All of lore.kernel.org
 help / color / mirror / Atom feed
* xhci 8086:7a60 (HM770) skips ~9.5% of isoc OUT packets when hot; HDA on same PCH underruns in step
@ 2026-11-16 21:05 Gabriel Ilie
  2026-11-20 20:30 ` Michal Pecio
  0 siblings, 1 reply; 2+ messages in thread
From: Gabriel Ilie @ 2026-11-16 21:05 UTC (permalink / raw)
  To: linux-usb; +Cc: mathias.nyman, linux-sound

[-- Attachment #1: Type: text/plain, Size: 4867 bytes --]

Hi,

Sorry in advance if this email does not reach the correct channels,
but I have been looking into this issue for a long time, and thanks to
having some AI assistance I was able to find a fix for it, and thought
it might help for other users too.

On an ASUS ROG Strix G18 (G814JZ, i9-13980HX, Intel HM770) the PCH xHCI
controller skips a large share of isochronous OUT packets to a USB audio
interface whenever the chassis is hot, which is audible as crackle. The
HD-audio controller on the same PCH reports FIFO errors at the same rate
at the same time. Both go to zero within about a minute when the fans are
run faster, with the load unchanged. I would like to know whether this is
a known behaviour of this PCH, and whether the xhci driver could deliver
late instead of skipping in this situation.

Hardware and software
---------------------
xHCI 0000:00:14.0 [8086:7a60], subsystem 1043:201f
hcc params 0x20007fc1 hci version 0x120
quirks 0x0000000200009810
HD-audio 0000:00:1f.3 [8086:7a50] rev 11, snd_hda_intel
Device Universal Audio Volt 2 [2b5a:0021], high speed, UAC2 async
with explicit feedback, 48 kHz S32_LE stereo, interval 125 us,
behind two USB 2 hubs (17ef:30ba -> 17ef:30b7)
Kernel 7.3.3-300.fc45.x86_64 (Fedora 45)
Taint 12288 (O, E): NVIDIA open kernel modules 620.70.06. The
affected controllers are both Intel PCH devices.
Userspace PipeWire 1.6.11; it reports zero xruns throughout.

What I see
----------
1. usbmon, 30 s of playback while the crackle is audible:

- 239,990 isoc OUT packets on ep 1; about 22,700 (9.5 %) complete
with status -18 (-EXDEV) and actual length 0.
- They come in 2,522 separate events (84 per second), a median of 9
consecutive packets each, i.e. about 1.1 ms.
- Gaps between events: minimum 3.9 ms, median 10 ms, 90th percentile
22 ms. Above the 4 ms floor they look random.
- URB status is always 0 and the iso error_count field stays 0.
- Completions arrive one URB period late at each event, then two
arrive together, the second fully skipped. Example:

C Zo:1:012:1 0:1:14116:0 7 0:0:48 0:48:48 0:96:48 0:144:48 0:192:48 336 >
C Zo:1:012:1 0:1:14123:0 6 -18:0:0 -18:48:0 -18:96:0 -18:144:0 -18:192:0 288 >
C Zo:1:012:1 0:1:14129:0 7 -18:0:0 -18:48:0 0:96:48 0:144:48 0:192:48 336 >

The URBs are queued about 10 ms ahead (submit-to-complete latency
9.9 ms median), so the ring is not running dry.

2. At the same time, playing silence straight to the HD-audio device
(aplay -D hw:0,0 /dev/zero, no sound server) produces

snd_hda_intel 0000:00:1f.3: stream 7 dma error: 0x28

about 75 times per second (counted from the controller's IRQ, since
the message is rate-limited). The rate does not depend on format or
period size. The same test on the NVIDIA HDMI audio controller
(01:00.1, not on the PCH) gives no errors.

3. The rate follows temperature and airflow, not load:

- With GPU power constant, slowing the fans raised the HD-audio error
rate from about 75/s to over 90/s within 20 s.
- Running all fans at full speed took it from 80/s to 0 within about
70 s, again at constant GPU power. An ACPI board sensor (thermal
zone SEN3) fell from 88 C to 82 C over that time.
- It scaled from about 20/s at 64 C to over 100/s at 91 C on that
sensor with the vendor fan curve.

No effect on the rate: intel_pstate max_perf_pct 55/70/100, no_turbo,
performance vs powersave governor, uncore frequency pinned min=max,
intel_idle.max_cstate=1, synthetic CPU / memory / NVMe load, a split-lock
generator, and stopping all sensor polling. MSR 0x34 shows no SMIs during
the events. Reducing CPU package thermal throttling from 98 to 17 events
per second did not change it.

Questions
---------
1. Is a periodic ~1 ms inability to service periodic endpoints a known
behaviour of the 700-series PCH when it is hot (PCH thermal
throttling, DMI, or something else)? Is there a way to read the PCH's
throttle state or temperature on this platform to confirm it?

