Pull tracing updates from Steven Rostedt:
- Allow kernel trace instance creation to specify what events are
created
Inside the kernel, a subsystem may create a tracing instance that it
can use to send events to user space. This sub-system may not care
about the thousands of events that exist in eventfs. Allow the
sub-system to specify what sub-systems of events it cares about, and
only those events are exposed to this instance.
- Allow the ring buffer to be broken up into bigger sub-buffers than
just the architecture page size.
A new tracefs file called "buffer_subbuf_size_kb" is created. The
user can now specify a minimum size the sub-buffer may be in
kilobytes. Note, that the implementation currently make the
sub-buffer size a power of 2 pages (1, 2, 4, 8, 16, ...) but the user
only writes in kilobyte size, and the sub-buffer will be updated to
the next size that it will can accommodate it. If the user writes in
10, it will change the size to be 4 pages on x86 (16K), as that is
the next available size that can hold 10K pages.
- Update the debug output when a corrupt time is detected in the ring
buffer. If the ring buffer detects inconsistent timestamps, there's a
debug config options that will dump the contents of the meta data of
the sub-buffer that is used for debugging. Add some more information
to this dump that helps with debugging.
- Add more timestamp debugging checks (only triggers when the config is
enabled)
- Increase the trace_seq iterator to 2 page sizes.
- Allow strings written into tracefs_marker to be larger. Up to just
under 2 page sizes (based on what trace_seq can hold).
- Increase the trace_maker_raw write to be as big as a sub-buffer can
hold.
- Remove 32 bit time stamp logic, now that the rb_time_cmpxchg() has
been removed.
- More selftests were added.
- Some code clean ups as well.
* tag 'trace-v6.8' of git://git.kernel.org/pub/scm/linux/kernel/git/trace/linux-trace: (29 commits)
ring-buffer: Remove stale comment from ring_buffer_size()
tracing histograms: Simplify parse_actions() function
tracing/selftests: Remove exec permissions from trace_marker.tc test
ring-buffer: Use subbuf_order for buffer page masking
tracing: Update subbuffer with kilobytes not page order
ringbuffer/selftest: Add basic selftest to test changing subbuf order
ring-buffer: Add documentation on the buffer_subbuf_order file
ring-buffer: Just update the subbuffers when changing their allocation order
ring-buffer: Keep the same size when updating the order
tracing: Stop the tracing while changing the ring buffer subbuf size
tracing: Update snapshot order along with main buffer order
ring-buffer: Make sure the spare sub buffer used for reads has same size
ring-buffer: Do no swap cpu buffers if order is different
ring-buffer: Clear pages on error in ring_buffer_subbuf_order_set() failure
ring-buffer: Read and write to ring buffers with custom sub buffer size
ring-buffer: Set new size of the ring buffer sub page
ring-buffer: Add interface for configuring trace sub buffer size
ring-buffer: Page size per ring buffer
ring-buffer: Have ring_buffer_print_page_header() be able to access ring_buffer_iter
ring-buffer: Check if absolute timestamp goes backwards
...
If an application blocks on the snapshot or snapshot_raw files, expecting
to be woken up when a snapshot occurs, it will not happen. Or it may
happen with an unexpected result.
That result is that the application will be reading the main buffer
instead of the snapshot buffer. That is because when the snapshot occurs,
the main and snapshot buffers are swapped. But the reader has a descriptor
still pointing to the buffer that it originally connected to.
This is fine for the main buffer readers, as they may be blocked waiting
for a watermark to be hit, and when a snapshot occurs, the data that the
main readers want is now on the snapshot buffer.
But for waiters of the snapshot buffer, they are waiting for an event to
occur that will trigger the snapshot and they can then consume it quickly
to save the snapshot before the next snapshot occurs. But to do this, they
need to read the new snapshot buffer, not the old one that is now
receiving new data.
Also, it does not make sense to have a watermark "buffer_percent" on the
snapshot buffer, as the snapshot buffer is static and does not receive new
data except all at once.
Link: https://lore.kernel.org/linux-trace-kernel/20231228095149.77f5b45d@gandalf.local.home
Cc: stable@vger.kernel.org
Cc: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Cc: Mark Rutland <mark.rutland@arm.com>
Acked-by: Masami Hiramatsu (Google) <mhiramat@kernel.org>
Fixes: debdd57f51 ("tracing: Make a snapshot feature available from userspace")
Signed-off-by: Steven Rostedt (Google) <rostedt@goodmis.org>
The tracefs file "buffer_percent" is to allow user space to set a
water-mark on how much of the tracing ring buffer needs to be filled in
order to wake up a blocked reader.
