- 07 Aug, 2015 14 commits
-
-
Wang Nan authored
This patch introduces bpf.c and bpf.h, which hold common functions issuing bpf syscall. The goal of these two files is to hide syscall completely from user. Note that bpf.c and bpf.h deal with kernel interface only. Things like structure of 'map' section in the ELF object is not cared by of bpf.[ch]. We first introduce bpf_create_map(). Note that, since functions in bpf.[ch] are wrapper of sys_bpf, they don't use OO style naming. Signed-off-by:
Wang Nan <wangnan0@huawei.com> Acked-by:
Alexei Starovoitov <ast@plumgrid.com> Cc: Brendan Gregg <brendan.d.gregg@gmail.com> Cc: Daniel Borkmann <daniel@iogearbox.net> Cc: David Ahern <dsahern@gmail.com> Cc: He Kuang <hekuang@huawei.com> Cc: Jiri Olsa <jolsa@kernel.org> Cc: Kaixu Xia <xiakaixu@huawei.com> Cc: Masami Hiramatsu <masami.hiramatsu.pt@hitachi.com> Cc: Namhyung Kim <namhyung@kernel.org> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Zefan Li <lizefan@huawei.com> Cc: pi3orama@163.com Link: http://lkml.kernel.org/r/1435716878-189507-16-git-send-email-wangnan0@huawei.comSigned-off-by:
Arnaldo Carvalho de Melo <acme@redhat.com>
-
Wang Nan authored
This patch records the indices of instructions which are needed to be relocated. That information is saved in the 'reloc_desc' field in 'struct bpf_program'. In the loading phase (this patch takes effect in the opening phase), the collected instructions will be replaced by map loading instructions. Since we are going to close the ELF file and clear all data at the end of the 'opening' phase, the ELF information will no longer be valid in the 'loading' phase. We have to locate the instructions before maps are loaded, instead of directly modifying the instruction. 'struct bpf_map_def' is introduced in this patch to let us know how many maps are defined in the object. This is the third part of map relocation. The principle of map relocation is described in commit message of 'bpf tools: Collect symbol table from SHT_SYMTAB section'. Signed-off-by:
-
Wang Nan authored
This patch collects relocation sections into 'struct object'. Such sections are used for connecting maps to bpf programs. 'reloc' field in 'struct bpf_object' is introduced for storing such information. This patch simply store the data into 'reloc' field. Following patch will parse them to know the exact instructions which are needed to be relocated. Note that the collected data will be invalid after ELF object file is closed. This is the second patch related to map relocation. The first one is 'bpf tools: Collect symbol table from SHT_SYMTAB section'. The principle of map relocation is described in its commit message. Signed-off-by:
-
Wang Nan authored
This patch collects all programs in an object file into an array of 'struct bpf_program' for further processing. That structure is for representing each eBPF program. 'bpf_prog' should be a better name, but it has been used by linux/filter.h. Although it is a kernel space name, I still prefer to call it 'bpf_program' to prevent possible confusion. bpf_object__add_program() creates a new 'struct bpf_program' object. It first init a variable in stack using bpf_program__init(), then if success, enlarges obj->programs array and copy the new object in. Signed-off-by:
-
Wang Nan authored
This patch collects symbols section. This section is useful when linking BPF maps. What 'bpf_map_xxx()' functions actually require are map's file descriptors (and the internal verifier converts fds into pointers to 'struct bpf_map'), which we don't know when compiling. Therefore, we should make compiler generate a 'ldr_64 r1, <imm>' instruction, and fill the 'imm' field with the actual file descriptor when loading in libbpf. BPF programs should be written in this way: struct bpf_map_def SEC("maps") my_map = { .type = BPF_MAP_TYPE_HASH, .key_size = sizeof(unsigned long), .value_size = sizeof(unsigned long), .max_entries = 1000000, }; SEC("my_func=sys_write") int my_func(void *ctx) { ... bpf_map_update_elem(&my_map, &key, &value, BPF_ANY); ... } Compiler should convert '&my_map' into a 'ldr_64, r1, <imm>' instruction, where imm should be the address of 'my_map'. According to the address, libbpf knows which map it actually referenced, and then fills the imm field with the 'fd' of that map created by