- 07 Apr, 2009 1 commit
-
-
Peter Zijlstra authored
Update the userspace read method. Paul noted that: - userspace cannot observe ->lock & 1 on the same cpu. - we need a barrier() between reading ->lock and ->index to ensure we read them in that prticular order. Signed-off-by:
Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Paul Mackerras <paulus@samba.org> Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com> LKML-Reference: <20090406094517.368446033@chello.nl> Signed-off-by:
Ingo Molnar <mingo@elte.hu>
-
- 06 Apr, 2009 39 commits
-
-
Peter Zijlstra authored
Paul noted that we don't need SMP barriers for the mmap() counter read because its always on the same cpu (otherwise you can't access the hw counter anyway). So remove the SMP barriers and replace them with regular compiler barriers. Further, update the comment to include a race free method of reading said hardware counter. The primary change is putting the pmc_read inside the seq-loop, otherwise we can still race and read rubbish. Noticed-by:
Paul Mackerras <paulus@samba.org> Signed-off-by:
-
Peter Zijlstra authored
Put in counts to tell which ips belong to what context. ----- | | hv | -- nr | | kernel | -- | | user ----- Signed-off-by:
-
Peter Zijlstra authored
Update to reflect the new record_type ABI changes. Signed-off-by:
-
Peter Zijlstra authored
By request, provide a way to request a wakeup every 'n' events instead of every page of output. Signed-off-by:
-
Peter Zijlstra authored
Per suggestion from Paul, move the event overflow bits to record_type and sanitize the enums a bit. Breaks the ABI -- again ;-) Suggested-by:
-
Mike Galbraith authored
Decouple kerneltop display from event acquisition by introducing a separate data acquisition thread. This fixes annnoying kerneltop display refresh jitter and missed events. Also add a -r <prio> option, to switch the data acquisition thread to real-time priority. Signed-off-by:
Mike Galbraith <efault@gmx.de> Cc: Paul Mackerras <paulus@samba.org> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Orig-LKML-Reference: <new-submission> Signed-off-by:
-
Peter Zijlstra authored
Follow the example set by powerpc and try to play nice with oprofile and the nmi watchdog. Signed-off-by:
-
Peter Zijlstra authored
Provide the x86 perf_callchain() implementation. Code based on the ftrace/sysprof code from Soeren Sandmann Pedersen. Signed-off-by:
-
Peter Zijlstra authored
Provide the generic callchain support bits. If hw_event->callchain is set the arch specific perf_callchain() function is called upon to provide a perf_callchain_entry structure filled with the current callchain. If it does so, it is added to the overflow output event. Signed-off-by:
-
Peter Zijlstra authored
Go along with the new perf_event_type ABI. Signed-off-by:
-
Peter Zijlstra authored
Breaks ABI yet again :-) Change the event type so that [0, 2^31-1] are regular event types, but [2^31, 2^32-1] forms a bitmask for overflow events. Signed-off-by:
-
Peter Zijlstra authored
Move the nmi argument to the _begin() function, so that _end() only needs the handle. This allows the _begin() function to generate a wakeup on event loss. Signed-off-by:
-
Paul Mackerras authored
Impact: new functionality This adds add an option to the perfstat mode of kerneltop to scale the reported counter values according to the fraction of time that each counter gets to count. This is invoked with the -l option (I used 'l' because s, c, a and e were all taken already.) This uses the new PERF_RECORD_TOTAL_TIME_{ENABLED,RUNNING} read format options. With this, we get output like this: $ ./perfstat -l -e 0:0,0:1,0:2,0:3,0:4,0:5 ./spin Performance counter stats for './spin': 4016072055 CPU cycles (events) (scaled from 66.53%) 2005887318 instructions (events) (scaled from 66.53%) 1762849 cache references (events) (scaled from 66.69%) 165229 cache misses (events) (scaled from 66.85%) 1001298009 branches (events) (scaled from 66.78%) 41566 branch misses (events) (scaled from 66.61%) Wall-clock time elapsed: 2438.227446 msecs This also lets us detect when a counter is zero because the counter never got to go on the CPU at all. In that case we print <not counted> rather than 0. Signed-off-by: -
Peter Zijlstra authored
Now that Paul cleaned up the error propagation paths, pass down the x86 error as well. Signed-off-by:
-
Paul Mackerras authored