0 - is to wait until any data is in the buffer
1 - is to wait for 1% of the sub buffers to be filled
50 - would be half of the sub buffers are filled with data
100 - is not to wake the waiter until the ring buffer is completely full
Unfortunately the test for being full was:
dirty = ring_buffer_nr_dirty_pages(buffer, cpu);
return (dirty * 100) > (full * nr_pages);
Where "full" is the value for "buffer_percent".
There is two issues with the above when full == 100.
1. dirty * 100 > 100 * nr_pages will never be true
That is, the above is basically saying that if the user sets
buffer_percent to 100, more pages need to be dirty than exist in the
ring buffer!
2. The page that the writer is on is never considered dirty, as dirty
pages are only those that are full. When the writer goes to a new
sub-buffer, it clears the contents of that sub-buffer.
That is, even if the check was ">=" it would still not be equal as the
most pages that can be considered "dirty" is nr_pages - 1.
To fix this, add one to dirty and use ">=" in the compare.
Link: https://lore.kernel.org/linux-trace-kernel/20231226125902.4a057f1d@gandalf.local.home
Cc: stable@vger.kernel.org
Cc: Mark Rutland <mark.rutland@arm.com>
Cc: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Acked-by: Masami Hiramatsu (Google) <mhiramat@kernel.org>
Fixes: 03329f9939 ("tracing: Add tracefs file buffer_percentage")
Signed-off-by: Steven Rostedt (Google) <rostedt@goodmis.org>
The ring_buffer_subbuf_order_set() was creating ring_buffer_per_cpu
cpu_buffers with the new subbuffers with the updated order, and if they
all successfully were created, then they the ring_buffer's per_cpu buffers
would be freed and replaced by them.
The problem is that the freed per_cpu buffers contains state that would be
lost. Running the following commands:
1. # echo 3 > /sys/kernel/tracing/buffer_subbuf_order
2. # echo 0 > /sys/kernel/tracing/tracing_cpumask
3. # echo 1 > /sys/kernel/tracing/snapshot
4. # echo ff > /sys/kernel/tracing/tracing_cpumask
5. # echo test > /sys/kernel/tracing/trace_marker
Would result in:
-bash: echo: write error: Bad file descriptor
That's because the state of the per_cpu buffers of the snapshot buffer is
lost when the order is changed (the order of a freed snapshot buffer goes
to 0 to save memory, and when the snapshot buffer is allocated again, it
goes back to what the main buffer is).
In operation 2, the snapshot buffers were set to "disable" (as all the
ring buffers CPUs were disabled).
In operation 3, the snapshot is allocated and a call to
ring_buffer_subbuf_order_set() replaced the per_cpu buffers losing the
"record_disable" count.
When it was enabled again, the atomic_dec(&cpu_buffer->record_disable) was
decrementing a zero, setting it to -1. Writing 1 into the snapshot would
swap the snapshot buffer with the main buffer, so now the main buffer is
"disabled", and nothing can write to the ring buffer anymore.
Instead of creating new per_cpu buffers and losing the state of the old
buffers, basically do what the resize does and just allocate new subbuf
pages into the new_pages link list of the per_cpu buffer and if they all
succeed, then replace the old sub buffers with the new ones. This keeps
the per_cpu buffer descriptor in tact and by doing so, keeps its state.
Link: https://lore.kernel.org/linux-trace-kernel/20231219185630.944104939@goodmis.org
Cc: Masami Hiramatsu <mhiramat@kernel.org>
Cc: Mark Rutland <mark.rutland@arm.com>
Cc: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Tzvetomir Stoyanov <tz.stoyanov@gmail.com>
Cc: Vincent Donnefort <vdonnefort@google.com>
Cc: Kent Overstreet <kent.overstreet@gmail.com>
Fixes: f9b94daa54 ("ring-buffer: Set new size of the ring buffer sub page")
Signed-off-by: Steven Rostedt (Google) <rostedt@goodmis.org>
The function ring_buffer_subbuf_order_set() just updated the sub-buffers
to the new size, but this also changes the size of the buffer in doing so.