2. The controller advertises CFC (hcc params bit 11), so as far as I
understand the isoc TDs carry explicit frame IDs and the ones whose
frame has passed are skipped after the Missed Service Error. For a
UAC2 async device with its own buffer, sending those packets late
would probably be inaudible, while dropping 1 ms is not. Would using
SIA instead of frame IDs be acceptable here, perhaps as a quirk or an
option I can test?

3. For the sound list: is "dma error: 0x28" (FIFOE with FIFORDY) on an
output stream consistent with the controller being unable to fetch
data for about a millisecond, or does it point at something else?

I can test patches, run with xhci dynamic debug, and provide the full
usbmon capture (3 MB compressed). I can also retest on a mainline kernel
without the NVIDIA modules if that is needed.

Short excerpts of the capture, the kernel messages and the measurements
are attached as small text files.

Thanks,
Gabriel Ilie

[-- Attachment #2: hardware.txt --]
[-- Type: text/plain, Size: 1259 bytes --]

Machine     ASUS ROG Strix G18 G814JZ, BIOS G814JZ.334 (02/11/2026)
CPU         Intel Core i9-13980HX
Chipset     Intel HM770 (eSPI controller 8086:7a0c rev 11)
xHCI        0000:00:14.0 Intel 700 Series Chipset Family USB 3.2 Gen 2x2 xHCI [8086:7a60], subsystem 1043:201f
            hcc params 0x20007fc1 hci version 0x120 quirks 0x0000000200009810
HD-audio    0000:00:1f.3 Intel 700 Series Chipset Family HD Audio [8086:7a50] rev 11, subsystem 1043:1cdf, snd_hda_intel
            codec ALC285, CS35L41 amplifiers
GPU         NVIDIA RTX 4080 Laptop (dGPU-only MUX mode); its HDMI audio function 01:00.1 shows no errors
Kernel      7.3.3-300.fc45.x86_64 (Fedora 45), PREEMPT_DYNAMIC (lazy), HZ=1000
            tainted 12288 (O, E): out-of-tree NVIDIA open kernel modules 620.70.06
            cmdline extras: intel_idle.max_cstate=1 processor.max_cstate=1
USB device  Universal Audio Volt 2 [2b5a:0021], high speed, UAC2 async with explicit feedback endpoint,
            48 kHz S32_LE stereo, data packet interval 125 us
            path: root port 1 -> hub 17ef:30ba -> hub 17ef:30b7 -> device (Lenovo Thunderbolt 4 dock, USB 2 lines)
            snd_usb_audio lowlatency=Y
Audio stack PipeWire 1.6.11, WirePlumber 0.5.19 (zero xruns reported throughout)

[-- Attachment #3: usbmon-excerpt.txt --]
[-- Type: text/plain, Size: 3039 bytes --]

usbmon text capture (bus 1, device 12 = Universal Audio Volt 2, 2b5a:0021), kernel 7.3.3.
Zo:1:012:1 = isochronous OUT ep 1 (playback, 8 packets/ms); Zi:1:012:1 = feedback endpoint.
In completions ('C'), each 'status:offset:length' is one packet; -18 = -EXDEV with length 0 = skipped.
Audio payload words removed. Timestamps are microseconds.

