perf, c2c: Add new tool to analyze cacheline contention on NUMA systems
v2 cover perf, c2c: Add new tool to analyze cacheline contention on NUMA systems
With the introduction of NUMA systems, came the possibility of remote memory accesses. Combine those remote memory accesses with contention on the remote node (ie a modified cacheline) and you have a possibility for very long latencies. These latencies can bottleneck a program.
The program added by these patches, helps detect the situation where two nodes are 'tugging' on the same _data_ cacheline. The term used through out this program and the various changelogs is called a HITM. This means nodeX went to read a cacheline and it was discovered to be loaded in nodeY's LLC cache (hence the cacheHIT). The remote cacheline was also in a 'M'odified state thus creating a 'HIT M' for hit in a modified state. HITMs can happen locally and remotely. This program's interest is mainly in remote HITMs as they cause the longest latencies.
Why a program has a remote HITM derives from how the two nodes are 'sharing' the cacheline. Is the sharing intentional ("true") or unintentional ("false"). We have seen lots of "false" sharing cases, which lead to simple solutions such as seperating the data onto different cachelines.
This tool does not distinguish between 'true' or 'false' sharing, instead it just points to the more expensive sharing situations under the current workload. It is up to the user to understand what the workload is doing to determine whether a problem exists or not and how to report it.
The data output is verbose and there are lots of data tables that interprit the latencies and data addresses in different ways to help see where bottlenecks might be lying.
Most of this idea, work and calculations were done by Dick Fowles. My work mainly includes porting it to perf. Joe Mario has contributed greatly with ideas to make the output more informative based on his usage of the tool. Joe has found a handful of bottlenecks using various industry benchmarks and has worked with developers to fix them.
I would also like to thank Stephane Eranian for his early help and guidance on navigating the differences between the current perf tool and how similar tools looked at HP. And also his tireless work in getting the MMAP2 interface to stick.
Also thanks to Arnaldo and Jiri Olso for their help in suggestions for this tool.
I also have a test program that generated a controlled number of HITMs that we used frequently to validate our early work (the Intel docs were not always clear which bits had to be set and some arches do not work well). I would like to add it, but didn't know how (nor did I spend any serious time looking either).
This program has been tested primarily on Intel's Ivy Bridge platforms. The Sandy Bridge platforms had some quirks that were fixed on Ivy Bridge. We haven't tried Haswell as that has a re-worked latency event implementation.
A handful of patches include re-enabling MMAP2 support and some fixes to perf itself. One in particular hacks up how standard deviation is calculated. It works with our calculations but may break other tools expectations. Feedback is welcomed.
Comemnts, feedback, anything else welcomed.
V2: updated to latest perf/core branch 1029f9fedf87fa6
switched to hist_entry based on Jiri O's suggestion dropped latency analyze for now until this patchset is accepted little fixes and tweaks
Arnaldo Carvalho de Melo (2):
perf c2c: Shared data analyser
perf c2c: Dump raw records, decode data_src bits
Don Zickus (19):
Revert "perf: Disable PERF_RECORD_MMAP2 support"
perf, machine: Use map as success in ip__resolve_ams
perf, session: Change header.misc dump from decimal to hex perf, stat: FIXME Stddev calculation is incorrect
perf, callchain: Add generic callchain print handler for stdio perf, c2c: Rework setup code to prepare for features
perf, c2c: Add rbtree sorted on mmap2 data
perf, c2c: Add stats to track data source bits and cpu to node maps perf, c2c: Sort based on hottest cache line
perf, c2c: Display cacheline HITM analysis to stdout
perf, c2c: Add callchain support
perf, c2c: Output summary stats
perf, c2c: Dump rbtree for debugging
perf, c2c: Fixup tid because of perf map is broken
perf, c2c: Add symbol count table
perf, c2c: Add shared cachline summary table
perf, c2c: Add framework to analyze latency and display summary stats perf, c2c: Add selected extreme latencies to output cacheline stats
table
perf, c2c: Add summary latency table for various parts of caches
kernel/events/core.c | 4 - tools/perf/Documentation/perf-c2c.c | 22 + tools/perf/Makefile.perf | 1 + tools/perf/builtin-c2c.c | 2963 +++++++++++++++++++++++++++++++++++ tools/perf/builtin.h | 1 + tools/perf/perf.c | 1 + tools/perf/ui/stdio/hist.c | 37 + tools/perf/util/event.c | 36 +- tools/perf/util/evlist.c | 37 + tools/perf/util/evlist.h | 7 + tools/perf/util/evsel.c | 1 + tools/perf/util/hist.h | 4 + tools/perf/util/machine.c | 2 +- tools/perf/util/session.c | 2 +- tools/perf/util/stat.c | 3 +- 15 files changed, 3097 insertions(+), 24 deletions(-) create mode 100644 tools/perf/Documentation/perf-c2c.c create mode 100644 tools/perf/builtin-c2c.c
Signature · 50 lines
1.7.11.7 Arnaldo Carvalho de Melo (2): perf c2c: Shared data analyser perf c2c: Dump raw records, decode data_src bits Don Zickus (17): Revert "perf: Disable PERF_RECORD_MMAP2 support" perf, sort: Add physid sorting based on mmap2 data perf, sort: Allow unique sorting instead of combining hist_entries perf: Allow ability to map cpus to nodes easily perf, kmem: Utilize the new generic cpunode_map perf: Fix stddev calculation perf, callchain: Add generic callchain print handler for stdio perf, c2c: Rework setup code to prepare for features perf, c2c: Add in sort on physid perf, c2c: Add stats to track data source bits and cpu to node maps perf, c2c: Sort based on hottest cache line perf, c2c: Display cacheline HITM analysis to stdout perf, c2c: Add callchain support perf, c2c: Output summary stats perf, c2c: Dump rbtree for debugging perf, c2c: Add symbol count table perf, c2c: Add shared cachline summary table kernel/events/core.c | 4 - tools/perf/Documentation/perf-c2c.c | 22 + tools/perf/Makefile.perf | 1 + tools/perf/builtin-c2c.c | 1787 +++++++++++++++++++++++++++++++++++ tools/perf/builtin-kmem.c | 78 +- tools/perf/builtin-report.c | 2 +- tools/perf/builtin.h | 1 + tools/perf/perf.c | 1 + tools/perf/ui/stdio/hist.c | 37 + tools/perf/util/cpumap.c | 150 +++ tools/perf/util/cpumap.h | 35 + tools/perf/util/event.c | 36 +- tools/perf/util/evsel.c | 1 + tools/perf/util/hist.c | 10 +- tools/perf/util/hist.h | 5 + tools/perf/util/sort.c | 149 +++ tools/perf/util/sort.h | 4 + tools/perf/util/stat.c | 13 + tools/perf/util/stat.h | 1 + 19 files changed, 2236 insertions(+), 101 deletions(-) create mode 100644 tools/perf/Documentation/perf-c2c.c create mode 100644 tools/perf/builtin-c2c.c -- 1.7.11.7
PATCH v2 1/19 Revert "perf: Disable PERF_RECORD_MMAP2 support"
This reverts commit 3090ffb5a2515990182f3f55b0688a7817325488.