As the size is determined by nr_pages * subbuf_size. If the subbuf_size is
increased without decreasing the nr_pages, this causes the total size of
the buffer to increase.
This broke the latency tracers as the snapshot needs to be the same size
as the main buffer. The size of the snapshot buffer is only expanded when
needed, and because the order is still the same, the size becomes out of
sync with the main buffer, as the main buffer increased in size without
the tracing system knowing.
Calculate the nr_pages to allocate with the new subbuf_size to be
buffer_size / new_subbuf_size.
Link: https://lore.kernel.org/linux-trace-kernel/20231219185630.649397785@goodmis.org
Cc: Masami Hiramatsu <mhiramat@kernel.org>
Cc: Mark Rutland <mark.rutland@arm.com>
Cc: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Tzvetomir Stoyanov <tz.stoyanov@gmail.com>
Cc: Vincent Donnefort <vdonnefort@google.com>
Cc: Kent Overstreet <kent.overstreet@gmail.com>
Fixes: f9b94daa54 ("ring-buffer: Set new size of the ring buffer sub page")
Signed-off-by: Steven Rostedt (Google) <rostedt@goodmis.org>
The check_buffer() which checks the timestamps of the ring buffer
sub-buffer page, when enabled, only checks if the adding of deltas of the
events from the last absolute timestamp or the timestamp of the sub-buffer
page adds up to the current event.
What it does not check is if the absolute timestamp causes the time of the
events to go backwards, as that can cause issues elsewhere.
Test for the timestamp going backwards too.
This also fixes a slight issue where if the warning triggers at boot up
(because of the resetting of the tsc), it will disable all further checks,
even those that are after boot Have it continue checking if the warning
was ignored during boot up.
Link: https://lore.kernel.org/linux-trace-kernel/20231219074732.18b092d4@gandalf.local.home
Cc: Masami Hiramatsu <mhiramat@kernel.org>
Cc: Mark Rutland <mark.rutland@arm.com>
Cc: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Signed-off-by: Steven Rostedt (Google) <rostedt@goodmis.org>
When the ring buffer timestamp verifier triggers, it dumps the content of
the sub-buffer. But currently it only dumps the timestamps and the offset
of the data as well as the deltas. It would be even more informative if
the event data also showed the interrupt context level it was in.
That is, if each event showed that the event was written in normal,
softirq, irq or NMI context. Then a better idea about how the events may
have been interrupted from each other.
As the payload of the ring buffer is really a black box of the ring
buffer, just assume that if the payload is larger than a trace entry, that
it is a trace entry. As trace entries have the interrupt context
information saved in a flags field, look at that location and report the
output of the flags.
If the payload is not a trace entry, there's no way to really know, and
the information will be garbage. But that's OK, because this is for
debugging only (this output is not used in production as the buffer check
that calls it causes a huge overhead to the tracing). This information,
when available, is crucial for debugging timestamp issues. If it's
garbage, it will also be pretty obvious that its garbage too.
As this output usually happens in kselftests of the tracing code, the user
will know what the payload is at the time.
Link: https://lore.kernel.org/linux-trace-kernel/20231219074542.6f304601@gandalf.local.home
Cc: Masami Hiramatsu <mhiramat@kernel.org>
Cc: Mark Rutland <mark.rutland@arm.com>
Cc: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Suggested-by: Joel Fernandes (Google) <joel@joelfernandes.org>
Signed-off-by: Steven Rostedt (Google) <rostedt@goodmis.org>
On bugs that have the ring buffer timestamp get out of sync, the config
CONFIG_RING_BUFFER_VALIDATE_TIME_DELTAS, that checks for it and if it is
detected it causes a dump of the bad sub buffer.
It shows each event and their timestamp as well as the delta in the event.
But it's also good to see the offset into the subbuffer for that event to
know if how close to the end it is.
Also print where the last event actually ended compared to where it was
expected to end.
Link: https://lore.kernel.org/linux-trace-kernel/20231211131623.59eaebd2@gandalf.local.home
Cc: Mark Rutland <mark.rutland@arm.com>
Cc: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Acked-by: Masami Hiramatsu (Google) <mhiramat@kernel.org>
Signed-off-by: Steven Rostedt (Google) <rostedt@goodmis.org>
To synchronize the timestamps with the ring buffer reservation, there are
two timestamps that are saved in the buffer meta data.
