Both rb_end_commit() and rb_set_commit_to_write() are in the fast path of
the ring buffer recording. Make sure they are always inlined.
Link: http://lkml.kernel.org/r/20161121183700.GW26852@two.firstfloor.org
Reported-by: Andi Kleen <andi@firstfloor.org>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
The function rb_update_write_stamp() is in the hotpath of the ring buffer
recording. Make sure that it is inlined as well. There's not many places
that call it.
Link: http://lkml.kernel.org/r/20161121183700.GW26852@two.firstfloor.org
Reported-by: Andi Kleen <andi@firstfloor.org>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
There's several small helper functions in ring_buffer.c that are used in the
hot path. For some reason, even though they are marked inline, gcc tends not
to enforce it. Make sure these functions are always inlined.
Link: http://lkml.kernel.org/r/20161121183700.GW26852@two.firstfloor.org
Reported-by: Andi Kleen <andi@firstfloor.org>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
The rb_event_data() is the fast path of getting the ring buffer data from an
event. Externally, ring_buffer_event_data() is used to access this function.
But unfortunately, rb_event_data() is not inlined, and calling
ring_buffer_event_data() causes that function to be called again. Force
rb_event_data() to be inlined to lower the number of operations needed when
calling ring_buffer_event_data().
Link: http://lkml.kernel.org/r/20161121183700.GW26852@two.firstfloor.org
Reported-by: Andi Kleen <andi@firstfloor.org>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
The function rb_reserved_next_event() is called by two functions:
ring_buffer_lock_reserve() and ring_buffer_write(). This is in a very hot
path of the tracing code, and it is best that they are not functions. The
two callers are basically wrapers for rb_reserver_next_event(). Removing the
function calls can save execution time in the hotpath of tracing.
Link: http://lkml.kernel.org/r/20161121183700.GW26852@two.firstfloor.org
Reported-by: Andi Kleen <andi@firstfloor.org>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
If the size passed to ring_buffer_resize() is greater than MAX_LONG - BUF_PAGE_SIZE
then the DIV_ROUND_UP() will return zero.
Here's the details:
# echo 18014398509481980 > /sys/kernel/debug/tracing/buffer_size_kb
tracing_entries_write() processes this and converts kb to bytes.
18014398509481980 << 10 = 18446744073709547520
and this is passed to ring_buffer_resize() as unsigned long size.
size = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
Where DIV_ROUND_UP(a, b) is (a + b - 1)/b
BUF_PAGE_SIZE is 4080 and here
18446744073709547520 + 4080 - 1 = 18446744073709551599
where 18446744073709551599 is still smaller than 2^64
2^64 - 18446744073709551599 = 17
But now 18446744073709551599 / 4080 = 4521260802379792
and size = size * 4080 = 18446744073709551360
This is checked to make sure its still greater than 2 * 4080,
which it is.
Then we convert to the number of buffer pages needed.
nr_page = DIV_ROUND_UP(size, BUF_PAGE_SIZE)
but this time size is 18446744073709551360 and
2^64 - (18446744073709551360 + 4080 - 1) = -3823
Thus it overflows and the resulting number is less than 4080, which makes
3823 / 4080 = 0
an nr_pages is set to this. As we already checked against the minimum that
nr_pages may be, this causes the logic to fail as well, and we crash the
kernel.
There's no reason to have the two DIV_ROUND_UP() (that's just result of
historical code changes), clean up the code and fix this bug.
Cc: stable@vger.kernel.org # 3.5+
Fixes: 83f40318da ("ring-buffer: Make removal of ring buffer pages atomic")
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
The size variable to change the ring buffer in ftrace is a long. The
nr_pages used to update the ring buffer based on the size is int. On 64 bit
machines this can cause an overflow problem.
For example, the following will cause the ring buffer to crash:
# cd /sys/kernel/debug/tracing
# echo 10 > buffer_size_kb
# echo 8556384240 > buffer_size_kb
Then you get the warning of:
WARNING: CPU: 1 PID: 318 at kernel/trace/ring_buffer.c:1527 rb_update_pages+0x22f/0x260
Which is:
RB_WARN_ON(cpu_buffer, nr_removed);
Note each ring buffer page holds 4080 bytes.
This is because:
1) 10 causes the ring buffer to have 3 pages.
(10kb requires 3 * 4080 pages to hold)
2) (2^31 / 2^10 + 1) * 4080 = 8556384240
The value written into buffer_size_kb is shifted by 10 and then passed
to ring_buffer_resize(). 8556384240 * 2^10 = 8761737461760
3) The size passed to ring_buffer_resize() is then divided by BUF_PAGE_SIZE
which is 4080. 8761737461760 / 4080 = 2147484672
4) nr_pages is subtracted from the current nr_pages (3) and we get:
2147484669. This value is saved in a signed integer nr_pages_to_update
5) 2147484669 is greater than 2^31 but smaller than 2^32, a signed int
turns into the value of -2147482627
6) As the value is a negative number, in update_pages_handler() it is
negated and passed to rb_remove_pages() and 2147482627 pages will
be removed, which is much larger than 3 and it causes the warning
because not all the pages asked to be removed were removed.
Link: https://bugzilla.kernel.org/show_bug.cgi?id=118001
Cc: stable@vger.kernel.org # 2.6.28+
Fixes: 7a8e76a382 ("tracing: unified trace buffer")
Reported-by: Hao Qin <QEver.cn@gmail.com>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
When crossing over to a new page, commit the current work. This will allow
readers to get data with less latency, and also simplifies the work to get
timestamps working for interrupted events.
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
The first commit of a buffer page updates the timestamp of that page. No
need to have the update to the next page add the timestamp too. It will only
be replaced by the first commit on that page anyway.
Only update to a page if it contains an event.
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
As cpu_buffer->tail_page may be modified by interrupts at almost any time,
the flow of logic is very important. Do not let gcc get smart with
re-reading cpu_buffer->tail_page by adding READ_ONCE() around most of its
accesses.
