gen8 actually changed the command layout, not just extended the
relocation value. Oh well.
References: https://bugs.freedesktop.org/show_bug.cgi?id=83915
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Since relocations are variable size, depending upon generation, it is
easier to handle the resizing of the batch request inside the
BEGIN_BATCH macro. This still leaves us with having to resize commands
in a few places - which still need adaption for gen8+.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
The load_helper isn't killed by the signal, but it exits gracefully. So
update the assertion to check for the successful exit.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
commit 745945546f7366a413a3a51a37f90caa3a227b1d
breaks the build under Android because some of the
macros used in pm_rps.c are defined in sys/wait.h
which is not included.
Signed-off-by: Tim Gore <tim.gore@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Reduce code duplication as the igt_stop_helper can reuse
igt_wait_helper() to replace its own waiting routine.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Note that the sleep(5); to fully idle the gpu is _really_ important.
Without it the bug is not exhibited.
The issue at hand is that after gem_quiescent_gpu we are at max
(expected, since the blocking waits peg to max), but then we never go
down to a lower freq again until we're fully idle. The tiny load is
sufficient to keep the gpu at max. I've played around with this a bit
and even ridiculously low loads (like one MI_STORE per 50ms) are
enough to keep the gpu at max freq.
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Apparently some rps chips are prone to stick to max once you're there.
And all the synchronous waits in gem_quiescent_gpu are _really_ good
at forcing the kernel to go full throttle.
And even a positively tiny load (one MI_STORE_IMM per 50 ms) is enough
to eternally keeep it there, which means the first idle measurement of
the freqs before we do the reset it completely busted.
Just add a wait before the tricky test and call it a day.
Option b) would have been to now have this tiny load, but I think a
good reminder how quirky rps is is always welcome. Just in case we
start sleeping too easy at night.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=74956
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
With the header cleanup we can now give this header a suitable name,
since it now really only contains register access and other I/O
functions and assorted definitions.
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
I've left unistd.h in it - it's not strictly required but most users
of drmtest.h want it for the open helpers, and then you kinda need to
close that file descriptor again ...
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Only the igt core and non-test tools should have asserts to catch
internal errors, tests and helper libraries should all user igt_asert
instead.
Fix things up where assert instead of igt_assert was used.
One tiny step towards header sanity.
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Also update the api docs a bit since the Returns: section was missing.
v2: Readd the accidentally lost line for @filename.
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Big cores seem to take forever to idle, at least my ivb here. Fairly
ridiculous, so maybe there's more room for our debooster to kick in.
Anyway, this gets the min-max-config-idle and reset subtests going
somewhat reliably on my ivb. They still occasionally fail with the
current frequency pegged to max (or close to it) for no apparent
reason at all. Rebooting tends to fix it.
Don't ask, don't tell.
Cc: Jeff McGee <jeff.mcgee@intel.com>
Cc: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Big core platforms need some seriuos omph to break a sweat.
This fixes min-max-config-loaded here on my ivb.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=75146
Cc: Jeff McGee <jeff.mcgee@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
We don't call cleanup handlers when exiting a subtest currently, only
when exiting the entire binary. Which means pm_rps falls over when it
fails more than one subtest.
Cc: Jeff McGee <jeff.mcgee@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
The kernel will round it, so if we don't we'll have a spurious
mismatch. Happens on my machine here with 650-1300MHz range, where the
midpoint is 975.
Cc: Jeff McGee <jeff.mcgee@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Test that RPS functions as expected after a gpu reset.
Signed-off-by: Jeff McGee <jeff.mcgee@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
The load helper can be set to HIGH or LOW. HIGH is the original mode
of sending continuous dword store commands. LOW adds a pause between
each command to reduce throughput.
Signed-off-by: Jeff McGee <jeff.mcgee@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
This ensures that gpu is quiescent when load helper exits.
Signed-off-by: Jeff McGee <jeff.mcgee@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Like subtest basic-api, but additionally requires that current
frequency is increasing to the configured maximum within reasonable
time since we are loaded.
Signed-off-by: Jeff McGee <jeff.mcgee@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
The load helper submits repeated dword store commands to keep the
gpu loaded while subtests running in the parent process check for
expected rps response.
Signed-off-by: Jeff McGee <jeff.mcgee@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Like subtest basic-api, but additionally requires that current
frequency is settling to the configured minimum within reasonable
time since we are idle.
Signed-off-by: Jeff McGee <jeff.mcgee@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Storing values avoids some unnecessary overhead but more importantly
allows all of our processing to be atomic.
Signed-off-by: Jeff McGee <jeff.mcgee@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Add a function that methodically varies min and max to exercise
several valid and invalid combinations. Allow the caller to
define what is to be checked between each step.
Signed-off-by: Jeff McGee <jeff.mcgee@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
The goal of the test is to confirm that gt_cur_freq_mhz can be forced
to the boundaries of the frequency range by collapsing gt_min_freq_mhz
and gt_max_freq_mhz to the target value. But we miss testing the upper
end of the range by targetting the current value of max after it has
been set equal to min. So fix by targetting orginal max instead of
current max.
This correction exposes a problem in setfreq where min is always set
to target before max, which should fail if the target value is greater
than max. So fix that too.
Signed-off-by: Jeff McGee <jeff.mcgee@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
do_writeval now always checks the return value, whether we expect
success or a specific error. Also add new macro writeval_inval to
simplify repeated use of do_writeval to test for EINVAL return code.
Signed-off-by: Jeff McGee <jeff.mcgee@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Bionic C library may not re-read a buffered, read-only file which
results in failure to monitor changes in gt_cur_freq_mhz.
Signed-off-by: Jeff McGee <jeff.mcgee@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Imo power management, power consumption and performance are tightly
enough coupled that we can throw them all into one bin.
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
