Re: [Intel-gfx] [PATCH v3 3/5] drm/print: RFC add choice to use dynamic debug in drm-debug
From: Daniel Vetter Hi Daniel, everyone, Ive substantially updated this patchset, and I thought it useful to reply here. Im noting the biggest changes in context below, trimming heavily otherwize. Also, Ive lost the msg in my gmail-cloud, so this lacks the usual "Daniel said: " attribution, and the "> " marks. Ive prefixed "< " to 1st line of my replies where it helps. The current patchset is here: https://patchwork.freedesktop.org/series/100290/ https://lore.kernel.org/lkml/20220217034829.64395-1-jim.cro...@gmail.com/ Its "failing" patchwork CI cuz of a macro warning in dyndbg, which would be nice for CI to "pass" because its out of DRM tree, and subject to a separate review process, and in the meantime, it would be nice to see it under test. Going item by item: On Wed, Jul 14, 2021 at 11:51:36AM -0600, Jim Cromie wrote: > drm's debug system uses distinct categories of debug messages, encoded > in an enum (DRM_UT_), enum is now renumbered to 0-15 (fits in 4 bits) drm_debug_enabled() does BIT(cat) 15 is unclassed/reserved > Dynamic debug has no concept of category, but we can map the DRM_UT_* now it has "class" keyword, and class_id:4 field. #> echo module drm class 1 +p > control # 1=DRM_UT_KMS iirc This interface is unused yet, DRM is its A-1 customer, are you shopping ? Since its unused, it cannot break anything not using it :-) > CONFIG_DRM_USE_DYNAMIC_DEBUG enables this, and is available if still true. > The indirection/switchover is layered into the macro scheme: Heres where the biggest DRM centric changes (vs old) are: The cDRM_UT_* => prefix-string mapping is gone; while it worked, it made the literal format strings config dependent, and was more complicated to explain. Fitting category into class_id is much cleaner. old way replaced drm*dbg with pr_debug (at surface level) new way adapts drm*dbg to reuse the class-Factory macro under pr_debug. This loses pr_debug's log-decorating feature, but DRM has its own conventions, and extra decorations are unlikely to really add anything. OTOH, drm could re-use those flags to optionalize some of its conventions. > 0. A new callback on drm.debug which calls dynamic_debug_exec_queries now in dyndbg, where it belonged. now just uses class query inside. This is really awesome. For merging I think we need to discuss with dyn debug folks whether they're all ok with this, but it's exported already should should be fine. < Your (fresh) endorsement should help :-) Particularly, the new dyndbg features need a user. > > 1. A "converted" or "classy" DRM_UT_* map > repeating, 2nd map is gone, DRM_UT_* is merely renumbered. >DRM_UT_* are unchanged, since theyre used in drm_debug_enabled() >and elsewhere. I think for the production version of these we need to retire/deprecate them, at least for drm core. Otherwise you have an annoying mismatch between drm.debug module option and dyn debug. < I think this isnt relevant now, since symbols are preserved. Also, the __drm_debug var is used directly by the CLASSBITS macro, (and therefore the callbacks), so /sys/module/drm/parameters/debug is preserved at an ABI-ish level. (__drm_debug now ulong, to work with BIT) > > 2. drm_dev_dbg & drm_debug are renamed (prefixed with '_') > >old names are now macros, calling either: > legacy: -> to renamed fn > enabled: -> dev_dbg & pr_debug, with cDRM-prefix # format. > For merging, can we pull out the renames and reorgs from this patch, and then maybe also the reorder the next patch in your series here to be before the dyn debug stuff? < the above adaptation is re-adapted to use dyndbg's new class-Factory macro, other stuff is gone. > Signed-off-by: Jim Cromie > --- > drivers/gpu/drm/Kconfig | 13 + > drivers/gpu/drm/drm_print.c | 75 -- > include/drm/drm_print.h | 102 ++-- > 3 files changed, 158 insertions(+), 32 deletions(-) update - drm parts are slightly smaller now :-) [jimc@frodo wk-next]$ git diff --stat master Documentation/admin-guide/dynamic-debug-howto.rst | 7 + drivers/gpu/drm/Kconfig | 12 drivers/gpu/drm/Makefile | 2 ++ drivers/gpu/drm/drm_print.c | 56 ++ include/drm/drm_print.h | 80 + include/linux/dynamic_debug.h | 113 + lib/dynamic_debug.c | 140 + 7 files changed, 323 insertions(+), 87 deletions(-) I really like this, I think you can drop the RFC. A few more things that I think we need: - An overview kerneldoc section which explains the interfaces and how it all works together. Essentially your commit message with some light markup to make it look good. < not sure anything worth documenting has survived. - I think it would
Re: [Intel-gfx] [PATCH v3 3/5] drm/print: RFC add choice to use dynamic debug in drm-debug
On Tue, Jul 27, 2021 at 10:03 AM Sean Paul wrote:
>
> On Thu, Jul 22, 2021 at 11:20 AM Sean Paul wrote:
> >
>
> Reply-all fail. Adding everyone else back to my response.
