[PATCH i-g-t v3 0/2] Integrate kmemleak scans in igt_runner

[Peter Senna Tschudin]

This patch series introduces a library to interact with the Linux
kernel's kmemleak feature and integrates it into igt_runner. If
kmemleaks are detected, they will be saved in the igt_runner results
directory in a file named kmemleak.txt.

During testing, the size of the kmemleak.txt file varied significantly.
Larger files, up to 2 MB, were observed when running i915-BAT on a
Tiger Lake DUT. Conversely, smaller files, typically under 100 KB, were
generated when running Xe BAT on the same DUT.

Large files often contain numerous false positives, with the e1000
driver being a frequent source of noise. The time required for the
Linux kernel to complete a kmemleak scan ranges from 5 to 60 seconds.
This variability can cause igt_runner to slow down by a factor of 4
when using the -keach option.

Transient leaks are a common phenomenon but are mostly undetected by
the current version of this library. A typical transient leak occurs
when pointers are reused, such as in linked lists. For example, if 10
calls to kmalloc are made, storing the address in the same variable
and freeing only the final allocation, the previous 9 allocations
become transient leaks. These leaks will go undetected unless the
kernel thread performs continuous scanning.

To enable continuous scanning:
 # echo scan=1 > /sys/kernel/debug/kmemleak

This configures the kmemleak kernel thread to scan the memory
continuously with 1 second pauses. While this may increase the
likelihood of detecting transient leaks, it can significantly impact
system performance. In most cases, the igt_runner slowdown remains in
the 4x range, but the kernel thread will consume 100% of a CPU core,
as observed with tools like top.

When using scan=1, transient leaks that exist during an active scan
will be detected. However, detection remains non-deterministic due
to timing. It is recommended to reset the scan interval to the
default value of 600 seconds after completing your tests.

v3:
 - Removed '<' from the end of one of the email addresses in the cc list
 - Removed email addresses that no longer exist

v2:
 - Pass igt_kmemleak_sync as a function variable to igt_kmemleak instead of
   keeping it stored as a global variable
 - igt_kmemleak_found_leaks(): Remove call to fseek() after close()
 - igt_kmemleak_write(): Increase retry counter when writing 0 bytes
 - igt_kmemleak_write(): change type to bool
 - Unit Testing: Move the call to igt_kmemleak_init() to a fixture.
 - igt_kmemleak_append_to(): Add brackets to the if statement for improved
   readability

CC: Rodrigo.Siqueira at amd.com
CC: ramadevi.gandi at intel.com
CC: ryszard.knop at intel.com
CC: sameer.lattannavar at intel.com
CC: lucas.demarchi at intel.com
CC: jani.saarinen at intel.com
CC: katarzyna.piecielska at intel.com
CC: matthew.d.roper at intel.com
CC: clinton.a.taylor at intel.com
CC: jianshui.yu at intel.com
CC: balasubramani.vivekanandan at intel.com
CC: jonathan.cavitt at intel.com

Peter Senna Tschudin (2):
  lib/igt_kmemleak: library to interact with kmemleak
  runner/executor: Integrate igt_kmemleak scans

 lib/igt_kmemleak.c       | 274 +++++++++++++++++++++++++++++++++++++++
 lib/igt_kmemleak.h       |  16 +++
 lib/meson.build          |   1 +
 lib/tests/igt_kmemleak.c | 267 ++++++++++++++++++++++++++++++++++++++
 lib/tests/meson.build    |   1 +
 runner/executor.c        |  25 +++-
 runner/runner_tests.c    |  18 ++-
 runner/settings.c        |  31 ++++-
 runner/settings.h        |   2 +
 9 files changed, 630 insertions(+), 5 deletions(-)
 create mode 100644 lib/igt_kmemleak.c
 create mode 100644 lib/igt_kmemleak.h
 create mode 100644 lib/tests/igt_kmemleak.c

-- 
2.34.1