Re: [PATCH 2/5] dyndbg: add class_id field and query support
On Mon, Mar 14, 2022 at 3:30 PM Jason Baron wrote: > > > > On 3/11/22 20:06, jim.cro...@gmail.com wrote: > > On Fri, Mar 11, 2022 at 12:06 PM Jason Baron wrote: > >> > >> > >> > >> On 3/10/22 23:47, Jim Cromie wrote: > >>> > >>> With the patch, one can enable current/unclassed callsites by: > >>> > >>> #> echo drm class 15 +p > /proc/dynamic_debug/control > >>> > >> > >> To me, this is hard to read, what the heck is '15'? I have to go look it > >> up in the control file and it's not descriptive. I think that using > >> classes/categories makes sense but I'm wondering if it can be a bit more > >> user friendly? Perhaps, we can pass an array of strings that is indexed > >> by the class id to each pr_debug() site that wants to use class. So > >> something like: hi Jason, Im now in basically full agreement with you 1. .class_id is a "global" space, in that all callsites have one. 2.0-15 is an exceedingly small range for a global space Fix that by 3. make it private (by removing "class N" parsing) now nobody can do echo module * class N +p >control 4. add private/per-module "string" -> class_id map each module will have to declare the class-strings they use/accept opt-in - want coordinated / shared names for DRM_UT_KMS etc. 5. validate input using the known class_string -> class_id then, this is knowably right or wrong, depending on DRM_FOO: echo module drm class DRM_FOO +p > control it also means that echo class DRM_UT_KMS +p >control is both wellformed and minimal; any module that has DRM_UT_KMS defined will know which class_id *they* have mapped it to. theres no global "DRM_UT_KMS" to be abused. So Ive been working towards that.. Heres my current biggest roadblock DEFINE_DYNAMIC_DEBUG_CLASSBITS creates the class-strings[] array declaratively, at compile-time. This array is attached to the kernel-param.args, so it can be used by the callbacks (to construct the command) But its not obviously available from outside the sysfs knob that its attached to, as is needed to apply command >control generally. If I can attach the class-strings[] to struct ddebug_table, then ddebug_change() can use it to validate >control input. So, how to attach ? its got to work for both builtin & loadable modules. (which precludes something in struct module ?) I looked for a kernel_param_register callback, came up empty. Trying to add it feels invasive / imposing. > > > > If that works, its cuz DEFINE_DYNAMIC_DEBUG_CLASSBITS() > > plucks class symbols out of its __VA_ARGS__, and #stringifes them. > > So that macro could then build the 1-per-module static constant string array > > and (only) the callbacks would be able to use it. > > So Ive been tinkering hard on this macro, since its pretty central to the interface defn. heres some choices: this is what Ive been working towards. using enum symbols directly like this associates them by grep, in contrast, class-strings are mealy-mouthed, milquetoast. DEFINE_DYNAMIC_DEBUG_CLASSBITS(debug, __drm_debug, "p", "enable drm.debug categories - 1 bit per category", DRM_UT_CORE, DRM_UT_DRIVER, DRM_UT_KMS, DRM_UT_PRIME, DRM_UT_ATOMIC, DRM_UT_VBL, DRM_UT_STATE, DRM_UT_LEASE, DRM_UT_DP, DRM_UT_DRMRES); I found a slick MAP ( ) macro to do this: #define DEFINE_DYNAMIC_DEBUG_CLASSBITS(fsname, _var, _flgs, desc, ...) \ MODULE_PARM_DESC(fsname, desc); \ static struct dyndbg_classbits_param ddcats_##_var = { \ .bits = &(_var), \ .flags = _flgs, \ .classes = { __VA_ARGS__, _DPRINTK_CLASS_DFLT }, \ .class_names = { mgMAP(__stringify, sCOMMA, \ __VA_ARGS__, _DPRINTK_CLASS_DFLT) } \ }; \ module_param_cb(fsname, ¶m_ops_dyndbg_classbits, \ &ddcats_##_var, 0644) __VA_ARGS__ is used 2x .class_names is available for validating command >control As much as I like the above, the MAP macro has a longer, more risky path to the kernel so