it. However, since we never really 'link' the object file, the imm field is only a record in relocation section. Therefore libbpf should do the relocation: 1. In relocation section (type == SHT_REL), positions of each such 'ldr_64' instruction are recorded with a reference of an entry in symbol table (SHT_SYMTAB); 2. From records in symbol table we can find the indics of map variables. Libbpf first record SHT_SYMTAB and positions of each instruction which required bu such operation. Then create file descriptor. Finally, after map creation complete, replace the imm field. This is the first patch of BPF map related stuff. It records SHT_SYMTAB into object's efile field for further use. Signed-off-by: -
Wang Nan authored
If maps are used by eBPF programs, corresponding object file(s) should contain a section named 'map'. Which contains map definitions. This patch copies the data of the whole section. Map data parsing should be acted just before map loading. Signed-off-by:
-
Wang Nan authored
Expand bpf_obj_elf_collect() to collect license and kernel version information in eBPF object file. eBPF object file should have a section named 'license', which contains a string. It should also have a section named 'version', contains a u32 LINUX_VERSION_CODE. bpf_obj_validate() is introduced to validate object file after loaded. Currently it only check existence of 'version' section. Signed-off-by:
-
Wang Nan authored
bpf_obj_elf_collect() is introduced to iterate over each elf sections to collection information in eBPF object files. This function will futher enhanced to collect license, kernel version, programs, configs and map information. Signed-off-by:
-
Wang Nan authored
Check endianness according to EHDR. Code is taken from tools/perf/util/symbol-elf.c. Libbpf doesn't magically convert missmatched endianness. Even if we swap eBPF instructions to correct byte order, we are unable to deal with endianness in code logical generated by LLVM. Therefore, libbpf should simply reject missmatched ELF object, and let LLVM to create good code. Signed-off-by:
-
Wang Nan authored
To support dynamic compiling, this patch allows caller to pass a in-memory buffer to libbpf by bpf_object__open_buffer(). libbpf calls elf_memory() to open it as ELF object file. Because __bpf_object__open() collects all required data and won't need that buffer anymore, libbpf uses that buffer directly instead of clone a new buffer. Caller of libbpf can free that buffer or use it do other things after bpf_object__open_buffer() return. Signed-off-by:
-
Wang Nan authored
This patch defines basic interface of libbpf. 'struct bpf_object' will be the handler of each object file. Its internal structure is hide to user. eBPF object files are compiled by LLVM as ELF format. In this patch, libelf is used to open those files, read EHDR and do basic validation according to e_type and e_machine. All elf related staffs are grouped together and reside in efile field of 'struct bpf_object'. bpf_object__elf_finish() is introduced to clear it. After all eBPF programs in an object file are loaded, related ELF information is useless. Close the object file and free those memory. The zfree() and zclose() functions are introduced to ensure setting NULL pointers and negative file descriptors after resources are released. Signed-off-by:
-
Wang Nan authored
By libbpf_set_print(), users of libbpf are allowed to register he/she own debug, info and warning printing functions. Libbpf will use those functions to print messages. If not provided, default info and warning printing functions are fprintf(stderr, ...); default debug printing is NULL. This API is designed to be used by perf, enables it to register its own logging functions to make all logs uniform, instead of separated logging level control. Signed-off-by:
-
Wang Nan authored
