From:   Jim Cromie <jim.cromie@gmail.com>
To:     linux-kernel@vger.kernel.org, dri-devel@lists.freedesktop.org,
        amd-gfx@lists.freedesktop.org, intel-gvt-dev@lists.freedesktop.org,
        intel-gfx@lists.freedesktop.org
Cc:     jani.nikula@intel.com, ville.syrjala@linux.intel.com,
        daniel.vetter@ffwll.ch, seanpaul@chromium.org, robdclark@gmail.com,
        Jim Cromie <jim.cromie@gmail.com>
Subject: [PATCH v3 00/19] fix DRM_USE_DYNAMIC_DEBUG regression
Date:   Wed, 25 Jan 2023 13:37:24 -0700
Message-Id: <20230125203743.564009-1-jim.cromie@gmail.com>
MIME-Version: 1.0
Content-Transfer-Encoding: 8bit
Precedence: bulk

Hi everyone,

In v6.1 DRM_USE_DYNAMIC_DEBUG=y has a regression enabling drm.debug in
drivers at modprobe.

It is due to a chicken-egg problem loading modules; on `modprobe
i915`, drm is loaded 1st, and drm/parameters/debug is set.  When
drm_debug_enabled() tested __drm_debug at runtime, this just worked.

But with DRM_USE_DYNAMIC_DEBUG=y, the runtime test is replaced with a
static_key for each drm_dbg/dyndbg callsite, enabled by dyndbg's
kparam callback on __drm_debug.  So with drm.ko loaded and initialized
before the dependent modules, their debug callsites aren't yet present
to be enabled.

STATUS - v3

not quite ready.
rebased on -rc5, hopefully applies to patchwork head 
still has RFC patch -> CI_ONLY temporary, to avoid panics
boots on my amdgpu box, drm.debug=0x3ff works at boot-time
the "toggled" warning is repeatable with test_dynamic_debug*.ko
it also occurs on amdgpu, so not just artificial.
v2 is https://lore.kernel.org/lkml/20230113193016.749791-1-jim.cromie@gmail.com/

OVERVIEW

As Jani Nikula noted rather more gently, DECLARE_DYNDBG_CLASSMAP is
error-prone enough to call broken: sharing of a common classmap
required identical classmap definitions in all modules using DRM_UT_*,
which is inherently error-prone.  IOW, it muddled the K&R distinction
between a (single) definition, and multiple references.

So patches 10-13 split it into:

DYNDBG_CLASSMAP_DEFINE	used once per subsystem to define each classmap.
DYNDBG_CLASSMAP_USE	declare dependence on a DEFINEd classmap.

DYNDBG_CLASSMAP_DEFINE initializes the classmap, stores it into the
(existing) __dyndbg_classes section, and exports the struct var
(unlike DECLARE_DYNDBG_CLASSMAP).

DYNDBG_CLASSMAP_USE initializes a class-ref struct, containing the
user-module-name, and a ref to the exported classmap var.

The distinction allows separate treatment of classmaps and
classmap-refs, the latter getting additional behavior to propagate
parent's kparam settings to USEr. (forex: drm.debug to drm-drivers) 

. lookup the classmap defn being referenced, and its module
. find the module's kernel-params using the classmap
. propagate kparam vals into the prdgs in module being added.

It also makes the weird coordinated-changes-by-identical-classmaps
"feature" unnecessary.

Patch-10 splits the DECLARE macro into DEFINE & USE, and updates uses.

Patch-11 is the core of it; the separate treatment begins in
ddebug_add_module().  It calls ddebug_attach_module_classes(1) to
handle class-defns; this adds ddebug_attach_client_module_classes(2)
to handle class-refs, as they are found while modprobing drm
drivers. (2) calls ddebug_apply_parents_params(3) on each USEr's
referred classmap definition.

(3) scans kernel-params owned by the module DEFINEing the classmap,
either builtin or loadable, calls ddebug_match_apply_kparam(4) on each.

(4) looks for kparams which are wired to dyndbg's param-ops.  Those
params have a struct ddebug_class_param attached, which has a classmap
and a ref to a state-var (__drm_debug for DRM case).  If the kparam's
classmap is the same as from (2), then apply its state-var to the
client module by calling ddebug_apply_class_bitmap().