a more modest alternative: module user defines class-strings in interface, but they *must* align manually with the enum values they correspond to; the order determines the bit-pos in the sysfs node, since the interface no longer provides the enum values themselves. DEFINE_DYNAMIC_DEBUG_CLASS_STRINGS(debug, __drm_debug, "p", "enable drm.debug categories - 1 bit per category", "DRM_UT_CORE", "DRM_UT_DRIVER", "DRM_UT_KMS", different name allows CLASSBITs or similar in future, if MAP works out. class-strings are completely defined by users, drm can drop UT naming TLDR: FWIW iSTM that the same macro will support the coordinated use of class-strings across multiple modules. drm_print.c - natural home for exposed sysfs node amdgpu/, drm_helper/ i915/ nouveau/ will all need a DEFINE_DD added, so that ddebug_change() can allow those .class_ids to be controlled. sysfs perm inits can disable their nodes, since theyre coordinated. > > > > Ok,
Re: [PATCH 2/5] dyndbg: add class_id field and query support
On 3/11/22 20:06, jim.cro...@gmail.com wrote: > On Fri, Mar 11, 2022 at 12:06 PM Jason Baron wrote: >> >> >> >> On 3/10/22 23:47, Jim Cromie wrote: >>> DRM defines/uses 10 enum drm_debug_category's to create exclusive >>> classes of debug messages. To support this directly in dynamic-debug, >>> add the following: >>> >>> - struct _ddebug.class_id:4 - 4 bits is enough >>> - define _DPRINTK_SITE_UNCLASSED 15 - see below >>> >>> and the query support: >>> - struct _ddebug_query.class_id >>> - ddebug_parse_query - looks for "class" keyword, then calls.. >>> - parse_class - accepts uint: 0-15, sets query.class_id and marker >>> - vpr_info_dq - displays new field >>> - ddebug_proc_show- append column with "cls:%d" if !UNCLASSED >>> >>> With the patch, one can enable current/unclassed callsites by: >>> >>> #> echo drm class 15 +p > /proc/dynamic_debug/control >>> >> >> To me, this is hard to read, what the heck is '15'? I have to go look it >> up in the control file and it's not descriptive. I think that using >> classes/categories makes sense but I'm wondering if it can be a bit more >> user friendly? Perhaps, we can pass an array of strings that is indexed >> by the class id to each pr_debug() site that wants to use class. So >> something like: >> > > Im not at all averse to nice names, but as something added on. > > 1st, the interface to make friendlier is DRM's > > echo 0x04 > /sys/module/drm/parameters/debug # which category ? > > parm: debug:Enable debug output, where each bit enables a > debug category. > Bit 0 (0x01) will enable CORE messages (drm core code) > Bit 1 (0x02) will enable DRIVER messages (drm controller code) > Bit 2 (0x04) will enable KMS messages (modesetting code) > > echo DRM_UT_DRIVER,DRM_UT_KMS > /sys/module/drm/parameters/debug # > now its pretty clear > > If that works, its cuz DEFINE_DYNAMIC_DEBUG_CLASSBITS() > plucks class symbols out of its __VA_ARGS__, and #stringifes them. > So that macro could then build the 1-per-module static constant string array > and (only) the callbacks would be able to use it. > > from there, it shouldnt be hard to ref that from the module's ddebug_table, > so parse_query could validate class args against the module given (or > fail if none given) > > Speaking strictly, theres a gap: >echo module * class DRM_UT_KMS +p > control > is nonsense for * other than drm + drivers, > but its fair/safe to just disallow wildcards on modname for this purpose. > > it does however imply that module foo must exist for FOO_CAT_1 to be usable. > thats not currently the case: > bash-5.1# echo module foo +p > /proc/dynamic_debug/control > [ 15.403749] dyndbg: read 14 bytes from userspace > [ 15.405413] dyndbg: query 0: "module foo +p" mod:* > [ 15.406486] dyndbg: split into words: "module" "foo" "+p" > [ 15.407070] dyndbg: op='+' > [ 15.407388] dyndbg: flags=0x1 > [ 