>
> > On Tue, Jul 20, 2021 at 03:29:34PM +0200, Daniel Vetter wrote:
> > > On Wed, Jul 14, 2021 at 11:51:36AM -0600, Jim Cromie wrote:
> > > > drm's debug system uses distinct categories of debug messages, encoded
> > > > in an enum (DRM_UT_), which are mapped to bits in drm.debug.
> > > > drm_debug_enabled() does a lot of unlikely bit-mask checks on
> > > > drm.debug; we can use dynamic debug instead, and get all that
> > > > static_key/jump_label goodness.
> >
> > Hi Jim,
> > Thanks for your patches! Daniel pointed me at them in response to my
> > drm_trace
> > patchset (https://patchwork.freedesktop.org/series/78133/). I'd love to get
> > your
> > input on it. I think the 2 sets are mostly compatible, we'd just need to
> > keep
> > drm_dev_dbg and do the CONFIG check in the function beside the trace_enabled
> > checks.
> >
> > > >
> > > > Dynamic debug has no concept of category, but we can map the DRM_UT_*
> > > > to a set of distinct prefixes; "drm:core:", "drm:kms:" etc, and
> > > > prepend them to the given formats.
> > > >
> > > > Then we can use:
> > > > `echo module drm format ^drm:core: +p > control`
> > > >
> > > > to enable every such "prefixed" pr_debug with one query. This new
> > > > prefix changes pr_debug's output, so is user visible, but it seems
> > > > unlikely to cause trouble for log watchers; they're not relying on the
> > > > absence of class prefix strings.
> > > >
> > > > This conversion yields ~2100 new callsites on my i7/i915 laptop:
> > > >
> > > > dyndbg: 195 debug prints in module drm_kms_helper
> > > > dyndbg: 298 debug prints in module drm
> > > > dyndbg: 1630 debug prints in module i915
> > > >
> > > > CONFIG_DRM_USE_DYNAMIC_DEBUG enables this, and is available if
> > > > CONFIG_DYNAMIC_DEBUG or CONFIG_DYNAMIC_DEBUG_CORE is chosen, and if
> > > > CONFIG_JUMP_LABEL is enabled; this because its required to get the
> > > > promised optimizations.
> > > >
> > > > The indirection/switchover is layered into the macro scheme:
> > > >
> > > > 0. A new callback on drm.debug which calls dynamic_debug_exec_queries
> > > >to map those bits to specific query/commands
> > > >dynamic_debug_exec_queries("format ^drm:kms: +p", "drm*");
> > > >here for POC, this should be in dynamic_debug.c
> > > >with a MODULE_PARAM_DEBUG_BITMAP(__drm_debug, { "prefix-1", "desc-1"
> > > > }+)
> > >
> > > This is really awesome.
> >
> >
> > Agreed, this is a very clever way of merging the 2 worlds!
> >
> >
> > > For merging I think we need to discuss with dyn
> > > debug folks whether they're all ok with this, but it's exported already
> > > should should be fine.
> >
> > I wonder if this is a good time to reconsider our drm categories. IMO
> > they're
> > overly broad and it's hard to get the right information without subscribing
> > to
> > the firehose. It seems like dyndbg might be a good opportunity to unlock
> > subcategories of log messages.
> >
> > More concretely, on CrOS we can't subscribe to atomic or state categories
> > since
> > they're too noisy. However if there was a "fail" subcategory which dumped
> > state/atomic logs on check failures, that would be really compelling.
> > Something
> > like:
> >
> > drm:atomic:fail vs. drm:atomic
> >
YES
> > Both would be picked up if (drm.debug & DRM_DBG_ATOMIC), however it would
> > allow
> > dyndbg-aware clients to get better logs without having a huge table of
> > individual log signatures.
> >
> > I'm not sure how tightly we'd want to control the subcategories. It could be
> > strict like the categories spelled out in drm_print.h, or an open prefix
> > arg to
> > drm_dev_dbg. I suspect we'd want the former, but would want to be careful to
> > provide enough flexibility to properly
formalizing categories and subcategories is where the
practical selectivity of format ^prefix is determined.
While endless bikeshedding is a possible downside,
there are a few subtleties to note, so we can pick a harmonious 3-color scheme:
"drm:kms: " and "drm:kms:" are different (2nd, w/o trailing space,
allows subcats)
"drm:kms" is also different (includes "drm:kmsmart", whatever that would be)
"drm.kms" again. dot names anyone ?
ASIDE(s):
"drm:kms" and "drm:kms*" are different,
the latter does not work as you might reasonably expect.