Impact: better error reporting At present, if hw_perf_counter_init encounters an error, all it can do is return NULL, which causes sys_perf_counter_open to return an EINVAL error to userspace. This isn't very informative for userspace; it means that userspace can't tell the difference between "sorry, oprofile is already using the PMU" and "we don't support this CPU" and "this CPU doesn't support the requested generic hardware event". This commit uses the PTR_ERR/ERR_PTR/IS_ERR set of macros to let hw_perf_counter_init return an error code on error rather than just NULL if it wishes. If it does so, that error code will be returned from sys_perf_counter_open to userspace. If it returns NULL, an EINVAL error will be returned to userspace, as before. This also adapts the powerpc hw_perf_counter_init to make use of this to return ENXIO, EINVAL, EBUSY, or EOPNOTSUPP as appropriate. It would be good to add extra error numbers in future to allow userspace to distinguish the various errors that are currently reported as EINVAL, i.e. irq_period < 0, too many events in a group, conflict between exclude_* settings in a group, and PMU resource conflict in a group. [ v2: fix a bug pointed out by Corey Ashford where error returns from hw_perf_counter_init were not handled correctly in the case of raw hardware events.] Signed-off-by: -
Paul Mackerras authored
Impact: cooperate with oprofile At present, on PowerPC, if you have perf_counters compiled in, oprofile doesn't work. There is code to allow the PMU to be shared between competing subsystems, such as perf_counters and oprofile, but currently the perf_counter subsystem reserves the PMU for itself at boot time, and never releases it. This makes perf_counter play nicely with oprofile. Now we keep a count of how many perf_counter instances are counting hardware events, and reserve the PMU when that count becomes non-zero, and release the PMU when that count becomes zero. This means that it is possible to have perf_counters compiled in and still use oprofile, as long as there are no hardware perf_counters active. This also means that if oprofile is active, sys_perf_counter_open will fail if the hw_event specifies a hardware event. To avoid races with other tasks creating and destroying perf_counters, we use a mutex. We use atomic_inc_not_zero and atomic_add_unless to avoid having to take the mutex unless there is a possibility of the count going between 0 and 1. Signed-off-by:
-
Peter Zijlstra authored
frob the kerneltop code to print the mmap data in the stream Better use would be collecting the IPs per PID and mapping them onto the provided userspace code.. TODO Signed-off-by:
-
Peter Zijlstra authored
Currently the profiling information returns userspace IPs but no way to correlate them to userspace code. Userspace could look into /proc/$pid/maps but that might not be current or even present anymore at the time of analyzing the IPs. Therefore provide means to track the mmap information and provide it in the output stream. XXX: only covers mmap()/munmap(), mremap() and mprotect() are missing. Signed-off-by:
-
Peter Zijlstra authored
Now that the kernel side changed, match up again. Signed-off-by:
-
Peter Zijlstra authored
It just occured to me it is possible to have multiple contending updates of the userpage (mmap information vs overflow vs counter). This would break the seqlock logic. It appear the arch code uses this from NMI context, so we cannot possibly serialize its use, therefore separate the data_head update from it and let it return to its original use. The arch code needs to make sure there are no contending callers by disabling the counter before using it -- powerpc appears to do this nicely. Signed-off-by:
-
Peter Zijlstra authored
While going over the wakeup code I noticed delayed wakeups only work for hardware counters but basically all software counters rely on them. This patch unifies and generalizes the delayed wakeup to fix this issue. Since we're dealing with NMI context bits here, use a cmpxchg() based single link list implementation to track counters that have pending wakeups. [ This should really be generic code for delayed wakeups, but since we cannot use cmpxchg()/xchg() in generic code, I've let it live in the perf_counter code. -- Eric Dumazet could use it to aggregate the network wakeups. ] Furthermore, the x86 method of using TIF flags was flawed in that its quite possible to end up setting the bit on the idle task, loosing the wakeup. The powerpc method uses per-cpu storage and does appear to be sufficient. Signed-off-by:
-
Paul Mackerras authored
Impact: new functionality Currently, if there are more counters enabled than can fit on the CPU, the kernel will multiplex the counters on to the hardware using round-robin scheduling. That isn't too bad for sampling counters, but for counting counters it means that the value read from a counter represents some unknown fraction of the true count of events that occurred while the counter was enabled. This remedies the situation by keeping track of how long each counter is enabled for, and how long it is actually on the cpu and counting events. These times are recorded in nanoseconds using the task clock for per-task counters and the cpu clock for per-cpu counters. These values can be supplied to userspace on a read from the counter. Userspace requests that they be supplied after the counter value by setting the PERF_FORMAT_TOTAL_TIME_ENABLED and/or PERF_FORMAT_TOTAL_TIME_RUNNING bits in the hw_event.read_format field when creating the counter. (There is no way to change the read format after the counter is created, though it would be possible to add some way to do that.) Using this information it is possible for userspace to scale the count it reads from the counter to get an estimate of the true count: true_count_estimate = count * total_time_enabled / total_time_running This also lets userspace detect the situation where the counter never got to go on the cpu: total_time_running == 0. This functionality has been requested by the PAPI developers, and will be generally needed for interpreting the count values from counting counters correctly. In the implementation, this keeps 5 time values (in nanoseconds) for each counter: total_time_enabled and total_time_running are used when the counter is in state OFF or ERROR and for reporting back to userspace. When the counter is in state INACTIVE or ACTIVE, it is the tstamp_enabled, tstamp_running and tstamp_stopped values that are relevant, and total_time_enabled and total_time_running are determined from them. (tstamp_stopped is only used in INACTIVE state.) The reason for doing it like this is that it means that only counters being enabled or disabled at sched-in and sched-out time need to be updated. There are no new loops that iterate over all counters to update total_time_enabled or total_time_running. This also keeps separate child_total_time_running and child_total_time_enabled fields that get added in when reporting the totals to userspace. They are separate fields so that they can be atomic. We don't want to use atomics for total_time_running, total_time_enabled etc., because then we would have to use atomic sequences to update them, which are slower than regular arithmetic and memory accesses. It is possible to measure total_time_running by adding a task_clock counter to each group of counters, and total_time_enabled can be measured approximately with a top-level task_clock counter (though inaccuracies will creep in if you need to disable and enable groups since it is not possible in general to disable/enable the top-level task_clock counter simultaneously with another group). However, that adds extra overhead - I measured around 15% increase in the context switch latency reported by lat_ctx (from lmbench) when a task_clock counter was added to each of 2 groups, and around 25% increase when a task_clock counter was added to each of 4 groups. (In both cases a top-level task-clock counter was also added.) In contrast, the code added in this commit gives better information with no overhead that I could measure (in fact in some cases I measured lower times with this code, but the differences were all less than one standard deviation). [ v2: address review comments by Andrew Morton. ] Signed-off-by:
-
Peter Zijlstra authored
A brainfart stopped single page mmap()s working. The rest of the code should be perfectly fine with not having any data pages. Reported-by:
-
Peter Zijlstra authored
Teach kerneltop about the new output ABI. XXX: anybody fancy integrating the PID/TID data into the output? Bump the mmap_data pages a little because we bloated the output and have to be more careful about overruns with structured data. Signed-off-by:
-
Peter Zijlstra authored
provide a knob to set the number of mmap data pages. Signed-off-by:
-
Peter Zijlstra authored
Allow cpu wide counters to profile userspace by providing what process the sample belongs to. This raises the first issue with the output type, lots of these options: group, tid, callchain, etc.. are non-exclusive and could be combined, suggesting a bitfield. However, things like the mmap() data stream doesn't fit in that. How to split the type field... Signed-off-by:
-
Peter Zijlstra authored
Ensure we never write more than we said we would. Signed-off-by:
-
Peter Zijlstra authored
Provide a {type,size} header for each output entry. This should provide extensible output, and the ability to mix multiple streams. Signed-off-by: -
Peter Zijlstra authored
Provide a begin, copy, end interface to the output buffer. begin() reserves the space, copy() copies the data over, considering page boundaries, end() finalizes the event and does the wakeup. Signed-off-by:
-
Peter Zijlstra authored
Impact: fix kerneltop 100% CPU usage Only return a poll event when there's actually been one, poll_wait() doesn't actually wait for the waitq you pass it, it only enqueues you on it. Only once all FDs have been iterated and none of thm returned a poll-event will it schedule(). Also make it return POLL_HUP when there's not mmap() area to read from. Further, fix a silly bug in the write code. Reported-by:
-
Paul Mackerras authored
Impact: documentation fix This updates the perfcounter documentation to reflect recent changes. Signed-off-by: -
Peter Zijlstra authored
Paul Mackerras wrote: > I noticed the poll stuff is bogus - we have a 2D array of struct > pollfds (MAX_NR_CPUS x MAX_COUNTERS), we fill in a sub-array (with the > rest being uninitialized, since the array is on the stack) and then > pass the first nr_cpus elements to poll. Not what we really meant, I > suspect. :) Not even if we only have one counter, since it's the > counter dimension that varies fastest. This should fix the most obvious poll fubar.. not enough to fix the full problem though.. Reported-by:
-
Paul Mackerras authored
The glib dependency in kerneltop.c is only for a little bit of list manipulation, and I find it inconvenient. This adds a 'next' field to struct source_line, which lets us link them together into a list. The code to do the linking ourselves turns out to be no longer or more difficult than using glib. This also fixes a few other problems: - We need to #include <limits.h> to get PATH_MAX on powerpc. - We need to #include <linux/types.h> rather than have our own definitions of __u64 and __s64; on powerpc the installed headers define them to be unsigned long and long respectively, and if we have our own, different definition here that causes a compile error. - This takes out the x86 setting of errno from -ret in sys_perf_counter_open. My experiments on x86 indicate that the glibc syscall() does this for us already. - We had two CPU migration counters in the default set, which seems unnecessary; I changed one of them to a context switch counter. - In perfstat mode we were printing CPU cycles and instructions as milliseconds, and the cpu clock and task clock counters as events. This fixes that. - In perfstat mode we were still printing a blank line after the first counter, which was a holdover from when a task clock counter was automatically included as the first counter. This removes the blank line. - On a test machine here, parse_symbols() and parse_vmlinux() were taking long enough (almost 0.5 seconds) for the mmap buffer to overflow before we got to the first mmap_read() call, so this moves them before we open all the counters. - The error message if sys_perf_counter_open fails needs to use errno, not -fd[i][counter]. Signed-off-by:
-
Ingo Molnar authored
Commit b7368fdd7d decreased the CPU cycles interval 100-fold, but this is causig kerneltop failures on my Nehalem box: aldebaran:/home/mingo/linux/linux/Documentation/perf_counter> ./kerneltop KernelTop refresh period: 2 seconds ERROR: failed to keep up with mmap data 10,000 cycles is way too short. What we should do instead on mostly-idle systems is some sort of read/poll timeout, so that we display something every 2 seconds for sure. Cc: Wu Fengguang <fengguang.wu@intel.com> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Paul Mackerras <paulus@samba.org> Signed-off-by: -
Ingo Molnar authored
Fix: kerneltop.c: In function ‘record_ip’: kerneltop.c:1005: warning: format ‘%016llx’ expects type ‘long long unsigned int’, but argument 2 has type ‘uint64_t’ Cc: Wu Fengguang <fengguang.wu@intel.com> Cc: Paul Mackerras <paulus@samba.org> Acked-by:
-
Ingo Molnar authored
Remove now unified perfstat.c and perf_counter.h, and link to the in-kernel perf_counter.h. Cc: Wu Fengguang <fengguang.wu@intel.com> Cc: Paul Mackerras <paulus@samba.org> Acked-by:
-
Peter Zijlstra authored
update kerneltop to use the mmap() output to gather overflow information Signed-off-by:
-
Peter Zijlstra authored
update the kerneltop userspace to work with the latest syscall ABI Signed-off-by:
-
Peter Zijlstra authored
Impact: Rework the perfcounter output ABI use sys_read() only for instant data and provide mmap() output for all async overflow data. The first mmap() determines the size of the output buffer. The mmap() size must be a PAGE_SIZE multiple of 1+pages, where pages must be a power of 2 or 0. Further mmap()s of the same fd must have the same size. Once all maps are gone, you can again mmap() with a new size. In case of 0 extra pages there is no data output and the first page only contains meta data. When there are data pages, a poll() event will be generated for each full page of data. Furthermore, the output is circular. This means that although 1 page is a valid configuration, its useless, since we'll start overwriting it the instant we report a full page. Future work will focus on the output format (currently maintained) where we'll likey want each entry denoted by a header which includes a type and length. Further future work will allow to splice() the fd, also containing the async overflow data -- splice() would be mutually exclusive with mmap() of the data. Signed-off-by:
-