Conflicts:
tools/perf/util/event.c
--- kernel/events/core.c | 4 ---- tools/perf/util/event.c | 36 +++++++++++++++++++----------------- tools/perf/util/evsel.c | 1 + 3 files changed, 20 insertions(+), 21 deletions(-)
kernel/events/core.c +0 -4
@@ -6851,10 +6851,6 @@ static int perf_copy_attr(struct perf_event_attr __user *uattr,if (ret)return -EFAULT ; -/* disabled for now */ -if (attr->mmap2) -return -EINVAL ; -if (attr->__reserved_1)return -EINVAL ;
tools/perf/util/event.c +19 -17
@@ -155,13 +155,14 @@ int perf_event__synthesize_mmap_events(struct perf_tool *tool,return -1 ; } - event->header.type =PERF_RECORD_MMAP ; + event->header.type =PERF_RECORD_MMAP2 ;while (1 ) {char bf[BUFSIZ ];char prot[5 ];char execname[PATH_MAX ];char anonstr[] ="//anon" ; +unsigned int ino;size_t size;ssize_t n;
@@ -172,14 +173,15 @@ int perf_event__synthesize_mmap_events(struct perf_tool *tool,strcpy (execname,"" );/* 00400000-0040c000 r-xp 00000000 fd:01 41038 /bin/cat */ - n =sscanf (bf,"%" PRIx64"-%" PRIx64" %s %" PRIx64" %*x:%*x %*u %s\n" , - &event->mmap.start, &event->mmap.len, prot, - &event->mmap.pgoff, - execname); -/* -* Anon maps don't have the execname. -*/ -if (n <4 ) + n =sscanf (bf,"%" PRIx64"-%" PRIx64" %s %" PRIx64" %x:%x %u %s\n" , + &event->mmap2.start, &event->mmap2.len, prot, + &event->mmap2.pgoff, &event->mmap2.maj, + &event->mmap2.min, + &ino, execname); + + event->mmap2.ino = (u64 )ino; + +if (n <7 )continue ;/* * Just like the kernel, see __perf_event_mmap in kernel/perf_event.c
@@ -200,15 +202,15 @@ int perf_event__synthesize_mmap_events(struct perf_tool *tool,strcpy (execname, anonstr); size =strlen (execname) +1 ; -memcpy (event->mmap.filename, execname, size); +memcpy (event->mmap2.filename, execname, size); size =PERF_ALIGN (size,sizeof (u64)); - event->mmap.len -= event->mmap.start; - event->mmap.header.size = (sizeof (event->mmap) - - (sizeof (event->mmap.filename) - size)); -memset (event->mmap.filename + size,0 , machine->id_hdr_size); - event->mmap.header.size += machine->id_hdr_size; - event->mmap.pid = tgid; - event->mmap.tid = pid; + event->mmap2.len -= event->mmap.start; + event->mmap2.header.size = (sizeof (event->mmap2) - + (sizeof (event->mmap2.filename) - size)); +memset (event->mmap2.filename + size,0 , machine->id_hdr_size); + event->mmap2.header.size += machine->id_hdr_size; + event->mmap2.pid = tgid; + event->mmap2.tid = pid;if (process (tool, event, &synth_sample, machine) !=0 ) { rc =-1 ;
tools/perf/util/evsel.c +1 -0
@@ -640,6 +640,7 @@ void perf_evsel__config(struct perf_evsel *evsel, struct record_opts *opts)perf_evsel__set_sample_bit (evsel,WEIGHT ); attr->mmap = track; + attr->mmap2 = track && !perf_missing_features.mmap2; attr->comm = track;if (opts->sample_transaction)
Signature · 1 line
1.7.11.7
PATCH v2 2/19 perf, sort: Add physid sorting based on mmap2 data
In reply to Don ZickusIn order for the c2c tool to work correctly, it needs to properly sort all the records on uniquely identifiable data addresses. These unique addresses are converted from virtual addresses provided by the hardware into a kernel address using an mmap2 record as the decoder.
Once a unique address is converted, we can sort on them based on various rules. Then it becomes clear which address are overlapping with each other across mmap regions or pid spaces.
This patch just creates the rules and inserts the records into a sort entry for safe keeping until later patches process them.
The general sorting rule is:
o group cpumodes together
o group similar major, minor, inode, inode generation numbers togther
o if (nonzero major/minor number - ie mmap'd areas)
o sort on data addresses
o sort on instruction address
o sort on pid
o sort on tid
o if cpumode is kernel
o sort on data addresses
o sort on instruction address
o sort on pid
o sort on tid
o else (private to pid space)
o sort on pid
o sort on tid
o sort on data addresses
o sort on instruction address
I also hacked in the concept of 'color'. The purpose of that bit is to provides hints later when processing these records that indicate a new unique address has been encountered. Because later processing only checks the data addresses, there can be a theoretical scenario that similar sequential data addresses (when walking the rbtree) could be misinterpreted as overlapping when in fact they are not.
--- tools/perf/builtin-report.c | 2 +- tools/perf/util/hist.c | 7 ++- tools/perf/util/hist.h | 1 + tools/perf/util/sort.c | 148 ++++++++++++++++++++++++++++++++++++++++++++ tools/perf/util/sort.h | 3 + 5 files changed, 157 insertions(+), 4 deletions(-)
tools/perf/builtin-report.c +1 -1
@@ -755,7 +755,7 @@ int cmd_report(int argc, const char **argv, const char *prefix __maybe_unused)"sort by key(s): pid, comm, dso, symbol, parent, cpu, srcline," " dso_to, dso_from, symbol_to, symbol_from, mispredict," " weight, local_weight, mem, symbol_daddr, dso_daddr, tlb, " - "snoop, locked, abort, in_tx, transaction"), + "snoop, locked, abort, in_tx, transaction, physid"),OPT_BOOLEAN (0 ,"showcpuutilization" , &symbol_conf.show_cpu_utilization,"Show sample percentage for different cpu modes" ), OPT_STRING('p', "parent", &parent_pattern, "regex",
tools/perf/util/hist.c +4 -3
@@ -420,9 +420,10 @@ struct hist_entry *__hists__add_entry(struct hists *hists, .map = al->map, .sym = al->sym, }, - .cpu = al->cpu, - .ip = al->addr, - .level = al->level, + .cpu = al->cpu, + .cpumode = al->cpumode, + .ip = al->addr, + .level = al->level, .stat = { .nr_events = 1, .period = period,
tools/perf/util/hist.h +1 -0
@@ -62,6 +62,7 @@ enum hist_column {HISTC_MEM_LVL ,HISTC_MEM_SNOOP ,HISTC_TRANSACTION , +HISTC_PHYSID , HISTC_NR_COLS, /* Last entry */ };
tools/perf/util/sort.c +148 -0
166 lines not on this page · raw message
tools/perf/util/sort.h +3 -0
@@ -87,11 +87,13 @@ struct hist_entry {u64 ip;u64 transaction;s32 cpu; +u8 cpumode;struct hist_entry_diff diff;/* We are added by hists__add_dummy_entry. */ bool dummy; +bool color;/* XXX These two should move to some tree widget lib */ u16 row_offset;
@@ -166,6 +168,7 @@ enum sort_type {SORT_MEM_TLB ,SORT_MEM_LVL ,SORT_MEM_SNOOP , + SORT_MEM_PHYSID, };/*
Signature · 1 line
1.7.11.7
Re: [PATCH 02/19] perf, sort: Add physid sorting based on mmap2 data
On Fri, Feb 28, 2014 at 12:42:51PM -0500, Don Zickus wrote:
SNIP
Quoted 87 lines
tools/perf/util/sort.c +87
-0+static int64_t +sort__physid_cmp (struct hist_entry *left,struct hist_entry *right) +{ +u64 l, r; +struct map *l_map = left->mem_info->daddr.map; +struct map *r_map = right->mem_info->daddr.map; + +/* store all NULL mem maps at the bottom */ +/* shouldn't even need this check, should have stubs */ +if (!left->mem_info->daddr.map || !right->mem_info->daddr.map) +return 1 ; + +/* group event types together */ +if (left->cpumode > right->cpumode)return -1 ; +if (left->cpumode < right->cpumode)return 1 ; + +if (l_map->maj > r_map->maj)return -1 ; +if (l_map->maj < r_map->maj)return 1 ; + +if (l_map->min > r_map->min)return -1 ; +if (l_map->min < r_map->min)return 1 ; + +if (l_map->ino > r_map->ino)return -1 ; +if (l_map->ino < r_map->ino)return 1 ; + +if (l_map->ino_generation > r_map->ino_generation)return -1 ; +if (l_map->ino_generation < r_map->ino_generation)return 1 ; + +/* +* Addresses with no major/minor numbers are assumed to be +* anonymous in userspace. Sort those on pid then address. +* +* The kernel and non-zero major/minor mapped areas are +* assumed to be unity mapped. Sort those on address then pid. +*/ + +/* al_addr does all the right addr - start + offset calculations */ + l = left->mem_info->daddr.al_addr; + r = right->mem_info->daddr.al_addr; + +if (l_map->maj || l_map->min || l_map->ino || l_map-> ino_generation) { +/* mmapped areas */ + +/* hack to mark similar regions, 'right' is new entry */ +/* entries with same maj/min/ino/inogen are in same address space */ + right->color = TRUE; + +if (l > r)return -1 ; +if (l < r)return 1 ; + +/* sorting by iaddr makes calculations easier later */ +if (left->mem_info->iaddr.al_addr > right->mem_info->iaddr.al_addr)return -1 ; +if (left->mem_info->iaddr.al_addr < right->mem_info->iaddr.al_addr)return 1 ; + +if (left->thread->pid_ > right->thread->pid_)return -1 ; +if (left->thread->pid_ < right->thread->pid_)return 1 ; + +if (left->thread->tid > right->thread->tid)return -1 ; +if (left->thread->tid < right->thread->tid)return 1 ; + }else if (left->cpumode ==PERF_RECORD_MISC_KERNEL ) { +/* kernel mapped areas where 'start' doesn't matter */ + +/* hack to mark similar regions, 'right' is new entry */ +/* whole kernel region is in the same address space */ + right->color = TRUE; + +if (l > r)return -1 ; +if (l < r)return 1 ; + +/* sorting by iaddr makes calculations easier later */ +if (left->mem_info->iaddr.al_addr > right->mem_info->iaddr.al_addr)return -1 ; +if (left->mem_info->iaddr.al_addr < right->mem_info->iaddr.al_addr)return 1 ; + +if (left->thread->pid_ > right->thread->pid_)return -1 ; +if (left->thread->pid_ < right->thread->pid_)return 1 ; + +if (left->thread->tid > right->thread->tid)return -1 ; +if (left->thread->tid < right->thread->tid)return 1 ; + }else { +/* userspace anonymous */ +if (left->thread->pid_ > right->thread->pid_)return -1 ; +if (left->thread->pid_ < right->thread->pid_)return 1 ; + +if (left->thread->tid > right->thread->tid)return -1 ; +if (left->thread->tid < right->thread->tid)return 1 ; + +/* hack to mark similar regions, 'right' is new entry */
tools/perf/util/sort.c struct sort_entry sort_transaction = { · line 1076
Quoted 10 lines
tools/perf/util/sort.c +10
-0+/* userspace anonymous address space is contained within pid */ + right->color = TRUE; + +if (l > r)return -1 ; +if (l < r)return 1 ; + +/* sorting by iaddr makes calculations easier later */ +if (left->mem_info->iaddr.al_addr > right->mem_info->iaddr.al_addr)return -1 ; +if (left->mem_info->iaddr.al_addr < right->mem_info->iaddr.al_addr)return 1 ; + }
do you need single column for 'physid' ?