Patch-12 cleans up DYNDBG_CLASSMAP_USE, dropping now unneeded args.

Patch-13 improves DYNDBG_CLASSMAP_DEFINE, by accepting DRM_UT_*
symbols directly, not "DRM_UT_*" (their strings).  It adds new
include/linux/map.h to support this.

Patches 1-9 are prep, refactor, cleanup, tighten interfaces

Patches 15-18 extend test_dynamic_debug to recreate DRM's multi-module
regression; it builds both test_dynamic_debug.ko and _submod.ko, with
an ifdef to _DEFINE in the main module, and _USE in the submod.  This
gives both modules identical set of prdbgs, which is helpful for
comparing results.

here it is, working properly:

doing class DRM_UT_CORE -p
[ 9904.961750] dyndbg: read 21 bytes from userspace
[ 9904.962286] dyndbg: query 0: "class DRM_UT_CORE -p" mod:*
[ 9904.962848] dyndbg: split into words: "class" "DRM_UT_CORE" "-p"
[ 9904.963444] dyndbg: op='-' flags=0x0 maskp=0xfffffffe
[ 9904.963945] dyndbg: parsed: func="" file="" module="" format="" lineno=0-0 class=DRM_UT_CORE
[ 9904.964781] dyndbg: good-class: drm.DRM_UT_CORE  module:drm nd:302 nc:1 nu:0
[ 9904.966411] dyndbg: class-ref: drm_kms_helper.DRM_UT_CORE  module:drm_kms_helper nd:95 nc:0 nu:1
[ 9904.967265] dyndbg: class-ref: drm_display_helper.DRM_UT_CORE  module:drm_display_helper nd:150 nc:0 nu:1
[ 9904.968349] dyndbg: class-ref: i915.DRM_UT_CORE  module:i915 nd:1659 nc:0 nu:1
[ 9904.969801] dyndbg: class-ref: amdgpu.DRM_UT_CORE  module:amdgpu nd:4425 nc:0 nu:1
[ 9904.977079] dyndbg: class-ref: nouveau.DRM_UT_CORE  module:nouveau nd:103 nc:0 nu:1
[ 9904.977830] dyndbg: processed 1 queries, with 507 matches, 0 errs
doing class DRM_UT_DRIVER +p
[ 9906.151761] dyndbg: read 23 bytes from userspace
[ 9906.152241] dyndbg: query 0: "class DRM_UT_DRIVER +p" mod:*
[ 9906.152793] dyndbg: split into words: "class" "DRM_UT_DRIVER" "+p"
[ 9906.153388] dyndbg: op='+' flags=0x1 maskp=0xffffffff
[ 9906.153896] dyndbg: parsed: func="" file="" module="" format="" lineno=0-0 class=DRM_UT_DRIVER
[ 9906.154746] dyndbg: good-class: drm.DRM_UT_DRIVER  module:drm nd:302 nc:1 nu:0
[ 9906.155433] dyndbg: class-ref: drm_kms_helper.DRM_UT_DRIVER  module:drm_kms_helper nd:95 nc:0 nu:1
[ 9906.156267] dyndbg: class-ref: drm_display_helper.DRM_UT_DRIVER  module:drm_display_helper nd:150 nc:0 nu:1
[ 9906.157365] dyndbg: class-ref: i915.DRM_UT_DRIVER  module:i915 nd:1659 nc:0 nu:1
[ 9906.163848] dyndbg: class-ref: amdgpu.DRM_UT_DRIVER  module:amdgpu nd:4425 nc:0 nu:1
[ 9906.178963] dyndbg: class-ref: nouveau.DRM_UT_DRIVER  module:nouveau nd:103 nc:0 nu:1
[ 9906.179934] dyndbg: processed 1 queries, with 1286 matches, 0 errs


Patch-19 is a *workaround* for a panic: __jump_label_patch can "crash
the box" when the jump-entry is in the wrong state.  The current code
makes no distinction between a well-formed "toggled" state and an
"insane" state.  Not for keeps.

