_trb \
- -e xhci-hcd:xhci_ring_ep_doorbell -- sleep 2
- Decode the isoc TRBs from xhci_queue_trb: TRB type (field3 bits 15:10)
- == 5, Frame ID = field3 bits 30:20. The endpoint is identified by the
- xhci_ring_ep_doorbell event that follows the TRBs of each URB
- (DCI = 2 * endpoint number + 1 for IN). Count the TDs per Frame ID for
- each endpoint. Skip the first and last frame of the trace, because the
- trace window cuts them part-way.
+ trace-cmd record -b 16384 -e xhci-hcd:xhci_queue_trb \
+ -e xhci-hcd:xhci_ring_ep_doorbell -- sleep 2
+ Decode the isoc TRBs from xhci_queue_trb: TRB type (field3 bits 15:10)
+ == 5, Frame ID = field3 bits 30:20. The endpoint is identified by the
+ xhci_ring_ep_doorbell event that follows the TRBs of each URB
+ (DCI = 2 * endpoint number + 1 for IN). Count the TDs per Frame ID for
+ each endpoint. Skip the first and last frame of the trace, because the
+ trace window cuts them part-way.
4. Expected with the fix: 8 TDs per Frame ID for a 125 us endpoint, Frame IDs
- strictly consecutive, and no idle microframes on a bus analyzer. Without
- the fix: 4-7 TDs per Frame ID and periodic idle microframes.
+ strictly consecutive, and no idle microframes on a bus analyzer. Without
+ the fix: 4-7 TDs per Frame ID and periodic idle microframes.
Results on the hardware above (2 s trace during playback; histogram of TDs
per Frame ID, {TDs per frame: number of frames}):
- - 7.0.0-43-generic (stock, unpatched):
- EP 0x01 OUT {4: 352, 5: 816, 6: 651, 7: 183}
- EP 0x81 IN (feedback) {1: 1001, 3: 1001}
- Periodic idle microframes on the bus.
- - 7.0.0-43.43 source + these three patches (built locally, 7.0.14-xhcibp):
- EP 0x01 OUT {8: 2004}
- EP 0x81 IN (feedback) {8: 2004}
- Frame IDs strictly consecutive. No idle microframes on the bus.
- - resolute master-next 70d93713221b (7.0.0-44.44) + these three patches
- (built locally, 7.0.14-master):
- EP 0x01 OUT {8: 2003}
- EP 0x81 IN (feedback) {8: 2003}
- Frame IDs strictly consecutive. No idle microframes on the bus.
- - Mainline v7.4-rc2, which already contains the three commits:
- no idle microframes on the bus.
+ - 7.0.0-43-generic (stock, unpatched):
+ EP 0x01 OUT {4: 352, 5: 816, 6: 651, 7: 183}
+ EP 0x81 IN (feedback) {1: 1001, 3: 1001}
+ Periodic idle microframes on the bus.
+ - 7.0.0-43.43 source + these three patches (built locally, 7.0.14-xhcibp):
+ EP 0x01 OUT {8: 2004}
+ EP 0x81 IN (feedback) {8: 2004}
+ Frame IDs strictly consecutive. No idle microframes on the bus.
+ - resolute master-next 70d93713221b (7.0.0-44.44) + these three patches
+ (built locally, 7.0.14-master):
+ EP 0x01 OUT {8: 2003}
+ EP 0x81 IN (feedback) {8: 2003}
+ Frame IDs strictly consecutive. No idle microframes on the bus.
+ - Mainline v7.4-rc2, which already contains the three commits:
+ no idle microframes on the bus.
[ Where problems could occur ]
The changes affect scheduling of all isochronous transfers on xHCI hosts:
USB audio, UVC webcams, and other isoc users. Hosts with CFC are the most
affected, because they use an explicit Frame ID for every TD.
- - Patch 1 changes how the first URB of a stream is placed: it is aligned
- to a full frame, or to the interval start if the interval is longer than
- a frame. It also stops falling back to SIA or inserting Frame ID gaps
- mid-stream. A device that relied on those gaps for recovery, or a driver
- that submits isoc URBs late, could now see TDs scheduled into the past,
- which complete with -EXDEV (Missed Service), instead of being silently
- moved forward.
- - Patch 2 also sets the Frame ID for the first TD of a stream on hosts
- without CFC. A controller that mishandles Frame ID on non-CFC hosts could
- start streams at a different time than before.
- - Patch 3 fixes a regression introduced by patch 1. Without it, a stream
- restarted after an OUT underrun (e.g. snd-usb-audio in duplex mode) could
- be scheduled into the past and loop in -EXDEV recovery. It is required
- together with patch 1.