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
Commit fcc742eaad "ring-buffer: Add event descriptor to simplify passing
data" added a descriptor that holds various data instead of passing around
several variables through parameters. The problem was that one of the
parameters was modified in a function and the code was designed not to have
an effect on that modified parameter. Now that the parameter is a
descriptor and any modifications to it are non-volatile, the size of the
data could be unnecessarily expanded.
Remove the extra space added if a timestamp was added and the event went
across the page.
Cc: stable@vger.kernel.org # 4.3+
Fixes: fcc742eaad "ring-buffer: Add event descriptor to simplify passing data"
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
Do not update the read stamp after swapping out the reader page from the
write buffer. If the reader page is swapped out of the buffer before an
event is written to it, then the read_stamp may get an out of date
timestamp, as the page timestamp is updated on the first commit to that
page.
rb_get_reader_page() only returns a page if it has an event on it, otherwise
it will return NULL. At that point, check if the page being returned has
events and has not been read yet. Then at that point update the read_stamp
to match the time stamp of the reader page.
Cc: stable@vger.kernel.org # 2.6.30+
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
Make rb_event_is_commit() return bool to improve readability
due to this particular function only using either one or zero as its
return value.
No functional change.
Link: http://lkml.kernel.org/r/1443537816-5788-7-git-send-email-bywxiaobai@163.com
Signed-off-by: Yaowei Bai <bywxiaobai@163.com>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
Make rb_is_reader_page() return bool to improve readability due to this
particular function only using either true or false as its return value.
No functional change.
Link: http://lkml.kernel.org/r/1443537816-5788-4-git-send-email-bywxiaobai@163.com
Signed-off-by: Yaowei Bai <bywxiaobai@163.com>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
The commit a4543a2fa9 "ring-buffer: Get timestamp after event is
allocated" is needed for some future work. But after adding it, there is a
race somewhere that causes the saved timestamp to have a slight shift, and
get ahead of the actual timestamp and make it look like time goes backwards.
I'm still looking into why this happens, but in the mean time, this is
holding up other work to get in. I'm reverting the change for now (which
makes the problem go away), and will add it back after I know what is wrong
and fix it.
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
Functions in ring-buffer.c have gotten interleaved between different
use cases. Move the functions around to get like functions closer
together. This may or may not help gcc keep cache locality, but it
makes it a little easier to work with the code.
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
Now that events only add time extends after it is committed, in case
an event comes in before it can discard the allocated event, the time
extend needs to be stored within the event. If the event is bigger
than then size needed for the time extend, padding must be added.
The minimum padding size is 8 bytes. Thus if the event is 12 bytes
(size of time extend + 4), there will not be enough room to add both
the time extend and padding. Make sure all events are either 8 bytes
or 16 or more bytes.
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
Move the capturing of the timestamp to after an event is allocated.
If the event is not a commit (where it is an event that preempted
another event), then no timestamp is needed, because the delta of
nested events is always zero.
If the event starts on a new page, no delta needs to be calculated
as the full timestamp will be added to the page header, and the
event will have a delta of zero.
Now if the event requires a time extend (the delta does not fit
in the 27 bit delta slot in the header), then the event is discarded,
the length is extended to hold the TIME_EXTEND event that allows for
a 59 bit delta, and the commit is tried again.
If the event can't be discarded (another event came in after it),
then the TIME_EXTEND is added directly to the allocated event and
the rest of the event is given padding.
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
Requiring a extended time stamp is an uncommon occurrence, and it is
best to do it out of line when needed.
Add a noinline function that handles the extended timestamp and
have it called with an unlikely to completely move it out of the
fast path.
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
Add rb_event_info descriptor to pass event info to functions a bit
easier than using a bunch of parameters. This will also allow for
changing the code around a bit to find better fast paths.
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
The tracing_off_permanent() call is a way to disable all ring_buffers.
Nothing uses it and nothing should use it, as tracing_off() and
friends are better, as they disable the ring buffers related to
tracing. The tracing_off_permanent() even disabled non tracing
ring buffers. This is a bit drastic, and was added to handle NMIs
doing outputs that could corrupt the ring buffer when only tracing
used them. It is now obsolete and adds a little overhead, it should
be removed.
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
Currently, if an NMI does a dump of a ring buffer, it disables
all ring buffers from ever doing any writes again. This is because
it wont take the locks for the cpu_buffer and this can cause
corruption if it preempted a read, or a read happens on another
CPU for the current cpu buffer. This is a bit overkill.
First, it should at least try to take the lock, and if it fails
then disable it. Also, there's no need to disable all ring
buffers, even those that are unrelated to what is being read.
Only disable the per cpu ring buffer that is being read if
it can not get the lock for it.
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
The ring_buffer_write() function isn't protected by the trace recursive
writes. Luckily, this function is not used as much and is unlikely
to ever recurse. But it should still have the protection, because
even a call to ring_buffer_lock_reserve() could cause ring buffer
corruption if called when ring_buffer_write() is being used.
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
Currently the trace_recursive checks are only done if CONFIG_TRACING
is enabled. That was because there use to be a dependency with tracing
for the recursive checks (it used the task_struct trace recursive
variable). But now it uses its own variable and there is no dependency.
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
Instead of using a global per_cpu variable to perform the recursive
checks into the ring buffer, use the already existing per_cpu descriptor
that is part of the ring buffer itself.
Not only does this simplify the code, it also allows for one ring buffer
to be used within the guts of the use of another ring buffer. For example
trace_printk() can now be used within the ring buffer to record changes
done by an instance into the main ring buffer. The recursion checks
will prevent the trace_printk() itself from causing recursive issues
with the main ring buffer (it is just ignored), but the recursive
checks wont prevent the trace_printk() from recording other ring buffers.
