The ring buffer is broken up into sub buffers (currently of page size).
Each sub buffer has a pointer to its "tail" (the last event written to the
sub buffer). When a new event is requested, the tail is locally
incremented to cover the size of the new event. This is done in a way that
there is no need for locking.
If the tail goes past the end of the sub buffer, the process of moving to
the next sub buffer takes place. After setting the current sub buffer to
the next one, the previous one that had the tail go passed the end of the
sub buffer needs to be reset back to the original tail location (before
the new event was requested) and the rest of the sub buffer needs to be
"padded".
The race happens when a reader takes control of the sub buffer. As readers
do a "swap" of sub buffers from the ring buffer to get exclusive access to
the sub buffer, it replaces the "head" sub buffer with an empty sub buffer
that goes back into the writable portion of the ring buffer. This swap can
happen as soon as the writer moves to the next sub buffer and before it
updates the last sub buffer with padding.
Because the sub buffer can be released to the reader while the writer is
still updating the padding, it is possible for the reader to see the event
that goes past the end of the sub buffer. This can cause obvious issues.
To fix this, add a few memory barriers so that the reader definitely sees
the updates to the sub buffer, and also waits until the writer has put
back the "tail" of the sub buffer back to the last event that was written
on it.
To be paranoid, it will only spin for 1 second, otherwise it will
warn and shutdown the ring buffer code. 1 second should be enough as
the writer does have preemption disabled. If the writer doesn't move
within 1 second (with preemption disabled) something is horribly
wrong. No interrupt should last 1 second!
Link: https://lore.kernel.org/all/20220830120854.7545-1-jiazi.li@transsion.com/
Link: https://bugzilla.kernel.org/show_bug.cgi?id=216369
Link: https://lkml.kernel.org/r/20220929104909.0650a36c@gandalf.local.home
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: stable@vger.kernel.org
Fixes: c7b0930857 ("ring-buffer: prevent adding write in discarded area")
Reported-by: Jiazi.Li <jiazi.li@transsion.com>
Signed-off-by: Steven Rostedt (Google) <rostedt@goodmis.org>
The wake up waiters only checks the "wakeup_full" variable and not the
"full_waiters_pending". The full_waiters_pending is set when a waiter is
added to the wait queue. The wakeup_full is only set when an event is
triggered, and it clears the full_waiters_pending to avoid multiple calls
to irq_work_queue().
The irq_work callback really needs to check both wakeup_full as well as
full_waiters_pending such that this code can be used to wake up waiters
when a file is closed that represents the ring buffer and the waiters need
to be woken up.
Link: https://lkml.kernel.org/r/20220927231824.209460321@goodmis.org
Cc: stable@vger.kernel.org
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Andrew Morton <akpm@linux-foundation.org>
Fixes: 15693458c4 ("tracing/ring-buffer: Move poll wake ups into ring buffer code")
Signed-off-by: Steven Rostedt (Google) <rostedt@goodmis.org>
If a page is partially read, and then the splice system call is run
against the ring buffer, it will always fail to read, no matter how much
is in the ring buffer. That's because the code path for a partial read of
the page does will fail if the "full" flag is set.
The splice system call wants full pages, so if the read of the ring buffer
is not yet full, it should return zero, and the splice will block. But if
a previous read was done, where the beginning has been consumed, it should
still be given to the splice caller if the rest of the page has been
written to.
This caused the splice command to never consume data in this scenario, and
let the ring buffer just fill up and lose events.
Link: https://lkml.kernel.org/r/20220927144317.46be6b80@gandalf.local.home
Cc: stable@vger.kernel.org
Fixes: 8789a9e7df ("ring-buffer: read page interface")
Signed-off-by: Steven Rostedt (Google) <rostedt@goodmis.org>
When the new logic was made to handle deltas of events from interrupts
that interrupted other events, it required 64 bit local atomics.
Unfortunately, 64 bit local atomics are expensive on 32 bit architectures.
Thus, commit 10464b4aa6 ("ring-buffer: Add rb_time_t 64 bit operations
for speeding up 32 bit") created a type of seq lock timer for 32 bits.
It used two 32 bit local atomics, but required 2 bits from them each for
synchronization, making it only 60 bits.
Add a new "msb" field to hold the extra 4 bits that are cut off.