ffff899786233500 3445746947 C Zi:1:012:1 0:8:14104:0 1 0:0:4 4 = <payload removed>
ffff899786233500 3445746950 S Zi:1:012:1 -115:8:14104 1 -18:0:4 4 <
ffff89997ee12c00 3445747696 C Zo:1:012:1 0:1:14104:0 7 0:0:48 0:48:48 0:96:48 0:144:48 0:192:48 336 >
ffff89997ee12c00 3445747699 S Zo:1:012:1 -115:1:14104 6 -18:0:48 -18:48:48 -18:96:48 -18:144:48 -18:192:48 288 = <payload removed>
ffff899786232200 3445747942 C Zi:1:012:1 0:8:14112:0 1 0:0:4 4 = <payload removed>
ffff899786232200 3445747943 S Zi:1:012:1 -115:8:14112 1 -18:0:4 4 <
ffff89974350ae00 3445748316 C Zo:1:012:1 0:1:14111:0 5 0:0:48 0:48:48 0:96:48 0:144:48 0:192:48 240 >
ffff89974350ae00 3445748319 S Zo:1:012:1 -115:1:14111 7 -18:0:48 -18:48:48 -18:96:48 -18:144:48 -18:192:48 336 = <payload removed>
ffff899786232400 3445748943 C Zi:1:012:1 0:8:14120:0 1 0:0:4 4 = <payload removed>
ffff899786232400 3445748944 S Zi:1:012:1 -115:8:14120 1 -18:0:4 4 <
ffff89974fdcb600 3445750125 C Zo:1:012:1 0:1:14116:0 7 0:0:48 0:48:48 0:96:48 0:144:48 0:192:48 336 >
ffff89974fdcb600 3445750131 S Zo:1:012:1 -115:1:14116 5 -18:0:48 -18:48:48 -18:96:48 -18:144:48 -18:192:48 240 = <payload removed>
ffff899786233200 3445750136 C Zi:1:012:1 0:8:14128:0 1 0:0:4 4 = <payload removed>
ffff899786233200 3445750136 S Zi:1:012:1 -115:8:14128 1 -18:0:4 4 <
ffff899fd6381e00 3445750188 C Zo:1:012:1 0:1:14123:0 6 -18:0:0 -18:48:0 -18:96:0 -18:144:0 -18:192:0 288 >
ffff899fd6381e00 3445750191 S Zo:1:012:1 -115:1:14123 7 -18:0:48 -18:48:48 -18:96:48 -18:144:48 -18:192:48 336 = <payload removed>
ffff899f8144f600 3445750820 C Zo:1:012:1 0:1:14129:0 7 -18:0:0 -18:48:0 0:96:48 0:144:48 0:192:48 336 >
ffff899f8144f600 3445750823 S Zo:1:012:1 -115:1:14129 7 -18:0:48 -18:48:48 -18:96:48 -18:144:48 -18:192:48 336 = <payload removed>
ffff899786233500 3445750947 C Zi:1:012:1 0:8:14136:0 1 0:0:4 4 = <payload removed>
ffff899786233500 3445750948 S Zi:1:012:1 -115:8:14136 1 -18:0:4 4 <
ffff899da7b5be00 3445751706 C Zo:1:012:1 0:1:14136:0 7 0:0:48 0:48:48 0:96:48 0:144:48 0:192:48 336 >
ffff899da7b5be00 3445751709 S Zo:1:012:1 -115:1:14136 6 -18:0:48 -18:48:48 -18:96:48 -18:144:48 -18:192:48 288 = <payload removed>
ffff899786232200 3445751947 C Zi:1:012:1 0:8:14144:0 1 0:0:4 4 = <payload removed>
ffff899786232200 3445751948 S Zi:1:012:1 -115:8:14144 1 -18:0:4 4 <
ffff899f2b8e6000 3445752453 C Zo:1:012:1 0:1:14143:0 6 0:0:48 0:48:48 0:96:48 0:144:48 0:192:48 288 >
ffff899f2b8e6000 3445752459 S Zo:1:012:1 -115:1:14143 7 -18:0:48 -18:48:48 -18:96:48 -18:144:48 -18:192:48 336 = <payload removed>
ffff899786232400 3445752947 C Zi:1:012:1 0:8:14152:0 1 0:0:4 4 = <payload removed>
ffff899786232400 3445752948 S Zi:1:012:1 -115:8:14152 1 -18:0:4 4 <

[-- Attachment #4: dmesg-excerpts.txt --]
[-- Type: text/plain, Size: 2471 bytes --]

Kernel messages, boot of 2026-11-16 10:09 (kernel 7.3.3-300.fc45.x86_64). Hostname replaced.

--- xHCI probe ---
[    1.845377] host kernel: xhci_hcd 0000:00:14.0: xHCI Host Controller
[    1.845415] host kernel: xhci_hcd 0000:00:14.0: new USB bus registered, assigned bus number 1
[    1.845456] host kernel: xhci_hcd 0000:00:14.0: hcc params 0x20007fc1 hci version 0x120 quirks 0x0000000200009810
[    1.845495] host kernel: xhci_hcd 0000:00:14.0: xHCI Host Controller
[    1.845533] host kernel: xhci_hcd 0000:00:14.0: new USB bus registered, assigned bus number 2
[    1.845571] host kernel: xhci_hcd 0000:00:14.0: Host supports USB 3.2 Enhanced SuperSpeed