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
The term "ftrace" is really the infrastructure of the function hooks,
and not the trace events. Rename ftrace_event.h to trace_events.h to
represent the trace_event infrastructure and decouple the term ftrace
from it.
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
A clean up of the recursive protection code changed
val = this_cpu_read(current_context);
val--;
val &= this_cpu_read(current_context);
to
val = this_cpu_read(current_context);
val &= val & (val - 1);
Which has a duplicate use of '&' as the above is the same as
val = val & (val - 1);
Actually, it would be best to remove that line altogether and
just add it to where it is used.
And Christoph even mentioned that it can be further compacted to
just a single line:
__this_cpu_and(current_context, __this_cpu_read(current_context) - 1);
Link: http://lkml.kernel.org/alpine.DEB.2.11.1503271423580.23114@gentwo.org
Suggested-by: Christoph Lameter <cl@linux.com>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
It has come to my attention that this_cpu_read/write are horrible on
architectures other than x86. Worse yet, they actually disable
preemption or interrupts! This caused some unexpected tracing results
on ARM.
101.356868: preempt_count_add <-ring_buffer_lock_reserve
101.356870: preempt_count_sub <-ring_buffer_lock_reserve
The ring_buffer_lock_reserve has recursion protection that requires
accessing a per cpu variable. But since preempt_disable() is traced, it
too got traced while accessing the variable that is suppose to prevent
recursion like this.
The generic version of this_cpu_read() and write() are:
#define this_cpu_generic_read(pcp) \
({ typeof(pcp) ret__; \
preempt_disable(); \
ret__ = *this_cpu_ptr(&(pcp)); \
preempt_enable(); \
ret__; \
})
#define this_cpu_generic_to_op(pcp, val, op) \
do { \
unsigned long flags; \
raw_local_irq_save(flags); \
*__this_cpu_ptr(&(pcp)) op val; \
raw_local_irq_restore(flags); \
} while (0)
Which is unacceptable for locations that know they are within preempt
disabled or interrupt disabled locations.
Paul McKenney stated that __this_cpu_() versions produce much better code on
other architectures than this_cpu_() does, if we know that the call is done in
a preempt disabled location.
I also changed the recursive_unlock() to use two local variables instead
of accessing the per_cpu variable twice.
Link: http://lkml.kernel.org/r/20150317114411.GE3589@linux.vnet.ibm.com
Link: http://lkml.kernel.org/r/20150317104038.312e73d1@gandalf.local.home
Cc: stable@vger.kernel.org
Acked-by: Christoph Lameter <cl@linux.com>
Reported-by: Uwe Kleine-Koenig <u.kleine-koenig@pengutronix.de>
Tested-by: Uwe Kleine-Koenig <u.kleine-koenig@pengutronix.de>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
When an application connects to the ring buffer via splice, it can only
read full pages. Splice does not work with partial pages. If there is
not enough data to fill a page, the splice command will either block
or return -EAGAIN (if set to nonblock).
Code was added where if the page is not full, to just sleep again.
The problem is, it will get woken up again on the next event. That
is, when something is written into the ring buffer, if there is a waiter
it will wake it up. The waiter would then check the buffer, see that
it still does not have enough data to fill a page and go back to sleep.
To make matters worse, when the waiter goes back to sleep, it could
cause another event, which would wake it back up again to see it
doesn't have enough data and sleep again. This produces a tremendous
overhead and fills the ring buffer with noise.
For example, recording sched_switch on an idle system for 10 seconds
produces 25,350,475 events!!!
Create another wait queue for those waiters wanting full pages.
When an event is written, it only wakes up waiters if there's a full
page of data. It does not wake up the waiter if the page is not yet
full.
After this change, recording sched_switch on an idle system for 10
seconds produces only 800 events. Getting rid of 25,349,675 useless
events (99.9969% of events!!), is something to take seriously.
Cc: stable@vger.kernel.org # 3.16+
Cc: Rabin Vincent <rabin@rab.in>
Fixes: e30f53aad2 "tracing: Do not busy wait in buffer splice"
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
The creation of tracing files and directories is for the most part
encapsulated in helper functions in trace.c. Other files do not need to
include debugfs.h or fs.h, as they may have needed to in the past.
Remove them from the files that do not need them.
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
to the trace_seq code. It also removed the return values to the
trace_seq_*() functions and use trace_seq_has_overflowed() to see if
the buffer filled up or not. This is similar to work being done to the
seq_file code as well in another tree.
Some of the other goodies include:
o Added some "!" (NOT) logic to the tracing filter.
o Fixed the frame pointer logic to the x86_64 mcount trampolines
o Added the logic for dynamic trampolines on !CONFIG_PREEMPT systems.
That is, the ftrace trampoline can be dynamically allocated
and be called directly by functions that only have a single hook
to them.
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1
iQEcBAABAgAGBQJUhbLGAAoJEEjnJuOKh9ldRV4H/3NcLbgGB2iu96la1zdYE6pG
Q7cDJMxXK80YIIL70h9G0IItcD4t62LMb72lfBnMGRj3msgFb3AgISW57EuI0Pxk
xk24wuIPoTG2S7v9sc3SboNFwO8qbtIjxD2OBmqIUrGo2sZIiGjyj3gX7mCY3uzL
WB2bUOSFz/22OgaANinR5EELHA3pZZCf54Vz1K9ndmtK0xp0j1a7xJShD6TrMdYv
mZ3zH5ViIhW4A3mdcMceh6fy2JLQAiEKF0uPTvcMMz7NlVul0mxyL/+10P7AE/3R
Ehw4fzmm4NDshPDtBOkKH0LsppgXzuItFuQUTpact3JlqTg++bV6onSsrkt1hlY=
=Z7Cm
-----END PGP SIGNATURE-----
Merge tag 'trace-3.19' of git://git.kernel.org/pub/scm/linux/kernel/git/rostedt/linux-trace
Pull tracing updates from Steven Rostedt:
"There was a lot of clean ups and minor fixes. One of those clean ups
was to the trace_seq code. It also removed the return values to the
trace_seq_*() functions and use trace_seq_has_overflowed() to see if
the buffer filled up or not. This is similar to work being done to
the seq_file code as well in another tree.