1. before_stamp
2. write_stamp
When the two are equal, the write_stamp is considered valid, as in, it may
be used to calculate the delta of the next event as the write_stamp is the
timestamp of the previous reserved event on the buffer.
This is done by the following:
/*A*/ w = current position on the ring buffer
before = before_stamp
after = write_stamp
ts = read current timestamp
if (before != after) {
write_stamp is not valid, force adding an absolute
timestamp.
}
/*B*/ before_stamp = ts
/*C*/ write = local_add_return(event length, position on ring buffer)
if (w == write - event length) {
/* Nothing interrupted between A and C */
/*E*/ write_stamp = ts;
delta = ts - after
/*
* If nothing interrupted again,
* before_stamp == write_stamp and write_stamp
* can be used to calculate the delta for
* events that come in after this one.
*/
} else {
/*
* The slow path!
* Was interrupted between A and C.
*/
This is the place that there's a bug. We currently have:
after = write_stamp
ts = read current timestamp
/*F*/ if (write == current position on the ring buffer &&
after < ts && cmpxchg(write_stamp, after, ts)) {
delta = ts - after;
} else {
delta = 0;
}
The assumption is that if the current position on the ring buffer hasn't
moved between C and F, then it also was not interrupted, and that the last
event written has a timestamp that matches the write_stamp. That is the
write_stamp is valid.
But this may not be the case:
If a task context event was interrupted by softirq between B and C.
And the softirq wrote an event that got interrupted by a hard irq between
C and E.
and the hard irq wrote an event (does not need to be interrupted)
We have:
/*B*/ before_stamp = ts of normal context
---> interrupted by softirq
/*B*/ before_stamp = ts of softirq context
---> interrupted by hardirq
/*B*/ before_stamp = ts of hard irq context
/*E*/ write_stamp = ts of hard irq context
/* matches and write_stamp valid */
<----
/*E*/ write_stamp = ts of softirq context
/* No longer matches before_stamp, write_stamp is not valid! */
<---
w != write - length, go to slow path
// Right now the order of events in the ring buffer is:
//
// |-- softirq event --|-- hard irq event --|-- normal context event --|
//
after = write_stamp (this is the ts of softirq)
ts = read current timestamp
if (write == current position on the ring buffer [true] &&
after < ts [true] && cmpxchg(write_stamp, after, ts) [true]) {
delta = ts - after [Wrong!]
The delta is to be between the hard irq event and the normal context
event, but the above logic made the delta between the softirq event and
the normal context event, where the hard irq event is between the two. This
will shift all the remaining event timestamps on the sub-buffer
incorrectly.
The write_stamp is only valid if it matches the before_stamp. The cmpxchg
does nothing to help this.
Instead, the following logic can be done to fix this:
before = before_stamp
ts = read current timestamp
before_stamp = ts
after = write_stamp
if (write == current position on the ring buffer &&
after == before && after < ts) {
delta = ts - after
} else {
delta = 0;
}
The above will only use the write_stamp if it still matches before_stamp
and was tested to not have changed since C.
As a bonus, with this logic we do not need any 64-bit cmpxchg() at all!
This means the 32-bit rb_time_t workaround can finally be removed. But
that's for a later time.
Link: https://lore.kernel.org/linux-trace-kernel/20231218175229.58ec3daf@gandalf.local.home/
Link: https://lore.kernel.org/linux-trace-kernel/20231218230712.3a76b081@gandalf.local.home
Cc: stable@vger.kernel.org
Cc: Masami Hiramatsu <mhiramat@kernel.org>
Cc: Mark Rutland <mark.rutland@arm.com>
Cc: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Fixes: dd93942570 ("ring-buffer: Do not try to put back write_stamp")
Signed-off-by: Steven Rostedt (Google) <rostedt@goodmis.org>
The following race can cause rb_time_read() to observe a corrupted time
stamp:
rb_time_cmpxchg()
[...]
if (!rb_time_read_cmpxchg(&t->msb, msb, msb2))
return false;
if (!rb_time_read_cmpxchg(&t->top, top, top2))
return false;
<interrupted before updating bottom>
__rb_time_read()
[...]
do {
c = local_read(&t->cnt);
top = local_read(&t->top);
bottom = local_read(&t->bottom);
msb = local_read(&t->msb);
} while (c != local_read(&t->cnt));
*cnt = rb_time_cnt(top);
/* If top and msb counts don't match, this interrupted a write */
if (*cnt != rb_time_cnt(msb))
return false;
^ this check fails to catch that "bottom" is still not updated.