This is the first patch of libbpf. The goal of libbpf is to create a standard way for accessing eBPF object files. This patch creates 'Makefile' and 'Build' for it, allows 'make' to build libbpf.a and libbpf.so, 'make install' to put them into proper directories. Most part of Makefile is borrowed from traceevent. Before building, it checks the existence of libelf in Makefile, and deny to build if not found. Instead of throwing an error if libelf not found, the error raises in a phony target "elfdep". This design is to ensure 'make clean' still workable even if libelf is not found. Because libbpf requires 'kern_version' field set for 'union bpf_attr' (bpfdep" is used for that dependency), Kernel BPF API is also checked by intruducing a new feature check 'bpf' into tools/build/feature, which checks the existence and version of linux/bpf.h. When building libbpf, it searches that file from include/uapi/linux in kernel source tree (controlled by FEATURE_CHECK_CFLAGS-bpf). Since it searches kernel source tree it reside, installing of newest kernel headers is not required, except we are trying to port these files to an old kernel. To avoid checking that file when perf building, the newly introduced 'bpf' feature check doesn't added into FEATURE_TESTS and FEATURE_DISPLAY by default in tools/build/Makefile.feature, but added into libbpf's specific. Signed-off-by:
-
Ingo Molnar authored
Merge tag 'perf-core-for-mingo' of git://git.kernel.org/pub/scm/linux/kernel/git/acme/linux into perf/core Pull perf/core improvements and fixes from Arnaldo Carvalho de Melo: User visible changes: - IPC and cycle accounting in 'perf annotate'. (Andi Kleen) - Display cycles in branch sort mode in 'perf report'. (Andi Kleen) - Add total time column to 'perf trace' syscall stats summary. (Milian Woff) Infrastructure changes: - PMU helpers to use in Intel PT. (Adrian Hunter) - Fix perf-with-kcore script not to split args with spaces. (Adrian Hunter) - Add empty Build files for some more architectures. (Ben Hutchings) - Move 'perf stat' config variables to a struct to allow using some of its functions in more places. (Jiri Olsa) - Add DWARF register names for 'xtensa' arch. (Max Filippov) - Implement BPF programs attached to uprobes. (Wang Nan) Signed-off-by:
Ingo Molnar <mingo@kernel.org>
-
- 06 Aug, 2015 26 commits
-
-
Adrian Hunter authored
Extend the event parser maximum error index from 10 to 13. That allows PMU config terms of up to 10 characters to display un-truncated in the error message. Signed-off-by:
Adrian Hunter <adrian.hunter@intel.com> Cc: Jiri Olsa <jolsa@redhat.com> Link: http://lkml.kernel.org/r/1437150840-31811-17-git-send-email-adrian.hunter@intel.comSigned-off-by:
-
Adrian Hunter authored
Currently the value of a PMU config term is silently truncated if it is too big. This is an impediment to validating the value for other criteria later on i.e. the user provides an invalid value that gets truncated to a valid one. The maximum value validation is only done for the parser where the error is passed back to the user. In other cases the silent truncation continues so as not to affect tools that perhaps rely on it. Signed-off-by:
-
Adrian Hunter authored
Add perf_pmu__format_bits() to get the format bits for a PMU config term. Intel PT will use this to validate terms and to record format bits to enable later interpreting the config from the attribute stored in the perf.data file. Signed-off-by:
-
Adrian Hunter authored
Fix the perf-with-kcore script so that it doesn't split arguments that contain spaces. Signed-off-by:
-
Adrian Hunter authored
PERF_ITRACE_PERIOD_INSTRUCTIONS is zero so it got overwritten by the default period type. Fix by checking if the period type was set rather than if the value was zero when applying the default. Signed-off-by:
-
Max Filippov authored
Signed-off-by:
Max Filippov <jcmvbkbc@gmail.com> Cc: Chris Zankel <chris@zankel.net> Cc: Marc Gauthier <marc@cadence.com> Cc: Paul Mackerras <paulus@samba.org> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: linux-xtensa@linux-xtensa.org Link: http://lkml.kernel.org/r/1437208216-15729-9-git-send-email-jcmvbkbc@gmail.comSigned-off-by:
-
Andi Kleen authored
Display the cycles by default in branch sort mode. To make enough room for the new column I removed dso_to. It is usually redundant with dso_from. Signed-off-by:
Andi Kleen <ak@linux.intel.com> Cc: Jiri Olsa <jolsa@kernel.org> Cc: Namhyung Kim <namhyung@kernel.org> Link: http://lkml.kernel.org/r/1437233094-12844-9-git-send-email-andi@firstfloor.orgSigned-off-by:
-
Andi Kleen authored