+ - Patch 1 changes how the first URB of a stream is placed: it is aligned
+ to a full frame, or to the interval start if the interval is longer than
+ a frame. It also stops falling back to SIA or inserting Frame ID gaps
+ mid-stream. A device that relied on those gaps for recovery, or a driver
+ that submits isoc URBs late, could now see TDs scheduled into the past,
+ which complete with -EXDEV (Missed Service), instead of being silently
+ moved forward.
+ - Patch 2 also sets the Frame ID for the first TD of a stream on hosts
+ without CFC. A controller that mishandles Frame ID on non-CFC hosts could
+ start streams at a different time than before.
+ - Patch 3 fixes a regression introduced by patch 1. Without it, a stream
+ restarted after an OUT underrun (e.g. snd-usb-audio in duplex mode) could
+ be scheduled into the past and loop in -EXDEV recovery. It is required
+ together with patch 1.
[ Other Info ]
- - The issue was diagnosed with xhci trace events and an OpenVizsla capture.
- snd-usb-audio is not at fault: URBs are resubmitted promptly, and the
- gaps come from the Frame ID assignment in xhci-ring.c.
+ - The issue was diagnosed with xhci trace events and an OpenVizsla capture.
+ snd-usb-audio is not at fault: URBs are resubmitted promptly, and the
+ gaps come from the Frame ID assignment in xhci-ring.c.
ProblemType: Bug
DistroRelease: Ubuntu 26.04
Package: linux-image-7.0.0-43-generic 7.0.0-43.43
ProcVersionSignature: Ubuntu 7.0.0-43.43-generic 7.0.14
Uname: Linux 7.0.0-43-generic x86_64
ApportVersion: 2.34.1-0ubuntu0.1
Architecture: amd64
AudioDevicesInUse:
- USER PID ACCESS COMMAND
- /dev/snd/controlC1: tomo-local 5797 F.... wireplumber
- /dev/snd/controlC0: tomo-local 5797 F.... wireplumber
- /dev/snd/seq: tomo-local 5780 F.... pipewire
+ USER PID ACCESS COMMAND
+ /dev/snd/controlC1: tomo-local 5797 F.... wireplumber
+ /dev/snd/controlC0: tomo-local 5797 F.... wireplumber
+ /dev/snd/seq: tomo-local 5780 F.... pipewire
CasperMD5CheckResult: pass
CurrentDesktop: XFCE
Date: Tue Nov 17 15:26:13 2026
InstallationDate: Installed on 2026-06-23 (147 days ago)
InstallationMedia: Ubuntu 26.04 "Resolute Raccoon" - Release amd64
(20260423.1)
MachineType: Dell Inc. Precision 3490
ProcEnviron:
- LANG=en_US.UTF-8
- PATH=(custom, no user)
- SHELL=/bin/bash
- TERM=xterm-256color
- XDG_RUNTIME_DIR=<set>
+ LANG=en_US.UTF-8
+ PATH=(custom, no user)
+ SHELL=/bin/bash
+ TERM=xterm-256color
+ XDG_RUNTIME_DIR=<set>
ProcFB: 0 i915drmfb
ProcKernelCmdLine: BOOT_IMAGE=/vmlinuz-7.0.0-43-generic
root=/dev/mapper/ubuntu--vg-ubuntu--lv ro quiet splash
crashkernel=2G-4G:320M,4G-32G:512M,32G-64G:1024M,64G-128G:2048M,128G-:4096M
RebootRequiredPkgs: Error: path contained symlinks.
SourcePackage: linux
UpgradeStatus: No upgrade log present (probably fresh install)
dmi.bios.date: 08/25/2026
dmi.bios.release: 1.27
dmi.bios.vendor: Dell Inc.
dmi.bios.version: 1.27.1
dmi.board.name: 0FNG8D
dmi.board.vendor: Dell Inc.
dmi.board.version: A00
dmi.chassis.type: 10
dmi.chassis.vendor: Dell Inc.
dmi.ec.firmware.release: 1.23
dmi.modalias:
dmi:bvnDellInc.:bvr1.27.1:bd08/25/2026:br1.27:efr1.23:svnDellInc.:pnPrecision3490:pvr:rvnDellInc.:rn0FNG8D:rvrA00:cvnDellInc.:ct10:cvr:sku0CB3:pfaPrecision:
dmi.product.family: Precision
dmi.product.name: Precision 3490
dmi.product.sku: 0CB3
dmi.sys.vendor: Dell Inc.
--
You received this bug notification because you are a member of Ubuntu
Bugs, which is subscribed to Ubuntu.
https://bugs.launchpad.net/bugs/2174743
Title:
xhci: High-Speed isochronous OUT stream has periodic gaps on CFC-
capable hosts (USB Audio 2 explicit feedback, bInterval=1)
To manage notifications about this bug go to:
https://bugs.launchpad.net/ubuntu/+source/linux/+bug/2174743/+subscriptions
--
ubuntu-bugs mailing list
[email protected]
https://lists.ubuntu.com/mailman/listinfo/ubuntu-bugs