15.407809] dyndbg: *flagsp=0x1 *maskp=0x > [ 15.408300] dyndbg: parsed: func="" file="" module="foo" format="" > lineno=0-0 class=15 > [ 15.409151] dyndbg: no matches for query > [ 15.409591] dyndbg: no-match: func="" file="" module="foo" > format="" lineno=0-0 class=15 > [ 15.410524] dyndbg: processed 1 queries, with 0 matches, 0 errs > bash-5.1# > > ISTM we can keep that "0 errs" response for that case, but still reject this: > >echo module foo class FOO_NOT_HERE +p > /proc/dynamic_debug/control > > Ok, yeah so I guess I'm thinking about the 'class' more as global namespace, so that for example, it could span modules, or be specific to certain modules. I'm also thinking of it as a 'string' which is maybe hierarchical, so that it's clear what sub-system, or sub-sub-system it belongs to. So for DRM for example, it could be a string like "DRM:CORE". The index num I think is still helpful for implementation so we don't have to store a pointer size, but I don't think it's really exposed (except perhaps in module params b/c drm is doing that already?). >> enum levels { >> LOW, >> MEDIUM, >> HIGH >> }; > > I want to steer clear of "level" anything, > since 2>1 implies non independence of the categories > > Agreed, that was a bad example on my part. I've put together a rough patch on top of your series, to make my thinking hopefully clear. I also think that the module param callback thing could be implemented in this 'global' space with the 0-14 low numbers, by adding some sort of offset into the table. IE you would add the low number + the offset to get the 'string' to add to the ddebug query. I commented it out in my patch below b/c I didn't implement that part. Thoughts? diff --git a/include/linux/dynamic_debug.h b/include/linux/dynamic_debug.h index 664bb83778d2..b0bc1b536d54 100644 --- a/include/linux/dynamic_debug.h +++ b/include/linux/dynamic_debug.h @@ -8,6 +8,14 @@ #include +enum ddebug_categories { +FOO_BAR = 0, +DRM_A = 1, +DRM_B = 2, +DRM_C = 3
Re: [PATCH 2/5] dyndbg: add class_id field and query support
On 3/10/22 23:47, Jim Cromie wrote: > DRM defines/uses 10 enum drm_debug_category's to create exclusive > classes of debug messages. To support this directly in dynamic-debug, > add the following: > > - struct _ddebug.class_id:4 - 4 bits is enough > - define _DPRINTK_SITE_UNCLASSED 15 - see below > > and the query support: > - struct _ddebug_query.class_id > - ddebug_parse_query - looks for "class" keyword, then calls.. > - parse_class - accepts uint: 0-15, sets query.class_id and marker > - vpr_info_dq - displays new field > - ddebug_proc_show- append column with "cls:%d" if !UNCLASSED > > With the patch, one can enable current/unclassed callsites by: > > #> echo drm class 15 +p > /proc/dynamic_debug/control > To me, this is hard to read, what the heck is '15'? I have to go look it up in the control file and it's not descriptive. I think that using classes/categories makes sense but I'm wondering if it can be a bit more user friendly? Perhaps, we can pass an array of strings that is indexed by the class id to each pr_debug() site that wants to use class. So something like: enum levels { LOW, MEDIUM, HIGH }; static const char * const level_to_strings[] = { [LOW] = "low", [MEDIUM] = "medium", [HIGH] = "high", }; And then you'd have a wrapper macros in your driver: #define module_foo_pr_debug_class(level, fmt, args...) pr_debug_class(level, level_to_strings, fmt, args); Such that call sites look like: module_foo_pr_debug_class(LOW, fmt, args...); Such that you're not always passing the strings array around. Now, this does mean another pointer for struct _ddebug and most wouldn't have it. Maybe we could just add another linker section for these so as to save space. > parse_class() accepts 0 .. _DPRINTK_SITE_UNCLASSED, which allows the >> control interface to explicitly manipulate