Link: https://lore.kernel.org/all/20220426175338.3807ca4f@gandalf.local.home/
Link: https://lkml.kernel.org/r/20220427170812.53cc7139@gandalf.local.home
Signed-off-by: Steven Rostedt (Google) <rostedt@goodmis.org>
There's an absolute timestamp event in the ring buffer, but this only
saves 59 bits of the timestamp, as the 5 MSB is used for meta data
(stating it is an absolute time stamp). This was never an issue as all the
clocks currently in use never used those 5 MSB. But now there's a new
clock (TAI) that does.
To handle this case, when reading an absolute timestamp, a previous full
timestamp is passed in, and the 5 MSB of that timestamp is OR'd to the
absolute timestamp (if any of the 5 MSB are set), and then to test for
overflow, if the new result is smaller than the passed in previous
timestamp, then 1 << 59 is added to it.
All the extra processing is done on the reader "slow" path, with the
exception of the "too big delta" check, and the reading of timestamps
for histograms.
Note, libtraceevent will need to be updated to handle this case as well.
But this is not a user space regression, as user space was never able to
handle any timestamps that used more than 59 bits.
Link: https://lore.kernel.org/all/20220426175338.3807ca4f@gandalf.local.home/
Link: https://lkml.kernel.org/r/20220427153339.16c33f75@gandalf.local.home
Cc: Tom Zanussi <zanussi@kernel.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Kurt Kanzenbach <kurt@linutronix.de>
Signed-off-by: Steven Rostedt (Google) <rostedt@goodmis.org>
The resetting of the entire ring buffer use to simply go through and reset
each individual CPU buffer that had its own protection and synchronization.
But this was very slow, due to performing a synchronization for each CPU.
The code was reshuffled to do one disabling of all CPU buffers, followed
by a single RCU synchronization, and then the resetting of each of the CPU
buffers. But unfortunately, the mutex that prevented multiple occurrences
of resetting the buffer was not moved to the upper function, and there is
nothing to protect from it.
Take the ring buffer mutex around the global reset.
Cc: stable@vger.kernel.org
Fixes: b23d7a5f4a ("ring-buffer: speed up buffer resets by avoiding synchronize_rcu for each CPU")
Reported-by: "Tzvetomir Stoyanov (VMware)" <tz.stoyanov@gmail.com>
Signed-off-by: Steven Rostedt (VMware) <rostedt@goodmis.org>
Now that there are three different instances of doing the addition trick
to the preempt_count() and NMI_MASK, HARDIRQ_MASK and SOFTIRQ_OFFSET
macros, it deserves a helper function defined in the preempt.h header.
Add the interrupt_context_level() helper and replace the three instances
that do that logic with it.
Link: https://lore.kernel.org/all/20211015142541.4badd8a9@gandalf.local.home/
Signed-off-by: Steven Rostedt (VMware) <rostedt@goodmis.org>
The "rb_per_cpu_empty()" misinterpret the condition (as not-empty) when
"head_page" and "commit_page" of "struct ring_buffer_per_cpu" points to
the same buffer page, whose "buffer_data_page" is empty and "read" field
is non-zero.
An error scenario could be constructed as followed (kernel perspective):
1. All pages in the buffer has been accessed by reader(s) so that all of
them will have non-zero "read" field.
2. Read and clear all buffer pages so that "rb_num_of_entries()" will
return 0 rendering there's no more data to read. It is also required
that the "read_page", "commit_page" and "tail_page" points to the same
page, while "head_page" is the next page of them.
3. Invoke "ring_buffer_lock_reserve()" with large enough "length"
so that it shot pass the end of current tail buffer page. Now the
"head_page", "commit_page" and "tail_page" points to the same page.
4. Discard current event with "ring_buffer_discard_commit()", so that
"head_page", "commit_page" and "tail_page" points to a page whose buffer
data page is now empty.
When the error scenario has been constructed, "tracing_read_pipe" will
be trapped inside a deadloop: "trace_empty()" returns 0 since
"rb_per_cpu_empty()" returns 0 when it hits the CPU containing such
constructed ring buffer. Then "trace_find_next_entry_inc()" always
return NULL since "rb_num_of_entries()" reports there's no more entry
to read. Finally "trace_seq_to_user()" returns "-EBUSY" spanking
"tracing_read_pipe" back to the start of the "waitagain" loop.