It fixes mis-initialization problems like this:

[ 1594.032504] dyndbg: query 0: "class D2_DRIVER -p" mod:*
[ 1594.032823] dyndbg: split into words: "class" "D2_DRIVER" "-p"
[ 1594.033183] dyndbg: op='-' flags=0x0 maskp=0xfffffffe
[ 1594.033507] dyndbg: parsed: func="" file="" module="" format="" lineno=0-0 class=D2_DRIVER
[ 1594.034014] dyndbg: good-class: test_dynamic_debug.D2_DRIVER  module:test_dynamic_debug nd:32 nc:4 nu:0
[ 1594.034695] dyndbg: changed lib/test_dynamic_debug.c:156 [test_dynamic_debug]do_cats p => _
[ 1594.035304] dyndbg: class-ref: test_dynamic_debug_submod.D2_DRIVER  module:test_dynamic_debug_submod nd:32 nc:0 nu:4
[ 1594.036052] jump_label: found toggled op at do_cats+0x16/0x180 [test_dynamic_debug_submod] [00000000ff2582ac] (0f 1f 44 00 00 != e9 e1 00 00 00)) size:5 type:0
[ 1594.037036] dyndbg: changed lib/test_dynamic_debug.c:156 [test_dynamic_debug_submod]do_cats p => _
[ 1594.037604] dyndbg: processed 1 queries, with 2 matches, 0 errs
[ 1594.037968] dyndbg: bit_1: 2 matches on class: D2_DRIVER -> 0x0

These errors are reliably reproduced by a shell-func which modprobes
(with the right args) the test mod & submod.ko (in the commit message).

So this isnt really ready for inclusion, but Id like to send the whole
set to the CI-gym for a workout.  The RFC/for-TESTING patch will
mitigate panics, and still be detectable.

Besides, Murphys law requires I publish some error before I can make progress.


Jim Cromie (19):
  test-dyndbg: fixup CLASSMAP usage error
  test-dyndbg: show that DEBUG enables prdbgs at compiletime
  dyndbg: replace classmap list with a vector
  dyndbg: make ddebug_apply_class_bitmap more selective
  dyndbg: split param_set_dyndbg_classes to inner/outer fns
  dyndbg: drop NUM_TYPE_ARRAY
  dyndbg: reduce verbose/debug clutter
  dyndbg: tighten ddebug_class_name() 1st arg
  dyndbg: constify ddebug_apply_class_bitmap args
  dyndbg-API: split DECLARE_(DYNDBG_CLASSMAP) to $1(_DEFINE|_USE)
  dyndbg-API: specialize DYNDBG_CLASSMAP_(DEFINE|USE)
  dyndbg-API: DYNDBG_CLASSMAP_USE drop extra args
  dyndbg-API: DYNDBG_CLASSMAP_DEFINE() improvements
  drm_print: fix stale macro-name in comment
  test-dyndbg: build test_dynamic_debug_submod
  test-dyndbg: rename DD_SYS_WRAP to DYNDBG_CLASSMAP_PARAM
  test-dyndbg: disable WIP dyndbg-trace params
  test-dyndbg: tune sub-module behavior
  jump_label: RFC / temporary for CI - tolerate toggled state

 arch/x86/kernel/jump_label.c            |  24 ++-
 drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c |  14 +-
 drivers/gpu/drm/display/drm_dp_helper.c |  14 +-
 drivers/gpu/drm/drm_crtc_helper.c       |  14 +-
 drivers/gpu/drm/drm_print.c             |  22 +-
 drivers/gpu/drm/i915/i915_params.c      |  14 +-
 drivers/gpu/drm/nouveau/nouveau_drm.c   |  14 +-
 include/asm-generic/vmlinux.lds.h       |   1 +
 include/drm/drm_print.h                 |   6 +-
 include/linux/dynamic_debug.h           |  57 ++++--
 include/linux/map.h                     |  55 +++++
 kernel/module/main.c                    |   3 +
 lib/Makefile                            |   3 +-
 lib/dynamic_debug.c                     | 258 ++++++++++++++++++------
 lib/test_dynamic_debug.c                | 118 +++++++----
 lib/test_dynamic_debug_submod.c         |  10 +
 16 files changed, 437 insertions(+), 190 deletions(-)
 create mode 100644 include/linux/map.h
 create mode 100644 lib/test_dynamic_debug_submod.c

-- 
2.39.1