The mgag200_bmc_stop_scanout() function is called by the .atomic_disable()
handler for the MGA G200 VGA BMC encoder. This function performs a few
register writes to inform the BMC of an upcoming mode change, and then
polls to wait until the BMC actually stops.
The polling is implemented using a busy loop with udelay() and an iteration
timeout of 300, resulting in the function blocking for 300 milliseconds.
The function gets called ultimately by the output_poll_execute work thread
for the DRM output change polling thread of the mgag200 driver:
kworker/0:0-mm_ 3528 [000] 4555.315364:
ffffffffaa0e25b3 delay_halt.part.0+0x33
ffffffffc03f6188 mgag200_bmc_stop_scanout+0x178
ffffffffc087ae7a disable_outputs+0x12a
ffffffffc087c12a drm_atomic_helper_commit_tail+0x1a
ffffffffc03fa7b6 mgag200_mode_config_helper_atomic_commit_tail+0x26
ffffffffc087c9c1 commit_tail+0x91
ffffffffc087d51b drm_atomic_helper_commit+0x11b
ffffffffc0509694 drm_atomic_commit+0xa4
ffffffffc05105e8 drm_client_modeset_commit_atomic+0x1e8
ffffffffc0510ce6 drm_client_modeset_commit_locked+0x56
ffffffffc0510e24 drm_client_modeset_commit+0x24
ffffffffc088a743 __drm_fb_helper_restore_fbdev_mode_unlocked+0x93
ffffffffc088a683 drm_fb_helper_hotplug_event+0xe3
ffffffffc050f8aa drm_client_dev_hotplug+0x9a
ffffffffc088555a output_poll_execute+0x29a
ffffffffa9b35924 process_one_work+0x194
ffffffffa9b364ee worker_thread+0x2fe
ffffffffa9b3ecad kthread+0xdd
ffffffffa9a08549 ret_from_fork+0x29
On a server running ptp4l with the mgag200 driver loaded, we found that
ptp4l would sometimes get blocked from execution because of this busy
waiting loop.
Every so often, approximately once every 20 minutes -- though with large
variance -- the output_poll_execute() thread would detect some sort of
change that required performing a hotplug event which results in attempting
to stop the BMC scanout, resulting in a 300msec delay on one CPU.
On this system, ptp4l was pinned to a single CPU. When the
output_poll_execute() thread ran on that CPU, it blocked ptp4l from
executing for its 300 millisecond duration.
This resulted in PTP service disruptions such as failure to send a SYNC
message on time, failure to handle ANNOUNCE messages on time, and clock
check warnings from the application. All of this despite the application
being configured with FIFO_RT and a higher priority than the background
workqueue tasks. (However, note that the kernel did not use
CONFIG_PREEMPT...)
It is unclear if the event is due to a faulty VGA connection, another bug,
or actual events causing a change in the connection. At least on the system
under test it is not a one-time event and consistently causes disruption to
the time sensitive applications.
The function has some helpful comments explaining what steps it is
attempting to take. In particular, step 3a and 3b are explained as such:
3a - The third step is to verify if there is an active scan. We are
waiting on a 0 on remhsyncsts (<XSPAREREG<0>.
3b - This step occurs only if the remove is actually scanning. We are
waiting for the end of the frame which is a 1 on remvsyncsts
(<XSPAREREG<1>).
The actual steps 3a and 3b are implemented as while loops with a
non-sleeping udelay(). The first step iterates while the tmp value at
position 0 is *not* set. That is, it keeps iterating as long as the bit is
zero. If the bit is already 0 (because there is no active scan), it will
iterate the entire 300 attempts which wastes 300 milliseconds in total.
This is opposite of what the description claims.
The step 3b logic only executes if we do not iterate over the entire 300
attempts in the first loop. If it does trigger, it is trying to check and
wait for a 1 on the remvsyncsts. However, again the condition is actually
inverted and it will loop as long as the bit is 1, stopping once it hits
zero (rather than the explained attempt to wait until we see a 1).
Worse, both loops are implemented using non-sleeping waits which spin
instead of allowing the scheduler to run other processes. If the kernel is
not configured to allow arbitrary preemption, it will waste valuable CPU
time doing nothing.
There does not appear to be any documentation for the BMC register
interface, beyond what is in the comments here. It seems more probable that
the comment here is correct and the implementation accidentally got
inverted from the intended logic.
Reading through other DRM driver implementations, it does not appear that
the .atomic_enable or .atomic_disable handlers need to delay instead of
sleep. For example, the ast_astdp_encoder_helper_atomic_disable() function
calls ast_dp_set_phy_sleep() which uses msleep(). The "atomic" in the name
is referring to the atomic modesetting support, which is the support to
enable atomic configuration from userspace, and not to the "atomic context"
of the kernel. There is no reason to use udelay() here if a sleep would be
sufficient.
Replace the while loops with a read_poll_timeout() based implementation
that will sleep between iterations, and which stops polling once the
condition is met (instead of looping as long as the condition is met). This
aligns with the commented behavior and avoids blocking on the CPU while
doing nothing.
Note the RREG_DAC is implemented using a statement expression to allow
working properly with the read_poll_timeout family of functions. The other
RREG_<TYPE> macros ought to be cleaned up to have better semantics, and
several places in the mgag200 driver could make use of RREG_DAC or similar
RREG_* macros should likely be cleaned up for better semantics as well, but
that task has been left as a future cleanup for a non-bugfix.
Fixes: 414c45310625 ("mgag200: initial g200se driver (v2)")
Suggested-by: Thomas Zimmermann <tzimmermann@suse.de>
Signed-off-by: Jacob Keller <jacob.e.keller@intel.com>
Reviewed-by: Thomas Zimmermann <tzimmermann@suse.de>
Reviewed-by: Jocelyn Falempe <jfalempe@redhat.com>
Signed-off-by: Thomas Zimmermann <tzimmermann@suse.de>
Link: https://patch.msgid.link/20260202-jk-mgag200-fix-bad-udelay-v2-1-ce1e9665987d@intel.com
projects / thirdparty / kernel / linux.git / commitdiff