This is because
commit 578b1e0701af34f9ef69daabda4431f1e8501109
Author: Changbin Du
Date: Thu Jan 23 15:54:14 2014 -0800
dynamic_debug: add wildcard support to filter files/functions/modules
Add wildcard '*'(matches zero or more characters) and '?' (matches one
character) support when qurying debug flags.
specifically left format out of the wildcarded query terms.
And that was rational since the format search was already special
(floating substring, not exact match or basename match)
and format
Re: [Intel-gfx] [PATCH v3 3/5] drm/print: RFC add choice to use dynamic debug in drm-debug
On Thu, Jul 22, 2021 at 11:20 AM Sean Paul wrote:
>
Reply-all fail. Adding everyone else back to my response.
> On Tue, Jul 20, 2021 at 03:29:34PM +0200, Daniel Vetter wrote:
> > On Wed, Jul 14, 2021 at 11:51:36AM -0600, Jim Cromie wrote:
> > > drm's debug system uses distinct categories of debug messages, encoded
> > > in an enum (DRM_UT_), which are mapped to bits in drm.debug.
> > > drm_debug_enabled() does a lot of unlikely bit-mask checks on
> > > drm.debug; we can use dynamic debug instead, and get all that
> > > static_key/jump_label goodness.
>
> Hi Jim,
> Thanks for your patches! Daniel pointed me at them in response to my drm_trace
> patchset (https://patchwork.freedesktop.org/series/78133/). I'd love to get
> your
> input on it. I think the 2 sets are mostly compatible, we'd just need to keep
> drm_dev_dbg and do the CONFIG check in the function beside the trace_enabled
> checks.
>
> > >
> > > Dynamic debug has no concept of category, but we can map the DRM_UT_*
> > > to a set of distinct prefixes; "drm:core:", "drm:kms:" etc, and
> > > prepend them to the given formats.
> > >
> > > Then we can use:
> > > `echo module drm format ^drm:core: +p > control`
> > >
> > > to enable every such "prefixed" pr_debug with one query. This new
> > > prefix changes pr_debug's output, so is user visible, but it seems
> > > unlikely to cause trouble for log watchers; they're not relying on the
> > > absence of class prefix strings.
> > >
> > > This conversion yields ~2100 new callsites on my i7/i915 laptop:
> > >
> > > dyndbg: 195 debug prints in module drm_kms_helper
> > > dyndbg: 298 debug prints in module drm
> > > dyndbg: 1630 debug prints in module i915
> > >
> > > CONFIG_DRM_USE_DYNAMIC_DEBUG enables this, and is available if
> > > CONFIG_DYNAMIC_DEBUG or CONFIG_DYNAMIC_DEBUG_CORE is chosen, and if
> > > CONFIG_JUMP_LABEL is enabled; this because its required to get the
> > > promised optimizations.
> > >
> > > The indirection/switchover is layered into the macro scheme:
> > >
> > > 0. A new callback on drm.debug which calls dynamic_debug_exec_queries
> > >to map those bits to specific query/commands
> > >dynamic_debug_exec_queries("format ^drm:kms: +p", "drm*");
> > >here for POC, this should be in dynamic_debug.c
> > >with a MODULE_PARAM_DEBUG_BITMAP(__drm_debug, { "prefix-1", "desc-1"
> > > }+)
> >
> > This is really awesome.
>
>
> Agreed, this is a very clever way of merging the 2 worlds!
>
>
> > For merging I think we need to discuss with dyn
> > debug folks whether they're all ok with this, but it's exported already
> > should should be fine.
>
> I wonder if this is a good time to reconsider our drm categories. IMO they're
> overly broad and it's hard to get the right information without subscribing to
> the firehose. It seems like dyndbg might be a good opportunity to unlock
> subcategories of log messages.
>
> More concretely, on CrOS we can't subscribe to atomic or state categories
> since
> they're too noisy. However if there was a "fail" subcategory which dumped
> state/atomic logs on check failures, that would be really compelling.
> Something
> like:
>
> drm:atomic:fail vs. drm:atomic
>
> Both would be picked up if (drm.debug & DRM_DBG_ATOMIC), however it would
> allow
> dyndbg-aware clients to get better logs without having a huge table of
> individual log signatures.
>
> I'm not sure how tightly we'd want to control the subcategories. It could be
> strict like the categories spelled out in drm_print.h, or an open prefix arg
> to
> drm_dev_dbg. I suspect we'd want the former, but would want to be careful to
> provide enough flexibility to properly
>
> Of course, none of this needs to be decided to land this initial support, it
> can
> be bolted on later easily enough (I think).