my first idea was to have separate sort entries for all checked entries:
(same way like for memory memory_sort_dimensions)
- mem_info->daddr.al_addr (we already have 'addr' check)
- mem_info->iaddr.al_addr
- thread->pid_ (we have only 'tid' check so far)
- l_map->maj, l_map->min, l_map->ino, l_map, ino_generation (we could probably group these)
and init sort order with:
sort_order = "physid,pid,...";
'...' is whatever name you choose for above entries
you can check the sort entries cmp function call
logic in hist_entry__cmp function
jirka
Re: [PATCH 02/19] perf, sort: Add physid sorting based on mmap2 data
Quoted 108 lines · Jiri Olsa
Quoted 87 lines
tools/perf/util/sort.c +87
-0+static int64_t +sort__physid_cmp (struct hist_entry *left,struct hist_entry *right) +{ +u64 l, r; +struct map *l_map = left->mem_info->daddr.map; +struct map *r_map = right->mem_info->daddr.map; + +/* store all NULL mem maps at the bottom */ +/* shouldn't even need this check, should have stubs */ +if (!left->mem_info->daddr.map || !right->mem_info->daddr.map) +return 1 ; + +/* group event types together */ +if (left->cpumode > right->cpumode)return -1 ; +if (left->cpumode < right->cpumode)return 1 ; + +if (l_map->maj > r_map->maj)return -1 ; +if (l_map->maj < r_map->maj)return 1 ; + +if (l_map->min > r_map->min)return -1 ; +if (l_map->min < r_map->min)return 1 ; + +if (l_map->ino > r_map->ino)return -1 ; +if (l_map->ino < r_map->ino)return 1 ; + +if (l_map->ino_generation > r_map->ino_generation)return -1 ; +if (l_map->ino_generation < r_map->ino_generation)return 1 ; + +/* +* Addresses with no major/minor numbers are assumed to be +* anonymous in userspace. Sort those on pid then address. +* +* The kernel and non-zero major/minor mapped areas are +* assumed to be unity mapped. Sort those on address then pid. +*/ + +/* al_addr does all the right addr - start + offset calculations */ + l = left->mem_info->daddr.al_addr; + r = right->mem_info->daddr.al_addr; + +if (l_map->maj || l_map->min || l_map->ino || l_map-> ino_generation) { +/* mmapped areas */ + +/* hack to mark similar regions, 'right' is new entry */ +/* entries with same maj/min/ino/inogen are in same address space */ + right->color = TRUE; + +if (l > r)return -1 ; +if (l < r)return 1 ; + +/* sorting by iaddr makes calculations easier later */ +if (left->mem_info->iaddr.al_addr > right->mem_info->iaddr.al_addr)return -1 ; +if (left->mem_info->iaddr.al_addr < right->mem_info->iaddr.al_addr)return 1 ; + +if (left->thread->pid_ > right->thread->pid_)return -1 ; +if (left->thread->pid_ < right->thread->pid_)return 1 ; + +if (left->thread->tid > right->thread->tid)return -1 ; +if (left->thread->tid < right->thread->tid)return 1 ; + }else if (left->cpumode ==PERF_RECORD_MISC_KERNEL ) { +/* kernel mapped areas where 'start' doesn't matter */ + +/* hack to mark similar regions, 'right' is new entry */ +/* whole kernel region is in the same address space */ + right->color = TRUE; + +if (l > r)return -1 ; +if (l < r)return 1 ; + +/* sorting by iaddr makes calculations easier later */ +if (left->mem_info->iaddr.al_addr > right->mem_info->iaddr.al_addr)return -1 ; +if (left->mem_info->iaddr.al_addr < right->mem_info->iaddr.al_addr)return 1 ; + +if (left->thread->pid_ > right->thread->pid_)return -1 ; +if (left->thread->pid_ < right->thread->pid_)return 1 ; + +if (left->thread->tid > right->thread->tid)return -1 ; +if (left->thread->tid < right->thread->tid)return 1 ; + }else { +/* userspace anonymous */ +if (left->thread->pid_ > right->thread->pid_)return -1 ; +if (left->thread->pid_ < right->thread->pid_)return 1 ; + +if (left->thread->tid > right->thread->tid)return -1 ; +if (left->thread->tid < right->thread->tid)return 1 ; + +/* hack to mark similar regions, 'right' is new entry */ tools/perf/util/sort.c struct sort_entry sort_transaction = { · line 1076
Quoted 10 lines
tools/perf/util/sort.c +10
-0+/* userspace anonymous address space is contained within pid */ + right->color = TRUE; + +if (l > r)return -1 ; +if (l < r)return 1 ; + +/* sorting by iaddr makes calculations easier later */ +if (left->mem_info->iaddr.al_addr > right->mem_info->iaddr.al_addr)return -1 ; +if (left->mem_info->iaddr.al_addr < right->mem_info->iaddr.al_addr)return 1 ; + }do you need single column for 'physid' ?
my first idea was to have separate sort entries for all checked entries:
(same way like for memorymemory_sort_dimensions)-
mem_info->daddr.al_addr(we already have 'addr' check)
-mem_info->iaddr.al_addr
-thread->pid_(we have only 'tid' check so far)
-l_map->maj,l_map->min,l_map->ino,l_map,ino_generation(we could probably group these)
Quoted 5 lines
and init sort order with:
sort_order = "physid,pid,...";'...' is whatever name you choose for above entries
The problem is (as you can see in the code above), physid is _not_ just one piece of data. It is dependent on multiple things. It is mainly dependent on four things:
- cpumode
- major, minor, inode, inode_generation (call it mmap2 data)
- data address
- pid
If cpumode == KERNEL, sort in this order
- mmap2 data
- data address
- pid (optional)
If cpumode == USERSPACE and mmap2 data != 0, sort in this order
- mmap2 data
- data address
- pid (optional)
If cpumode == USERSAPCE and mmap2 data == 0, sort in this order
- pid
- data address
Notice how sorting on the pid is different depending on the scenario.
I'll agree that sorting on iaddr and tid can be filtered out as a nice to have.
But in order to sort on physid, you really need lots of pieces. Otherwise, what do you consider the definition of 'physid'? :-)
Cheers,
Don
PATCH v2 3/19 perf, sort: Allow unique sorting instead of combining hist_entries
In reply to Don ZickusThe cache contention tools needs to keep all the perf records unique in order to properly parse all the data. Currently add_hist_entry() will combine the duplicate record and add the weight/period to the existing record.
This throws away the unique data the cache contention tool needs (mainly the data source). Create a flag to force the records to stay unique.