Some of the other goodies include:
- Added some "!" (NOT) logic to the tracing filter.
- Fixed the frame pointer logic to the x86_64 mcount trampolines
- Added the logic for dynamic trampolines on !CONFIG_PREEMPT systems.
That is, the ftrace trampoline can be dynamically allocated and be
called directly by functions that only have a single hook to them"
* tag 'trace-3.19' of git://git.kernel.org/pub/scm/linux/kernel/git/rostedt/linux-trace: (55 commits)
tracing: Truncated output is better than nothing
tracing: Add additional marks to signal very large time deltas
Documentation: describe trace_buf_size parameter more accurately
tracing: Allow NOT to filter AND and OR clauses
tracing: Add NOT to filtering logic
ftrace/fgraph/x86: Have prepare_ftrace_return() take ip as first parameter
ftrace/x86: Get rid of ftrace_caller_setup
ftrace/x86: Have save_mcount_regs macro also save stack frames if needed
ftrace/x86: Add macro MCOUNT_REG_SIZE for amount of stack used to save mcount regs
ftrace/x86: Simplify save_mcount_regs on getting RIP
ftrace/x86: Have save_mcount_regs store RIP in %rdi for first parameter
ftrace/x86: Rename MCOUNT_SAVE_FRAME and add more detailed comments
ftrace/x86: Move MCOUNT_SAVE_FRAME out of header file
ftrace/x86: Have static tracing also use ftrace_caller_setup
ftrace/x86: Have static function tracing always test for function graph
kprobes: Add IPMODIFY flag to kprobe_ftrace_ops
ftrace, kprobes: Support IPMODIFY flag to find IP modify conflict
kprobes/ftrace: Recover original IP if pre_handler doesn't change it
tracing/trivial: Fix typos and make an int into a bool
tracing: Deletion of an unnecessary check before iput()
...
Remove checking the return value of all trace_seq_puts(). It was wrong
anyway as only the last return value mattered. But as the trace_seq_puts()
is going to be a void function in the future, we should not be checking
the return value of it anyway.
Just return !trace_seq_has_overflowed() instead.
Reviewed-by: Petr Mladek <pmladek@suse.cz>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
On a !PREEMPT kernel, attempting to use trace-cmd results in a soft
lockup:
# trace-cmd record -e raw_syscalls:* -F false
NMI watchdog: BUG: soft lockup - CPU#0 stuck for 22s! [trace-cmd:61]
...
Call Trace:
[<ffffffff8105b580>] ? __wake_up_common+0x90/0x90
[<ffffffff81092e25>] wait_on_pipe+0x35/0x40
[<ffffffff810936e3>] tracing_buffers_splice_read+0x2e3/0x3c0
[<ffffffff81093300>] ? tracing_stats_read+0x2a0/0x2a0
[<ffffffff812d10ab>] ? _raw_spin_unlock+0x2b/0x40
[<ffffffff810dc87b>] ? do_read_fault+0x21b/0x290
[<ffffffff810de56a>] ? handle_mm_fault+0x2ba/0xbd0
[<ffffffff81095c80>] ? trace_event_buffer_lock_reserve+0x40/0x80
[<ffffffff810951e2>] ? trace_buffer_lock_reserve+0x22/0x60
[<ffffffff81095c80>] ? trace_event_buffer_lock_reserve+0x40/0x80
[<ffffffff8112415d>] do_splice_to+0x6d/0x90
[<ffffffff81126971>] SyS_splice+0x7c1/0x800
[<ffffffff812d1edd>] tracesys_phase2+0xd3/0xd8
The problem is this: tracing_buffers_splice_read() calls
ring_buffer_wait() to wait for data in the ring buffers. The buffers
are not empty so ring_buffer_wait() returns immediately. But
tracing_buffers_splice_read() calls ring_buffer_read_page() with full=1,
meaning it only wants to read a full page. When the full page is not
available, tracing_buffers_splice_read() tries to wait again with
ring_buffer_wait(), which again returns immediately, and so on.
Fix this by adding a "full" argument to ring_buffer_wait() which will
make ring_buffer_wait() wait until the writer has left the reader's
page, i.e. until full-page reads will succeed.
Link: http://lkml.kernel.org/r/1415645194-25379-1-git-send-email-rabin@rab.in
Cc: stable@vger.kernel.org # 3.16+
Fixes: b1169cc69b ("tracing: Remove mock up poll wait function")
Signed-off-by: Rabin Vincent <rabin@rab.in>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
Commit 651e22f270 "ring-buffer: Always reset iterator to reader page"
fixed one bug but in the process caused another one. The reset is to
update the header page, but that fix also changed the way the cached
reads were updated. The cache reads are used to test if an iterator
needs to be updated or not.
A ring buffer iterator, when created, disables writes to the ring buffer
but does not stop other readers or consuming reads from happening.
Although all readers are synchronized via a lock, they are only
synchronized when in the ring buffer functions. Those functions may
be called by any number of readers. The iterator continues down when
its not interrupted by a consuming reader. If a consuming read
occurs, the iterator starts from the beginning of the buffer.
The way the iterator sees that a consuming read has happened since
its last read is by checking the reader "cache". The cache holds the
last counts of the read and the reader page itself.
Commit 651e22f270 changed what was saved by the cache_read when
the rb_iter_reset() occurred, making the iterator never match the cache.
Then if the iterator calls rb_iter_reset(), it will go into an
infinite loop by checking if the cache doesn't match, doing the reset
and retrying, just to see that the cache still doesn't match! Which
should never happen as the reset is suppose to set the cache to the
current value and there's locks that keep a consuming reader from
having access to the data.
Fixes: 651e22f270 "ring-buffer: Always reset iterator to reader page"
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
Epoll on trace_pipe can sometimes hang in a weird case. If the ring buffer is
empty when we set waiters_pending but an event shows up exactly at that moment
we can miss being woken up by the ring buffers irq work. Since
ring_buffer_empty() is inherently racey we will sometimes think that the buffer
is not empty. So we don't get woken up and we don't think there are any events
even though there were some ready when we added the watch, which makes us hang.
This patch fixes this by making sure that we are actually on the wait list
before we set waiters_pending, and add a memory barrier to make sure
ring_buffer_empty() is going to be correct.
Link: http://lkml.kernel.org/p/1408989581-23727-1-git-send-email-jbacik@fb.com
Cc: stable@vger.kernel.org # 3.10+
Cc: Martin Lau <kafai@fb.com>
Signed-off-by: Josef Bacik <jbacik@fb.com>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
rarely ever hit, and requires the user to do something that users rarely
do. It took a few special test cases to even trigger this bug,
and one of them was just one test in the process of finishing up as another
one started.
Both bugs have to do with the ring buffer iterator rb_iter_peek(), but one
is more indirect than the other.
The fist bug fix is simply an increase in the safety net loop counter.
The counter makes sure that the rb_iter_peek() only iterates the number
of times we expect it can, and no more. Well, there was one way it could
iterate one more than we expected, and that caused the ring buffer
to shutdown with a nasty warning. The fix was simply to up that counter by
one.
The other bug has to be with rb_iter_reset() (called by rb_iter_peek()).
This happens when a user reads both the trace_pipe and trace files.
The trace_pipe is a consuming read and does not use the ring buffer
iterator, but the trace file is not a consuming read and does use the
ring buffer iterator. When the trace file is being read, if it detects
that a consuming read occurred, it resets the iterator and starts over.
But the reset code that does this (rb_iter_reset()), checks if the
reader_page is linked to the ring buffer or not, and will look into
the ring buffer itself if it is not. This is wrong, as it should always
try to read the reader page first. Not to mention, the code that looked
into the ring buffer did it wrong, and used the header_page "read" offset
to start reading on that page. That offset is bogus for pages in the
writable ring buffer, and was corrupting the iterator, and it would start
returning bogus events.
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1
iQEcBAABAgAGBQJT44tRAAoJEKQekfcNnQGuMVIH/3evbjKT+w6Kon4S0LfLSejl
YDsXYkeO/lGiO3MnUveqq1jfw2+yHtyBVUipvfG0A61urMUhyUvjveph8Ec2cQ4Q
qHl0J28vDfF5tpKiYzygRN01wHD6GXYh+XZSChkA4ItzzD8K51lsZT1Yd+I2pTy2
DgH01EEEYiwYJcih+T4vlbKqYju6pwgxqKNCTv0RdVXUPya/tG9X2Nf8VGQTbmKS
FIO73qObYR+P9iXGIuPfyOxk2EvBiWS15WownZmfhZZxOIKx9IrDYwTsiV1+AJw+
sJFER1ulobYGDtGDa9yyrNJQr6wgbrfCDELnNKmdLUTlSwgZjLXpE2HEmlelY/s=
=5mQl
-----END PGP SIGNATURE-----
Merge tag 'trace-fixes-3.16' of git://git.kernel.org/pub/scm/linux/kernel/git/rostedt/linux-trace
Pull trace file read iterator fixes from Steven Rostedt:
"This contains a fix for two long standing bugs. Both of which are
rarely ever hit, and requires the user to do something that users
rarely do. It took a few special test cases to even trigger this bug,
and one of them was just one test in the process of finishing up as
another one started.
Both bugs have to do with the ring buffer iterator rb_iter_peek(), but
one is more indirect than the other.
The fist bug fix is simply an increase in the safety net loop counter.
The counter makes sure that the rb_iter_peek() only iterates the
number of times we expect it can, and no more. Well, there was one
way it could iterate one more than we expected, and that caused the
ring buffer to shutdown with a nasty warning. The fix was simply to
up that counter by one.
The other bug has to be with rb_iter_reset() (called by
rb_iter_peek()). This happens when a user reads both the trace_pipe
and trace files. The trace_pipe is a consuming read and does not use
the ring buffer iterator, but the trace file is not a consuming read
and does use the ring buffer iterator. When the trace file is being
read, if it detects that a consuming read occurred, it resets the
iterator and starts over. But the reset code that does this
(rb_iter_reset()), checks if the reader_page is linked to the ring
buffer or not, and will look into the ring buffer itself if it is not.
This is wrong, as it should always try to read the reader page first.
Not to mention, the code that looked into the ring buffer did it
wrong, and used the header_page "read" offset to start reading on that
page. That offset is bogus for pages in the writable ring buffer, and
was corrupting the iterator, and it would start returning bogus
events"
* tag 'trace-fixes-3.16' of git://git.kernel.org/pub/scm/linux/kernel/git/rostedt/linux-trace:
ring-buffer: Always reset iterator to reader page
ring-buffer: Up rb_iter_peek() loop count to 3
When performing a consuming read, the ring buffer swaps out a
page from the ring buffer with a empty page and this page that
was swapped out becomes the new reader page. The reader page
is owned by the reader and since it was swapped out of the ring
buffer, writers do not have access to it (there's an exception
to that rule, but it's out of scope for this commit).
When reading the "trace" file, it is a non consuming read, which
means that the data in the ring buffer will not be modified.
When the trace file is opened, a ring buffer iterator is allocated
and writes to the ring buffer are disabled, such that the iterator
will not have issues iterating over the data.
Although the ring buffer disabled writes, it does not disable other
reads, or even consuming reads. If a consuming read happens, then
the iterator is reset and starts reading from the beginning again.
My tests would sometimes trigger this bug on my i386 box:
WARNING: CPU: 0 PID: 5175 at kernel/trace/trace.c:1527 __trace_find_cmdline+0x66/0xaa()
Modules linked in:
CPU: 0 PID: 5175 Comm: grep Not tainted 3.16.0-rc3-test+ #8
Hardware name: /DG965MQ, BIOS MQ96510J.86A.0372.2006.0605.1717 06/05/2006
00000000 00000000 f09c9e1c c18796b3 c1b5d74c f09c9e4c c103a0e3 c1b5154b
f09c9e78 00001437 c1b5d74c 000005f7 c10bd85a c10bd85a c1cac57c f09c9eb0
ed0e0000 f09c9e64 c103a185 00000009 f09c9e5c c1b5154b f09c9e78 f09c9e80^M
Call Trace:
[<c18796b3>] dump_stack+0x4b/0x75
[<c103a0e3>] warn_slowpath_common+0x7e/0x95
[<c10bd85a>] ? __trace_find_cmdline+0x66/0xaa
[<c10bd85a>] ? __trace_find_cmdline+0x66/0xaa
[<c103a185>] warn_slowpath_fmt+0x33/0x35
[<c10bd85a>] __trace_find_cmdline+0x66/0xaa^M
[<c10bed04>] trace_find_cmdline+0x40/0x64
[<c10c3c16>] trace_print_context+0x27/0xec
[<c10c4360>] ? trace_seq_printf+0x37/0x5b
[<c10c0b15>] print_trace_line+0x319/0x39b
[<c10ba3fb>] ? ring_buffer_read+0x47/0x50
[<c10c13b1>] s_show+0x192/0x1ab
[<c10bfd9a>] ? s_next+0x5a/0x7c
[<c112e76e>] seq_read+0x267/0x34c
[<c1115a25>] vfs_read+0x8c/0xef
[<c112e507>] ? seq_lseek+0x154/0x154
[<c1115ba2>] SyS_read+0x54/0x7f
[<c188488e>] syscall_call+0x7/0xb
---[ end trace 3f507febd6b4cc83 ]---
>>>> ##### CPU 1 buffer started ####
Which was the __trace_find_cmdline() function complaining about the pid
in the event record being negative.
After adding more test cases, this would trigger more often. Strangely
enough, it would never trigger on a single test, but instead would trigger
only when running all the tests. I believe that was the case because it
required one of the tests to be shutting down via delayed instances while
a new test started up.
After spending several days debugging this, I found that it was caused by
the iterator becoming corrupted. Debugging further, I found out why
the iterator became corrupted. It happened with the rb_iter_reset().
As consuming reads may not read the full reader page, and only part
of it, there's a "read" field to know where the last read took place.
The iterator, must also start at the read position. In the rb_iter_reset()
code, if the reader page was disconnected from the ring buffer, the iterator
would start at the head page within the ring buffer (where writes still
happen). But the mistake there was that it still used the "read" field
to start the iterator on the head page, where it should always start
at zero because readers never read from within the ring buffer where
writes occur.
I originally wrote a patch to have it set the iter->head to 0 instead
of iter->head_page->read, but then I questioned why it wasn't always
setting the iter to point to the reader page, as the reader page is
still valid. The list_empty(reader_page->list) just means that it was
successful in swapping out. But the reader_page may still have data.
There was a bug report a long time ago that was not reproducible that
had something about trace_pipe (consuming read) not matching trace
(iterator read). This may explain why that happened.
Anyway, the correct answer to this bug is to always use the reader page
an not reset the iterator to inside the writable ring buffer.
Cc: stable@vger.kernel.org # 2.6.28+
Fixes: d769041f86 "ring_buffer: implement new locking"
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
After writting a test to try to trigger the bug that caused the
ring buffer iterator to become corrupted, I hit another bug:
WARNING: CPU: 1 PID: 5281 at kernel/trace/ring_buffer.c:3766 rb_iter_peek+0x113/0x238()
Modules linked in: ipt_MASQUERADE sunrpc [...]
CPU: 1 PID: 5281 Comm: grep Tainted: G W 3.16.0-rc3-test+ #143
Hardware name: To Be Filled By O.E.M. To Be Filled By O.E.M./To be filled by O.E.M., BIOS SDBLI944.86P 05/08/2007
0000000000000000 ffffffff81809a80 ffffffff81503fb0 0000000000000000
ffffffff81040ca1 ffff8800796d6010 ffffffff810c138d ffff8800796d6010
ffff880077438c80 ffff8800796d6010 ffff88007abbe600 0000000000000003
Call Trace:
[<ffffffff81503fb0>] ? dump_stack+0x4a/0x75
[<ffffffff81040ca1>] ? warn_slowpath_common+0x7e/0x97
[<ffffffff810c138d>] ? rb_iter_peek+0x113/0x238
[<ffffffff810c138d>] ? rb_iter_peek+0x113/0x238
[<ffffffff810c14df>] ? ring_buffer_iter_peek+0x2d/0x5c
[<ffffffff810c6f73>] ? tracing_iter_reset+0x6e/0x96
[<ffffffff810c74a3>] ? s_start+0xd7/0x17b
[<ffffffff8112b13e>] ? kmem_cache_alloc_trace+0xda/0xea
[<ffffffff8114cf94>] ? seq_read+0x148/0x361
[<ffffffff81132d98>] ? vfs_read+0x93/0xf1
[<ffffffff81132f1b>] ? SyS_read+0x60/0x8e
[<ffffffff8150bf9f>] ? tracesys+0xdd/0xe2
Debugging this bug, which triggers when the rb_iter_peek() loops too
many times (more than 2 times), I discovered there's a case that can
cause that function to legitimately loop 3 times!
rb_iter_peek() is different than rb_buffer_peek() as the rb_buffer_peek()
only deals with the reader page (it's for consuming reads). The
rb_iter_peek() is for traversing the buffer without consuming it, and as
such, it can loop for one more reason. That is, if we hit the end of
the reader page or any page, it will go to the next page and try again.