I've also written a proof-of-concept script to construct the scenario
and trigger the bug automatically, you can use it to trace and validate
my reasoning above:
https://github.com/aegistudio/RingBufferDetonator.git
Tests has been carried out on linux kernel 5.14-rc2
(2734d6c1b1), my fixed version
of kernel (for testing whether my update fixes the bug) and
some older kernels (for range of affected kernels). Test result is
also attached to the proof-of-concept repository.
Link: https://lore.kernel.org/linux-trace-devel/YPaNxsIlb2yjSi5Y@aegistudio/
Link: https://lore.kernel.org/linux-trace-devel/YPgrN85WL9VyrZ55@aegistudio
Cc: stable@vger.kernel.org
Fixes: bf41a158ca ("ring-buffer: make reentrant")
Suggested-by: Linus Torvalds <torvalds@linuxfoundation.org>
Signed-off-by: Haoran Luo <www@aegistudio.net>
Signed-off-by: Steven Rostedt (VMware) <rostedt@goodmis.org>
The ring_buffer_event_time_stamp() must be only called by an event that has
not been committed yet, and is on the buffer that is passed in. This was
used to help debug converting the histogram logic over to using the new
time stamp code, and was proven to be very useful.
Add a verifier that can check that this is the case, and extra WARN_ONs to
catch unexpected use cases.
Link: https://lkml.kernel.org/r/20210316164113.987294354@goodmis.org
Reviewed-by: Tom Zanussi <zanussi@kernel.org>
Signed-off-by: Steven Rostedt (VMware) <rostedt@goodmis.org>
Currently, ring_buffer_event_time_stamp() only returns an accurate time
stamp of the event if it has an absolute extended time stamp attached to
it. To make it more robust, use the event_stamp() in case the event does
not have an absolute value attached to it.
This will allow ring_buffer_event_time_stamp() to be used in more cases
than just histograms, and it will also allow histograms to not require
including absolute values all the time.
Link: https://lkml.kernel.org/r/20210316164113.704830885@goodmis.org
Reviewed-by: Tom Zanussi <zanussi@kernel.org>
Signed-off-by: Steven Rostedt (VMware) <rostedt@goodmis.org>
The exported use of ring_buffer_event_time_stamp() is going to become
different than how it is used internally. Move the internal logic out into a
static function called rb_event_time_stamp(), and have the internal callers
call that instead.
Link: https://lkml.kernel.org/r/20210316164113.257790481@goodmis.org
Reviewed-by: Tom Zanussi <zanussi@kernel.org>
Signed-off-by: Steven Rostedt (VMware) <rostedt@goodmis.org>
When the CONFIG_RING_BUFFER_VALIDATE_TIME_DELTAS is enabled, and the time
stamps are detected as not being valid, it reports information about the
write stamp, but does not show the before_stamp which is still useful
information. Also, it should give a warning once, such that tests detect
this happening.
Signed-off-by: Steven Rostedt (VMware) <rostedt@goodmis.org>
Part of the logic of the new time stamp code depends on the before_stamp and
the write_stamp to be different if the write_stamp does not match the last
event on the buffer, as it will be used to calculate the delta of the next
event written on the buffer.
The discard logic depends on this, as the next event to come in needs to
inject a full timestamp as it can not rely on the last event timestamp in
the buffer because it is unknown due to events after it being discarded. But
by changing the write_stamp back to the time before it, it forces the next
event to use a full time stamp, instead of relying on it.
The issue came when a full time stamp was used for the event, and
rb_time_delta() returns zero in that case. The update to the write_stamp
(which subtracts delta) made it not change. Then when the event is removed
from the buffer, because the before_stamp and write_stamp still match, the
next event written would calculate its delta from the write_stamp, but that
would be wrong as the write_stamp is of the time of the event that was
discarded.
In the case that the delta change being made to write_stamp is zero, set the
before_stamp to zero as well, and this will force the next event to inject a
full timestamp and not use the current write_stamp.
Cc: stable@vger.kernel.org
Fixes: a389d86f7f ("ring-buffer: Have nested events still record running time stamp")
Signed-off-by: Steven Rostedt (VMware) <rostedt@goodmis.org>