--- tools/perf/util/hist.c | 3 ++- tools/perf/util/sort.c | 1 + tools/perf/util/sort.h | 1 + 3 files changed, 4 insertions(+), 1 deletion(-)
tools/perf/util/hist.c +2 -1
@@ -365,7 +365,8 @@ static struct hist_entry *add_hist_entry(struct hists *hists,*/ cmp =hist_entry__cmp (he, entry); -if (!cmp) { +if (!cmp && !sort__wants_unique) { +he_stat__add_period (&he->stat, period, weight);/*
tools/perf/util/sort.c +1 -0
@@ -14,6 +14,7 @@ int sort__need_collapse = 0;int sort__has_parent =0 ;int sort__has_sym =0 ;int sort__has_dso =0 ; +int sort__wants_unique =0 ;enum sort_mode sort__mode =SORT_MODE__NORMAL ;enum sort_type sort__first_dimension;
tools/perf/util/sort.h +1 -0
@@ -33,6 +33,7 @@ extern int have_ignore_callees;extern int sort__need_collapse;extern int sort__has_parent;extern int sort__has_sym; +extern int sort__wants_unique;extern enum sort_mode sort__mode;extern struct sort_entry sort_comm;extern struct sort_entry sort_dso;
Signature · 1 line
1.7.11.7
PATCH v2 4/19 perf: Allow ability to map cpus to nodes easily
In reply to Don ZickusThis patch figures out the max number of cpus and nodes that are on the system and creates a map of cpu to node. This allows us to provide a cpu and quickly get the node associated with it.
It was mostly copied from builtin-kmem.c and tweaked slightly to use less memory (use possible cpus instead of max). It also calculates the max number of nodes.
--- tools/perf/util/cpumap.c | 150 +++++++++++++++++++++++++++++++++++++++++++++++ tools/perf/util/cpumap.h | 35 +++++++++++ 2 files changed, 185 insertions(+)
tools/perf/util/cpumap.c +150 -0
@@ -317,3 +317,153 @@ int cpu_map__build_core_map(struct cpu_map *cpus, struct cpu_map **corep) {return cpu_map__build_map (cpus, corep, cpu_map__get_core); } + +/* setup simple routines to easily access node numbers given a cpu number */ + + +#define PATH_SYS_NODE "/sys/devices/system/node" + +/* Determine highest possible cpu in the system for sparse allocation */ +static void set_max_cpu_num (void ) +{ +FILE *fp; +char buf[256 ]; +int num; + +/* set up default */ + max_cpu_num =4096 ; + +/* get the highest possible cpu number for a sparse allocation */ + fp =fopen ("/sys/devices/system/cpu/possible" ,"r" ); +if (!fp) + goto out; + + num =fread (&buf,1 ,sizeof (buf), fp); +if (!num) + goto out_close; + buf[num] ='\0' ; + +/* start on the right, to find highest cpu num */ +while (--num) { +if ((buf[num] ==',' ) || (buf[num] =='-' )) { + num++; +break ; + } + } +if (sscanf (&buf[num],"%d" , &max_cpu_num) <1 ) + goto out_close; + + max_cpu_num++; + +fclose (fp); +return ; + +out_close: +fclose (fp); +out: +pr_err ("Failed to read max cpus, using default of %d\n" , + max_cpu_num); +return ; +} + +/* Determine highest possible node in the system for sparse allocation */ +static void set_max_node_num (void ) +{ +FILE *fp; +char buf[256 ]; +int num; + +/* set up default */ + max_node_num =8 ; + +/* get the highest possible cpu number for a sparse allocation */ + fp =fopen ("/sys/devices/system/node/possible" ,"r" ); +if (!fp) + goto out; + + num =fread (&buf,1 ,sizeof (buf), fp); +if (!num) + goto out_close; + buf[num] ='\0' ; + +/* start on the right, to find highest node num */ +while (--num) { +if ((buf[num] ==',' ) || (buf[num] =='-' )) { + num++; +break ; + } + } +if (sscanf (&buf[num],"%d" , &max_node_num) <1 ) + goto out_close; + + max_node_num++; + +fclose (fp); +return ; + +out_close: +fclose (fp); +out: +pr_err ("Failed to read max nodes, using default of %d\n" , + max_node_num); +return ; +} + +static int init_cpunode_map (void ) +{ +int i; + +set_max_cpu_num (); +set_max_node_num (); + + cpunode_map =calloc (max_cpu_num,sizeof (int )); +if (!cpunode_map) { +pr_err ("%s: calloc failed\n" , __func__); + goto out; + } + +for (i =0 ; i < max_cpu_num; i++) + cpunode_map[i] =-1 ; + +return 0 ; +out: +return -1 ; +} + +int cpu_map__setup_cpunode_map (void ) +{ +struct dirent *dent1, *dent2; +DIR *dir1, *dir2; +unsigned int cpu, mem; +char buf[PATH_MAX ]; + +/* initialize globals */ +if (init_cpunode_map ()) +return -1 ; + + dir1 =opendir (PATH_SYS_NODE ); +if (!dir1) +return 0 ; + +/* walk tree and setup map */ +while ((dent1 =readdir (dir1)) !=NULL ) { +if (dent1->d_type !=DT_DIR || +sscanf (dent1->d_name,"node%u" , &mem) <1 ) +continue ; + +snprintf (buf,PATH_MAX ,"%s/%s" ,PATH_SYS_NODE , dent1->d_name); + dir2 =opendir (buf); +if (!dir2) +continue ; +while ((dent2 =readdir (dir2)) !=NULL ) { +if (dent2->d_type !=DT_LNK || +sscanf (dent2->d_name,"cpu%u" , &cpu) <1 ) +continue ; + cpunode_map[cpu] = mem; + } +closedir (dir2); + } +closedir (dir1); +return 0 ; +} +
tools/perf/util/cpumap.h +35 -0
@@ -4,6 +4,9 @@#include <stdio.h> #include <stdbool.h> +#include "perf.h" +#include "util/debug.h" +struct cpu_map {int nr;int map[];
@@ -46,4 +49,36 @@ static inline bool cpu_map__empty(const struct cpu_map *map) return map ? map->map[0] == -1 : true; } +int max_cpu_num; +int max_node_num; +int *cpunode_map; + +int cpu_map__setup_cpunode_map (void ); + +static inline int cpu_map__max_node (void ) +{ +if (unlikely (!max_node_num)) +pr_debug ("cpu_map not initiailzed\n" ); + +return max_node_num; +} + +static inline int cpu_map__max_cpu (void ) +{ +if (unlikely (!max_cpu_num)) +pr_debug ("cpu_map not initiailzed\n" ); + +return max_cpu_num; +} + +static inline int cpu_map__get_node (int cpu) +{ +if (unlikely (cpunode_map ==NULL )) { +pr_debug ("cpu_map not initialized\n" ); +return -1 ; + } + +return cpunode_map[cpu]; +} +#endif /* __PERF_CPUMAP_H */
Signature · 1 line
1.7.11.7
Re: [PATCH 04/19] perf: Allow ability to map cpus to nodes easily
Quoted 1 line · Don Zickus
This patch figures out the max number of cpus and nodes that are on the
SNIP
Quoted 77 lines
tools/perf/util/cpumap.c +38
-0+ +int cpu_map__setup_cpunode_map (void ) +{ +struct dirent *dent1, *dent2; +DIR *dir1, *dir2; +unsigned int cpu, mem; +char buf[PATH_MAX ]; + +/* initialize globals */ +if (init_cpunode_map ()) +return -1 ; + + dir1 =opendir (PATH_SYS_NODE ); +if (!dir1) +return 0 ; + +/* walk tree and setup map */ +while ((dent1 =readdir (dir1)) !=NULL ) { +if (dent1->d_type !=DT_DIR || +sscanf (dent1->d_name,"node%u" , &mem) <1 ) +continue ; + +snprintf (buf,PATH_MAX ,"%s/%s" ,PATH_SYS_NODE , dent1->d_name); + dir2 =opendir (buf); +if (!dir2) +continue ; +while ((dent2 =readdir (dir2)) !=NULL ) { +if (dent2->d_type !=DT_LNK || +sscanf (dent2->d_name,"cpu%u" , &cpu) <1 ) +continue ; + cpunode_map[cpu] = mem; + } +closedir (dir2); + } +closedir (dir1); +return 0 ; +} +tools/perf/util/cpumap.h +24
-0@@ -4,6 +4,9 @@#include <stdio.h> #include <stdbool.h> +#include "perf.h" +#include "util/debug.h" +struct cpu_map {int nr;int map[];@@ -46,4 +49,36 @@ static inline bool cpu_map__empty(const struct cpu_map *map)return map ? map->map[0 ] ==-1 : true; } +int max_cpu_num; +int max_node_num; +int *cpunode_map; + +int cpu_map__setup_cpunode_map (void ); + +static inline int cpu_map__max_node (void ) +{ +if (unlikely (!max_node_num)) +pr_debug ("cpu_map not initiailzed\n" ); + +return max_node_num; +} + +static inline int cpu_map__max_cpu (void ) +{ +if (unlikely (!max_cpu_num)) +pr_debug ("cpu_map not initiailzed\n" ); + +return max_cpu_num; +}