That is, we have this:
1. iter->head > iter->head_page->page->commit
(rb_inc_iter() which moves the iter to the next page)
try again
2. event = rb_iter_head_event()
event->type_len == RINGBUF_TYPE_TIME_EXTEND
rb_advance_iter()
try again
3. read the event.
But we never get to 3, because the count is greater than 2 and we
cause the WARNING and return NULL.
Up the counter to 3.
Cc: stable@vger.kernel.org # 2.6.37+
Fixes: 69d1b839f7 "ring-buffer: Bind time extend and data events together"
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
to the ftrace function callback infrastructure. It's introducing a
way to allow different functions to call directly different trampolines
instead of all calling the same "mcount" one.
The only user of this for now is the function graph tracer, which always
had a different trampoline, but the function tracer trampoline was called
and did basically nothing, and then the function graph tracer trampoline
was called. The difference now, is that the function graph tracer
trampoline can be called directly if a function is only being traced by
the function graph trampoline. If function tracing is also happening on
the same function, the old way is still done.
The accounting for this takes up more memory when function graph tracing
is activated, as it needs to keep track of which functions it uses.
I have a new way that wont take as much memory, but it's not ready yet
for this merge window, and will have to wait for the next one.
Another big change was the removal of the ftrace_start/stop() calls that
were used by the suspend/resume code that stopped function tracing when
entering into suspend and resume paths. The stop of ftrace was done
because there was some function that would crash the system if one called
smp_processor_id()! The stop/start was a big hammer to solve the issue
at the time, which was when ftrace was first introduced into Linux.
Now ftrace has better infrastructure to debug such issues, and I found
the problem function and labeled it with "notrace" and function tracing
can now safely be activated all the way down into the guts of suspend
and resume.
Other changes include clean ups of uprobe code.
Clean up of the trace_seq() code.
And other various small fixes and clean ups to ftrace and tracing.
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1
iQEcBAABAgAGBQJT35zXAAoJEKQekfcNnQGuOz0H/38zqM0nLFhrgvz3EPk2UOjn
xqpX8qyb2V7TJZL+IqeXU2a5cQZl5ba0D4WtBGpxbTae3CJYiuQ87iKUNFoH0om5
FDpn80igb368k8V3qRdRsziKVCCf0XBd/NkHJXc0ZkfXGyzB2Ga4bBxALxp2gj9y
bnO+vKo6+tWYKG4hyQb4P3LRXUrK8/LWEsPr39cH2QH1Rdj69Lx9CgrCdUVJmwcb
Bj8hEiLXL/RYCFNn79A3wNTUvW0rG/AOIf4SLqXtasSRZ0ToaU0ZyDnrNv+0Ol47
rX8tSk+LfXchL9hpIvjCf1vlAYq3pO02favteR/jip3lx/dTjEDE4RJ9qtJzZ4Q=
=fwQY
-----END PGP SIGNATURE-----
Merge tag 'trace-3.17' of git://git.kernel.org/pub/scm/linux/kernel/git/rostedt/linux-trace
Pull tracing updates from Steven Rostedt:
"This pull request has a lot of work done. The main thing is the
changes to the ftrace function callback infrastructure. It's
introducing a way to allow different functions to call directly
different trampolines instead of all calling the same "mcount" one.
The only user of this for now is the function graph tracer, which
always had a different trampoline, but the function tracer trampoline
was called and did basically nothing, and then the function graph
tracer trampoline was called. The difference now, is that the
function graph tracer trampoline can be called directly if a function
is only being traced by the function graph trampoline. If function
tracing is also happening on the same function, the old way is still
done.
The accounting for this takes up more memory when function graph
tracing is activated, as it needs to keep track of which functions it
uses. I have a new way that wont take as much memory, but it's not
ready yet for this merge window, and will have to wait for the next
one.
Another big change was the removal of the ftrace_start/stop() calls
that were used by the suspend/resume code that stopped function
tracing when entering into suspend and resume paths. The stop of
ftrace was done because there was some function that would crash the
system if one called smp_processor_id()! The stop/start was a big
hammer to solve the issue at the time, which was when ftrace was first
introduced into Linux. Now ftrace has better infrastructure to debug
such issues, and I found the problem function and labeled it with
"notrace" and function tracing can now safely be activated all the way
down into the guts of suspend and resume
Other changes include clean ups of uprobe code, clean up of the
trace_seq() code, and other various small fixes and clean ups to
ftrace and tracing"
* tag 'trace-3.17' of git://git.kernel.org/pub/scm/linux/kernel/git/rostedt/linux-trace: (57 commits)
ftrace: Add warning if tramp hash does not match nr_trampolines
ftrace: Fix trampoline hash update check on rec->flags
ring-buffer: Use rb_page_size() instead of open coded head_page size
ftrace: Rename ftrace_ops field from trampolines to nr_trampolines
tracing: Convert local function_graph functions to static
ftrace: Do not copy old hash when resetting
tracing: let user specify tracing_thresh after selecting function_graph
ring-buffer: Always run per-cpu ring buffer resize with schedule_work_on()
tracing: Remove function_trace_stop and HAVE_FUNCTION_TRACE_MCOUNT_TEST
s390/ftrace: remove check of obsolete variable function_trace_stop
arm64, ftrace: Remove check of obsolete variable function_trace_stop
Blackfin: ftrace: Remove check of obsolete variable function_trace_stop
metag: ftrace: Remove check of obsolete variable function_trace_stop
microblaze: ftrace: Remove check of obsolete variable function_trace_stop
MIPS: ftrace: Remove check of obsolete variable function_trace_stop
parisc: ftrace: Remove check of obsolete variable function_trace_stop
sh: ftrace: Remove check of obsolete variable function_trace_stop
sparc64,ftrace: Remove check of obsolete variable function_trace_stop
tile: ftrace: Remove check of obsolete variable function_trace_stop
ftrace: x86: Remove check of obsolete variable function_trace_stop
...
There's a helper function to get a ring buffer page size (the number
of bytes of data recorded on the page), called rb_page_size().
Use that instead of open coding it.
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
The code for resizing the trace ring buffers has to run the per-cpu
resize on the CPU itself. The code was using preempt_off() and
running the code for the current CPU directly, otherwise calling
schedule_work_on().
At least on RT this could result in the following:
|BUG: sleeping function called from invalid context at kernel/rtmutex.c:673
|in_atomic(): 1, irqs_disabled(): 0, pid: 607, name: bash
|3 locks held by bash/607:
|CPU: 0 PID: 607 Comm: bash Not tainted 3.12.15-rt25+ #124
|(rt_spin_lock+0x28/0x68)
|(free_hot_cold_page+0x84/0x3b8)
|(free_buffer_page+0x14/0x20)
|(rb_update_pages+0x280/0x338)
|(ring_buffer_resize+0x32c/0x3dc)
|(free_snapshot+0x18/0x38)
|(tracing_set_tracer+0x27c/0x2ac)
probably via
|cd /sys/kernel/debug/tracing/
|echo 1 > events/enable ; sleep 2
|echo 1024 > buffer_size_kb
If we just always use schedule_work_on(), there's no need for the
preempt_off(). So do that.
Link: http://lkml.kernel.org/p/1405537633-31518-1-git-send-email-cminyard@mvista.com
Reported-by: Stanislav Meduna <stano@meduna.org>
Signed-off-by: Corey Minyard <cminyard@mvista.com>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
ring_buffer_poll_wait() should always put the poll_table to its wait_queue
even there is immediate data available. Otherwise, the following epoll and
read sequence will eventually hang forever:
1. Put some data to make the trace_pipe ring_buffer read ready first
2. epoll_ctl(efd, EPOLL_CTL_ADD, trace_pipe_fd, ee)
3. epoll_wait()
4. read(trace_pipe_fd) till EAGAIN
5. Add some more data to the trace_pipe ring_buffer
6. epoll_wait() -> this epoll_wait() will block forever
~ During the epoll_ctl(efd, EPOLL_CTL_ADD,...) call in step 2,
ring_buffer_poll_wait() returns immediately without adding poll_table,
which has poll_table->_qproc pointing to ep_poll_callback(), to its
wait_queue.
~ During the epoll_wait() call in step 3 and step 6,
ring_buffer_poll_wait() cannot add ep_poll_callback() to its wait_queue
because the poll_table->_qproc is NULL and it is how epoll works.
~ When there is new data available in step 6, ring_buffer does not know
it has to call ep_poll_callback() because it is not in its wait queue.
Hence, block forever.
Other poll implementation seems to call poll_wait() unconditionally as the very
first thing to do. For example, tcp_poll() in tcp.c.
Link: http://lkml.kernel.org/p/20140610060637.GA14045@devbig242.prn2.facebook.com
Cc: stable@vger.kernel.org # 2.6.27
Fixes: 2a2cc8f7c4 "ftrace: allow the event pipe to be polled"
Reviewed-by: Chris Mason <clm@fb.com>
Signed-off-by: Martin Lau <kafai@fb.com>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
The per_cpu buffers are created one per possible CPU. But these do
not mean that those CPUs are online, nor do they even exist.
With the addition of the ring buffer polling, it assumes that the
caller polls on an existing buffer. But this is not the case if
the user reads trace_pipe from a CPU that does not exist, and this
causes the kernel to crash.
Simple fix is to check the cpu against buffer bitmask against to see
if the buffer was allocated or not and return -ENODEV if it is
not.
More updates were done to pass the -ENODEV back up to userspace.
Link: http://lkml.kernel.org/r/5393DB61.6060707@oracle.com
Reported-by: Sasha Levin <sasha.levin@oracle.com>
Cc: stable@vger.kernel.org # 3.10+
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
Subsystems that want to register CPU hotplug callbacks, as well as perform
initialization for the CPUs that are already online, often do it as shown
below:
get_online_cpus();
for_each_online_cpu(cpu)
init_cpu(cpu);
register_cpu_notifier(&foobar_cpu_notifier);
put_online_cpus();
This is wrong, since it is prone to ABBA deadlocks involving the
cpu_add_remove_lock and the cpu_hotplug.lock (when running concurrently
with CPU hotplug operations).
Instead, the correct and race-free way of performing the callback
registration is:
cpu_notifier_register_begin();
for_each_online_cpu(cpu)
init_cpu(cpu);
/* Note the use of the double underscored version of the API */
__register_cpu_notifier(&foobar_cpu_notifier);
cpu_notifier_register_done();
Fix the tracing ring-buffer code by using this latter form of callback
registration.
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Ingo Molnar <mingo@kernel.org>
Acked-by: Steven Rostedt <rostedt@goodmis.org>
Signed-off-by: Srivatsa S. Bhat <srivatsa.bhat@linux.vnet.ibm.com>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Each sub-buffer (buffer page) has a full 64 bit timestamp. The events on
that page use a 27 bit delta against that timestamp in order to save on
bits written to the ring buffer. If the time between events is larger than
what the 27 bits can hold, a "time extend" event is added to hold the
entire 64 bit timestamp again and the events after that hold a delta from
that timestamp.
As a "time extend" is always paired with an event, it is logical to just
allocate the event with the time extend, to make things a bit more efficient.
Unfortunately, when the pairing code was written, it removed the "delta = 0"
from the first commit on a page, causing the events on the page to be
slightly skewed.
Fixes: 69d1b839f7 "ring-buffer: Bind time extend and data events together"
Cc: stable@vger.kernel.org # 2.6.37+
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>