mirror of
https://git.kernel.org/pub/scm/linux/kernel/git/stable/linux.git
synced 2024-09-26 20:38:12 +00:00
ca8a66738a
- Fix broken direct trampolines being called when another callback is attached the same function. ARM 64 does not support FTRACE_WITH_REGS, and when it added direct trampoline calls from ftrace, it removed the "WITH_REGS" flag from the ftrace_ops for direct trampolines. This broke x86 as x86 requires direct trampolines to have WITH_REGS. This wasn't noticed because direct trampolines work as long as the function it is attached to is not shared with other callbacks (like the function tracer). When there's other callbacks, a helper trampoline is called, to call all the non direct callbacks and when it returns, the direct trampoline is called. For x86, the direct trampoline sets a flag in the regs field to tell the x86 specific code to call the direct trampoline. But this only works if the ftrace_ops had WITH_REGS set. ARM does things differently that does not require this. For now, set WITH_REGS if the arch supports WITH_REGS (which ARM does not), and this makes it work for both ARM64 and x86. - Fix wasted memory in the saved_cmdlines logic. The saved_cmdlines is a cache that maps PIDs to COMMs that tracing can use. Most trace events only save the PID in the event. The saved_cmdlines file lists PIDs to COMMs so that the tracing tools can show an actual name and not just a PID for each event. There's an array of PIDs that map to a small set of saved COMM strings. The array is set to PID_MAX_DEFAULT which is usually set to 32768. When a PID comes in, it will add itself to this array along with the index into the COMM array (note if the system allows more than PID_MAX_DEFAULT, this cache is similar to cache lines as an update of a PID that has the same PID_MAX_DEFAULT bits set will flush out another task with the same matching bits set). A while ago, the size of this cache was changed to be dynamic and the array was moved into a structure and created with kmalloc(). But this new structure had the size of 131104 bytes, or 0x20020 in hex. As kmalloc allocates in powers of two, it was actually allocating 0x40000 bytes (262144) leaving 131040 bytes of wasted memory. The last element of this structure was a pointer to the COMM string array which defaulted to just saving 128 COMMs. By changing the last field of this structure to a variable length string, and just having it round up to fill the allocated memory, the default size of the saved COMM cache is now 8190. This not only uses the wasted space, but actually saves space by removing the extra allocation for the COMM names. -----BEGIN PGP SIGNATURE----- iIoEABYIADIWIQRRSw7ePDh/lE+zeZMp5XQQmuv6qgUCZcYi8RQccm9zdGVkdEBn b29kbWlzLm9yZwAKCRAp5XQQmuv6qqENAQD6xGE9EPkbHArElKfgpSuQOfGhcyyP LjgVhqVgmIoqUwD8CeVpxk3VwZIOQYvPn5XictcZgkYSeEWUZcKYg4c/3gs= =iIBv -----END PGP SIGNATURE----- Merge tag 'trace-v6.8-rc3' of git://git.kernel.org/pub/scm/linux/kernel/git/trace/linux-trace Pull tracing fixes from Steven Rostedt: - Fix broken direct trampolines being called when another callback is attached the same function. ARM 64 does not support FTRACE_WITH_REGS, and when it added direct trampoline calls from ftrace, it removed the "WITH_REGS" flag from the ftrace_ops for direct trampolines. This broke x86 as x86 requires direct trampolines to have WITH_REGS. This wasn't noticed because direct trampolines work as long as the function it is attached to is not shared with other callbacks (like the function tracer). When there are other callbacks, a helper trampoline is called, to call all the non direct callbacks and when it returns, the direct trampoline is called. For x86, the direct trampoline sets a flag in the regs field to tell the x86 specific code to call the