Intel-XE Archive on lore.kernel.org
 help / color / mirror / Atom feed
From: James R Taylor <jrtaylor11@gmail.com>
To: Karthik Poosa <karthik.poosa@intel.com>
Cc: Soham Purkait <soham.purkait@intel.com>,
	Rodrigo Vivi <rodrigo.vivi@intel.com>,
	Victor Ferreira <dev.victorferreira@gmail.com>,
	intel-xe@lists.freedesktop.org
Subject: Re: [PATCH v3 0/9] Add fan control support
Date: Sun, 15 Nov 2026 17:30:33 +0000	[thread overview]
Message-ID: <1794763833.7934.xefan.hermes@gmail.com> (raw)
In-Reply-To: <20261015190751.1268916-1-karthik.poosa@intel.com>

Hi Karthik,

I've tested v3 on a consumer Arc B580. Victor already covered the basics
on the same board with v2, and I saw the same on v3: the stock table
reads back, pwm1 writes are refused outside mode 1, 200 reads back as
199, and pwm1=0 stops the fan. (Full speed reached 3490 RPM here, so it
stayed under fan1_max = 3600 on this card.) So I'll only add what's new.

Hardware: ASRock Arc B580 Steel Legend 12GB (8086:e20b, BMG G21),
          in a USB4 eGPU dock
Kernel:   7.3.4 with v3 applied, plus 8721a51d4 ("drm/xe/pcode:
          Increase default pcode mailbox timeout to 10ms") and
          1a925ad51, which v3 needs on 7.3. One hunk of 8721a51d4
          needed applying by hand because 7.3's xe_pcode_read() names
          its arguments val/val1.
dmesg:    "Number of fans detected: 1"

1. Fan curves under sustained load. I compared three curves under the
   same load: six parallel 4K HDR HEVC -> 1080p HEVC VAAPI transcodes
   for five minutes each, sampled every 5 seconds.

     curve                      package temp     fan          throughput
     stock (pwm1_enable=2)      rose to 52 C     0 RPM until  721 fps
                                                 near the end,
                                                 then 215-300
     user table, 24% -> 100%    flat 33 C        ~1190 RPM    720 fps
     user table, 10% -> 90%     rose to 44 C     200-310 RPM  721 fps

   The user tables committed cleanly once the tenth point was written,
   and switching back to mode 2 restored the stock table each time. One
   thing to note is that the stock curve keeps the fan stopped under
   real sustained load until about 50 C, not only at idle. In table
   mode, duties of 26-31 kept the fan at only 200-310 RPM.

2. Temperatures freeze while the GPU is idle. Every xe temperature
   channel (pkg, vram, mctrl, pcie) held exactly the same value for at
   least 40 seconds while the card was idle, then dropped on the next
   GPU activity. pkg went from 44 C to 32 C and vram from 52 C to 36 C
   after a one-second decode. This isn't from this series, but it
   matters for anyone writing a curve or a userspace controller on top
   of it: an idle card can look much hotter than it is. It may be worth
   a line in the ABI documentation.

3. pwm1 while in firmware mode. Before pwm1_enable has ever been set to
   1, reading pwm1 fails with -ENXIO and logs "pwm not set for fan 0".
   After using mode 1 and going back to mode 2, pwm1 returns the last
   user value instead. Here it read 61 while the firmware curve was
   running the fan at its 10% point, so monitoring tools show a duty
   the fan isn't actually running at. Would it be better for pwm1 to
   report the current firmware duty in mode 2, or consistently return
   an error there?

I tried to test patch 8 with an s2idle suspend, but on this host the
eGPU doesn't come back from suspend at all: "Unable to change power
state from D3cold to D0, device inaccessible", then pci_pm_resume
returns -5 (the dock's USB controller fails the same way). That's a
USB4/eGPU resume problem, not something in this series, but it means I
couldn't test whether the user table is preserved. What I could see is
that the restore path ran and failed cleanly with the device gone:

  xe 0000:6b:00.0: [drm] failed to activate user fan table for fan 0, ret=-11
  xe 0000:6b:00.0: [drm] failed to disable user fan table for fan 0, ret=-11
  xe 0000:6b:00.0: [drm] *ERROR* failed to read fan 0 user point count, ret=-11

There was no crash, and after a reboot the card and the boot-time curve
came back normally. One small thing from the dead-device state, which
is in the existing fan code rather than this series: fan1_input read
4793553658, apparently the all-ones value from the inaccessible device
turned into RPM. Monitoring tools would show that as a real speed.