So the old "bottom" value is returned, which is wrong.
Fix this by checking that all three of msb, top, and bottom 2-bit cnt
values match.
The reason to favor checking all three fields over requiring a specific
update order for both rb_time_set() and rb_time_cmpxchg() is because
checking all three fields is more robust to handle partial failures of
rb_time_cmpxchg() when interrupted by nested rb_time_set().
Link: https://lore.kernel.org/lkml/20231211201324.652870-1-mathieu.desnoyers@efficios.com/
Link: https://lore.kernel.org/linux-trace-kernel/20231212193049.680122-1-mathieu.desnoyers@efficios.com
Fixes: f458a14534 ("ring-buffer: Test last update in 32bit version of __rb_time_read()")
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Signed-off-by: Steven Rostedt (Google) <rostedt@goodmis.org>
Mathieu Desnoyers pointed out an issue in the rb_time_cmpxchg() for 32 bit
architectures. That is:
static bool rb_time_cmpxchg(rb_time_t *t, u64 expect, u64 set)
{
unsigned long cnt, top, bottom, msb;
unsigned long cnt2, top2, bottom2, msb2;
u64 val;
/* The cmpxchg always fails if it interrupted an update */
if (!__rb_time_read(t, &val, &cnt2))
return false;
if (val != expect)
return false;
<<<< interrupted here!
cnt = local_read(&t->cnt);
The problem is that the synchronization counter in the rb_time_t is read
*after* the value of the timestamp is read. That means if an interrupt
were to come in between the value being read and the counter being read,
it can change the value and the counter and the interrupted process would
be clueless about it!
The counter needs to be read first and then the value. That way it is easy
to tell if the value is stale or not. If the counter hasn't been updated,
then the value is still good.
Link: https://lore.kernel.org/linux-trace-kernel/20231211201324.652870-1-mathieu.desnoyers@efficios.com/
Link: https://lore.kernel.org/linux-trace-kernel/20231212115301.7a9c9a64@gandalf.local.home
Cc: stable@vger.kernel.org
Cc: Masami Hiramatsu <mhiramat@kernel.org>
Cc: Mark Rutland <mark.rutland@arm.com>
Fixes: 10464b4aa6 ("ring-buffer: Add rb_time_t 64 bit operations for speeding up 32 bit")
Reported-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Reviewed-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Signed-off-by: Steven Rostedt (Google) <rostedt@goodmis.org>
When filtering is enabled, a temporary buffer is created to place the
content of the trace event output so that the filter logic can decide
from the trace event output if the trace event should be filtered out or
not. If it is to be filtered out, the content in the temporary buffer is
simply discarded, otherwise it is written into the trace buffer.
But if an interrupt were to come in while a previous event was using that
temporary buffer, the event written by the interrupt would actually go
into the ring buffer itself to prevent corrupting the data on the
temporary buffer. If the event is to be filtered out, the event in the
ring buffer is discarded, or if it fails to discard because another event
were to have already come in, it is turned into padding.
The update to the write_stamp in the rb_try_to_discard() happens after a
fix was made to force the next event after the discard to use an absolute
timestamp by setting the before_stamp to zero so it does not match the
write_stamp (which causes an event to use the absolute timestamp).
But there's an effort in rb_try_to_discard() to put back the write_stamp
to what it was before the event was added. But this is useless and
wasteful because nothing is going to be using that write_stamp for
calculations as it still will not match the before_stamp.
Remove this useless update, and in doing so, we remove another
cmpxchg64()!
Also update the comments to reflect this change as well as remove some
extra white space in another comment.
Link: https://lore.kernel.org/linux-trace-kernel/20231215081810.1f4f38fe@rorschach.local.home
Cc: Masami Hiramatsu <mhiramat@kernel.org>
Cc: Mark Rutland <mark.rutland@arm.com>
Cc: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Cc: Joel Fernandes <joel@joelfernandes.org>
Cc: Vincent Donnefort <vdonnefort@google.com>
Fixes: b2dd797543 ("ring-buffer: Force absolute timestamp on discard of event")
Signed-off-by: Steven Rostedt (Google) <rostedt@goodmis.org>