tools/perf/util/cpumap.h static inline bool cpu_map__empty(const struct cpu_map *map) · line 82
Quoted 10 lines
tools/perf/util/cpumap.h +10
-0+ +static inline int cpu_map__get_node (int cpu) +{ +if (unlikely (cpunode_map ==NULL )) { +pr_debug ("cpu_map not initialized\n" ); +return -1 ; + } + +return cpunode_map[cpu]; +}
cool, maybe above function names should not have 'cpu_map..', it's like pure cpu-ish stuff, maybe:
setup_cpunode_map
cpu__max_node
cpu__max_cpu
cpu__get_node
or something like that ;-)
jirka
Re: [PATCH 04/19] perf: Allow ability to map cpus to nodes easily
Quoted 96 lines · Jiri Olsa
Quoted 1 line · Don Zickus
This patch figures out the max number of cpus and nodes that are on the
SNIP
Quoted 77 lines
tools/perf/util/cpumap.c +38
-0+ +int cpu_map__setup_cpunode_map (void ) +{ +struct dirent *dent1, *dent2; +DIR *dir1, *dir2; +unsigned int cpu, mem; +char buf[PATH_MAX ]; + +/* initialize globals */ +if (init_cpunode_map ()) +return -1 ; + + dir1 =opendir (PATH_SYS_NODE ); +if (!dir1) +return 0 ; + +/* walk tree and setup map */ +while ((dent1 =readdir (dir1)) !=NULL ) { +if (dent1->d_type !=DT_DIR || +sscanf (dent1->d_name,"node%u" , &mem) <1 ) +continue ; + +snprintf (buf,PATH_MAX ,"%s/%s" ,PATH_SYS_NODE , dent1->d_name); + dir2 =opendir (buf); +if (!dir2) +continue ; +while ((dent2 =readdir (dir2)) !=NULL ) { +if (dent2->d_type !=DT_LNK || +sscanf (dent2->d_name,"cpu%u" , &cpu) <1 ) +continue ; + cpunode_map[cpu] = mem; + } +closedir (dir2); + } +closedir (dir1); +return 0 ; +} +tools/perf/util/cpumap.h +24
-0@@ -4,6 +4,9 @@#include <stdio.h> #include <stdbool.h> +#include "perf.h" +#include "util/debug.h" +struct cpu_map {int nr;int map[];@@ -46,4 +49,36 @@ static inline bool cpu_map__empty(const struct cpu_map *map)return map ? map->map[0 ] ==-1 : true; } +int max_cpu_num; +int max_node_num; +int *cpunode_map; + +int cpu_map__setup_cpunode_map (void ); + +static inline int cpu_map__max_node (void ) +{ +if (unlikely (!max_node_num)) +pr_debug ("cpu_map not initiailzed\n" ); + +return max_node_num; +} + +static inline int cpu_map__max_cpu (void ) +{ +if (unlikely (!max_cpu_num)) +pr_debug ("cpu_map not initiailzed\n" ); + +return max_cpu_num; +}tools/perf/util/cpumap.h static inline bool cpu_map__empty(const struct cpu_map *map) · line 82
Quoted 10 lines
tools/perf/util/cpumap.h +10
-0+ +static inline int cpu_map__get_node (int cpu) +{ +if (unlikely (cpunode_map ==NULL )) { +pr_debug ("cpu_map not initialized\n" ); +return -1 ; + } + +return cpunode_map[cpu]; +}cool, maybe above function names should not have '
cpu_map..', it's like pure cpu-ish stuff, maybe:
Quoted 6 lines
setup_cpunode_map
cpu__max_node
cpu__max_cpu
cpu__get_nodeor something like that ;-)
Sure. That works for me. :-)
Cheers,
Don
Re: [PATCH 04/19] perf: Allow ability to map cpus to nodes easily
In reply to Don ZickusOn Fri, Feb 28, 2014 at 12:42:53PM -0500, Don Zickus wrote:
SNIP
Quoted 25 lines
tools/perf/util/cpumap.c +25
-0+ +/* Determine highest possible node in the system for sparse allocation */ +static void set_max_node_num (void ) +{ +FILE *fp; +char buf[256 ]; +int num; + +/* set up default */ + max_node_num =8 ; + +/* get the highest possible cpu number for a sparse allocation */ + fp =fopen ("/sys/devices/system/node/possible" ,"r" ); +if (!fp) + goto out; + + num =fread (&buf,1 ,sizeof (buf), fp); +if (!num) + goto out_close; + buf[num] ='\0' ; + +/* start on the right, to find highest node num */ +while (--num) { +if ((buf[num] ==',' ) || (buf[num] =='-' )) { + num++;
tools/perf/util/cpumap.c int cpu_map__build_core_map(struct cpu_map *cpus, struct cpu_map **corep) · line 402
Quoted 10 lines
tools/perf/util/cpumap.c +10
-0+break ; + } + } +if (sscanf (&buf[num],"%d" , &max_node_num) <1 ) + goto out_close; + + max_node_num++; + +fclose (fp); +return ;
these ^^^ inners should go into single function and be used in both set_max_cpu_num and set_max_node_num functions
jirka
PATCH v2 5/19 perf, kmem: Utilize the new generic cpunode_map
In reply to Don ZickusUse the previous patch implementation of cpunode_map for builtin-kmem.c Should not be any functional difference.
--- tools/perf/builtin-kmem.c | 78 ++--------------------------------------------- 1 file changed, 3 insertions(+), 75 deletions(-)
tools/perf/builtin-kmem.c +3 -75
@@ -14,6 +14,7 @@#include "util/parse-options.h" #include "util/trace-event.h" #include "util/data.h" +#include "util/cpumap.h" #include "util/debug.h"
@@ -31,9 +32,6 @@ static int caller_lines = -1;static bool raw_ip; -static int *cpunode_map; -static int max_cpu_num; -struct alloc_stat {u64 call_site;u64 ptr;
@@ -55,76 +53,6 @@ static struct rb_root root_caller_sorted;static unsigned long total_requested, total_allocated;static unsigned long nr_allocs, nr_cross_allocs; -#define PATH_SYS_NODE "/sys/devices/system/node" - -static int init_cpunode_map (void ) -{ -FILE *fp; -int i, err =-1 ; - - fp =fopen ("/sys/devices/system/cpu/kernel_max" ,"r" ); -if (!fp) { - max_cpu_num =4096 ; -return 0 ; - } - -if (fscanf (fp,"%d" , &max_cpu_num) <1 ) { -pr_err ("Failed to read 'kernel_max' from sysfs" ); - goto out_close; - } - - max_cpu_num++; - - cpunode_map =calloc (max_cpu_num,sizeof (int )); -if (!cpunode_map) { -pr_err ("%s: calloc failed\n" , __func__); - goto out_close; - } - -for (i =0 ; i < max_cpu_num; i++) - cpunode_map[i] =-1 ; - - err =0 ; -out_close: -fclose (fp); -return err; -} - -static int setup_cpunode_map (void ) -{ -struct dirent *dent1, *dent2; -DIR *dir1, *dir2; -unsigned int cpu, mem; -char buf[PATH_MAX ]; - -if (init_cpunode_map ()) -return -1 ; - - dir1 =opendir (PATH_SYS_NODE ); -if (!dir1) -return 0 ; - -while ((dent1 =readdir (dir1)) !=NULL ) { -if (dent1->d_type !=DT_DIR || -sscanf (dent1->d_name,"node%u" , &mem) <1 ) -continue ; - -snprintf (buf,PATH_MAX ,"%s/%s" ,PATH_SYS_NODE , dent1->d_name); - dir2 =opendir (buf); -if (!dir2) -continue ; -while ((dent2 =readdir (dir2)) !=NULL ) { -if (dent2->d_type !=DT_LNK || -sscanf (dent2->d_name,"cpu%u" , &cpu) <1 ) -continue ; - cpunode_map[cpu] = mem; - } -closedir (dir2); - } -closedir (dir1); -return 0 ; -} -static int insert_alloc_stat (unsigned long call_site,unsigned long ptr,int bytes_req,int bytes_alloc,int cpu) {
@@ -235,7 +163,7 @@ static int perf_evsel__process_alloc_node_event(struct perf_evsel *evsel,int ret =perf_evsel__process_alloc_event (evsel, sample);if (!ret) { -int node1 = cpunode_map[sample->cpu], +int node1 =cpu_map__get_node (sample->cpu), node2 =perf_evsel__intval (evsel, sample,"node" );if (node1 != node2)
@@ -770,7 +698,7 @@ int cmd_kmem(int argc, const char **argv, const char *prefix __maybe_unused)if (!strncmp (argv[0 ],"rec" ,3 )) {return __cmd_record (argc, argv); }else if (!strcmp (argv[0 ],"stat" )) { -if (setup_cpunode_map ()) +if (cpu_map__setup_cpunode_map ())return -1 ;if (list_empty (&caller_sort))
Signature · 1 line
1.7.11.7
PATCH v2 6/19 perf: Fix stddev calculation
In reply to Don ZickusThe stddev calculation written matched standard error. As a result when using this result to find the relative stddev between runs, it was not accurate.
