ftrace can fail to allocate per-CPU ring buffer on systems with a large
number of CPUs coupled while large amounts of cache happening in the
page cache. Currently the ring buffer allocation doesn't retry in the VM
implementation even if direct-reclaim made some progress but still
wasn't able to find a free page. On retrying I see that the allocations
almost always succeed. The retry doesn't happen because __GFP_NORETRY is
used in the tracer to prevent the case where we might OOM, however if we
drop __GFP_NORETRY, we risk destabilizing the system if OOM killer is
triggered. To prevent this situation, use the __GFP_RETRY_MAYFAIL flag
introduced recently [1].
Tested the following still succeeds without destabilizing a system with
1GB memory.
echo 300000 > /sys/kernel/debug/tracing/buffer_size_kb
[1] https://marc.info/?l=linux-mm&m=149820805124906&w=2
Link: http://lkml.kernel.org/r/20170713021416.8897-1-joelaf@google.com
Cc: Tim Murray <timmurray@google.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Joel Fernandes <joelaf@google.com>
Signed-off-by: Steven Rostedt (VMware) <rostedt@goodmis.org>
o Pretty much a full rewrite of the processing of function plugins.
i.e. echo do_IRQ:stacktrace > set_ftrace_filter
o The rewrite was needed to add plugins to be unique to tracing instances.
i.e. mkdir instance/foo; cd instances/foo; echo do_IRQ:stacktrace > set_ftrace_filter
The old way was written very hacky. This removes a lot of those hacks.
o New "function-fork" tracing option. When set, pids in the set_ftrace_pid
will have their children added when the processes with their pids
listed in the set_ftrace_pid file forks.
o Exposure of "maxactive" for kretprobe in kprobe_events
o Allow for builtin init functions to be traced by the function tracer
(via the kernel command line). Module init function tracing will come
in the next release.
o Added more selftests, and have selftests also test in an instance.
-----BEGIN PGP SIGNATURE-----
iQExBAABCAAbBQJZCRchFBxyb3N0ZWR0QGdvb2RtaXMub3JnAAoJEMm5BfJq2Y3L
zuIH/RsLUb8Hj6GmhAvn/tblUDzWyqlXX2h79VVlo/XrWayHYNHnKOmua1WwMZC6
xESXb/AffAc89VWTkKsrwaK7yfRPG6+w8zTZOcFuXSBpqSGG/oey9Fxj5Wqqpche
oJ2UY7ngxANAipkP5GxdYTafFSoWhGZGfUUtW+5tAHoFHzqO2lOjO8olbXP69sON
kVX/b461S20cVvRe5H/F0klXLSc37Tlp5YznXy4H4V4HcJSN1Fb6/uozOXALZ4se
SBpVMWmVVoGJorzj+ic7gVOeohvC8RnR400HbeMVwaI0Lj50noidDj/5Hv8F7T+D
h1B8vATNZLFAFUOSHINCBIu6Vj0=
=t8mg
-----END PGP SIGNATURE-----
Merge tag 'trace-v4.12' of git://git.kernel.org/pub/scm/linux/kernel/git/rostedt/linux-trace
Pull tracing updates from Steven Rostedt:
"New features for this release:
- Pretty much a full rewrite of the processing of function plugins.
i.e. echo do_IRQ:stacktrace > set_ftrace_filter
- The rewrite was needed to add plugins to be unique to tracing
instances. i.e. mkdir instance/foo; cd instances/foo; echo
do_IRQ:stacktrace > set_ftrace_filter The old way was written very
hacky. This removes a lot of those hacks.
- New "function-fork" tracing option. When set, pids in the
set_ftrace_pid will have their children added when the processes
with their pids listed in the set_ftrace_pid file forks.
- Exposure of "maxactive" for kretprobe in kprobe_events
- Allow for builtin init functions to be traced by the function
tracer (via the kernel command line). Module init function tracing
will come in the next release.
- Added more selftests, and have selftests also test in an instance"
* tag 'trace-v4.12' of git://git.kernel.org/pub/scm/linux/kernel/git/rostedt/linux-trace: (60 commits)
ring-buffer: Return reader page back into existing ring buffer
selftests: ftrace: Allow some event trigger tests to run in an instance
selftests: ftrace: Have some basic tests run in a tracing instance too
selftests: ftrace: Have event tests also run in an tracing instance
selftests: ftrace: Make func_event_triggers and func_traceonoff_triggers tests do instances
selftests: ftrace: Allow some tests to be run in a tracing instance
tracing/ftrace: Allow for instances to trigger their own stacktrace probes
tracing/ftrace: Allow for the traceonoff probe be unique to instances
tracing/ftrace: Enable snapshot function trigger to work with instances
tracing/ftrace: Allow instances to have their own function probes
tracing/ftrace: Add a better way to pass data via the probe functions
ftrace: Dynamically create the probe ftrace_ops for the trace_array
tracing: Pass the trace_array into ftrace_probe_ops functions
tracing: Have the trace_array hold the list of registered func probes
ftrace: If the hash for a probe fails to update then free what was initialized
ftrace: Have the function probes call their own function
ftrace: Have each function probe use its own ftrace_ops
ftrace: Have unregister_ftrace_function_probe_func() return a value
ftrace: Add helper function ftrace_hash_move_and_update_ops()
ftrace: Remove data field from ftrace_func_probe structure
...
When reading the ring buffer for consuming, it is optimized for splice,
where a page is taken out of the ring buffer (zero copy) and sent to the
reading consumer. When the read is finished with the page, it calls
ring_buffer_free_read_page(), which simply frees the page. The next time the
reader needs to get a page from the ring buffer, it must call
ring_buffer_alloc_read_page() which allocates and initializes a reader page
for the ring buffer to be swapped into the ring buffer for a new filled page
for the reader.
The problem is that there's no reason to actually free the page when it is
passed back to the ring buffer. It can hold it off and reuse it for the next
iteration. This completely removes the interaction with the page_alloc
mechanism.
Using the trace-cmd utility to record all events (causing trace-cmd to
require reading lots of pages from the ring buffer, and calling
ring_buffer_alloc/free_read_page() several times), and also assigning a
stack trace trigger to the mm_page_alloc event, we can see how many times
the ring_buffer_alloc_read_page() needed to allocate a page for the ring
buffer.
Before this change:
# trace-cmd record -e all -e mem_page_alloc -R stacktrace sleep 1
# trace-cmd report |grep ring_buffer_alloc_read_page | wc -l
9968
After this change:
# trace-cmd record -e all -e mem_page_alloc -R stacktrace sleep 1
# trace-cmd report |grep ring_buffer_alloc_read_page | wc -l
4
Signed-off-by: Steven Rostedt (VMware) <rostedt@goodmis.org>
I noticed that reading the snapshot file when it is empty no longer gives a
status. It suppose to show the status of the snapshot buffer as well as how
to allocate and use it. For example:
># cat snapshot
# tracer: nop
#
#
# * Snapshot is allocated *
#
# Snapshot commands:
# echo 0 > snapshot : Clears and frees snapshot buffer
# echo 1 > snapshot : Allocates snapshot buffer, if not already allocated.
# Takes a snapshot of the main buffer.
# echo 2 > snapshot : Clears snapshot buffer (but does not allocate or free)
# (Doesn't have to be '2' works with any number that
# is not a '0' or '1')
But instead it just showed an empty buffer:
># cat snapshot
# tracer: nop
#
# entries-in-buffer/entries-written: 0/0 #P:4
#
# _-----=> irqs-off
# / _----=> need-resched
# | / _---=> hardirq/softirq
# || / _--=> preempt-depth
# ||| / delay
# TASK-PID CPU# |||| TIMESTAMP FUNCTION
# | | | |||| | |
What happened was that it was using the ring_buffer_iter_empty() function to
see if it was empty, and if it was, it showed the status. But that function
was returning false when it was empty. The reason was that the iter header
page was on the reader page, and the reader page was empty, but so was the
buffer itself. The check only tested to see if the iter was on the commit
page, but the commit page was no longer pointing to the reader page, but as
all pages were empty, the buffer is also.
Cc: stable@vger.kernel.org
Fixes: 651e22f270 ("ring-buffer: Always reset iterator to reader page")
Signed-off-by: Steven Rostedt (VMware) <rostedt@goodmis.org>
In case of error, the function kthread_run() returns ERR_PTR()
and never returns NULL. The NULL test in the return value check
should be replaced with IS_ERR().
Link: http://lkml.kernel.org/r/1466184839-14927-1-git-send-email-weiyj_lk@163.com
Cc: stable@vger.kernel.org
Fixes: 6c43e554a ("ring-buffer: Add ring buffer startup selftest")
Signed-off-by: Wei Yongjun <yongjun_wei@trendmicro.com.cn>
Signed-off-by: Steven Rostedt (VMware) <rostedt@goodmis.org>
We are going to split <linux/sched/clock.h> out of <linux/sched.h>, which
will have to be picked up from other headers and .c files.
Create a trivial placeholder <linux/sched/clock.h> file that just
maps to <linux/sched.h> to make this patch obviously correct and
bisectable.
Include the new header in the files that are going to need it.
Acked-by: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
o STM can hook into the function tracer
o Function filtering now supports more advance glob matching
o Ftrace selftests updates and added tests
o Softirq tag in traces now show only softirqs
o ARM nop added to non traced locations at compile time
o New trace_marker_raw file that allows for binary input
o Optimizations to the ring buffer
o Removal of kmap in trace_marker
o Wakeup and irqsoff tracers now adhere to the set_graph_notrace file
o Other various fixes and clean ups
Note, there are two patches marked for stable. These were discovered
near the end of the 4.9 rc release cycle. By the time I had them tested
it was just a matter of days before 4.9 would be released, and I
figured I would just submit them in the merge window. They are old
bugs and not critical. Nothing non-root could abuse.