>
>
> >
> > >
> > > 1. A "converted" or "classy" DRM_UT_* map
> > >
> > >based on: DRM_UT_* ( symbol => bit-mask )
> > >named it: cDRM_UT_* ( symbol => format-class-prefix-string )
> > >
> > >So cDRM_UT_* is either:
> > >legacy: cDRM_UT_* <-- DRM_UT_* ( !CONFIG_DRM_USE_DYNAMIC_DEBUG )
> > >enabled:
> > > #define cDRM_UT_KMS"drm:kms: "
> > > #define cDRM_UT_PRIME "drm:prime: "
> > > #define cDRM_UT_ATOMIC "drm:atomic: "
> >
> > the cDRM looks a bit funny, plus I don't eve have an idea what _UT_ means
> > (and git history isn't helpful either). What about just using
> > DRM_DBG_CLASS_ as the prefix here for these indirection macros, i.e.
> > DRM_DBG_CLASS_KMS.
> >
> > Also would be really nice if we could make these a table or something, but
> > I guess with the macro magic that's not possible.
> >
> > >
> > >DRM_UT_* are unchanged, since theyre used in drm_debug_enabled()
> > >and elsewhere.
> >
> > I think for the production version of these we need to retire/deprecate
> > them, at least for drm core. Otherwise you have an annoying mismatch
> > between drm.debug module option and dyn debug.
> >
> > >
> > > 2. drm_dev_dbg &
Re: [Intel-gfx] [PATCH v3 3/5] drm/print: RFC add choice to use dynamic debug in drm-debug
On Tue, Jul 20, 2021 at 03:29:34PM +0200, Daniel Vetter wrote:
> On Wed, Jul 14, 2021 at 11:51:36AM -0600, Jim Cromie wrote:
> > drm's debug system uses distinct categories of debug messages, encoded
> > in an enum (DRM_UT_), which are mapped to bits in drm.debug.
> > drm_debug_enabled() does a lot of unlikely bit-mask checks on
> > drm.debug; we can use dynamic debug instead, and get all that
> > static_key/jump_label goodness.
Hi Jim,
Thanks for your patches! Daniel pointed me at them in response to my drm_trace
patchset (https://patchwork.freedesktop.org/series/78133/). I'd love to get your
input on it. I think the 2 sets are mostly compatible, we'd just need to keep
drm_dev_dbg and do the CONFIG check in the function beside the trace_enabled
checks.
> >
> > Dynamic debug has no concept of category, but we can map the DRM_UT_*
> > to a set of distinct prefixes; "drm:core:", "drm:kms:" etc, and
> > prepend them to the given formats.
> >
> > Then we can use:
> > `echo module drm format ^drm:core: +p > control`
> >
> > to enable every such "prefixed" pr_debug with one query. This new
> > prefix changes pr_debug's output, so is user visible, but it seems
> > unlikely to cause trouble for log watchers; they're not relying on the
> > absence of class prefix strings.
> >
> > This conversion yields ~2100 new callsites on my i7/i915 laptop:
> >
> > dyndbg: 195 debug prints in module drm_kms_helper
> > dyndbg: 298 debug prints in module drm
> > dyndbg: 1630 debug prints in module i915
> >
> > CONFIG_DRM_USE_DYNAMIC_DEBUG enables this, and is available if
> > CONFIG_DYNAMIC_DEBUG or CONFIG_DYNAMIC_DEBUG_CORE is chosen, and if
> > CONFIG_JUMP_LABEL is enabled; this because its required to get the
> > promised optimizations.
> >
> > The indirection/switchover is layered into the macro scheme:
> >
> > 0. A new callback on drm.debug which calls dynamic_debug_exec_queries
> >to map those bits to specific query/commands
> >dynamic_debug_exec_queries("format ^drm:kms: +p", "drm*");
> >here for POC, this should be in dynamic_debug.c
> >with a MODULE_PARAM_DEBUG_BITMAP(__drm_debug, { "prefix-1", "desc-1" }+)
>
> This is really awesome.
Agreed, this is a very clever way of merging the 2 worlds!
> For merging I think we need to discuss with dyn
> debug folks whether they're all ok with this, but it's exported already
> should should be fine.
I wonder if this is a good time to reconsider our drm categories. IMO they're
overly broad and it's hard to get the right information without subscribing to
the firehose. It seems like dyndbg might be a good opportunity to unlock
subcategories of log messages.
More concretely, on CrOS we can't subscribe to atomic or state categories since
they're too noisy. However if there was a "fail" subcategory which dumped
state/atomic logs on check failures, that would be really compelling. Something
like:
drm:atomic:fail vs. drm:atomic
Both would be picked up if (drm.debug & DRM_DBG_ATOMIC), however it would allow
dyndbg-aware clients to get better logs without having a huge table of
individual log signatures.