direct trampoline. But this only works if the ftrace_ops had WITH_REGS set. ARM does things differently that does not require this. For now, set WITH_REGS if the arch supports WITH_REGS (which ARM does not), and this makes it work for both ARM64 and x86. - Fix wasted memory in the saved_cmdlines logic. The saved_cmdlines is a cache that maps PIDs to COMMs that tracing can use. Most trace events only save the PID in the event. The saved_cmdlines file lists PIDs to COMMs so that the tracing tools can show an actual name and not just a PID for each event. There's an array of PIDs that map to a small set of saved COMM strings. The array is set to PID_MAX_DEFAULT which is usually set to 32768. When a PID comes in, it will add itself to this array along with the index into the COMM array (note if the system allows more than PID_MAX_DEFAULT, this cache is similar to cache lines as an update of a PID that has the same PID_MAX_DEFAULT bits set will flush out another task with the same matching bits set). A while ago, the size of this cache was changed to be dynamic and the array was moved into a structure and created with kmalloc(). But this new structure had the size of 131104 bytes, or 0x20020 in hex. As kmalloc allocates in powers of two, it was actually allocating 0x40000 bytes (262144) leaving 131040 bytes of wasted memory. The last element of this structure was a pointer to the COMM string array which defaulted to just saving 128 COMMs. By changing the last field of this structure to a variable length string, and just having it round up to fill the allocated memory, the default size of the saved COMM cache is now 8190. This not only uses the wasted space, but actually saves space by removing the extra allocation for the COMM names. * tag 'trace-v6.8-rc3' of git://git.kernel.org/pub/scm/linux/kernel/git/trace/linux-trace: tracing: Fix wasted memory in saved_cmdlines logic ftrace: Fix DIRECT_CALLS to use SAVE_REGS by default |
||
|---|---|---|
| .. | ||
| rv | ||
| blktrace.c | ||
| bpf_trace.c | ||
| bpf_trace.h | ||
| error_report-traces.c | ||
| fgraph.c | ||
| fprobe.c | ||
| ftrace.c | ||
| ftrace_internal.h | ||
| Kconfig | ||
| kprobe_event_gen_test.c | ||
| Makefile | ||
| pid_list.c | ||
| pid_list.h | ||
| power-traces.c | ||
| preemptirq_delay_test.c | ||
| rethook.c | ||
| ring_buffer.c | ||
| ring_buffer_benchmark.c | ||
| rpm-traces.c | ||
| synth_event_gen_test.c | ||
| trace.c | ||
| trace.h | ||
| trace_benchmark.c | ||
| trace_benchmark.h | ||
| trace_boot.c | ||
| trace_branch.c | ||
| trace_btf.c | ||
| trace_btf.h | ||
| trace_clock.c | ||
| trace_dynevent.c | ||
| trace_dynevent.h | ||
| trace_entries.h | ||
| trace_eprobe.c | ||
| trace_event_perf.c | ||
| trace_events.c | ||
| trace_events_filter.c | ||
| trace_events_filter_test.h | ||
| trace_events_hist.c | ||
| trace_events_inject.c | ||
| trace_events_synth.c | ||
| trace_events_trigger.c | ||
| trace_events_user.c | ||
| trace_export.c | ||
| trace_fprobe.c | ||
| trace_functions.c | ||
| trace_functions_graph.c | ||
| trace_hwlat.c | ||
| trace_irqsoff.c | ||
| trace_kdb.c | ||
| trace_kprobe.c | ||
| trace_kprobe_selftest.c | ||
| trace_kprobe_selftest.h | ||
| trace_mmiotrace.c | ||
| trace_nop.c | ||
| trace_osnoise.c | ||
| trace_output.c | ||
| trace_output.h | ||
| trace_preemptirq.c | ||
| trace_printk.c | ||
| trace_probe.c | ||
| trace_probe.h | ||
| trace_probe_kernel.h | ||
| trace_probe_tmpl.h | ||
| trace_recursion_record.c | ||
| trace_sched_switch.c | ||
| trace_sched_wakeup.c | ||
| trace_selftest.c | ||
| trace_selftest_dynamic.c | ||
| trace_seq.c | ||
| trace_stack.c | ||
| trace_stat.c | ||
| trace_stat.h | ||
| trace_synth.h | ||
| trace_syscalls.c | ||
| trace_uprobe.c | ||
| tracing_map.c | ||
| tracing_map.h | ||