unclassed callsites. > > After parsing keywords, ddebug_parse_query() sets .class_id=15, iff it > wasnt explicitly set. This allows future classed/categorized > callsites to be untouched by legacy (class unaware) queries. > > DEFINE_DYNAMIC_DEBUG_METADATA gets _CLS(cls,) suffix and arg, and > initializes the new .class_id=cls field. The old name gets the default. > > Then, these _CLS(cls,...) modifications are repeated up through the > stack of *dynamic_func_call* macros that use the METADATA initializer, > so as to actually supply the category into it. > > NOTES: > > _DPRINTK_SITE_UNCLASSED: this symbol is used to initialize all > existing/unclassed pr-debug callsites. Normally, the default would be > zero, but DRM_UT_CORE "uses" that value, in the sense that 0 is > exposed as a bit position in drm.debug. Using 15 allows identity > mapping from category to class, avoiding fiddly offsets. > > Default .class_id = 15 means that ``echo +p > control`` no longer > toggles ALL the callsites, only the unclassed ones. This was only > useful for static-branch toggle load testing anyway. > I think that # echo +p > control should continue to work as is, why should the introduction of classes change that ? > RFC: > > The new _CLS macro flavor gets a warning from DRM/dri-devel's CI, > but not from checkpatch (on this subject). > > a8f6c71f283e dyndbg: add class_id field and query support > -:141: CHECK:MACRO_ARG_REUSE: Macro argument reuse 'id' - possible > side-effects? > +#define __dynamic_func_call_cls(id, cls, fmt, func, ...) do {\ > + DEFINE_DYNAMIC_DEBUG_METADATA_CLS(id, cls, fmt);\ > + if (DYNAMIC_DEBUG_BRANCH(id)) \ > + func(&id, ##__VA_ARGS__); \ > } while (0) > > I couldn't fix it with a ``typeof(id) _id = id`` construct. I haven't > seen the warning myself, on the _CLS extended macro, nor the original. > > CC: Rasmus Villemoes > Signed-off-by: Jim Cromie > --- > .../admin-guide/dynamic-debug-howto.rst | 7 +++ > include/linux/dynamic_debug.h | 54 ++- > lib/dynamic_debug.c | 48 ++--- > 3 files changed, 88 insertions(+), 21 deletions(-) > > diff --git a/Documentation/admin-guide/dynamic-debug-howto.rst > b/Documentation/admin-guide/dynamic-debug-howto.rst > index a89cfa083155..8ef8d7dcd140 100644 > --- a/Documentation/admin-guide/dynamic-debug-howto.rst > +++ b/Documentation/admin-guide/dynamic-debug-howto.rst > @@ -35,6 +35,7 @@ Dynamic debug has even more useful features: > - line number (including ranges of line numbers) > - module name > - format string > + - class number:0-15 > > * Provides a debugfs control file: ``/dynamic_debug/control`` > which can be read to display the complete list of known debug > @@ -143,6 +144,7 @@ against. Possible keywords are::: >'module' string | >'format' string | >'line' line-range > + 'class' integer:[0-15] > >line-range ::= linen
Re: [PATCH 2/5] dyndbg: add class_id field and query support
On Fri, Mar 11, 2022 at 12:06 PM Jason Baron wrote: > > > > On 3/10/22 23:47, Jim Cromie wrote: > > DRM defines/uses 10 enum drm_debug_category's to create exclusive > > classes of debug messages. To support this directly in dynamic-debug, > > add the following: > > > > - struct _ddebug.class_id:4 - 4 bits is enough > > - define _DPRINTK_SITE_UNCLASSED 15 - see below > > > > and the query support: > > - struct _ddebug_query.class_id > > - ddebug_parse_query - looks for "class" keyword, then calls.. > > - parse_class - accepts uint: 0-15, sets query.class_id and marker > > - vpr_info_dq - displays new field > > - ddebug_proc_show- append column with "cls:%d" if !UNCLASSED > > > > With the patch, one can enable current/unclassed callsites by: > > > > #> echo drm class 15 +p > /proc/dynamic_debug/control > > > > To me, this is hard to read, what the heck is '15'? I have to go look it > up in the control file and it's not descriptive. I think that using > classes/categories