I'm not sure how tightly we'd want to control the subcategories. It could be
strict like the categories spelled out in drm_print.h, or an open prefix arg to
drm_dev_dbg. I suspect we'd want the former, but would want to be careful to
provide enough flexibility to properly
Of course, none of this needs to be decided to land this initial support, it can
be bolted on later easily enough (I think).
>
> >
> > 1. A "converted" or "classy" DRM_UT_* map
> >
> >based on: DRM_UT_* ( symbol => bit-mask )
> >named it: cDRM_UT_* ( symbol => format-class-prefix-string )
> >
> >So cDRM_UT_* is either:
> >legacy: cDRM_UT_* <-- DRM_UT_* ( !CONFIG_DRM_USE_DYNAMIC_DEBUG )
> >enabled:
> > #define cDRM_UT_KMS"drm:kms: "
> > #define cDRM_UT_PRIME "drm:prime: "
> > #define cDRM_UT_ATOMIC "drm:atomic: "
>
> the cDRM looks a bit funny, plus I don't eve have an idea what _UT_ means
> (and git history isn't helpful either). What about just using
> DRM_DBG_CLASS_ as the prefix here for these indirection macros, i.e.
> DRM_DBG_CLASS_KMS.
>
> Also would be really nice if we could make these a table or something, but
> I guess with the macro magic that's not possible.
>
> >
> >DRM_UT_* are unchanged, since theyre used in drm_debug_enabled()
> >and elsewhere.
>
> I think for the production version of these we need to retire/deprecate
> them, at least for drm core. Otherwise you have an annoying mismatch
> between drm.debug module option and dyn debug.
>
> >
> > 2. drm_dev_dbg & drm_debug are renamed (prefixed with '_')
> >
> >old names are now macros, calling either:
> > legacy: -> to renamed fn
> > enabled: -> dev_dbg & pr_debug, with cDRM-prefix # format.
> >
> >these names are used in a fat layer of macros (3) which supply the
> >category; those macros are used throughout drm
Re: [Intel-gfx] [PATCH v3 3/5] drm/print: RFC add choice to use dynamic debug in drm-debug
Thanks for the feedback!
On Tue, Jul 20, 2021 at 9:29 AM Daniel Vetter wrote:
>
> On Wed, Jul 14, 2021 at 11:51:36AM -0600, Jim Cromie wrote:
> > drm's debug system uses distinct categories of debug messages, encoded
> > in an enum (DRM_UT_), which are mapped to bits in drm.debug.
> > drm_debug_enabled() does a lot of unlikely bit-mask checks on
> > drm.debug; we can use dynamic debug instead, and get all that
> > static_key/jump_label goodness.
> >
> > Dynamic debug has no concept of category, but we can map the DRM_UT_*
> > to a set of distinct prefixes; "drm:core:", "drm:kms:" etc, and
> > prepend them to the given formats.
> >
> > Then we can use:
> > `echo module drm format ^drm:core: +p > control`
> >
> > to enable every such "prefixed" pr_debug with one query. This new
> > prefix changes pr_debug's output, so is user visible, but it seems
> > unlikely to cause trouble for log watchers; they're not relying on the
> > absence of class prefix strings.
> >
> > This conversion yields ~2100 new callsites on my i7/i915 laptop:
> >
> > dyndbg: 195 debug prints in module drm_kms_helper
> > dyndbg: 298 debug prints in module drm
> > dyndbg: 1630 debug prints in module i915
> >
> > CONFIG_DRM_USE_DYNAMIC_DEBUG enables this, and is available if
> > CONFIG_DYNAMIC_DEBUG or CONFIG_DYNAMIC_DEBUG_CORE is chosen, and if
> > CONFIG_JUMP_LABEL is enabled; this because its required to get the
> > promised optimizations.
> >
> > The indirection/switchover is layered into the macro scheme:
> >
> > 0. A new callback on drm.debug which calls dynamic_debug_exec_queries
> >to map those bits to specific query/commands
> >dynamic_debug_exec_queries("format ^drm:kms: +p", "drm*");
> >here for POC, this should be in dynamic_debug.c
> >with a MODULE_PARAM_DEBUG_BITMAP(__drm_debug, { "prefix-1", "desc-1" }+)
>
> This is really awesome. For merging I think we need to discuss with dyn
> debug folks whether they're all ok with this, but it's exported already
> should should be fine.
>
Yay. FWIW, Im to blame for that export, with this use case in mind.
That said, I dont know if that macro can be written as I described,
but if not, then { BIT(0), "prefix-0", "description-0" }, { BIT(1)
} should work.
If its been done elsewhere, Id copy it, or imitate it.
> >
> > 1. A "converted" or "classy" DRM_UT_* map
> >
> >based on: DRM_UT_* ( symbol => bit-mask )
> >named it: cDRM_UT_* ( symbol => format-class-prefix-string )
> >
> >So cDRM_UT_* is either:
> >legacy: cDRM_UT_* <-- DRM_UT_* ( !CONFIG_DRM_USE_DYNAMIC_DEBUG )
> >enabled:
> > #define cDRM_UT_KMS"drm:kms: "
> > #define cDRM_UT_PRIME "drm:prime: "
> > #define cDRM_UT_ATOMIC "drm:atomic: "
>
> the cDRM looks a bit funny, plus I don't eve have an idea what _UT_ means
> (and git history isn't helpful either). What about just using
> DRM_DBG_CLASS_ as the prefix here for these indirection macros, i.e.
> DRM_DBG_CLASS_KMS.
>
yes.
> Also would be really nice if we could make these a table or something, but
> I guess with the macro magic that's not possible.
>
not obvious to me, I'll watch for an opportunity.
> >
> >DRM_UT_* are unchanged, since theyre used in drm_debug_enabled()
> >and elsewhere.
>
> I think for the production version of these we need to retire/deprecate
> them, at least for drm core. Otherwise you have an annoying mismatch
> between drm.debug module option and dyn debug.
>
I will look at renaming it : __drm_debug_enabled
and making a macro for the old name.
so enabled, it would end up like
if (unlikely(1) && ... )
drivers/gpu/drm/drm_atomic_uapi.c
1457: if (drm_debug_enabled(DRM_UT_STATE))
drivers/gpu/drm/drm_dp_mst_topology.c
1358: if (unlikely(ret == -EIO) && drm_debug_enabled(DRM_UT_DP)) {
2875: if (unlikely(ret) && drm_debug_enabled(DRM_UT_DP)) {
2919: if (drm_debug_enabled(DRM_UT_DP)) {
> >
> > 2. drm_dev_dbg & drm_debug are renamed (prefixed with '_')
> >
> >old names are now macros, calling either:
> > legacy: -> to renamed fn
> > enabled: -> dev_dbg & pr_debug, with cDRM-prefix # format.
> >
> >these names are used in a fat layer of macros (3) which supply the
> >category; those macros are used throughout drm code, yielding the
> >~2100 new prdbgs reported above.
> >
> > 3. names in (2) are invoked by DRM_DEBUG_, drm_dbg_.
> >
> >all these macros get "converted" to use cDRM_UT_*
> >to get right token type for both !/!! DRM_USE_DYNAMIC_DEBUG
> >
> > 4. simplification of __DRM_DEFINE_DBG_RATELIMITED macro
> >
> >remove DRM_UT_ ## KMS as extra indirection
> >pass both DRM_UT & cDRM_UT, for drm_debug_enabled & drm_dev_dbg
>
> For merging, can we pull out the renames and reorgs from this patch, and
> then maybe also the reorder the next patch in your series here to be
> before the dyn debug stuff?
>
I will drop the 4. RATELIMITED tweaks.
FWIW, I have semi-working code to implement rate limiting
in dynamic debug, controlled by
Re: [Intel-gfx] [PATCH v3 3/5] drm/print: RFC add choice to use dynamic debug in drm-debug
On Wed, Jul 14, 2021 at 11:51:36AM -0600, Jim Cromie wrote:
> drm's debug system uses distinct categories of debug messages, encoded
> in an enum (DRM_UT_), which are mapped to bits in drm.debug.
> drm_debug_enabled() does a lot of unlikely bit-mask checks on
> drm.debug; we can use dynamic debug instead, and get all that
> static_key/jump_label goodness.
>
> Dynamic debug has no concept of category, but we can map the DRM_UT_*
> to a set of distinct prefixes; "drm:core:", "drm:kms:" etc, and
> prepend them to the given formats.
>
> Then we can use:
> `echo module drm format ^drm:core: +p > control`
>
> to enable every such "prefixed" pr_debug with one query. This new
> prefix changes pr_debug's output, so is user visible, but it seems
> unlikely to cause trouble for log watchers; they're not relying on the
> absence of class prefix strings.
>
> This conversion yields ~2100 new callsites on my i7/i915 laptop:
>
> dyndbg: 195 debug prints in module drm_kms_helper
> dyndbg: 298 debug prints in module drm
> dyndbg: 1630 debug prints in module i915
>
> CONFIG_DRM_USE_DYNAMIC_DEBUG enables this, and is available if
> CONFIG_DYNAMIC_DEBUG or CONFIG_DYNAMIC_DEBUG_CORE is chosen, and if
> CONFIG_JUMP_LABEL is enabled; this because its required to get the
> promised optimizations.
>
> The indirection/switchover is layered into the macro scheme:
>
> 0. A new callback on drm.debug which calls dynamic_debug_exec_queries
>to map those bits to specific query/commands
>dynamic_debug_exec_queries("format ^drm:kms: +p", "drm*");
>here for POC, this should be in dynamic_debug.c
>with a MODULE_PARAM_DEBUG_BITMAP(__drm_debug, { "prefix-1", "desc-1" }+)
This is really awesome. For merging I think we need to discuss with dyn
debug folks whether they're all ok with this, but it's exported already
should should be fine.
>
> 1. A "converted" or "classy" DRM_UT_* map
>
>based on: DRM_UT_* ( symbol => bit-mask )
>named it: cDRM_UT_* ( symbol => format-class-prefix-string )
>
>So cDRM_UT_* is either:
>legacy: cDRM_UT_* <-- DRM_UT_* ( !CONFIG_DRM_USE_DYNAMIC_DEBUG )
>enabled:
> #define cDRM_UT_KMS"drm:kms: "
> #define cDRM_UT_PRIME "drm:prime: "
> #define cDRM_UT_ATOMIC "drm:atomic: "
the cDRM looks a bit funny, plus I don't eve have an idea what _UT_ means
(and git history isn't helpful either). What about just using
DRM_DBG_CLASS_ as the prefix here for these indirection macros, i.e.
DRM_DBG_CLASS_KMS.
Also would be really nice if we could make these a table or something, but
I guess with the macro magic that's not possible.
>
>DRM_UT_* are unchanged, since theyre used in drm_debug_enabled()
>and elsewhere.
I think for the production version of these we need to retire/deprecate
them, at least for drm core. Otherwise you have an annoying mismatch
between drm.debug module option and dyn debug.
>
> 2. drm_dev_dbg & drm_debug are renamed (prefixed with '_')
>
>old names are now macros, calling either:
> legacy: -> to renamed fn
> enabled: -> dev_dbg & pr_debug, with cDRM-prefix # format.
>
>these names are used in a fat layer of macros (3) which supply the
>category; those macros are used throughout drm code, yielding the
>~2100 new prdbgs reported above.
>
> 3. names in (2) are invoked by DRM_DEBUG_, drm_dbg_.
>
>all these macros get "converted" to use cDRM_UT_*
>to get right token type for both !/!! DRM_USE_DYNAMIC_DEBUG
>
> 4. simplification of __DRM_DEFINE_DBG_RATELIMITED macro
>
>remove DRM_UT_ ## KMS as extra indirection
>pass both DRM_UT & cDRM_UT, for drm_debug_enabled & drm_dev_dbg
For merging, can we pull out the renames and reorgs from this patch, and
then maybe also the reorder the next patch in your series here to be
before the dyn debug stuff?
> Signed-off-by: Jim Cromie
> ---
> drivers/gpu/drm/Kconfig | 13 +
> drivers/gpu/drm/drm_print.c | 75 --
> include/drm/drm_print.h | 102 ++--
> 3 files changed, 158 insertions(+), 32 deletions(-)
I really like this, I think you can drop the RFC. A few more things that I
think we need:
- An overview kerneldoc section which explains the interfaces and how it
all works together. Essentially your commit message with some light
markup to make it look good.
- I think it would be really good to review the driver docs for all this
and make sure it's complete. Some of the interface functions aren't
documented yet (or maybe the ones that drivers shouldn't used need more
__ prefixes to denote them as internal, dunno).
- I guess deprecation notice for drm_debug_enabled() and all that, so that
we have a consistent interface. Doing the conversion will probably
highlight the need for a bit more infrastructure and tooling, e.g. the
bigger dump functions (like edid hex dump, or also the various decode
helpers we have for dp, hdmi infoframes and all that)
[Intel-gfx] [PATCH v3 3/5] drm/print: RFC add choice to use dynamic debug in drm-debug
drm's debug system uses distinct categories of debug messages, encoded
in an enum (DRM_UT_), which are mapped to bits in drm.debug.
drm_debug_enabled() does a lot of unlikely bit-mask checks on
drm.debug; we can use dynamic debug instead, and get all that
static_key/jump_label goodness.
Dynamic debug has no concept of category, but we can map the DRM_UT_*
to a set of distinct prefixes; "drm:core:", "drm:kms:" etc, and
prepend them to the given formats.
Then we can use:
`echo module drm format ^drm:core: +p > control`
to enable every such "prefixed" pr_debug with one query. This new
prefix changes pr_debug's output, so is user visible, but it seems
unlikely to cause trouble for log watchers; they're not relying on the
absence of class prefix strings.
This conversion yields ~2100 new callsites on my i7/i915 laptop:
dyndbg: 195 debug prints in module drm_kms_helper
dyndbg: 298 debug prints in module drm
dyndbg: 1630 debug prints in module i915
CONFIG_DRM_USE_DYNAMIC_DEBUG enables this, and is available if
CONFIG_DYNAMIC_DEBUG or CONFIG_DYNAMIC_DEBUG_CORE is chosen, and if
CONFIG_JUMP_LABEL is enabled; this because its required to get the
promised optimizations.
The indirection/switchover is layered into the macro scheme:
0. A new callback on drm.debug which calls dynamic_debug_exec_queries
to map those bits to specific query/commands
dynamic_debug_exec_queries("format ^drm:kms: +p", "drm*");
here for POC, this should be in dynamic_debug.c
with a MODULE_PARAM_DEBUG_BITMAP(__drm_debug, { "prefix-1", "desc-1" }+)
1. A "converted" or "classy" DRM_UT_* map
based on: DRM_UT_* ( symbol => bit-mask )
named it: cDRM_UT_* ( symbol => format-class-prefix-string )
So cDRM_UT_* is either:
legacy: cDRM_UT_* <-- DRM_UT_* ( !CONFIG_DRM_USE_DYNAMIC_DEBUG )
enabled:
#define cDRM_UT_KMS"drm:kms: "
#define cDRM_UT_PRIME "drm:prime: "
#define cDRM_UT_ATOMIC "drm:atomic: "
DRM_UT_* are unchanged, since theyre used in drm_debug_enabled()
and elsewhere.
2. drm_dev_dbg & drm_debug are renamed (prefixed with '_')
old names are now macros, calling either:
legacy: -> to renamed fn
enabled: -> dev_dbg & pr_debug, with cDRM-prefix # format.
these names are used in a fat layer of macros (3) which supply the
category; those macros are used throughout drm code, yielding the
~2100 new prdbgs reported above.
3. names in (2) are invoked by DRM_DEBUG_, drm_dbg_.
all these macros get "converted" to use cDRM_UT_*
to get right token type for both !/!! DRM_USE_DYNAMIC_DEBUG
4. simplification of __DRM_DEFINE_DBG_RATELIMITED macro
remove DRM_UT_ ## KMS as extra indirection
pass both DRM_UT & cDRM_UT, for drm_debug_enabled & drm_dev_dbg
Signed-off-by: Jim Cromie
---
drivers/gpu/drm/Kconfig | 13 +
drivers/gpu/drm/drm_print.c | 75 --
include/drm/drm_print.h | 102 ++--
3 files changed, 158 insertions(+), 32 deletions(-)
diff --git a/drivers/gpu/drm/Kconfig b/drivers/gpu/drm/Kconfig
index 7ff89690a976..e4524ccba040 100644
--- a/drivers/gpu/drm/Kconfig
+++ b/drivers/gpu/drm/Kconfig
@@ -57,6 +57,19 @@ config DRM_DEBUG_MM
If in doubt, say "N".
+config DRM_USE_DYNAMIC_DEBUG
+ bool "use dynamic debug to implement drm.debug"
+ default n
+ depends on DRM
+ depends on DYNAMIC_DEBUG || DYNAMIC_DEBUG_CORE
+ depends on JUMP_LABEL
+ help
+ The drm debug category facility does a lot of unlikely bit-field
+ tests at runtime; while cheap individually, the cost accumulates.
+ This option uses dynamic debug facility (if configured and
+ using jump_label) to avoid those runtime checks, patching
+ the kernel when those debugs are desired.
+
config DRM_DEBUG_SELFTEST
tristate "kselftests for DRM"
depends on DRM
diff --git a/drivers/gpu/drm/drm_print.c b/drivers/gpu/drm/drm_print.c
index 111b932cf2a9..e2acdfc7088b 100644
--- a/drivers/gpu/drm/drm_print.c
+++ b/drivers/gpu/drm/drm_print.c
@@ -52,8 +52,75 @@ MODULE_PARM_DESC(debug, "Enable debug output, where each bit
enables a debug cat
"\t\tBit 5 (0x20) will enable VBL messages (vblank code)\n"
"\t\tBit 7 (0x80) will enable LEASE messages (leasing code)\n"
"\t\tBit 8 (0x100) will enable DP messages (displayport code)");
+
+#ifndef CONFIG_DRM_USE_DYNAMIC_DEBUG
module_param_named(debug, __drm_debug, int, 0600);
+#else
+static char *format_class_prefixes[] = {
+ cDRM_UT_CORE,
+ cDRM_UT_DRIVER,
+ cDRM_UT_KMS,
+ cDRM_UT_PRIME,
+ cDRM_UT_ATOMIC,
+ cDRM_UT_VBL,
+ cDRM_UT_STATE,
+ cDRM_UT_LEASE,
+ cDRM_UT_DP,
+ cDRM_UT_DRMRES
+};
+
+#define OUR_QUERY_SIZE 64 /* > strlen "format '^%s' %cp" + longest prefix */
+
+static int param_set_dyndbg(const char *instr, const struct kernel_param *kp)
+{
+ unsigned int val;
+ unsigned long changes, result;
+ int rc,