Update the formula to match traditional stddev. Then rename the old stddev calculation to stderr_stats in case someone wants to use it.
--- tools/perf/util/stat.c | 13 +++++++++++++ tools/perf/util/stat.h | 1 + 2 files changed, 14 insertions(+)
tools/perf/util/stat.c +13 -0
@@ -33,6 +33,7 @@ double avg_stats(struct stats *stats)* http://en.wikipedia.org/wiki/Stddev * * The std dev of the mean is related to the std dev by: +* (also known as standard error) * * s * s_mean = -------
@@ -41,6 +42,18 @@ double avg_stats(struct stats *stats)*/ double stddev_stats (struct stats *stats) { +double variance; + +if (stats->n <2 ) +return 0.0 ; + + variance = stats->M2 / (stats->n -1 ); + +return sqrt (variance); +} + +double stderr_stats (struct stats *stats) +{double variance, variance_mean;if (stats->n <2 )
tools/perf/util/stat.h +1 -0
@@ -12,6 +12,7 @@ struct statsvoid update_stats (struct stats *stats,u64 val);double avg_stats (struct stats *stats);double stddev_stats (struct stats *stats); +double stderr_stats (struct stats *stats);double rel_stddev_stats (double stddev,double avg);static inline void init_stats (struct stats *stats)
Signature · 1 line
1.7.11.7
PATCH v2 7/19 perf, callchain: Add generic callchain print handler for stdio
In reply to Don ZickusMy initial implementation for rbtree sorting in the c2c tool does not use the normal history elements. As a result, adding callchain support (which is deeply integrated with history elements) is more challenging when trying to display its output.
To make things simpler for myself (and to avoid rewriting the same code into the c2c tool), I provided a generic interface that takes an unsorted callchain list along with its total and relative sample size, and sorts it locally based on period and calls the appropriate graph function (passing the correct sample size).
This makes things easier because the c2c tool can be dumber and just collect callchains and not worry about the magic needed to sort and display them correctly.
Unfortunately, this is assuming a stdio output only and does not use the other gui type outputs.
Regardless, this patch provides useful info for the tool right now. Tweaks and recommendations for a better approach are welcomed. :-)
--- tools/perf/ui/stdio/hist.c | 37 +++++++++++++++++++++++++++++++++++++ tools/perf/util/hist.h | 4 ++++ 2 files changed, 41 insertions(+)
tools/perf/ui/stdio/hist.c +37 -0
@@ -536,3 +536,40 @@ size_t events_stats__fprintf(struct events_stats *stats, FILE *fp)return ret; } + +size_t generic_entry_callchain__fprintf (struct callchain_root *unsorted_callchain, +u64 total_samples,u64 relative_samples, +int left_margin,FILE *fp) +{ +struct rb_root sorted_chain; +u64 min_callchain_hits; + +if (!symbol_conf.use_callchain) +return 0 ; + + min_callchain_hits = total_samples * (callchain_param.min_percent /100 ); + + callchain_param.sort (&sorted_chain, unsorted_callchain, + min_callchain_hits, &callchain_param); + +switch (callchain_param.mode) { +case CHAIN_GRAPH_REL : +return callchain__fprintf_graph (fp, &sorted_chain, relative_samples, + left_margin); +break ; +case CHAIN_GRAPH_ABS : +return callchain__fprintf_graph (fp, &sorted_chain, total_samples, + left_margin); +break ; +case CHAIN_FLAT : +return callchain__fprintf_flat (fp, &sorted_chain, total_samples); +break ; +case CHAIN_NONE : +break ; +default : +pr_err ("Bad callchain mode\n" ); + } + +return 0 ; +} +
tools/perf/util/hist.h +4 -0
@@ -112,6 +112,10 @@ size_t events_stats__fprintf(struct events_stats *stats, FILE *fp);size_t hists__fprintf (struct hists *hists,bool show_header,int max_rows,int max_cols,float min_pcnt,FILE *fp); +size_t generic_entry_callchain__fprintf (struct callchain_root *unsorted_callchain, +u64 total_samples,u64 relative_samples, +int left_margin,FILE *fp); +void hists__filter_by_dso (struct hists *hists);void hists__filter_by_thread (struct hists *hists);void hists__filter_by_symbol (struct hists *hists);
Signature · 1 line
1.7.11.7
PATCH v2 8/19 perf c2c: Shared data analyser
In reply to Don ZickusFrom: Arnaldo Carvalho de Melo <[email protected]>
This is the start of a new perf tool that will collect information about memory accesses and analyse it to find things like hot cachelines, etc.
This is basically trying to get a prototype written by Richard Fowles written using the tools/perf coding style and libraries.
Start it from 'perf sched', this patch starts the process by adding the 'record' subcommand to collect the needed mem loads and stores samples.
It also have the basic 'report' skeleton, resolving the sample address and hooking the events found in a perf.data file with methods to handle them, right now just printing the resolved perf_sample data structure after each event name.
[dcz: refreshed to latest upstream changes]
--- tools/perf/Documentation/perf-c2c.c | 22 +++++ tools/perf/Makefile.perf | 1 + tools/perf/builtin-c2c.c | 185 ++++++++++++++++++++++++++++++++++++ tools/perf/builtin.h | 1 + tools/perf/perf.c | 1 + 5 files changed, 210 insertions(+) create mode 100644 tools/perf/Documentation/perf-c2c.c create mode 100644 tools/perf/builtin-c2c.c
tools/perf/Documentation/perf-c2c.c new file +22 -0
@@ -0,0 +1,22 @@ +perf-c2c (1 ) +=========== + +NAME +---- +perf-c2c - Shared Data C2C/HITM Analyzer. + +SYNOPSIS +-------- +[verse] +'perf c2c' record + +DESCRIPTION +----------- +These are the variants of perf c2c: + + 'perf c2c record <command>' to record the memory accesses of an arbitrary + workload. + +SEE ALSO +-------- +linkperf:perf-record[1], linkperf:perf-mem[1]
tools/perf/Makefile.perf +1 -0
@@ -427,6 +427,7 @@ endif BUILTIN_OBJS += $(OUTPUT)bench/mem-memcpy.o BUILTIN_OBJS += $(OUTPUT)bench/mem-memset.o +BUILTIN_OBJS += $(OUTPUT)builtin-c2c.o BUILTIN_OBJS += $(OUTPUT)builtin-diff.o BUILTIN_OBJS += $(OUTPUT)builtin-evlist.o BUILTIN_OBJS += $(OUTPUT)builtin-help.o
tools/perf/builtin-c2c.c new file +185 -0
185 lines not on this page · raw message
tools/perf/builtin.h +1 -0
@@ -17,6 +17,7 @@ extern int cmd_annotate(int argc, const char **argv, const char *prefix);extern int cmd_bench (int argc,const char **argv,const char *prefix);extern int cmd_buildid_cache (int argc,const char **argv,const char *prefix);extern int cmd_buildid_list (int argc,const char **argv,const char *prefix); +extern int cmd_c2c (int argc,const char **argv,const char *prefix);extern int cmd_diff (int argc,const char **argv,const char *prefix);extern int cmd_evlist (int argc,const char **argv,const char *prefix);extern int cmd_help (int argc,const char **argv,const char *prefix);
tools/perf/perf.c +1 -0
@@ -35,6 +35,7 @@ struct cmd_struct {static struct cmd_struct commands[] = { {"buildid-cache" , cmd_buildid_cache,0 }, {"buildid-list" , cmd_buildid_list,0 }, + {"c2c" , cmd_c2c,0 }, {"diff" , cmd_diff,0 }, {"evlist" , cmd_evlist,0 }, {"help" , cmd_help,0 },
Signature · 1 line
1.7.11.7
Re: [PATCH 08/19] perf c2c: Shared data analyser
tools/perf/builtin-c2c.c line 46
Quoted 4 lines · Don Zickus
tools/perf/builtin-c2c.c new file +4
-0+ +static const struct perf_evsel_str_handler handlers[] = { + {"cpu/mem-loads,ldlat=30/pp" , perf_c2c__process_load, }, + {"cpu/mem-stores/pp" , perf_c2c__process_store, },
The 30 magic number should probably be configurable.