Everything else has been stable, and the card now runs a user curve
applied at boot.

Tested-by: James R Taylor <jrtaylor11@gmail.com>

Thanks for the series,
James


  parent reply	other threads:[~2026-11-16 14:38 UTC|newest]

Thread overview: 34+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2026-10-15 19:07 [PATCH v3 0/9] Add fan control support Karthik Poosa
2026-10-15 19:07 ` [PATCH v3 1/9] drm/xe/hwmon: initialize fan-control backend and table cache Karthik Poosa
2026-10-28 18:46   ` Purkait, Soham
2026-10-29 11:34     ` Poosa, Karthik
2026-10-29 12:02       ` Purkait, Soham
2026-10-29 12:04       ` Purkait, Soham
2026-11-05  5:58   ` Purkait, Soham
2026-11-05 16:56     ` Poosa, Karthik
2026-10-15 19:07 ` [PATCH v3 2/9] drm/xe/hwmon: expose fanN_max Karthik Poosa
2026-10-29  4:35   ` Purkait, Soham
2026-10-29 11:39     ` Poosa, Karthik
2026-10-15 19:07 ` [PATCH v3 3/9] drm/xe/pcode: add ordered write batch helper Karthik Poosa
2026-10-29  5:45   ` Purkait, Soham
2026-10-29 11:40     ` Poosa, Karthik
2026-10-15 19:07 ` [PATCH v3 4/9] drm/xe/hwmon: expose pwm[1-3] Karthik Poosa
2026-10-29 11:31   ` Purkait, Soham
2026-11-05 17:09     ` Poosa, Karthik
2026-10-15 19:07 ` [PATCH v3 5/9] drm/xe/hwmon: expose pwm[1-3]_enable Karthik Poosa
2026-10-29 18:33   ` Purkait, Soham
2026-11-05 17:41     ` Poosa, Karthik
2026-10-15 19:07 ` [PATCH v3 6/9] drm/xe/hwmon: Enable fan curve control support Karthik Poosa
2026-10-30  6:33   ` Purkait, Soham
2026-10-15 19:07 ` [PATCH v3 7/9] drm/xe/hwmon: Add kernel-doc for fan control Karthik Poosa
2026-10-30 13:42   ` Purkait, Soham
2026-10-15 19:07 ` [PATCH v3 8/9] drm/xe/hwmon: preserve fan user table across suspend resume Karthik Poosa
2026-11-02  5:58   ` Purkait, Soham
2026-10-15 19:07 ` [PATCH v3 9/9] drm/xe/hwmon: Update fan info after late binding Karthik Poosa
2026-11-02  6:26   ` Purkait, Soham
2026-10-15 19:46 ` ✗ CI.checkpatch: warning for Add fan control support Patchwork
2026-10-15 19:48 ` ✓ CI.KUnit: success " Patchwork
2026-10-15 21:01 ` ✗ Xe.CI.BAT: failure " Patchwork
2026-10-16  7:56 ` ✗ Xe.CI.FULL: " Patchwork
2026-11-15 17:30 ` James R Taylor [this message]
  -- strict thread matches above, loose matches on Subject: below --
2026-10-15 19:01 [PATCH v3 0/9] " Karthik Poosa

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=1794763833.7934.xefan.hermes@gmail.com \
    --to=jrtaylor11@gmail.com \
    --cc=dev.victorferreira@gmail.com \
    --cc=intel-xe@lists.freedesktop.org \
    --cc=karthik.poosa@intel.com \
    --cc=rodrigo.vivi@intel.com \
    --cc=soham.purkait@intel.com \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox