[for-next,02/11] tracing: Add __cpumask to denote a trace event field that is a cpumask_t

  From: "Steven Rostedt (Google)" <rostedt@goodmis.org>

The trace events have a __bitmask field that can be used for anything
that requires bitmasks. Although currently it is only used for CPU
masks, it could be used in the future for any type of bitmasks.

There is some user space tooling that wants to know if a field is a CPU
mask and not just some random unsigned long bitmask. Introduce
"__cpumask()" helper functions that work the same as the current
__bitmask() helpers but displays in the format file:

  field:__data_loc cpumask_t *[] mask;    offset:36;      size:4; signed:0;

Instead of:

  field:__data_loc unsigned long[] mask;  offset:32;      size:4; signed:0;

The main difference is the type. Instead of "unsigned long" it is
"cpumask_t *". Note, this type field needs to be a real type in the
__dynamic_array() logic that both __cpumask and__bitmask use, but the
comparison field requires it to be a scalar type whereas cpumask_t is a
structure (non-scalar). But everything works when making it a pointer.

Valentin added changes to remove the need of passing in "nr_bits" and the
__cpumask will always use nr_cpumask_bits as its size.

Link: https://lkml.kernel.org/r/20221014080456.1d32b989@rorschach.local.home

Requested-by: Valentin Schneider <vschneid@redhat.com>
Reviewed-by: Valentin Schneider <vschneid@redhat.com>
Signed-off-by: Valentin Schneider <vschneid@redhat.com>
Signed-off-by: Steven Rostedt (Google) <rostedt@goodmis.org>
---
 include/trace/bpf_probe.h                    |  6 ++++
 include/trace/perf.h                         |  6 ++++
 include/trace/stages/stage1_struct_define.h  |  6 ++++
 include/trace/stages/stage2_data_offsets.h   |  6 ++++
 include/trace/stages/stage3_trace_output.h   |  6 ++++
 include/trace/stages/stage4_event_fields.h   |  6 ++++
 include/trace/stages/stage5_get_offsets.h    |  6 ++++
 include/trace/stages/stage6_event_callback.h | 20 ++++++++++++
 include/trace/stages/stage7_class_define.h   |  2 ++
 samples/trace_events/trace-events-sample.c   |  2 +-
 samples/trace_events/trace-events-sample.h   | 34 +++++++++++++++-----
 11 files changed, 91 insertions(+), 9 deletions(-)
  

On 24-11-2022 14:50, Steven Rostedt wrote:
> From: "Steven Rostedt (Google)" <rostedt@goodmis.org>
> 
> The trace events have a __bitmask field that can be used for anything
> that requires bitmasks. Although currently it is only used for CPU
> masks, it could be used in the future for any type of bitmasks.
> 
> There is some user space tooling that wants to know if a field is a CPU
> mask and not just some random unsigned long bitmask. Introduce
> "__cpumask()" helper functions that work the same as the current
> __bitmask() helpers but displays in the format file:
> 
>    field:__data_loc cpumask_t *[] mask;    offset:36;      size:4; signed:0;
> 
> Instead of:
> 
>    field:__data_loc unsigned long[] mask;  offset:32;      size:4; signed:0;
> 
> The main difference is the type. Instead of "unsigned long" it is
> "cpumask_t *". Note, this type field needs to be a real type in the
> __dynamic_array() logic that both __cpumask and__bitmask use, but the
> comparison field requires it to be a scalar type whereas cpumask_t is a
> structure (non-scalar). But everything works when making it a pointer.

How is tooling expected to distinguish between a real dynamic array of pointers
from a type that is using dynamic arrays as an "implementation detail"
with a broken type description ? Any reasonable
interpretation of that type by the consuming tool will be broken
unless it specifically knows about __data_loc cpumask*[].
However, the set of types using that trick is unbounded so forward
compatibilty is impossible to ensure. On top of that, an actual
dynamic array of cpumask pointers becomes impossible to represent.

You might object that if the tool does not know about cpumask,
it does not matter "how it breaks" as the display will be useless anyway,
but that is not true. A parsing library might just parse up to
its knowledge limit and return the most elaborate it can for a given field.
It's acceptable for that representation to not be elaborated with the full
semantic expected by the end user, but it should not return
something that is lying on its nature. For example, it would be sane for
the user to assert the size of an array of pointers to be a multiple
of a pointer size. cpumask is currently an array of unsigned long but there is
nothing preventing a similar type to be based on an array of u8.
Such a type would also have different endianness handling and the resulting buffer
would be garbage.


To fix that issue, I propose to expose the following to userspace:
1. The binary representation type (unsigned long[] in cpumask case).
2. An (ordered list of) semantic type that may or may not be the same as 1.

Type (1) can be used to guarantee correct handling of endianness and a reasonable
default display, while (2) allows any sort of fancy interpretation, all that while preserving
forward compatibility. For cpumask, this would give:
1. unsigned long []
2. bitmask, cpumask

A consumer could know about bitmask as they are likely used in multiple places,
but not about cpumask specifically (e.g. assuming cpumask is a type recently introduced).
Displaying as a list of bits set in the mask would already allow proper formatting, and
knowing it's actually a cpumask can allow fancier behaviors.

 From an event format perspective, this could preserve reasonable backward compat
by simply adding another property:

   field:__data_loc unsigned long[] mask;    offset:36;      size:4; signed:0; semantic_type:bitmask,cpumask;

By default, "semantic_type" would simply have the same value as the normal type.

This applies to any type, not just dynamic arrays.

Thanks,

Douglas
  

On Mon, 12 Dec 2022 14:53:27 +0000
Douglas Raillard <douglas.raillard@arm.com> wrote:

> On 24-11-2022 14:50, Steven Rostedt wrote:
> > From: "Steven Rostedt (Google)" <rostedt@goodmis.org>
> > 
> > The trace events have a __bitmask field that can be used for anything
> > that requires bitmasks. Although currently it is only used for CPU
> > masks, it could be used in the future for any type of bitmasks.
> > 
> > There is some user space tooling that wants to know if a field is a CPU
> > mask and not just some random unsigned long bitmask. Introduce
> > "__cpumask()" helper functions that work the same as the current
> > __bitmask() helpers but displays in the format file:
> > 
> >    field:__data_loc cpumask_t *[] mask;    offset:36;      size:4; signed:0;

The current parsing tools break the above into:

 "field:" "__data_loc" <some-type> "[]" <var-name> ";" "offset:"
 <offset> ";" "size:" "<size>" ";" "signed:" <signed> ";"

Where the <some-type> really can be anything, and in lots of cases, it is.
Thus its only a hint for the tooling, and has never been limited to what
they are.

> > 
> > Instead of:
> > 
> >    field:__data_loc unsigned long[] mask;  offset:32;      size:4; signed:0;
> > 
> > The main difference is the type. Instead of "unsigned long" it is
> > "cpumask_t *". Note, this type field needs to be a real type in the
> > __dynamic_array() logic that both __cpumask and__bitmask use, but the
> > comparison field requires it to be a scalar type whereas cpumask_t is a
> > structure (non-scalar). But everything works when making it a pointer.  

The above is for the kernel to build.

> 
> How is tooling expected to distinguish between a real dynamic array of pointers
> from a type that is using dynamic arrays as an "implementation detail"
> with a broken type description ? Any reasonable
> interpretation of that type by the consuming tool will be broken
> unless it specifically knows about __data_loc cpumask*[].

I'm curious to what the tool does differently with the above. What tool are
you using? Does it just give up on how to print it?

> However, the set of types using that trick is unbounded so forward
> compatibilty is impossible to ensure. On top of that, an actual
> dynamic array of cpumask pointers becomes impossible to represent.

I never thought about a user case where we print out an array of cpumask
pointers.

> 
> You might object that if the tool does not know about cpumask,
> it does not matter "how it breaks" as the display will be useless anyway,
> but that is not true. A parsing library might just parse up to
> its knowledge limit and return the most elaborate it can for a given field.
> It's acceptable for that representation to not be elaborated with the full
> semantic expected by the end user, but it should not return
> something that is lying on its nature. For example, it would be sane for
> the user to assert the size of an array of pointers to be a multiple
> of a pointer size. cpumask is currently an array of unsigned long but there is
> nothing preventing a similar type to be based on an array of u8.
> Such a type would also have different endianness handling and the resulting buffer
> would be garbage.
> 
> 
> To fix that issue, I propose to expose the following to userspace:
> 1. The binary representation type (unsigned long[] in cpumask case).
> 2. An (ordered list of) semantic type that may or may not be the same as 1.
> 
> Type (1) can be used to guarantee correct handling of endianness and a reasonable
> default display, while (2) allows any sort of fancy interpretation, all that while preserving
> forward compatibility. For cpumask, this would give:
> 1. unsigned long []
> 2. bitmask, cpumask
> 
> A consumer could know about bitmask as they are likely used in multiple places,
> but not about cpumask specifically (e.g. assuming cpumask is a type recently introduced).
> Displaying as a list of bits set in the mask would already allow proper formatting, and
> knowing it's actually a cpumask can allow fancier behaviors.
> 
>  From an event format perspective, this could preserve reasonable backward compat
> by simply adding another property:
> 
>    field:__data_loc unsigned long[] mask;    offset:36;      size:4; signed:0; semantic_type:bitmask,cpumask;
> 
> By default, "semantic_type" would simply have the same value as the normal type.

The problem with the above is that it adds a new field, and I have to check
if that doesn't break existing tooling.

Another possibility is that I can add parsing to the format that is exposed
to user space and simply s/__cpumask *[]/__cpumask[]/

Which will get rid of the pointer array of cpu masks.

> 
> This applies to any type, not just dynamic arrays.
> 

Let me know if the above does break existing user space and I'll revert it.

-- Steve
  

On Mon, 12 Dec 2022 11:12:56 -0500
Steven Rostedt <rostedt@goodmis.org> wrote:

> Another possibility is that I can add parsing to the format that is exposed
> to user space and simply s/__cpumask *[]/__cpumask[]/
> 
> Which will get rid of the pointer array of cpu masks.

Actually, I found a better way to get rid of that "*". Would that work for
you? That is, if the above shows:

   field:__data_loc cpumask_t[] mask;    offset:36;      size:4; signed:0;
   
Instead?

-- Steve
  

On 12-12-2022 16:12, Steven Rostedt wrote:
> On Mon, 12 Dec 2022 14:53:27 +0000
> Douglas Raillard <douglas.raillard@arm.com> wrote:
> 
>> On 24-11-2022 14:50, Steven Rostedt wrote:
>>> From: "Steven Rostedt (Google)" <rostedt@goodmis.org>
>>>
>>> The trace events have a __bitmask field that can be used for anything
>>> that requires bitmasks. Although currently it is only used for CPU
>>> masks, it could be used in the future for any type of bitmasks.
>>>
>>> There is some user space tooling that wants to know if a field is a CPU
>>> mask and not just some random unsigned long bitmask. Introduce
>>> "__cpumask()" helper functions that work the same as the current
>>> __bitmask() helpers but displays in the format file:
>>>
>>>     field:__data_loc cpumask_t *[] mask;    offset:36;      size:4; signed:0;
> 
> The current parsing tools break the above into:
> 
>   "field:" "__data_loc" <some-type> "[]" <var-name> ";" "offset:"
>   <offset> ";" "size:" "<size>" ";" "signed:" <signed> ";"
> 
> Where the <some-type> really can be anything, and in lots of cases, it is.
> Thus its only a hint for the tooling, and has never been limited to what
> they are.

While we are having a look at that, what is the exact format of "<some-type>" ?
The only way it can truly be any type is if it's an declaration specifier followed by
an abstract declarator, e.g. like in a func prototype where parameters can be anonymous, as
shown by the "parameter_declaration" rule in:
http://www.lysator.liu.se/c/ANSI-C-grammar-y.html#declaration-specifiers

At least that is how I implemented the parsing and it works for sure for all what I saw in traces
plus more.

> 
>>>
>>> Instead of:
>>>
>>>     field:__data_loc unsigned long[] mask;  offset:32;      size:4; signed:0;
>>>
>>> The main difference is the type. Instead of "unsigned long" it is
>>> "cpumask_t *". Note, this type field needs to be a real type in the
>>> __dynamic_array() logic that both __cpumask and__bitmask use, but the
>>> comparison field requires it to be a scalar type whereas cpumask_t is a
>>> structure (non-scalar). But everything works when making it a pointer.
> 
> The above is for the kernel to build.

That was my understanding that the comparison issue is related to in-kernel filtering ?
If that's the case, I completely agree that the type kernel code sees does not _have_
to be the same thing that is exposed to userspace if that simplifies the problem.

>>
>> How is tooling expected to distinguish between a real dynamic array of pointers
>> from a type that is using dynamic arrays as an "implementation detail"
>> with a broken type description ? Any reasonable
>> interpretation of that type by the consuming tool will be broken
>> unless it specifically knows about __data_loc cpumask*[].
> 
> I'm curious to what the tool does differently with the above. What tool are
> you using? Does it just give up on how to print it?

I'm implementing the Rust libtraceevent/libtracecmd prototype. The use case I'm trying
to address in my first draft is:

* a parser for trace.dat exposing a stable API that does not need to be updated every day.
* A serde backend using that parser
* Any "fancy" type (i.e. that is not a basic scalar type or string) is handled by
   the consumer of the deserializer. This means the backend must be able to fallback
   on providing sensible raw data that can be processed by the user.

In a more general way, I'm trying to structure the code so that the generic trace.dat parsing
lib will be able to handle opaque types by returning a sensibly-typed buffer to the caller. The
caller then decodes it the way it wants. This would apply equally well to a C API.

That is especially important for user-defined fancy types, e.g. in a kernel module. We cannot
reasonably expect users to modify the parsing library itself when they could handle the
decoding on their side with an annotation and a custom decoder [1].

Beyond that, any tool that reads the trace and pretty prints it is bound to provide
broken printing for these types: there will be a date T1 where a new fancy type is introduced
in the trace. Until date T2 when the tool gets explicit support for the type, it will
display the thing as an array of pointers. Interpreting a bitmask as an array of pointer makes
no sense for anyone (machine or human). It so happens that unsigned long == ptr size so it
kind of makes sense for cpumask, but an opaque type might internally be u8[], in which case,
printing as void* [] is meaningless. It would likely be better to display a placeholder
or a plain u8[]. The "semantic_type" field property proposal is an improvement of that.


> )
>> However, the set of types using that trick is unbounded so forward
>> compatibilty is impossible to ensure. On top of that, an actual
>> dynamic array of cpumask pointers becomes impossible to represent.
> 
> I never thought about a user case where we print out an array of cpumask
> pointers.

That case (array of pointer) might be a bit far fetched but kernel already contains
weird stuff such as pointers to opaque struct, I suppose as a way of giving a unique
ID to something. The solution we settle on for cpumask should be applicable to any opaque
type present and future so baking such limitations in the foundations of the event system
does not sound like a terribly good idea.

After reflecting on how the caller of the deserializer can process opaque data, I came
to the conclusion that it will have to know the underlying type (unsigned long[]) and
ask for an enum such as:

enum UnsignedLongArray {
	Bit32(Vec<u32>),
	Bit64(Vec<u64>),
}

When asked for UnsignedLongArray, the serde backend will use the ABI info from the trace
to select the appropriate variant.

> 
>>
>> You might object that if the tool does not know about cpumask,
>> it does not matter "how it breaks" as the display will be useless anyway,
>> but that is not true. A parsing library might just parse up to
>> its knowledge limit and return the most elaborate it can for a given field.
>> It's acceptable for that representation to not be elaborated with the full
>> semantic expected by the end user, but it should not return
>> something that is lying on its nature. For example, it would be sane for
>> the user to assert the size of an array of pointers to be a multiple
>> of a pointer size. cpumask is currently an array of unsigned long but there is
>> nothing preventing a similar type to be based on an array of u8.
>> Such a type would also have different endianness handling and the resulting buffer
>> would be garbage.
>>
>>
>> To fix that issue, I propose to expose the following to userspace:
>> 1. The binary representation type (unsigned long[] in cpumask case).
>> 2. An (ordered list of) semantic type that may or may not be the same as 1.
>>
>> Type (1) can be used to guarantee correct handling of endianness and a reasonable
>> default display, while (2) allows any sort of fancy interpretation, all that while preserving
>> forward compatibility. For cpumask, this would give:
>> 1. unsigned long []
>> 2. bitmask, cpumask
>>
>> A consumer could know about bitmask as they are likely used in multiple places,
>> but not about cpumask specifically (e.g. assuming cpumask is a type recently introduced).
>> Displaying as a list of bits set in the mask would already allow proper formatting, and
>> knowing it's actually a cpumask can allow fancier behaviors.
>>
>>   From an event format perspective, this could preserve reasonable backward compat
>> by simply adding another property:
>>
>>     field:__data_loc unsigned long[] mask;    offset:36;      size:4; signed:0; semantic_type:bitmask,cpumask;
>>
>> By default, "semantic_type" would simply have the same value as the normal type.
> 
> The problem with the above is that it adds a new field, and I have to check
> if that doesn't break existing tooling.

