commit 524a38682573b2e15ab6317ccfe50280441514be upstream. Some archs (specifically PowerPC), are sensitive with the ordering of the enabling of the calls to function tracing and setting of the function to use to be traced. That is, update_ftrace_function() sets what function the ftrace_caller trampoline should call. Some archs require this to be set before calling ftrace_run_update_code(). Another bug was discovered, that ftrace_startup_sysctl() called ftrace_run_update_code() directly. If the function the ftrace_caller trampoline changes, then it will not be updated. Instead a call to ftrace_startup_enable() should be called because it tests to see if the callback changed since the code was disabled, and will tell the arch to update appropriately. Most archs do not need this notification, but PowerPC does. The problem could be seen by the following commands: # echo 0 > /proc/sys/kernel/ftrace_enabled # echo function > /sys/kernel/debug/tracing/current_tracer # echo 1 > /proc/sys/kernel/ftrace_enabled # cat /sys/kernel/debug/tracing/trace The trace will show that function tracing was not active. Signed-off-by: Steven Rostedt <rostedt@goodmis.org> Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
commit 1619dc3f8f555ee1cdd3c75db3885d5715442b12 upstream. When ftrace is enabled globally through the proc interface, we must check if ftrace_graph_active is set. If it is set, then we should also pass the FTRACE_START_FUNC_RET command to ftrace_run_update_code(). Similarly, when ftrace is disabled globally through the proc interface, we must check if ftrace_graph_active is set. If it is set, then we should also pass the FTRACE_STOP_FUNC_RET command to ftrace_run_update_code(). Consider the following situation. # echo 0 > /proc/sys/kernel/ftrace_enabled After this ftrace_enabled = 0. # echo function_graph > /sys/kernel/debug/tracing/current_tracer Since ftrace_enabled = 0, ftrace_enable_ftrace_graph_caller() is never called. # echo 1 > /proc/sys/kernel/ftrace_enabled Now ftrace_enabled will be set to true, but still ftrace_enable_ftrace_graph_caller() will not be called, which is not desired. Further if we execute the following after this: # echo nop > /sys/kernel/debug/tracing/current_tracer Now since ftrace_enabled is set it will call ftrace_disable_ftrace_graph_caller(), which causes a kernel warning on the ARM platform. On the ARM platform, when ftrace_enable_ftrace_graph_caller() is called, it checks whether the old instruction is a nop or not. If it's not a nop, then it returns an error. If it is a nop then it replaces instruction at that address with a branch to ftrace_graph_caller. ftrace_disable_ftrace_graph_caller() behaves just the opposite. Therefore, if generic ftrace code ever calls either ftrace_enable_ftrace_graph_caller() or ftrace_disable_ftrace_graph_caller() consecutively two times in a row, then it will return an error, which will cause the generic ftrace code to raise a warning. Note, x86 does not have an issue with this because the architecture specific code for ftrace_enable_ftrace_graph_caller() and ftrace_disable_ftrace_graph_caller() does not check the previous state, and calling either of these functions twice in a row has no ill effect. Link: http://lkml.kernel.org/r/e4fbe64cdac0dd0e86a3bf914b0f83c0b419f146.1425666454.git.panand@redhat.com Signed-off-by: Pratyush Anand <panand@redhat.com> [ removed extra if (ftrace_start_up) and defined ftrace_graph_active as 0 if CONFIG_FUNCTION_GRAPH_TRACER is not set. ] Signed-off-by: Steven Rostedt <rostedt@goodmis.org> Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
commit b24d443b8f17d9776f5fc1f6c780a0a21eb02913 upstream. When /proc/sys/kernel/ftrace_enabled is set to zero, all function tracing is disabled. But the records that represent the functions still hold information about the ftrace_ops that are hooked to them. ftrace_ops may request "REGS" (have a full set of pt_regs passed to the callback), or "TRAMP" (the ops has its own trampoline to use). When the record is updated to represent the state of the ops hooked to it, it sets "REGS_EN" and/or "TRAMP_EN" to state that the callback points to the correct trampoline (REGS has its own trampoline). When ftrace_enabled is set to zero, all ftrace locations are a nop, so they do not point to any trampoline. But the _EN flags are still set. This can cause the accounting to go wrong when ftrace_enabled is cleared and an ops that has a trampoline is registered or unregistered. For example, the following will cause ftrace to crash: # echo function_graph > /sys/kernel/debug/tracing/current_tracer # echo 0 > /proc/sys/kernel/ftrace_enabled # echo nop > /sys/kernel/debug/tracing/current_tracer # echo 1 > /proc/sys/kernel/ftrace_enabled # echo function_graph > /sys/kernel/debug/tracing/current_tracer As function_graph uses a trampoline, when ftrace_enabled is set to zero the updates to the record are not done. When enabling function_graph again, the record will still have the TRAMP_EN flag set, and it will look for an op that has a trampoline other than the function_graph ops, and fail to find one. Reported-by: Pratyush Anand <panand@redhat.com> Signed-off-by: Steven Rostedt <rostedt@goodmis.org> Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
commit 1e0d6714aceb770b04161fbedd7765d0e1fc27bd upstream. 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: Rabin Vincent <rabin@rab.in> Fixes: e30f53aad220 "tracing: Do not busy wait in buffer splice" Signed-off-by: Steven Rostedt <rostedt@goodmis.org> Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
commit 7215853e985a4bef1a6c14e00e89dfec84f1e457 upstream. Commit 6edb2a8a385f0cdef51dae37ff23e74d76d8a6ce introduced an array map_pages that contains the addresses returned by kmap_atomic. However, when unmapping those pages, map_pages[0] is unmapped before map_pages[1], breaking the nesting requirement as specified in the documentation for kmap_atomic/kunmap_atomic. This was caught by the highmem debug code present in kunmap_atomic. Fix the loop to do the unmapping properly. Link: http://lkml.kernel.org/r/1418871056-6614-1-git-send-email-markivx@codeaurora.org Reviewed-by: Stephen Boyd <sboyd@codeaurora.org> Reported-by: Lime Yang <limey@codeaurora.org> Signed-off-by: Vikram Mulukutla <markivx@codeaurora.org> Signed-off-by: Steven Rostedt <rostedt@goodmis.org> Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Pull ftrace fixes from Steven Rostedt:
"This holds a few fixes to the ftrace infrastructure as well as the
mixture of function graph tracing and kprobes.
When jprobes and function graph tracing is enabled at the same time it
will crash the system:
# modprobe jprobe_example
# echo function_graph > /sys/kernel/debug/tracing/current_tracer
After the first fork (jprobe_example probes it), the system will
crash.
This is due to the way jprobes copies the stack frame and does not do
a normal function return. This messes up with the function graph
tracing accounting which hijacks the return address from the stack and
replaces it with a hook function. It saves the return addresses in a
separate stack to put back the correct return address when done. But
because the jprobe functions do not do a normal return, their stack
addresses are not put back until the function they probe is called,
which means that the probed function will get the return address of
the jprobe handler instead of its own.
The simple fix here was to disable function graph tracing while the
jprobe handler is being called.
While debugging this I found two minor bugs with the function graph
tracing.
The first was about the function graph tracer sharing its function
hash with the function tracer (they both get filtered by the same
input). The changing of the set_ftrace_filter would not sync the
function recording records after a change if the function tracer was
disabled but the function graph tracer was enabled. This was due to
the update only checking one of the ops instead of the shared ops to
see if they were enabled and should perform the sync. This caused the
ftrace accounting to break and a ftrace_bug() would be triggered,
disabling ftrace until a reboot.
The second was that the check to update records only checked one of
the filter hashes. It needs to test both the "filter" and "notrace"
hashes. The "filter" hash determines what functions to trace where as
the "notrace" hash determines what functions not to trace (trace all
but these). Both hashes need to be passed to the update code to find
out what change is being done during the update. This also broke the
ftrace record accounting and triggered a ftrace_bug().
This patch set also include two more fixes that were reported
separately from the kprobe issue.
One was that init_ftrace_syscalls() was called twice at boot up. This
is not a major bug, but that call performed a rather large kmalloc
(NR_syscalls * sizeof(*syscalls_metadata)). The second call made the
first one a memory leak, and wastes memory.