makes sense but I'm wondering if it can be a bit more > user friendly? Perhaps, we can pass an array of strings that is indexed > by the class id to each pr_debug() site that wants to use class. So > something like: > Im not at all averse to nice names, but as something added on. 1st, the interface to make friendlier is DRM's echo 0x04 > /sys/module/drm/parameters/debug # which category ? parm: debug:Enable debug output, where each bit enables a debug category. Bit 0 (0x01) will enable CORE messages (drm core code) Bit 1 (0x02) will enable DRIVER messages (drm controller code) Bit 2 (0x04) will enable KMS messages (modesetting code) echo DRM_UT_DRIVER,DRM_UT_KMS > /sys/module/drm/parameters/debug # now its pretty clear If that works, its cuz DEFINE_DYNAMIC_DEBUG_CLASSBITS() plucks class symbols out of its __VA_ARGS__, and #stringifes them. So that macro could then build the 1-per-module static constant string array and (only) the callbacks would be able to use it. from there, it shouldnt be hard to ref that from the module's ddebug_table, so parse_query could validate class args against the module given (or fail if none given) Speaking strictly, theres a gap: echo module * class DRM_UT_KMS +p > control is nonsense for * other than drm + drivers, but its fair/safe to just disallow wildcards on modname for this purpose. it does however imply that module foo must exist for FOO_CAT_1 to be usable. thats not currently the case: bash-5.1# echo module foo +p > /proc/dynamic_debug/control [ 15.403749] dyndbg: read 14 bytes from userspace [ 15.405413] dyndbg: query 0: "module foo +p" mod:* [ 15.406486] dyndbg: split into words: "module" "foo" "+p" [ 15.407070] dyndbg: op='+' [ 15.407388] dyndbg: flags=0x1 [ 15.407809] dyndbg: *flagsp=0x1 *maskp=0x [ 15.408300] dyndbg: parsed: func="" file="" module="foo" format="" lineno=0-0 class=15 [ 15.409151] dyndbg: no matches for query [ 15.409591] dyndbg: no-match: func="" file="" module="foo" format="" lineno=0-0 class=15 [ 15.410524] dyndbg: processed 1 queries, with 0 matches, 0 errs bash-5.1# ISTM we can keep that "0 errs" response for that case, but still reject this: echo module foo class FOO_NOT_HERE +p > /proc/dynamic_debug/control > enum levels { > LOW, > MEDIUM, > HIGH > }; I want to steer clear of "level" anything, since 2>1 implies non independence of the categories > > static const char * const level_to_strings[] = { > [LOW] = "low", > [MEDIUM] = "medium", > [HIGH] = "high", > }; > > And then you'd have a wrapper macros in your driver: > > #define module_foo_pr_debug_class(level, fmt, args...) > pr_debug_class(level, level_to_strings, fmt, args); > > Such that call sites look like: > > module_foo_pr_debug_class(LOW, fmt, args...); > That macro, minus the "module_foo_" prefix, could go into dynamic_debug.h I didnt do that, for 2 reasons: DRM didnt need it - it had an enum var, and a set of macros to encapsulate the categories. - the "prototype" looks like this might be ok: define LOW "low" pr_debug_class(LOW, "mumble about something %p %p\n", foo, bar) ok thats a stretch, but... Basically, I didnt want to deal with creating a new interface. KIS > Such that you're not always passing the strings array around. Now, this > does mean another pointer for struct _ddebug and most wouldn't have it. > Maybe we could just add another linker section for these so as to save > space. >From a space consideration, adding DRM users means adding thousands of callsites ~2k for i915, ~4k for amdgpu Also increasing per/callsite memory seems counterproductive. see DEFINE_DYNAMIC_DEBUG_CLASSBITS for the way forward. > > parse_class() accepts 0 .. _DPRINTK_SITE_UNCLASSED, which allows the > >> control interface to explicitly manipulate unclassed callsites. > > > > After parsing keywords, ddebug_parse_query() sets .class_id=15, iff it > > wasnt explicitly se