Yes, I was hesitating to propose that as well. I think it would be reasonable to allow for this
sort of extension in man trace.dat in the future but maybe that's something for version 8.
In my implementation I parse that to a map anyway, so unknown properties will simply be ignored.

> Another possibility is that I can add parsing to the format that is exposed
> to user space and simply s/__cpumask *[]/__cpumask[]/
> 
> Which will get rid of the pointer array of cpu masks.

My main concern is if the item size info becomes available in the future as a field property,
we won't be able to make dynamic arrays of opaque types if the array notation is
already taken for single items.

So really ideally we want "__data_loc cpumask" and not "__data_loc cpumask[]", as the latter is an
array of cpumasks, and the former a single one. Or maybe something like "__data_loc SINGLE_cpumask[]".
That format would still be parseable by current userspace tools as an opaque type since SINGLE_cpumask
is a valid C identifier.

Beyond extensibility, it would make __data_loc inconsistent:
* for basic types like u32, it's completely expected to be an actual array.
* but for opaque types, it's expected to be a single item and not an array at all.

> 
>>
>> This applies to any type, not just dynamic arrays.
>>
> 
> Let me know if the above does break existing user space and I'll revert it.

I could experiment with trace-cmd but I'm not familiar with everything out there decoding that.
The parser I'm working on is 1.5 week old and unreleased so no change would be an issue as long
as it's possible to distinguish new formats from old ones (to process old traces).


> -- Steve

Douglas


[1] Vague example of how a user application would use serde and the trace parser:

// Serde doc: https://serde.rs/
use serde::Deserialize;

// The yet-to-exist trace parser, pulled from crates.io like any other dependency.
// The user is not expected to modify that ftraceparser code, just the same way you
// are not expected to modify libtraceevent to make use of it in your own application.
use ftraceparser::read_trace;


// Top-level enum that contains a variant for each event of interest.
#[derive(Deserialize)]
enum TraceEvent {
	#[serde(rename = "sched_switch")]
	EventSchedSwitch(EventSchedSwitchFields),

	// This includes a custom event, that is e.g. defined in a kernel module, or
	// maybe it's a brand new event that the userspace tooling does not handle yet.
	// No ftraceparser code has knowledge of that type.
	#[serde(rename = "my_custom_event")]
	MyCustomEvent(MyCustomEventFields),
}

#[derive(Deserialize)]
pub struct EventSchedSwitchFields {
     pub common_comm: Comm,
     pub common_pid: PID,
     pub common_cpu: CPU,
     pub prev_comm: Comm,
     pub next_comm: Comm,
     pub prev_state: u32,
     pub prev_pid: PID,
     pub next_pid: PID,
     pub next_prio: Prio,
     pub prev_prio: Prio,
}

// Custom event struct, with one field containing an opaque fancy type
// that the parser does not know about (e.g. the event is added by a kernel
// module, or a kernel dev has a custom junk patch with quick&dirty debugging aids)
pub struct MyCustomEventFields {
	#[serde(deserialize_with = "decode_fancy")]
	fancy_field: Fancy,
}


// If we use the automatic Deserialize implementation provided by #[derive(Deserialize)],
// it will be under the assumptions that the parser is able to split the trace in Fancy's
// fields (or a sequence if it was a simple wrapper around an array type) so they can be mapped
// on that struct's fields. This is by definition not possible because the parser have no clue about
// the content of Fancy. ftraceparser only knows how to split the fields of MyCustomEventFields because
// of the format in the header. No such format header exists for Fancy.
// We therefore use a custom "decode_fancy()" function where we ask for a byte buffer and decode it manually.
// We could also provide a manual Deserialize implementation to avoid needing "#[serde(deserialize_with=...)]".
struct Fancy {
	...
}

fn decode_fancy<D>(deserializer: D) {
	// Request a buffer of bytes and decode it into a Fancy value. All the trace parser needs to know
	// is how to get the chunk of data we asked, we handle the rest here.
}



fn process(event: TraceEvent) {
	...
}

fn main() {
	...

	// Read the trace, asking for an iterator of TraceEvent values.
	let events = read_trace::<TraceEvent>("path/to/trace.dat");
	for event in events {
		process(event);
	}
}
  

On Mon, 12 Dec 2022 22:19:17 +0000
Douglas Raillard <douglas.raillard@arm.com> wrote:

> >>>     field:__data_loc cpumask_t *[] mask;    offset:36;      size:4; signed:0;  
> > 
> > The current parsing tools break the above into:
> > 
> >   "field:" "__data_loc" <some-type> "[]" <var-name> ";" "offset:"
> >   <offset> ";" "size:" "<size>" ";" "signed:" <signed> ";"
> > 
> > Where the <some-type> really can be anything, and in lots of cases, it is.
> > Thus its only a hint for the tooling, and has never been limited to what
> > they are.  
> 
> While we are having a look at that, what is the exact format of "<some-type>" ?

It's literally anything that the kernel excepts and can be in a kernel
structure. As the macro in the kernel defines both the internal structure
and what gets printed in the field. This is why the parsing tooling never
was very strict on <some-type> because a lot of times its something that it
does not recognize. Any new type added in a trace event will show up here.


> > 
> > The above is for the kernel to build.  
> 
> That was my understanding that the comparison issue is related to in-kernel filtering ?
> If that's the case, I completely agree that the type kernel code sees does not _have_
> to be the same thing that is exposed to userspace if that simplifies the problem.

Yes, and note the patch I sent out to fix this.

> 
> >>
> >> How is tooling expected to distinguish between a real dynamic array of pointers
> >> from a type that is using dynamic arrays as an "implementation detail"
> >> with a broken type description ? Any reasonable
> >> interpretation of that type by the consuming tool will be broken
> >> unless it specifically knows about __data_loc cpumask*[].  
> > 
> > I'm curious to what the tool does differently with the above. What tool are
> > you using? Does it just give up on how to print it?  
> 
> I'm implementing the Rust libtraceevent/libtracecmd prototype. The use case I'm trying

Duh! Of course. This is what we discussed at the tracing summit. I just now
noticed your name ;-)

> to address in my first draft is:
> 
> * a parser for trace.dat exposing a stable API that does not need to be updated every day.
> * A serde backend using that parser
> * Any "fancy" type (i.e. that is not a basic scalar type or string) is handled by
>    the consumer of the deserializer. This means the backend must be able to fallback
>    on providing sensible raw data that can be processed by the user.

That third part may be difficult with the above issue I mentioned.

Just do:

 git grep '__field(' | cut -d',' -f1 | cut -d'(' -f2 | sed -e 's/[    ]*//'
 | sort -u

to see what's in the kernel.


> 
> In a more general way, I'm trying to structure the code so that the generic trace.dat parsing
> lib will be able to handle opaque types by returning a sensibly-typed buffer to the caller. The
> caller then decodes it the way it wants. This would apply equally well to a C API.
> 
> That is especially important for user-defined fancy types, e.g. in a kernel module. We cannot
> reasonably expect users to modify the parsing library itself when they could handle the
> decoding on their side with an annotation and a custom decoder [1].
> 
> Beyond that, any tool that reads the trace and pretty prints it is bound to provide
> broken printing for these types: there will be a date T1 where a new fancy type is introduced
> in the trace. Until date T2 when the tool gets explicit support for the type, it will
> display the thing as an array of pointers. Interpreting a bitmask as an array of pointer makes
> no sense for anyone (machine or human). It so happens that unsigned long == ptr size so it
> kind of makes sense for cpumask, but an opaque type might internally be u8[], in which case,
> printing as void* [] is meaningless. It would likely be better to display a placeholder
> or a plain u8[]. The "semantic_type" field property proposal is an improvement of that.

Note, I put much more effort into the offset, size and sign than the type.
But is this only for the arrays that you have these restrictions, or any
field type?

> 
> 
> > )  
> >> However, the set of types using that trick is unbounded so forward
> >> compatibilty is impossible to ensure. On top of that, an actual
> >> dynamic array of cpumask pointers becomes impossible to represent.  
> > 
> > I never thought about a user case where we print out an array of cpumask
> > pointers.  
> 
> That case (array of pointer) might be a bit far fetched but kernel already contains
> weird stuff such as pointers to opaque struct, I suppose as a way of giving a unique
> ID to something. The solution we settle on for cpumask should be applicable to any opaque
> type present and future so baking such limitations in the foundations of the event system
> does not sound like a terribly good idea.

The above should be moot by now, as my new patch removes the "*" pointer.

> 
> After reflecting on how the caller of the deserializer can process opaque data, I came
> to the conclusion that it will have to know the underlying type (unsigned long[]) and
> ask for an enum such as:
> 
> enum UnsignedLongArray {
> 	Bit32(Vec<u32>),
> 	Bit64(Vec<u64>),
> }
> 
> When asked for UnsignedLongArray, the serde backend will use the ABI info from the trace
> to select the appropriate variant.
> 
> >   
> >>
> >> You might object that if the tool does not know about cpumask,
> >> it does not matter "how it breaks" as the display will be useless anyway,
> >> but that is not true. A parsing library might just parse up to
> >> its knowledge limit and return the most elaborate it can for a given field.
> >> It's acceptable for that representation to not be elaborated with the full
> >> semantic expected by the end user, but it should not return
> >> something that is lying on its nature. For example, it would be sane for
> >> the user to assert the size of an array of pointers to be a multiple
> >> of a pointer size. cpumask is currently an array of unsigned long but there is
> >> nothing preventing a similar type to be based on an array of u8.
> >> Such a type would also have different endianness handling and the resulting buffer
> >> would be garbage.
> >>
> >>
> >> To fix that issue, I propose to expose the following to userspace:
> >> 1. The binary representation type (unsigned long[] in cpumask case).
> >> 2. An (ordered list of) semantic type that may or may not be the same as 1.
> >>
> >> Type (1) can be used to guarantee correct handling of endianness and a reasonable
> >> default display, while (2) allows any sort of fancy interpretation, all that while preserving
> >> forward compatibility. For cpumask, this would give:
> >> 1. unsigned long []
> >> 2. bitmask, cpumask
> >>
> >> A consumer could know about bitmask as they are likely used in multiple places,
> >> but not about cpumask specifically (e.g. assuming cpumask is a type recently introduced).
> >> Displaying as a list of bits set in the mask would already allow proper formatting, and
> >> knowing it's actually a cpumask can allow fancier behaviors.
> >>
> >>   From an event format perspective, this could preserve reasonable backward compat
> >> by simply adding another property:
> >>
> >>     field:__data_loc unsigned long[] mask;    offset:36;      size:4; signed:0; semantic_type:bitmask,cpumask;
> >>
> >> By default, "semantic_type" would simply have the same value as the normal type.  
> > 
> > The problem with the above is that it adds a new field, and I have to check
> > if that doesn't break existing tooling.  
> 
> Yes, I was hesitating to propose that as well. I think it would be reasonable to allow for this
> sort of extension in man trace.dat in the future but maybe that's something for version 8.
> In my implementation I parse that to a map anyway, so unknown properties will simply be ignored.
> 
> > Another possibility is that I can add parsing to the format that is exposed
> > to user space and simply s/__cpumask *[]/__cpumask[]/
> > 
> > Which will get rid of the pointer array of cpu masks.  
> 
> My main concern is if the item size info becomes available in the future as a field property,
> we won't be able to make dynamic arrays of opaque types if the array notation is
> already taken for single items.

I believe we already have opaque type arrays.

And the current libtraceevent allows for plugins that will override the
print-fmt parsing to allow for the plugin to know how to parse it properly.


> 
> So really ideally we want "__data_loc cpumask" and not "__data_loc cpumask[]", as the latter is an
> array of cpumasks, and the former a single one. Or maybe something like "__data_loc SINGLE_cpumask[]".

The [] just means that the size is not defined, and that current tools will
parse it correctly, even if it does not know what cpumask is.

An array of cpumask_t is redundant, as it really becomes one bigger
cpumask_t.


> That format would still be parseable by current userspace tools as an opaque type since SINGLE_cpumask
> is a valid C identifier.
> 
> Beyond extensibility, it would make __data_loc inconsistent:
> * for basic types like u32, it's completely expected to be an actual array.
> * but for opaque types, it's expected to be a single item and not an array at all.
> 
> >   
> >>
> >> This applies to any type, not just dynamic arrays.
> >>  
> > 
> > Let me know if the above does break existing user space and I'll revert it.  
> 
> I could experiment with trace-cmd but I'm not familiar with everything out there decoding that.
> The parser I'm working on is 1.5 week old and unreleased so no change would be an issue as long
> as it's possible to distinguish new formats from old ones (to process old traces).
> 

[ /me goes and does some testing ...]

Actually, the cpumask[] fails to parse regardless it seems, so I'm free to
change this anyway I want and still not cause regressions.

This means I can make it this:

  field:__data_loc cpumask_t mask;    offset:36;      size:4; signed:0;


and it should not cause any regression, as it already fails as this is new
anyway.

I get this from the sample module:

    event-sample-861   [000]    99.584942: foo_bar:              [FAILED TO PARSE] foo=hello bar=3 list=ARRAY[01, 00, 00, 00, 02, 00, 00, 00, 03, 00, 00, 00] str=Frodo cpus=ARRAY[ff, 00, 00, 00, 00, 00, 00, 00] cpum=ARRAY[ff, 00, 00, 00, 00, 00, 00, 00] vstr=iter=3

-- Steve
  

On 12-12-2022 23:53, Steven Rostedt wrote:
> On Mon, 12 Dec 2022 22:19:17 +0000
> Douglas Raillard <douglas.raillard@arm.com> wrote:
> 
>>>>>      field:__data_loc cpumask_t *[] mask;    offset:36;      size:4; signed:0;
>>>
>>> The current parsing tools break the above into:
>>>
>>>    "field:" "__data_loc" <some-type> "[]" <var-name> ";" "offset:"
>>>    <offset> ";" "size:" "<size>" ";" "signed:" <signed> ";"
>>>
>>> Where the <some-type> really can be anything, and in lots of cases, it is.
>>> Thus its only a hint for the tooling, and has never been limited to what
>>> they are.
>>
>> While we are having a look at that, what is the exact format of "<some-type>" ?
> 
> It's literally anything that the kernel excepts and can be in a kernel
> structure. As the macro in the kernel defines both the internal structure
> and what gets printed in the field. This is why the parsing tooling never
> was very strict on <some-type> because a lot of times its something that it
> does not recognize. Any new type added in a trace event will show up here.
> 
> 
>>>
>>> The above is for the kernel to build.
>>
>> That was my understanding that the comparison issue is related to in-kernel filtering ?
>> If that's the case, I completely agree that the type kernel code sees does not _have_
>> to be the same thing that is exposed to userspace if that simplifies the problem.
> 
> Yes, and note the patch I sent out to fix this.

I did, that's some seriously fast TAT :)

>>
>>>>
>>>> How is tooling expected to distinguish between a real dynamic array of pointers
>>>> from a type that is using dynamic arrays as an "implementation detail"
>>>> with a broken type description ? Any reasonable
>>>> interpretation of that type by the consuming tool will be broken
>>>> unless it specifically knows about __data_loc cpumask*[].
>>>
>>> I'm curious to what the tool does differently with the above. What tool are
>>> you using? Does it just give up on how to print it?
>>
>> I'm implementing the Rust libtraceevent/libtracecmd prototype. The use case I'm trying
> 
> Duh! Of course. This is what we discussed at the tracing summit. I just now
> noticed your name ;-)
> 
>> to address in my first draft is:
>>
>> * a parser for trace.dat exposing a stable API that does not need to be updated every day.
>> * A serde backend using that parser
>> * Any "fancy" type (i.e. that is not a basic scalar type or string) is handled by
>>     the consumer of the deserializer. This means the backend must be able to fallback
>>     on providing sensible raw data that can be processed by the user.
> 
> That third part may be difficult with the above issue I mentioned.
> 
> Just do:
> 
>   git grep '__field(' | cut -d',' -f1 | cut -d'(' -f2 | sed -e 's/[    ]*//'
>   | sort -u
> 
> to see what's in the kernel.

There are lots of types, but as long as the caller knows what to ask for, it shouldn't be an issue.
Pretty printing the trace is obviously an important aspect and ideally requires the parser to know
how to format everything.

But when it comes to other processing in a compiled language, it's not a big burden to let people
declare the events they require and the expected fields + types so they can get the data into their
own struct (e.g. as with serde or any equivalent technology).

> 
>>
>> In a more general way, I'm trying to structure the code so that the generic trace.dat parsing
>> lib will be able to handle opaque types by returning a sensibly-typed buffer to the caller. The
>> caller then decodes it the way it wants. This would apply equally well to a C API.
>>
>> That is especially important for user-defined fancy types, e.g. in a kernel module. We cannot
>> reasonably expect users to modify the parsing library itself when they could handle the
>> decoding on their side with an annotation and a custom decoder [1].
>>
>> Beyond that, any tool that reads the trace and pretty prints it is bound to provide
>> broken printing for these types: there will be a date T1 where a new fancy type is introduced
>> in the trace. Until date T2 when the tool gets explicit support for the type, it will
>> display the thing as an array of pointers. Interpreting a bitmask as an array of pointer makes
>> no sense for anyone (machine or human). It so happens that unsigned long == ptr size so it
>> kind of makes sense for cpumask, but an opaque type might internally be u8[], in which case,
>> printing as void* [] is meaningless. It would likely be better to display a placeholder
>> or a plain u8[]. The "semantic_type" field property proposal is an improvement of that.
> 
> Note, I put much more effort into the offset, size and sign than the type.
> But is this only for the arrays that you have these restrictions, or any
> field type?

In terms of borken pretty printing, it's a general issue not limited to dynamic arrays.
The only ways pretty printing for an opaque type can possibly work for new types the parser has no
specific knowledge of are:
1. The type is not actually opaque, i.e. it comes with some decoding schema (just like the events have
    a schema listing their fields + types)
2. The type is opaque, but also ships with an executable description of how to print it.
    E.g. if there was a WASM/eBPF/whatever bytecode printing routine made available to userspace.

Option (2) is not so appealing as it's both hard to achieve and only allows a fixed set of
behaviors for a type. Option (1) is a lot easier and allows the behaviors to be defined
on the user side.

Wild idea: include the BTF blob in the trace.dat header so no type is opaque anymore. The printing
issue is not entirely solved this way (e.g. cpumask still needs some plugin to be displayed as a list
of CPUs), but we could at least print all structs in "raw" mode and enum symbolically.

That could also allow creating a quick&dirty way of defining a proper event (aka not trace_printk()):

	#define SIMPLE_TRACE_EVENT(type, fields) \
	struct type fields;	
	TRACE_EVENT(type, \
		TP_PROTO(struct type *data), \
		TP_ARGS(data), \
		TP_STRUCT__entry(__field(struct type, data)), \
		TP_fast_assign(__entry->data = *data;), \
		TP_printk("print in raw mode to display the data"), \
	);
	#define SIMPLE_TRACE(type, fields) trace_struct_##type(&(struct type)fields)


	SIMPLE_TRACE_EVENT(myevent, {
		char name[11];
		int foobar;
	});
	
	SIMPLE_TRACE(myevent, {.name = "hello", .foobar = 42});

The format string could be either kernel-generated based on BTF or userspace could be expected
to make its own use of BTF.

> 
>>
>>
>>> )
>>>> However, the set of types using that trick is unbounded so forward
>>>> compatibilty is impossible to ensure. On top of that, an actual
>>>> dynamic array of cpumask pointers becomes impossible to represent.
>>>
>>> I never thought about a user case where we print out an array of cpumask
>>> pointers.
>>
>> That case (array of pointer) might be a bit far fetched but kernel already contains
>> weird stuff such as pointers to opaque struct, I suppose as a way of giving a unique
>> ID to something. The solution we settle on for cpumask should be applicable to any opaque
>> type present and future so baking such limitations in the foundations of the event system
>> does not sound like a terribly good idea.
> 
> The above should be moot by now, as my new patch removes the "*" pointer.
> 
>>
>> After reflecting on how the caller of the deserializer can process opaque data, I came
>> to the conclusion that it will have to know the underlying type (unsigned long[]) and
>> ask for an enum such as:
>>
>> enum UnsignedLongArray {
>> 	Bit32(Vec<u32>),
>> 	Bit64(Vec<u64>),
>> }
>>
>> When asked for UnsignedLongArray, the serde backend will use the ABI info from the trace
>> to select the appropriate variant.
>>
>>>    
>>>>
>>>> You might object that if the tool does not know about cpumask,
>>>> it does not matter "how it breaks" as the display will be useless anyway,
>>>> but that is not true. A parsing library might just parse up to
>>>> its knowledge limit and return the most elaborate it can for a given field.
>>>> It's acceptable for that representation to not be elaborated with the full
>>>> semantic expected by the end user, but it should not return
>>>> something that is lying on its nature. For example, it would be sane for
>>>> the user to assert the size of an array of pointers to be a multiple
>>>> of a pointer size. cpumask is currently an array of unsigned long but there is
>>>> nothing preventing a similar type to be based on an array of u8.
>>>> Such a type would also have different endianness handling and the resulting buffer
>>>> would be garbage.
>>>>
>>>>
>>>> To fix that issue, I propose to expose the following to userspace:
>>>> 1. The binary representation type (unsigned long[] in cpumask case).
>>>> 2. An (ordered list of) semantic type that may or may not be the same as 1.
>>>>
>>>> Type (1) can be used to guarantee correct handling of endianness and a reasonable
>>>> default display, while (2) allows any sort of fancy interpretation, all that while preserving
>>>> forward compatibility. For cpumask, this would give:
>>>> 1. unsigned long []
>>>> 2. bitmask, cpumask
>>>>
>>>> A consumer could know about bitmask as they are likely used in multiple places,
>>>> but not about cpumask specifically (e.g. assuming cpumask is a type recently introduced).
>>>> Displaying as a list of bits set in the mask would already allow proper formatting, and
>>>> knowing it's actually a cpumask can allow fancier behaviors.
>>>>
>>>>    From an event format perspective, this could preserve reasonable backward compat
>>>> by simply adding another property:
>>>>
>>>>      field:__data_loc unsigned long[] mask;    offset:36;      size:4; signed:0; semantic_type:bitmask,cpumask;
>>>>
>>>> By default, "semantic_type" would simply have the same value as the normal type.
>>>
>>> The problem with the above is that it adds a new field, and I have to check
>>> if that doesn't break existing tooling.
>>
>> Yes, I was hesitating to propose that as well. I think it would be reasonable to allow for this
>> sort of extension in man trace.dat in the future but maybe that's something for version 8.
>> In my implementation I parse that to a map anyway, so unknown properties will simply be ignored.
>>
>>> Another possibility is that I can add parsing to the format that is exposed
>>> to user space and simply s/__cpumask *[]/__cpumask[]/
>>>
>>> Which will get rid of the pointer array of cpu masks.
>>
>> My main concern is if the item size info becomes available in the future as a field property,
>> we won't be able to make dynamic arrays of opaque types if the array notation is
>> already taken for single items.
> 
> I believe we already have opaque type arrays.
> 
> And the current libtraceevent allows for plugins that will override the
> print-fmt parsing to allow for the plugin to know how to parse it properly.
> 
> 
>>
>> So really ideally we want "__data_loc cpumask" and not "__data_loc cpumask[]", as the latter is an
>> array of cpumasks, and the former a single one. Or maybe something like "__data_loc SINGLE_cpumask[]".
> 
> The [] just means that the size is not defined, and that current tools will
> parse it correctly, even if it does not know what cpumask is.
> 
> An array of cpumask_t is redundant, as it really becomes one bigger
> cpumask_t.

It's not. A list of 21 CPU affinity masks in a system with 5 CPUs is definitely not the same thing as one mask
of size 105. Yes they are isomorphic but so is a task's comm (15 char * 7 bits = 105 bits). We are trying to
express what the data represent, not merely the amount of info required to encode them.

Now I'm not saying that we can create such array of cpumasks using dynamic arrays in the current
state of the kernel macro infrastructure, but it's an implementation detail. That should be able to evolve
independently of the data model exposed to userspace.

> 
>> That format would still be parseable by current userspace tools as an opaque type since SINGLE_cpumask
>> is a valid C identifier.
>>
>> Beyond extensibility, it would make __data_loc inconsistent:
>> * for basic types like u32, it's completely expected to be an actual array.
>> * but for opaque types, it's expected to be a single item and not an array at all.
>>
>>>    
>>>>
>>>> This applies to any type, not just dynamic arrays.
>>>>   
>>>
>>> Let me know if the above does break existing user space and I'll revert it.
>>
>> I could experiment with trace-cmd but I'm not familiar with everything out there decoding that.
>> The parser I'm working on is 1.5 week old and unreleased so no change would be an issue as long
>> as it's possible to distinguish new formats from old ones (to process old traces).
>>
> 
> [ /me goes and does some testing ...]
> 
> Actually, the cpumask[] fails to parse regardless it seems, so I'm free to
> change this anyway I want and still not cause regressions.
> 
> This means I can make it this:
> 
>    field:__data_loc cpumask_t mask;    offset:36;      size:4; signed:0;
> 
> 
> and it should not cause any regression, as it already fails as this is new
> anyway.
> 
> I get this from the sample module:
> 
>      event-sample-861   [000]    99.584942: foo_bar:              [FAILED TO PARSE] foo=hello bar=3 list=ARRAY[01, 00, 00, 00, 02, 00, 00, 00, 03, 00, 00, 00] str=Frodo cpus=ARRAY[ff, 00, 00, 00, 00, 00, 00, 00] cpum=ARRAY[ff, 00, 00, 00, 00, 00, 00, 00] vstr=iter=3

Niiice. So we can have both real arrays of things like an "__data_loc pid_t[]" and at the same
time have types that happen to need dynamic amount of storage like cpumask_t. So __data_loc can
now be treated almost like a new type qualifier.

The only case I can think of where parsing would not follow regular C abstract declaration syntax is a type like that:

	__data_loc int [3][]

The outer-most array is by definition the dynamic one, so "[]". In normal C, [3] and [] would be swapped as
the outer-most array comes first. That's not too bad though as it is not ambiguous and easy to fixup directly
in the parse tree.

-- Douglas
  

On Tue, 13 Dec 2022 14:20:12 +0000
Douglas Raillard <douglas.raillard@arm.com> wrote:


> >>>
> >>> The above is for the kernel to build.  
> >>
> >> That was my understanding that the comparison issue is related to in-kernel filtering ?
> >> If that's the case, I completely agree that the type kernel code sees does not _have_
> >> to be the same thing that is exposed to userspace if that simplifies the problem.  
> > 
> > Yes, and note the patch I sent out to fix this.  
> 
> I did, that's some seriously fast TAT :)

I'm going to pull that one in and start testing, so I can push it out in
this merge window.



> > That third part may be difficult with the above issue I mentioned.
> > 
> > Just do:
> > 
> >   git grep '__field(' | cut -d',' -f1 | cut -d'(' -f2 | sed -e 's/[    ]*//'
> >   | sort -u
> > 
> > to see what's in the kernel.  
> 
> There are lots of types, but as long as the caller knows what to ask for, it shouldn't be an issue.
> Pretty printing the trace is obviously an important aspect and ideally requires the parser to know
> how to format everything.
> 
> But when it comes to other processing in a compiled language, it's not a big burden to let people
> declare the events they require and the expected fields + types so they can get the data into their
> own struct (e.g. as with serde or any equivalent technology).

That's what is done today. The size and offset is how the tools can get to
the data and it knows what to do with it.

I'm not sure how rust can handle this type of opaque type scheme.


> > Note, I put much more effort into the offset, size and sign than the type.
> > But is this only for the arrays that you have these restrictions, or any
> > field type?  
> 
> In terms of borken pretty printing, it's a general issue not limited to dynamic arrays.

Yeah, the pretty printing can easily fail, as it can have anything that the
kernel can do. Including calling functions that are not available to user
space. This is why the fallback is always back to size, offset and sign.

> The only ways pretty printing for an opaque type can possibly work for new types the parser has no
> specific knowledge of are:
> 1. The type is not actually opaque, i.e. it comes with some decoding schema (just like the events have
>     a schema listing their fields + types)
> 2. The type is opaque, but also ships with an executable description of how to print it.
>     E.g. if there was a WASM/eBPF/whatever bytecode printing routine made available to userspace.
> 
> Option (2) is not so appealing as it's both hard to achieve and only allows a fixed set of
> behaviors for a type. Option (1) is a lot easier and allows the behaviors to be defined
> on the user side.
> 
> Wild idea: include the BTF blob in the trace.dat header so no type is opaque anymore. The printing
> issue is not entirely solved this way (e.g. cpumask still needs some plugin to be displayed as a list
> of CPUs), but we could at least print all structs in "raw" mode and enum symbolically.

And how big is that blob?

I'm not against the idea, but I would like it to only hold what is needed.

> 
> That could also allow creating a quick&dirty way of defining a proper event (aka not trace_printk()):


I prefer not to have "quick&dirty" ;-)

> 
> 	#define SIMPLE_TRACE_EVENT(type, fields) \
> 	struct type fields;	
> 	TRACE_EVENT(type, \
> 		TP_PROTO(struct type *data), \
> 		TP_ARGS(data), \
> 		TP_STRUCT__entry(__field(struct type, data)), \
> 		TP_fast_assign(__entry->data = *data;), \
> 		TP_printk("print in raw mode to display the data"), \
> 	);
> 	#define SIMPLE_TRACE(type, fields) trace_struct_##type(&(struct type)fields)
> 
> 
> 	SIMPLE_TRACE_EVENT(myevent, {
> 		char name[11];
> 		int foobar;
> 	});
> 	
> 	SIMPLE_TRACE(myevent, {.name = "hello", .foobar = 42});


> 
> The format string could be either kernel-generated based on BTF or userspace could be expected
> to make its own use of BTF.

What's the use case for the above?

> 

> >>
> >> So really ideally we want "__data_loc cpumask" and not "__data_loc cpumask[]", as the latter is an
> >> array of cpumasks, and the former a single one. Or maybe something like "__data_loc SINGLE_cpumask[]".  
> > 
> > The [] just means that the size is not defined, and that current tools will
> > parse it correctly, even if it does not know what cpumask is.
> > 
> > An array of cpumask_t is redundant, as it really becomes one bigger
> > cpumask_t.  
> 
> It's not. A list of 21 CPU affinity masks in a system with 5 CPUs is definitely not the same thing as one mask
> of size 105. Yes they are isomorphic but so is a task's comm (15 char * 7 bits = 105 bits). We are trying to
> express what the data represent, not merely the amount of info required to encode them.
> 
> Now I'm not saying that we can create such array of cpumasks using dynamic arrays in the current
> state of the kernel macro infrastructure, but it's an implementation detail. That should be able to evolve
> independently of the data model exposed to userspace.
> 
> >   
> >> That format would still be parseable by current userspace tools as an opaque type since SINGLE_cpumask
> >> is a valid C identifier.
> >>
> >> Beyond extensibility, it would make __data_loc inconsistent:
> >> * for basic types like u32, it's completely expected to be an actual array.
> >> * but for opaque types, it's expected to be a single item and not an array at all.
> >>  
> >>>      
> >>>>
> >>>> This applies to any type, not just dynamic arrays.
> >>>>     
> >>>
> >>> Let me know if the above does break existing user space and I'll revert it.  
> >>
> >> I could experiment with trace-cmd but I'm not familiar with everything out there decoding that.
> >> The parser I'm working on is 1.5 week old and unreleased so no change would be an issue as long
> >> as it's possible to distinguish new formats from old ones (to process old traces).
> >>  
> > 
> > [ /me goes and does some testing ...]
> > 
> > Actually, the cpumask[] fails to parse regardless it seems, so I'm free to
> > change this anyway I want and still not cause regressions.
> > 
> > This means I can make it this:
> > 
> >    field:__data_loc cpumask_t mask;    offset:36;      size:4; signed:0;
> > 
> > 
> > and it should not cause any regression, as it already fails as this is new
> > anyway.
> > 
> > I get this from the sample module:
> > 
> >      event-sample-861   [000]    99.584942: foo_bar:              [FAILED TO PARSE] foo=hello bar=3 list=ARRAY[01, 00, 00, 00, 02, 00, 00, 00, 03, 00, 00, 00] str=Frodo cpus=ARRAY[ff, 00, 00, 00, 00, 00, 00, 00] cpum=ARRAY[ff, 00, 00, 00, 00, 00, 00, 00] vstr=iter=3  
> 
> Niiice. So we can have both real arrays of things like an "__data_loc pid_t[]" and at the same
> time have types that happen to need dynamic amount of storage like cpumask_t. So __data_loc can
> now be treated almost like a new type qualifier.
> 
> The only case I can think of where parsing would not follow regular C abstract declaration syntax is a type like that:
> 
> 	__data_loc int [3][]
> 
> The outer-most array is by definition the dynamic one, so "[]". In normal C, [3] and [] would be swapped as
> the outer-most array comes first. That's not too bad though as it is not ambiguous and easy to fixup directly
> in the parse tree.
> 

Well, currently we do not have a way to create the above, so I'm not
worried.

-- Steve
  

On 13-12-2022 15:11, Steven Rostedt wrote:
> On Tue, 13 Dec 2022 14:20:12 +0000
> Douglas Raillard <douglas.raillard@arm.com> wrote:
> 
> 
>>>>>
>>>>> The above is for the kernel to build.
>>>>
>>>> That was my understanding that the comparison issue is related to in-kernel filtering ?
>>>> If that's the case, I completely agree that the type kernel code sees does not _have_
>>>> to be the same thing that is exposed to userspace if that simplifies the problem.
>>>
>>> Yes, and note the patch I sent out to fix this.
>>
>> I did, that's some seriously fast TAT :)
> 
> I'm going to pull that one in and start testing, so I can push it out in
> this merge window.
> 
> 
> 
>>> That third part may be difficult with the above issue I mentioned.
>>>
>>> Just do:
>>>
>>>    git grep '__field(' | cut -d',' -f1 | cut -d'(' -f2 | sed -e 's/[    ]*//'
>>>    | sort -u
>>>
>>> to see what's in the kernel.
>>
>> There are lots of types, but as long as the caller knows what to ask for, it shouldn't be an issue.
>> Pretty printing the trace is obviously an important aspect and ideally requires the parser to know
>> how to format everything.
>>
>> But when it comes to other processing in a compiled language, it's not a big burden to let people
>> declare the events they require and the expected fields + types so they can get the data into their
>> own struct (e.g. as with serde or any equivalent technology).
> 
> That's what is done today. The size and offset is how the tools can get to
> the data and it knows what to do with it.
> 
> I'm not sure how rust can handle this type of opaque type scheme.

In a similar way, the user asks to parse e.g. an UnsignedLongArray and
the lib returns either the 32bit or 64bit variant, with mismatching endianness
fixed up (so the data are in host endianness). The caller has to handle both 32bit
and 64bits cases so everything is type safe.

> 
> 
>>> Note, I put much more effort into the offset, size and sign than the type.
>>> But is this only for the arrays that you have these restrictions, or any
>>> field type?
>>
>> In terms of borken pretty printing, it's a general issue not limited to dynamic arrays.
> 
> Yeah, the pretty printing can easily fail, as it can have anything that the
> kernel can do. Including calling functions that are not available to user
> space. This is why the fallback is always back to size, offset and sign.
> 
>> The only ways pretty printing for an opaque type can possibly work for new types the parser has no
>> specific knowledge of are:
>> 1. The type is not actually opaque, i.e. it comes with some decoding schema (just like the events have
>>      a schema listing their fields + types)
>> 2. The type is opaque, but also ships with an executable description of how to print it.
>>      E.g. if there was a WASM/eBPF/whatever bytecode printing routine made available to userspace.
>>
>> Option (2) is not so appealing as it's both hard to achieve and only allows a fixed set of
>> behaviors for a type. Option (1) is a lot easier and allows the behaviors to be defined
>> on the user side.
>>
>> Wild idea: include the BTF blob in the trace.dat header so no type is opaque anymore. The printing
>> issue is not entirely solved this way (e.g. cpumask still needs some plugin to be displayed as a list
>> of CPUs), but we could at least print all structs in "raw" mode and enum symbolically.
> 
> And how big is that blob?

Less than 3MB AFAIR. For comparison, /proc/kallsyms takes 11MB on my machine.
In one of my test traces, it's 17MB. Maybe it's compressed in trace.dat v7 though, haven't checked.
  

> I'm not against the idea, but I would like it to only hold what is needed

I suppose it's doable to derive a subset of the info with some efforts.

> 
>>
>> That could also allow creating a quick&dirty way of defining a proper event (aka not trace_printk()):
> 
> 
> I prefer not to have "quick&dirty" ;-)

I'm not saying that I would like to see such quick and dirty events upstream, but the reality around me is
that ftrace events is the only sane way of having an idea what the scheduler does. This means people need
to create experiments all the time with ad-hoc trace events, on top of the trace events that we attach to
tracepoints via a module. Currently, people use trace_printk() for that, which comes with some significant
amount of work and pain (mostly regex speed).

That said having just looked at bprint, I could probably support trace_printk() format strings with simple
struct member access (i.e. no __printflags shenanigans etc) as normal events relatively easily. It's even
possible to use the fmt string pointer as an "event ID". Still a shame that all the event field format infra
basically gets duplicated in a printf format string ...

> 
>>
>> 	#define SIMPLE_TRACE_EVENT(type, fields) \
>> 	struct type fields;	
>> 	TRACE_EVENT(type, \
>> 		TP_PROTO(struct type *data), \
>> 		TP_ARGS(data), \
>> 		TP_STRUCT__entry(__field(struct type, data)), \
>> 		TP_fast_assign(__entry->data = *data;), \
>> 		TP_printk("print in raw mode to display the data"), \
>> 	);
>> 	#define SIMPLE_TRACE(type, fields) trace_struct_##type(&(struct type)fields)
>>
>>
>> 	SIMPLE_TRACE_EVENT(myevent, {
>> 		char name[11];
>> 		int foobar;
>> 	});
>> 	
>> 	SIMPLE_TRACE(myevent, {.name = "hello", .foobar = 42});
> 
> 
>>
>> The format string could be either kernel-generated based on BTF or userspace could be expected
>> to make its own use of BTF.
> 
> What's the use case for the above?

An equivalent to trace_printk() that exposes its fields in the "normal" way rather than having to parse
the format string and a comma-separated list of C expressions. Life is too short to write C interpreters.
Parsing BTF is at least a finite amount of work. But I guess it would be easy to handle only "REC->field"
expressions.

-- Douglas
  

On Tue, 13 Dec 2022 17:40:06 +0000
Douglas Raillard <douglas.raillard@arm.com> wrote:

> > I prefer not to have "quick&dirty" ;-)  
> 
> I'm not saying that I would like to see such quick and dirty events upstream, but the reality around me is
> that ftrace events is the only sane way of having an idea what the scheduler does. This means people need
> to create experiments all the time with ad-hoc trace events, on top of the trace events that we attach to
> tracepoints via a module. Currently, people use trace_printk() for that, which comes with some significant
> amount of work and pain (mostly regex speed).

Have you seen custom events?

  https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/samples/trace_events/trace_custom_sched.h


> 
> That said having just looked at bprint, I could probably support trace_printk() format strings with simple
> struct member access (i.e. no __printflags shenanigans etc) as normal events relatively easily. It's even
> possible to use the fmt string pointer as an "event ID". Still a shame that all the event field format infra
> basically gets duplicated in a printf format string ...

It's the easiest thing to do in the kernel. The below is probably too much
work for people to use. The fact that it's just a string and does not have
any type information is one of the main reasons I force it not to be in
mainline (hence the nasty banner when it is added).

> 
> >   
> >>
> >> 	#define SIMPLE_TRACE_EVENT(type, fields) \
> >> 	struct type fields;	
> >> 	TRACE_EVENT(type, \
> >> 		TP_PROTO(struct type *data), \
> >> 		TP_ARGS(data), \
> >> 		TP_STRUCT__entry(__field(struct type, data)), \
> >> 		TP_fast_assign(__entry->data = *data;), \
> >> 		TP_printk("print in raw mode to display the data"), \
> >> 	);
> >> 	#define SIMPLE_TRACE(type, fields) trace_struct_##type(&(struct type)fields)
> >>
> >>
> >> 	SIMPLE_TRACE_EVENT(myevent, {
> >> 		char name[11];
> >> 		int foobar;
> >> 	});
> >> 	
> >> 	SIMPLE_TRACE(myevent, {.name = "hello", .foobar = 42});  
> > 
> >   
> >>
> >> The format string could be either kernel-generated based on BTF or userspace could be expected
> >> to make its own use of BTF.  
> > 
> > What's the use case for the above?  
> 
> An equivalent to trace_printk() that exposes its fields in the "normal" way rather than having to parse
> the format string and a comma-separated list of C expressions. Life is too short to write C interpreters.
> Parsing BTF is at least a finite amount of work. But I guess it would be easy to handle only "REC->field"
> expressions.

But the above isn't that much simpler than writing a trace event. When I
use trace_printk(), I seldom use it with tooling. And for the few times I
have written tools to parse printk, the printk formats are very easily
parsed, as I control them. Heck, I'd just do colon delimited string.

-- Steve
  

On 13-12-2022 14:20, Douglas Raillard wrote:
>
> The only case I can think of where parsing would not follow regular C abstract declaration syntax is a type like that:
> 
>      __data_loc int [3][]

After some experimentation, I came to the conclusion that "__data_loc <type> [] <id>" can indeed support any C type if they
need to be added in the future. I would heavily suggest that any future extension works as the following, to avoid messing up
the grammar in a subtle way that would prevent some types to be expressible, or make it a nightmare to implement.

The recipe to parse that with stock C parser is:
* consume "__data_loc"
* parse "<type> []" as a func prototype parameter declaration
  (declaration using an abstract declarator, i.e. not introducing any identifier)
* parse "<id>" as an identifier.

> The outer-most array is by definition the dynamic one, so "[]". In normal C, [3] and [] would be swapped as
> the outer-most array comes first. That's not too bad though as it is not ambiguous and easy to fixup directly
> in the parse tree.

Simply swapping is wrong in the general case. The correct modification of the "<type> []" parse tree is doing a
"barrel shift" on nested array sizes. If the type is "int [1][2][]", it needs to be turned into "int [][1][2]".
The now-top-level sizeless array is the dynamic array. Note that pointers level are transparent,
so "int (*[1])[2][]" needs to be turned into "int (*[])[1][2]"

-- Douglas
  

On 13-12-2022 19:44, Steven Rostedt wrote:
> On Tue, 13 Dec 2022 17:40:06 +0000
> Douglas Raillard <douglas.raillard@arm.com> wrote:
> 
>>> I prefer not to have "quick&dirty" ;-)
>>
>> I'm not saying that I would like to see such quick and dirty events upstream, but the reality around me is
>> that ftrace events is the only sane way of having an idea what the scheduler does. This means people need
>> to create experiments all the time with ad-hoc trace events, on top of the trace events that we attach to
>> tracepoints via a module. Currently, people use trace_printk() for that, which comes with some significant
>> amount of work and pain (mostly regex speed).
> 
> Have you seen custom events?
> 
>    https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/samples/trace_events/trace_custom_sched.h
> 

That's interesting, thanks. I currently define a new regular event along with a tp probe that I attach to the tp of interest.
The probe is then calling trace_mycustom_event(). However:
* Some events are emitted from multiple tracepoints. TRACE_CUSTOM_EVENT() can only attach to one tp, but maybe it's solvable
   with an event class.
* The tracepoint may or may not exist (we deal with all sorts of kernels including Android and our own dev kernel with extra tps).
   I use __find_tracepoint() and tracepoint_probe_register() to handle that.
* The module we have [1] is organized as a set of features that can be enabled dynamically by name as insmod parameter.
   The probe is registered only if the corresponding event feature is enabled. This allows an event feature to have
   side effects that would not be tolerable unless you want that specific event.

However TRACE_CUSTOM_EVENT() sounds great for simpler use cases, especially as (some) maintainers are much happier
with new tracepoints (no stable ABI) than new events (too easy for userspace to consume, ABI and tears when it breaks ensues)

[1] https://github.com/ARM-software/lisa/tree/master/lisa/_assets/kmodules/sched_tp


>>
>> That said having just looked at bprint, I could probably support trace_printk() format strings with simple
>> struct member access (i.e. no __printflags shenanigans etc) as normal events relatively easily. It's even
>> possible to use the fmt string pointer as an "event ID". Still a shame that all the event field format infra
>> basically gets duplicated in a printf format string ...
> 
> It's the easiest thing to do in the kernel. The below is probably too much
> work for people to use. The fact that it's just a string and does not have
> any type information is one of the main reasons I force it not to be in
> mainline (hence the nasty banner when it is added).
> 
>>
>>>    
>>>>
>>>> 	#define SIMPLE_TRACE_EVENT(type, fields) \
>>>> 	struct type fields;	
>>>> 	TRACE_EVENT(type, \
>>>> 		TP_PROTO(struct type *data), \
>>>> 		TP_ARGS(data), \
>>>> 		TP_STRUCT__entry(__field(struct type, data)), \
>>>> 		TP_fast_assign(__entry->data = *data;), \
>>>> 		TP_printk("print in raw mode to display the data"), \
>>>> 	);
>>>> 	#define SIMPLE_TRACE(type, fields) trace_struct_##type(&(struct type)fields)
>>>>
>>>>
>>>> 	SIMPLE_TRACE_EVENT(myevent, {
>>>> 		char name[11];
>>>> 		int foobar;
>>>> 	});
>>>> 	
>>>> 	SIMPLE_TRACE(myevent, {.name = "hello", .foobar = 42});
>>>
>>>    
>>>>
>>>> The format string could be either kernel-generated based on BTF or userspace could be expected
>>>> to make its own use of BTF.
>>>
>>> What's the use case for the above?
>>
>> An equivalent to trace_printk() that exposes its fields in the "normal" way rather than having to parse
>> the format string and a comma-separated list of C expressions. Life is too short to write C interpreters.
>> Parsing BTF is at least a finite amount of work. But I guess it would be easy to handle only "REC->field"
>> expressions.
> 
> But the above isn't that much simpler than writing a trace event.



The user would simply need this:

	SIMPLE_TRACE_EVENT(myevent, {
		char name[11];
		int foobar;
	});

	SIMPLE_TRACE(myevent, {.name = "hello", .foobar = 42});

Instead of currently that:

	TRACE_EVENT(myevent
		TP_PROTO(struct type *name[11], int foobar),
		TP_ARGS(name, foobar),
		TP_STRUCT__entry(
			__field(name[11], name),
			__field(int, foobar),
			),
		TP_fast_assign(
			__entry->name = *name;
			__entry->foobar = *foobar;
		),
		TP_printk("name=%s foobar=%i", __entry->name, __entry->foobar), \
	);

	trace_myevent(...);

We went from 3 lines of declaration to 12. "foobar" is repeated 7 times. While I agree that in the grand
scheme of things it's totally acceptable, people want to make that happen in 1 line, not have to repeat
themselves 7 times and then faff around with the #define CREATE_TRACE_POINTS etc.


> When I use trace_printk(), I seldom use it with tooling. And for the few times I
> have written tools to parse printk, the printk formats are very easily
> parsed, as I control them. Heck, I'd just do colon delimited string.

In the sort of things we do, people will want to put a trace_printk() to log a signal, and then get it as an overlay
on an existing per-task plot in a Jupyter notebook. This needs to take less than 15s (kernel compilation + reboot excluded).
So that means:

* We _always_ use tooling. In that case, tooling calls trace-cmd, parses output and loads it in a pandas dataframe, and from
   there either you do computations on it or you simply plot it (matplotlib, plotly, bokeh, whatever).

* As a result, the format must be machine-parsable in a robust way. The parser is written once and for all, no-one wants
   to have to write a regex to load a throw-away event. If whatever is provided requires customization to work on such a simple
   field-extraction case it's considered broken.

* The current format is ok-ish and works well for a human eye, but for all other purposes it's just a bad version of JSON:
	* untyped, you never know if a value is an integer or a string containing an integer. Any data pipeline
           will need to know that, even in dynamically typed language like Python (see numpy/pandas).

	* unquoted/escaped string. If I have a "comm" field, I'll end up with "comm=mytask foobar=1",
           but also "comm=mytask with spaces foobar=1" and why not "comm=mytask=1 foobar=1". So far, I never encountered
           task names with "=" in the wild but it would break parsing without hope of fixing. Android traces are full of tasks
           with spaces in their names, which makes the regex a lot less obvious and likely quite slower as well.
	
	* It's custom. That alone means (1) there is no off-the-shelve solution and (2) custom solutions will be slower, especially
           in languages like Python. You can find tons of JSON parsers, including ones using SIMD. Trying to match that speed with
	  a pure python implementation is not even remotely feasible, so it would require a C extension module, leading to having
	  to distribute a bunch of prebuilt binaries etc, which is _a lot_ more work than "import json; json.loads(a_line)".

So all that to say that concisely defining a structured event with a schema known by the tooling is _very_ valuable when you get
passed eyeballing a trace. The good news is that trace_printk() does provide a schema with its format string, it's just annoying
that it's a pure duplication of infrastructure wrt to other events that have their fields format reported in another way. Also
it's less powerful but we can live with that for quick experiments.

Maybe it would be feasible to write a function that takes a printk fmt string as input and creates a matching synthetic event. This
way we still get a regular event from userspace PoV, with the ease of definition of a format string in the code.

-- Douglas