-----BEGIN PGP SIGNATURE-----
iQExBAABCAAbBQJYUrFHFBxyb3N0ZWR0QGdvb2RtaXMub3JnAAoJEMm5BfJq2Y3L
2+AIAIr20kSQV/nA5htGAeCTobVk3WUxY6bvjd9mIJDKPP19akNLyREW0G3KnfCr
yhx4aFRZG98fRu/6F8qieRosyN36lADDVYHelMFHMpcTOpE2aZGjaaOuNGxOEA9v
FmMPTX+K3+dzKyFP4l68R3+5JuQ1/AqLTioTWeLW8IDQ2OOVsjD8+0BuXrNKMJDY
o6U4Hk5U/vn+zHc6BmgBzloAXemBd7iJ1t5V3FRRGvm8yv3HU85Twc5ofGeYTWvB
J8PboEywRlIzxg0Kd8mxnMI5PgaKZSEc2ub8E7cY/CZ5PYpDE2xDA2hJmJgfYp00
1VW+DHRpRZfElsCcya6S6P4bs5Y=
=MGZ/
-----END PGP SIGNATURE-----
Merge tag 'trace-v4.10' of git://git.kernel.org/pub/scm/linux/kernel/git/rostedt/linux-trace
Pull tracing updates from Steven Rostedt:
"This release has a few updates:
- STM can hook into the function tracer
- Function filtering now supports more advance glob matching
- Ftrace selftests updates and added tests
- Softirq tag in traces now show only softirqs
- ARM nop added to non traced locations at compile time
- New trace_marker_raw file that allows for binary input
- Optimizations to the ring buffer
- Removal of kmap in trace_marker
- Wakeup and irqsoff tracers now adhere to the set_graph_notrace file
- Other various fixes and clean ups"
* tag 'trace-v4.10' of git://git.kernel.org/pub/scm/linux/kernel/git/rostedt/linux-trace: (42 commits)
selftests: ftrace: Shift down default message verbosity
kprobes/trace: Fix kprobe selftest for newer gcc
tracing/kprobes: Add a helper method to return number of probe hits
tracing/rb: Init the CPU mask on allocation
tracing: Use SOFTIRQ_OFFSET for softirq dectection for more accurate results
tracing/fgraph: Have wakeup and irqsoff tracers ignore graph functions too
fgraph: Handle a case where a tracer ignores set_graph_notrace
tracing: Replace kmap with copy_from_user() in trace_marker writing
ftrace/x86_32: Set ftrace_stub to weak to prevent gcc from using short jumps to it
tracing: Allow benchmark to be enabled at early_initcall()
tracing: Have system enable return error if one of the events fail
tracing: Do not start benchmark on boot up
tracing: Have the reg function allow to fail
ring-buffer: Force rb_end_commit() and rb_set_commit_to_write() inline
ring-buffer: Froce rb_update_write_stamp() to be inlined
ring-buffer: Force inline of hotpath helper functions
tracing: Make __buffer_unlock_commit() always_inline
tracing: Make tracepoint_printk a static_key
ring-buffer: Always inline rb_event_data()
ring-buffer: Make rb_reserve_next_event() always inlined
...
Before commit b32614c034 ("tracing/rb: Convert to hotplug state
machine") the allocated cpumask was initialized to the mask of ONLINE or
POSSIBLE CPUs. After the CPU hotplug changes the buffer initialisation
moved to trace_rb_cpu_prepare() but I forgot to initially set the
cpumask to zero. This is done now.
Link: http://lkml.kernel.org/r/20161207133133.hzkcqfllxcdi3joz@linutronix.de
Fixes: b32614c034 ("tracing/rb: Convert to hotplug state machine")
Reported-by: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
Tested-by: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
Before commit b32614c034 ("tracing/rb: Convert to hotplug state machine")
the allocated cpumask was initialized to the mask of online or possible
CPUs. After the CPU hotplug changes the buffer initialization moved to
trace_rb_cpu_prepare() but the cpumask is allocated with alloc_cpumask()
and therefor has random content. As a consequence the cpu buffers are not
initialized and a later access dereferences a NULL pointer.
Use zalloc_cpumask() instead so trace_rb_cpu_prepare() initializes the
buffers properly.
Fixes: b32614c034 ("tracing/rb: Convert to hotplug state machine")
Reported-by: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Cc: rostedt@goodmis.org
Link: http://lkml.kernel.org/r/20161207133133.hzkcqfllxcdi3joz@linutronix.de
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Install the callbacks via the state machine. The notifier in struct
ring_buffer is replaced by the multi instance interface. Upon
__ring_buffer_alloc() invocation, cpuhp_state_add_instance() will invoke
the trace_rb_cpu_prepare() on each CPU.
This callback may now fail. This means __ring_buffer_alloc() will fail and
cleanup (like previously) and during a CPU up event this failure will not
allow the CPU to come up.
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: rt@linutronix.de
Link: http://lkml.kernel.org/r/20161126231350.10321-7-bigeasy@linutronix.de
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
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
Joel Fernandes