Using load-latency here rules out Atom, so at some point you would need to get rid of that.
I suspect on most systems you should rather use p instead of pp to get the overhead down (before Haswell pp is expensive)
tools/perf/builtin-c2c.c line 123
Quoted 7 lines
tools/perf/builtin-c2c.c new file +7
-0+static int perf_c2c__record (int argc,const char **argv) +{ +unsigned int rec_argc, i, j; +const char **rec_argv; +const char *const record_args[] = { +"record" , +/* "--phys-addr", */
So is that needed or not?
-Andi
Signature · 1 line
[email protected] -- Speaking for myself only
Re: [PATCH 08/19] perf c2c: Shared data analyser
Quoted 7 lines · Andi Kleen
tools/perf/builtin-c2c.c line 46
Quoted 4 lines · Don Zickus
tools/perf/builtin-c2c.c new file +4
-0+ +static const struct perf_evsel_str_handler handlers[] = { + {"cpu/mem-loads,ldlat=30/pp" , perf_c2c__process_load, }, + {"cpu/mem-stores/pp" , perf_c2c__process_store, },The 30 magic number should probably be configurable.
Yeah, I just didn't figure out how to make it configurable yet within this string.
Quoted 3 lines
Using load-latency here rules out Atom, so at some point you would need to get rid of that.
Oh. How do you get load-latency for Atom then?
Quoted 4 lines
I suspect on most systems you should rather use p instead of pp to get the overhead down (before Haswell pp is expensive)
Ok. Good to know.
Quoted 10 lines
tools/perf/builtin-c2c.c line 123
Quoted 8 lines
tools/perf/builtin-c2c.c new file +7
-0+static int perf_c2c__record (int argc,const char **argv) +{ +unsigned int rec_argc, i, j; +const char **rec_argv; +const char *const record_args[] = { +"record" , +/* "--phys-addr", */ So is that needed or not?
No. Legacy code before we had MMAP2 support. I'll remove it next respin.
Cheers,
Don
Re: [PATCH 08/19] perf c2c: Shared data analyser
Quoted 7 lines · Don Zickus
Quoted 6 lines · Andi Kleen
tools/perf/builtin-c2c.c line 46
Quoted 4 lines · Don Zickus
tools/perf/builtin-c2c.c new file +4
-0+ +static const struct perf_evsel_str_handler handlers[] = { + {"cpu/mem-loads,ldlat=30/pp" , perf_c2c__process_load, }, + {"cpu/mem-stores/pp" , perf_c2c__process_store, },
Hmm I'm getting this when running a simple record command.
invalid or unsupported event: 'cpu/mem-loads/pp'
This only occurs with c2c, other subcommands work normally. It's as if it were an old kernel, but it's Linus' latest. Is this an issue with the patch or something I'm missing?
Furthermore, I see:
ls /sys/bus/event_source/devices/cpu/events
branch-instructions branch-misses cache-misses cache-references cpu-cycles instructions mem-loads
Thanks!
Re: [PATCH 08/19] perf c2c: Shared data analyser
Apologies for the resend. My first msg contained html in it.
Quoted 13 lines · Davidlohr Bueso
Quoted 7 lines · Don Zickus
Quoted 6 lines · Andi Kleen
tools/perf/builtin-c2c.c line 46
Quoted 4 lines · Don Zickus
tools/perf/builtin-c2c.c new file +4
-0+ +static const struct perf_evsel_str_handler handlers[] = { + {"cpu/mem-loads,ldlat=30/pp" , perf_c2c__process_load, }, + {"cpu/mem-stores/pp" , perf_c2c__process_store, },Hmm I'm getting this when running a simple record command.
invalid or unsupported event: 'cpu/mem-loads/pp'
Quoted 7 lines
This only occurs with c2c, other subcommands work normally. It's as if it were an old kernel, but it's Linus' latest. Is this an issue with the patch or something I'm missing?
Furthermore, I see:
ls/sys/bus/event_source/devices/cpu/events
branch-instructions branch-misses cache-misses cache-references cpu-cycles instructions mem-loads
David:
It looks like you're running on an older Intel processor, which is missing necessary events for C2C to work. As Don noted in his patch 00/19, this was primarily developed and tested on Intel's Ivy Bridge platform. If you rerun this on an Ivy Bridge, it should work fine.
We should add a runtime check for supported platforms.
Joe
Quoted 4 lines
Thanks!
Re: [PATCH 08/19] perf c2c: Shared data analyser
Quoted 2 lines
David:
It looks like you're running on an older Intel processor, which is missing necessary events for C2C to work.
mem-loads should be supported Nehalem and up.
mem-stores is Sandy Bridge and up
You can check in perf list
-Andi
Re: [PATCH 08/19] perf c2c: Shared data analyser
In reply to Davidlohr BuesoQuoted 9 lines · Davidlohr Bueso
Quoted 7 lines · Don Zickus
Quoted 6 lines · Andi Kleen
tools/perf/builtin-c2c.c line 46
Quoted 4 lines · Don Zickus
tools/perf/builtin-c2c.c new file +4
-0+ +static const struct perf_evsel_str_handler handlers[] = { + {"cpu/mem-loads,ldlat=30/pp" , perf_c2c__process_load, }, + {"cpu/mem-stores/pp" , perf_c2c__process_store, },
Quoted 7 lines
Hmm I'm getting this when running a simple record command.
invalid or unsupported event: 'cpu/mem-loads/pp'
This only occurs with c2c, other subcommands work normally. It's as if it were an old kernel, but it's Linus' latest. Is this an issue with the patch or something I'm missing?
Does rebooting or 'echo 100000 > /proc/sys/kernel/perf_event_max_sample_rate'
fix the problem? Perf throttling might have lowered your sample rate so low that you can't do anything any more. Though if you are running the latest from Linus a lot of fixes should be in there to minimize that problem.
What processor are you on?
Cheers,
Don
Quoted 9 lines
Furthermore, I see:
ls/sys/bus/event_source/devices/cpu/events
branch-instructions branch-misses cache-misses cache-references cpu-cycles instructions mem-loadsThanks!
Re: [PATCH 08/19] perf c2c: Shared data analyser
In reply to Davidlohr BuesoQuoted 13 lines · Davidlohr Bueso
Quoted 7 lines · Don Zickus
Quoted 6 lines · Andi Kleen
tools/perf/builtin-c2c.c line 46
Quoted 4 lines · Don Zickus
tools/perf/builtin-c2c.c new file +4
-0+ +static const struct perf_evsel_str_handler handlers[] = { + {"cpu/mem-loads,ldlat=30/pp" , perf_c2c__process_load, }, + {"cpu/mem-stores/pp" , perf_c2c__process_store, },Hmm I'm getting this when running a simple record command.
invalid or unsupported event: 'cpu/mem-loads/pp'
Quoted 7 lines
This only occurs with c2c, other subcommands work normally. It's as if it were an old kernel, but it's Linus' latest. Is this an issue with the patch or something I'm missing?
Furthermore, I see:
ls/sys/bus/event_source/devices/cpu/events
branch-instructions branch-misses cache-misses cache-references cpu-cycles instructions mem-loads
Hmm, so based on Andi's reply, I am assuming you are running on a Westmere (or Nehalem) due to the lack of mem-stores.
If you don't have mem-stores, this tool isn't going to work. The tool can only detect contention when sampling reads _and_ writes to the same addresses.
Cheers,
Don
Re: [PATCH 08/19] perf c2c: Shared data analyser
Quoted 6 lines
Hmm, so based on Andi's reply, I am assuming you are running on a Westmere (or Nehalem) due to the lack of mem-stores.
If you don't have mem-stores, this tool isn't going to work. The tool can only detect contention when sampling reads _and_ writes to the same addresses.
On these CPUs you could simply sample on HITM. You won't get addresses, but at least IPs and call stacks.
-Andi
Re: [PATCH 08/19] perf c2c: Shared data analyser
Quoted 12 lines · Andi Kleen
Quoted 6 lines
Hmm, so based on Andi's reply, I am assuming you are running on a Westmere (or Nehalem) due to the lack of mem-stores.