Now that we can process branch data in annotate it makes sense to support enabling branch recording from top too. Most of the code needed for this is already in shared code with report. But we need to add: - The option parsing code (using shared code from the previous patch) - Document the options - Set up the IPC/cycles accounting state in the top session - Call the accounting code in the hist iter callback Signed-off-by:
-
Andi Kleen authored
Add two new columns to the annotate display and display the average cycles and the compute IPC if available. When the LBR was not in any branch mode the IPC computation is automatically disabled. We still display the cycle information. Example output (with made up numbers): The second column is the IPC and third average cycles. │ __attribute__((noinline)) f2() │ { 5.15 0.07 │ push %rbp 0.01 0.07 │ mov %rsp,%rbp │ c = a / b; 9.87 0.07 │ mov a,%eax 0.07 │ mov b,%ecx 0.07 │ cltd 4.92 0.07 123│ idiv %ecx 70.79 0.07 │ mov %eax,__TMC_END__ │ } 9.25 0.07 │ pop %rbp 0.01 0.07 123│ ← retq v2: Fix display problems. Signed-off-by: -
Andi Kleen authored
Compute the IPC and the basic block cycles for the annotate display. IPC is computed by counting the instructions, and then dividing the accounted cycles by that count. The actual IPC computation can only be done at annotate time, because we need to parse the objdump output first to know the number of instructions in the basic block. The cycles/IPC are also put into the perf function annotation so that the display code can show them. Again basic block overlaps are not handled, with the longest winning, but there are some heuristics to hide the IPC when the longest is not the most common. v2: Compute IPC correctly. Signed-off-by:
-
Andi Kleen authored
Call the earlier added cycle histogram infrastructure from the perf report hist iter callback. For this we walk the branch records. This allows to use cycle histograms when browsing perf report annotate. v2: Rename flag Signed-off-by:
-
Andi Kleen authored
This adds the basic infrastructure to keep track of cycle counts per basic block for annotate. We allocate an array similar to the normal accounting, and then account branch cycles there. We handle two cases: cycles per basic block with start and cycles per branch (these are later used for either IPC or just cycles per BB) In the start case we cannot handle overlaps, so always the longest basic block wins. For the cycles per branch case everything is accurately accounted. v2: Remove unnecessary checks. Slight restructure. Move symbol__get_annotation to another patch. Move histogram allocation. v3: Merged with current tree Signed-off-by:
Jiri Olsa <jolsa@kernel.org> Cc: Namhyung Kim <namhyung@kernel.org> Link: http://lkml.kernel.org/r/1437233094-12844-4-git-send-email-andi@firstfloor.orgSigned-off-by:
-
Andi Kleen authored
Later patches need to cheaply check that the branch mode is in ANY. Add a new function to check all event attrs and add a flag to the report state, which is then initialized. v2: Rename flag Signed-off-by:
-
Andi Kleen authored
cycles is a new branch_info field available on some CPUs that indicates the time deltas between branches in the LBR. Add a sort key and output code for the cycles to allow to display the basic block cycles individually in perf report. We also pass in the cycles for weight when LBRs are processed, which allows to get global and local weight, to get an estimate of the total cost. And also print the cycles information for perf report -D. I also added printing for the previously missing LBR flags (mispredict etc.) Signed-off-by:
-
Ben Hutchings authored
perf currently fails to build on MIPS as there is no tools/perf/arch/mips/Build file. Adding an empty file fixes this as there are no MIPS-specific sources to build. It looks like the same is needed for Alpha and PA-RISC, though I haven't been able to test those. Signed-off-by:
Ben Hutchings <ben@decadent.org.uk> Fixes: 5e8c0fb6 ("perf build: Add arch x86 objects building") Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Link: http://lkml.kernel.org/r/1438704627.7315.2.camel@decadent.org.ukSigned-off-by:
-
Jiri Olsa authored
Moving counter processing code into stat object as perf_stat__process_counter. Signed-off-by:
-
Jiri Olsa authored
Passing 'struct perf_stat_config' into process_counter(), so that we can make process_counter() non static and use it from other places. Signed-off-by:
-
Jiri Olsa authored
Moving 'interval' into struct perf_stat_config. The point is to centralize the base stat config so it could be used localy together with other stat routines in other parts of perf code. Signed-off-by:
-
Jiri Olsa authored
Moving 'output' into struct perf_stat_config. The point is to centralize the base stat config so it could be used localy together with other stat routines in other parts of perf code. Signed-off-by:
-
Jiri Olsa authored
Moving 'scale' into struct perf_stat_config. The point is to centralize the base stat config so it could be used localy together with other stat routines in other parts of perf code. Signed-off-by:
-
Jiri Olsa authored
Moving 'aggr_mode' into new struct. The point is to centralize the base stat config so it could be used localy together with other stat routines in other parts of perf code. Signed-off-by:
-
Wang Nan authored
Commit 7b6ff0bd ("perf probe ppc64le: Fixup function entry if using kallsyms lookup") adds 'struct map' into probe-event.h but not forward declares it. This patch fixes it. Signed-off-by:
-
Wang Nan authored
va_args alternative to eprintf(). Signed-off-by:
-
Wang Nan authored
By copying BPF related operation to uprobe processing path, this patch allow users attach BPF programs to uprobes like what they are already doing on kprobes. After this patch, users are allowed to use PERF_EVENT_IOC_SET_BPF on a uprobe perf event. Which make it possible to profile user space programs and kernel events together using BPF. Because of this patch, CONFIG_BPF_EVENTS should be selected by CONFIG_UPROBE_EVENT to ensure trace_call_bpf() is compiled even if KPROBE_EVENT is not set. Signed-off-by:
-
Wang Nan authored
Commit e1abf2cc ("bpf: Fix the build on BPF_SYSCALL=y && !CONFIG_TRACING kernels, make it more configurable") updated the building condition of bpf_trace.o from CONFIG_BPF_SYSCALL to CONFIG_BPF_EVENTS, but the corresponding #ifdef controller in trace_events.h for trace_call_bpf() was not changed. Which, in theory, is incorrect. With current Kconfigs, we can create a .config with CONFIG_BPF_SYSCALL=y and CONFIG_BPF_EVENTS=n by unselecting CONFIG_KPROBE_EVENT and selecting CONFIG_BPF_SYSCALL. With these options, trace_call_bpf() will be defined as an extern function, but if anyone calls it a symbol missing error will be triggered since bpf_trace.o was not built. This patch changes the #ifdef controller for trace_call_bpf() from CONFIG_BPF_SYSCALL to CONFIG_BPF_EVENTS. I'll show its correctness: Before this patch: BPF_SYSCALL BPF_EVENTS trace_call_bpf bpf_trace.o y y normal compiled n n inline not compiled y n normal not compiled (incorrect) n y impossible (BPF_EVENTS depends on BPF_SYSCALL) After this patch: BPF_SYSCALL BPF_EVENTS trace_call_bpf bpf_trace.o y y normal compiled n n inline not compiled y n inline not compiled (fixed) n y impossible (BPF_EVENTS depends on BPF_SYSCALL) So this patch doesn't break anything. QED. Signed-off-by:
-
Milian Wolff authored
It is cumbersome to manually calculate the total time spent in a given syscall by multiplying the average value with the number of calls. Instead, we now do this directly inside perf trace. Note that this is also done by 'strace', which even adds a column with relative numbers - something we could do in the future. Example: perf trace -s find /some/folder > /dev/null Summary of events: find (19976), 700123 events, 100.0%, 0.000 msec syscall calls total min avg max stddev (msec) (msec) (msec) (msec) (%) --------------- -------- --------- --------- --------- --------- ------ read 4 0.006 0.001 0.002 0.003 27.42% write 8046 9.617 0.001 0.001 0.035 0.56% open 34196 40.384 0.001 0.001 0.071 0.30% close 68375 57.104 0.001 0.001 0.076 0.25% stat 4 0.004 0.001 0.001 0.001 3.14% fstat 34189 27.518 0.001 0.001 0.060 0.34% mmap 13 0.029 0.001 0.002 0.003 10.74% mprotect 6 0.018 0.002 0.003 0.005 17.04% munmap 3 0.014 0.003 0.005 0.006 24.87% brk 87 0.490 0.001 0.006 0.016 6.50% ioctl 3 0.004 0.001 0.001 0.003 36.39% access 1 0.004 0.004 0.004 0.004 0.00% uname 1 0.001 0.001 0.001 0.001 0.00% getdents 68393 143.600 0.001 0.002 0.187 0.95% fchdir 68371 56.980 0.001 0.001 0.111 0.39% arch_prctl 1 0.001 0.001 0.001 0.001 0.00% openat 34184 41.737 0.001 0.001 0.102 0.41% newfstatat 34184 41.180 0.001 0.001 0.064 0.34% Signed-off-by:Milian Wolff <milian.wolff@kdab.com> Tested-by:
-