The other fix is a regression caused by an update in the v3.19 merge
window. The moving to enable events early, moved the enabling before
PID 1 was created. The syscall events require setting the
TIF_SYSCALL_TRACEPOINT for all tasks. But for_each_process_thread()
does not include the swapper task (PID 0), and ended up being a nop.
A suggested fix was to add the init_task() to have its flag set, but I
didn't really want to mess with PID 0 for this minor bug. Instead I
disable and re-enable events again at early_initcall() where it use to
be enabled. This also handles any other event that might have its own
reg function that could break at early boot up"
* tag 'trace-fixes-v3.19-rc3' of git://git.kernel.org/pub/scm/linux/kernel/git/rostedt/linux-trace:
tracing: Fix enabling of syscall events on the command line
tracing: Remove extra call to init_ftrace_syscalls()
ftrace/jprobes/x86: Fix conflict between jprobes and function graph tracing
ftrace: Check both notrace and filter for old hash
ftrace: Fix updating of filters for shared global_ops filters
Commit 5f893b2639b2 "tracing: Move enabling tracepoints to just after rcu_init()" broke the enabling of system call events from the command line. The reason was that the enabling of command line trace events was moved before PID 1 started, and the syscall tracepoints require that all tasks have the TIF_SYSCALL_TRACEPOINT flag set. But the swapper task (pid 0) is not part of that. Since the swapper task is the only task that is running at this early in boot, no task gets the flag set, and the tracepoint never gets reached. Instead of setting the swapper task flag (there should be no reason to do that), re-enabled trace events again after the init thread (PID 1) has been started. It requires disabling all command line events and re-enabling them, as just enabling them again will not reset the logic to set the TIF_SYSCALL_TRACEPOINT flag, as the syscall tracepoint will be fooled into thinking that it was already set, and wont try setting it again. For this reason, we must first disable it and re-enable it. Link: http://lkml.kernel.org/r/1421188517-18312-1-git-send-email-mpe@ellerman.id.au Link: http://lkml.kernel.org/r/20150115040506.216066449@goodmis.org Reported-by: Michael Ellerman <mpe@ellerman.id.au> Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
trace_init() calls init_ftrace_syscalls() and then calls trace_event_init() which also calls init_ftrace_syscalls(). It makes more sense to only call it from trace_event_init(). Calling it twice wastes memory, as it allocates the syscall events twice, and loses the first copy of it. Link: http://lkml.kernel.org/r/54AF53BD.5070303@huawei.com Link: http://lkml.kernel.org/r/20150115040505.930398632@goodmis.org Reported-by: Wang Nan <wangnan0@huawei.com> Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
Using just the filter for checking for trampolines or regs is not enough when updating the code against the records that represent all functions. Both the filter hash and the notrace hash need to be checked. To trigger this bug (using trace-cmd and perf): # perf probe -a do_fork # trace-cmd start -B foo -e probe # trace-cmd record -p function_graph -n do_fork sleep 1 The trace-cmd record at the end clears the filter before it disables function_graph tracing and then that causes the accounting of the ftrace function records to become incorrect and causes ftrace to bug. Link: http://lkml.kernel.org/r/20150114154329.358378039@goodmis.org Cc: stable@vger.kernel.org [ still need to switch old_hash_ops to old_ops_hash ] Reviewed-by: Masami Hiramatsu <masami.hiramatsu.pt@hitachi.com> Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
As the set_ftrace_filter affects both the function tracer as well as the function graph tracer, the ops that represent each have a shared ftrace_ops_hash structure. This allows both to be updated when the filter files are updated. But if function graph is enabled and the global_ops (function tracing) ops is not, then it is possible that the filter could be changed without the update happening for the function graph ops. This will cause the changes to not take place and may even cause a ftrace_bug to occur as it could mess with the trampoline accounting. The solution is to check if the ops uses the shared global_ops filter and if the ops itself is not enabled, to check if there's another ops that is enabled and also shares the global_ops filter. In that case, the modification still needs to be executed. Link: http://lkml.kernel.org/r/20150114154329.055980438@goodmis.org Cc: stable@vger.kernel.org # 3.17+ Reviewed-by: Masami Hiramatsu <masami.hiramatsu.pt@hitachi.com> Signed-off-by: Steven Rostedt <rostedt@goodmis.org>