--- HD-audio FIFO errors while playing silence on hw:0,0 (rate-limited by the kernel) ---
Nov 16 10:46:51 host kernel: snd_hda_intel 0000:00:1f.3: stream 7 dma error: 0x28
Nov 16 10:46:51 host kernel: snd_hda_intel 0000:00:1f.3: stream 7 dma error: 0x28
Nov 16 10:46:51 host kernel: snd_hda_intel 0000:00:1f.3: stream 7 dma error: 0x28
Nov 16 10:46:51 host kernel: snd_hda_intel 0000:00:1f.3: stream 7 dma error: 0x28
Nov 16 10:46:51 host kernel: snd_hda_intel 0000:00:1f.3: stream 7 dma error: 0x28
Nov 16 10:46:51 host kernel: snd_hda_intel 0000:00:1f.3: stream 7 dma error: 0x28
Nov 16 10:46:51 host kernel: snd_hda_intel 0000:00:1f.3: stream 7 dma error: 0x28
Nov 16 10:46:51 host kernel: snd_hda_intel 0000:00:1f.3: stream 7 dma error: 0x28
Nov 16 10:46:51 host kernel: snd_hda_intel 0000:00:1f.3: stream 7 dma error: 0x28
Nov 16 10:46:51 host kernel: snd_hda_intel 0000:00:1f.3: stream 7 dma error: 0x28
Nov 16 10:46:56 host kernel: snd_hdac_bus_handle_stream_irq: 312 callbacks suppressed
Nov 16 10:46:56 host kernel: snd_hda_intel 0000:00:1f.3: stream 7 dma error: 0x28
Nov 16 10:46:56 host kernel: snd_hda_intel 0000:00:1f.3: stream 7 dma error: 0x28
Nov 16 10:46:56 host kernel: snd_hda_intel 0000:00:1f.3: stream 7 dma error: 0x28
[...]
total 'dma error' lines that boot: 4013; suppressed (sum of 'callbacks suppressed'): 107690

--- xHCI warnings, seen once (10:48:49, while the audio interface's profile was being switched) ---
Nov 16 10:48:49 host kernel: xhci_hcd 0000:00:14.0: WARN: buffer overrun event for slot 11 ep 4 on endpoint   (x8)
Nov 16 10:48:49 host kernel: xhci_hcd 0000:00:14.0: Event dma <addr> for ep 1 status 1 not part of TD at <addr> - <addr>   (x10)
Nov 16 10:48:49 host kernel: xhci_hcd 0000:00:14.0: ERROR Unknown event condition 198 for slot 15 ep 1 , HC probably busted

[-- Attachment #5: measurements.txt --]
[-- Type: text/plain, Size: 2082 bytes --]

All figures from 2026-11-16, game running (GPU 130-160 W). "Stalls/s" = HD-audio FIFO error
interrupts per second on 00:1f.3 while aplay sends silence to hw:0,0 (IRQ count minus period IRQs).
"Board" = ACPI thermal zone SEN3. Fan speeds are CPU / GPU / middle fan in rpm.

USB capture (usbmon, 30 s, Volt 2 playback endpoint, board 87-90 C, fans 5500/4900/7400):
  239,990 isochronous OUT packets; about 22,700 (9.5 %) completed with -EXDEV and length 0
  2,522 separate events = 84 per second; median 9 consecutive packets (about 1.1 ms) per event
  time between events: min 3.9 ms, median 10.1 ms, 90th percentile 22 ms (looks random above 4 ms)
  feedback endpoint: 114 of 29,999 packets -EXDEV
  HD-audio stalls measured one minute earlier under the same conditions: about 75 per second

Airflow, GPU power held roughly constant:
  fans 5500 / 5400 / 4300 (vendor curve)         board 88-91 C   75-90 stalls/s
  fans slowed to 4300 / 4500 / 4300              board 89-91 C   rises to 90+ within 20 s
  fans 6200 / 6900 / 7400                        board 81-83 C   0 within about 70 s
  fans 5500 / 4900 / 7400                        board 79-85 C   0 at 128 W; about 16/s at 153 W, board 87 C
  fans 4270 / 6090 / 6100                        board 79 C      about 5/s and rising
  fans 4270 / 6090 / 7400                        board 79-84 C   0 (also at 154 W)

No effect on the stall rate (each tested for 8-15 s against a baseline of 64-87/s):
  intel_pstate max_perf_pct 55 / 70 / 100; no_turbo=1; governor performance vs powersave
  uncore frequency pinned (min=max) at 4.9 GHz and at 3.0 GHz
  PCM format S16 vs S32, period 128 to 4096 frames (direct ALSA, no PipeWire)
  about 112,000 split-lock operations per second from a test program
  pausing all sensor polling (nvidia-smi, KDE system monitor)
  synthetic CPU, memory-copy and NVMe-read load
  SMI count (MSR 0x34): 0 per second during stalls
  CPU package thermal throttling reduced from 98 to 17 events/s: stall rate unchanged

Same silent-playback test on the NVIDIA HDMI audio controller (01:00.1): no errors.

^ permalink raw reply	[flat|nested] 2+ messages in thread

end of thread, other threads:[~2026-11-20 20:30 UTC | newest]

Thread overview: 2+ messages (download: mbox.gz follow: Atom feed
-- links below jump to the message on this page --
2026-11-16 21:05 xhci 8086:7a60 (HM770) skips ~9.5% of isoc OUT packets when hot; HDA on same PCH underruns in step Gabriel Ilie
2026-11-20 20:30 ` Michal Pecio

This is an external index of several public inboxes,
see mirroring instructions on how to clone and mirror
all data and code used by this external index.