If you don't have mem-stores, this tool isn't going to work. The tool can only detect contention when sampling reads _and_ writes to the same addresses.
On these CPUs you could simply sample on HITM. You won't get addresses, but at least IPs and call stacks.
-Andi
If you only sample on the HITMs then you don't get the store misses. That means you'll not be able to detect who is simultaneously tugging on the same cache lines. That gives up much of the value of "perf c2c".
As we developed this, we ended up settling on Ivy Bridge to get the behavior we wanted.
Joe
Peter Zijlstra
maintainerRe: [PATCH 08/19] perf c2c: Shared data analyser
Quoted 4 lines · Joe Mario
If you only sample on the HITMs then you don't get the store misses. That means you'll not be able to detect who is simultaneously tugging on the same cache lines. That gives up much of the value of "perf c2c".
As long as you know which lines are hurting bringing in (loads) you can often figure out who is doing the stores on them.
Quoted 2 lines
As we developed this, we ended up settling on Ivy Bridge to get the behavior we wanted.
Wouldn't SNB also work?
Re: [PATCH 08/19] perf c2c: Shared data analyser
Quoted 8 lines · Peter Zijlstra
Quoted 4 lines · Joe Mario
If you only sample on the HITMs then you don't get the store misses. That means you'll not be able to detect who is simultaneously tugging on the same cache lines. That gives up much of the value of "perf c2c".
As long as you know which lines are hurting bringing in (loads) you can often figure out who is doing the stores on them.
Yes, especially since every store is a load too (unless you're talking WC)
The method c2c uses is more exact, but keep in mind it's a sampling heuristic in any cases, with some potential bias. load-latency tags the loads randomly and there's no guarantee that tagging is fully uniform. Also you only see a subset in any case.
Quoted 5 lines
Quoted 3 lines
As we developed this, we ended up settling on Ivy Bridge to get the behavior we wanted.
Wouldn't SNB also work?
Yes.
Haswell is best however because it can report addresses on far more events.
-Andi
Signature · 1 line
[email protected] -- Speaking for myself only.
Peter Zijlstra
maintainerRe: [PATCH 08/19] perf c2c: Shared data analyser
Quoted 2 lines · Andi Kleen
Haswell is best however because it can report addresses on far more events.
Sure; however I think the code to data structure mapping might be more useful still. Its very hard to map a random address into a heap and find a fitting data structure. You'd need some allocator with a reverse map, which to me seems like too much overhead for normal runtime use.
Re: [PATCH 08/19] perf c2c: Shared data analyser
In reply to Andi KleenQuoted 7 lines · Andi Kleen
Quoted 6 lines · Peter Zijlstra
Quoted 4 lines · Joe Mario
If you only sample on the HITMs then you don't get the store misses. That means you'll not be able to detect who is simultaneously tugging on the same cache lines. That gives up much of the value of "perf c2c".
Quoted 5 lines
Quoted 2 lines
As long as you know which lines are hurting bringing in (loads) you can often figure out who is doing the stores on them.
Yes, especially since every store is a load too (unless you're talking WC)
Thoughts on how to determine which load is a potential store? I agree every store needs to load the cacheline, but I wasn't sure if there was an approach that could be applied to determine anything useful.
Cheers,
Don
Quoted 21 lines
The method c2c uses is more exact, but keep in mind it's a sampling heuristic in any cases, with some potential bias. load-latency tags the loads randomly and there's no guarantee that tagging is fully uniform. Also you only see a subset in any case.
Quoted 5 lines
Quoted 3 lines
As we developed this, we ended up settling on Ivy Bridge to get the behavior we wanted.
Wouldn't SNB also work?
Yes.
Haswell is best however because it can report addresses on far more events.
-Andi
Signature · 1 line
[email protected] -- Speaking for myself only.
Re: [PATCH 08/19] perf c2c: Shared data analyser
Quoted 3 lines
Thoughts on how to determine which load is a potential store? I agree every store needs to load the cacheline, but I wasn't sure if there was an approach that could be applied to determine anything useful.
HITM covers both load and stores.
But it may have some skid. The main advantage of the address is that it is skidless.
-Andi
Re: [PATCH 08/19] perf c2c: Shared data analyser
Quoted 8 lines · Andi Kleen
Quoted 3 lines
Thoughts on how to determine which load is a potential store? I agree every store needs to load the cacheline, but I wasn't sure if there was an approach that could be applied to determine anything useful.
HITM covers both load and stores.
But it may have some skid. The main advantage of the address is that it is skidless.
I guess I don't follow, how does having a skidless address help me when I don't have any stores (on a Westmere for example)?
Cheers,
Don
Re: [PATCH 08/19] perf c2c: Shared data analyser
In reply to Peter ZijlstraQuoted 8 lines · Peter Zijlstra
Quoted 4 lines · Joe Mario
If you only sample on the HITMs then you don't get the store misses. That means you'll not be able to detect who is simultaneously tugging on the same cache lines. That gives up much of the value of "perf c2c".
As long as you know which lines are hurting bringing in (loads) you can often figure out who is doing the stores on them.
If you are familar with the code I guess. One scenario that would be difficult for is shared memory (ie databases) as it may not be obvious who is writing to a cacheline.
Quoted 5 lines
Quoted 3 lines
As we developed this, we ended up settling on Ivy Bridge to get the behavior we wanted.
Wouldn't SNB also work?
We started with SNB, but because latency info was not working (due to a GO bug) and some other event quirks, we found ourselves migrating to a more stable IVB platform instead.
Now that we stablized our test case and tool, SNB may work. We just
haven't tried in along time.
Cheers,
Don
Re: [PATCH 08/19] perf c2c: Shared data analyser
In reply to Andi KleenQuoted 9 lines · Andi Kleen
Quoted 6 lines
Hmm, so based on Andi's reply, I am assuming you are running on a Westmere (or Nehalem) due to the lack of mem-stores.
If you don't have mem-stores, this tool isn't going to work. The tool can only detect contention when sampling reads _and_ writes to the same addresses.
On these CPUs you could simply sample on HITM. You won't get addresses, but at least IPs and call stacks.
Heh. I never thought about that. And sure enough a quick test with mem-stores commented out produced the same results (minus the stores).
One would just have to 'figure' out what cacheline offsets are causing the HITMs.
Cheers,
Don
Re: [PATCH 08/19] perf c2c: Shared data analyser
Quoted 5 lines
Heh. I never thought about that. And sure enough a quick test with mem-stores commented out produced the same results (minus the stores).
One would just have to 'figure' out what cacheline offsets are causing the HITMs.
Often that can be determined statically from the instruction (register-offset)
However keep in mind that there is some skid so the instruction may not be correct.
-Andi
Signature · 1 line
[email protected] -- Speaking for myself only.
Peter Zijlstra
maintainerRe: [PATCH 08/19] perf c2c: Shared data analyser
Quoted 10 lines · Andi Kleen
Quoted 5 lines
Heh. I never thought about that. And sure enough a quick test with mem-stores commented out produced the same results (minus the stores).
One would just have to 'figure' out what cacheline offsets are causing the HITMs.
Often that can be determined statically from the instruction (register-offset)
However keep in mind that there is some skid so the instruction may not be correct.
Even with the real_ip pebs field? I thought the whole purpose of that was to tag the actual instruction matching the event.
Re: [PATCH 08/19] perf c2c: Shared data analyser
In reply to Don ZickusQuoted 18 lines · Don Zickus
Quoted 17 lines · Davidlohr Bueso
Quoted 7 lines · Don Zickus
Quoted 6 lines · Andi Kleen
tools/perf/builtin-c2c.c line 46
Quoted 4 lines · Don Zickus
tools/perf/builtin-c2c.c new file +4
-0+ +static const struct perf_evsel_str_handler handlers[] = { + {"cpu/mem-loads,ldlat=30/pp" , perf_c2c__process_load, }, + {"cpu/mem-stores/pp" , perf_c2c__process_store, },Hmm I'm getting this when running a simple record command.
invalid or unsupported event: 'cpu/mem-loads/pp'
This only occurs with c2c, other subcommands work normally. It's as if it were an old kernel, but it's Linus' latest. Is this an issue with the patch or something I'm missing?
Quoted 6 lines
Quoted 3 lines
Furthermore, I see:
ls/sys/bus/event_source/devices/cpu/events
branch-instructions branch-misses cache-misses cache-references cpu-cycles instructions mem-loadsHmm, so based on Andi's reply, I am assuming you are running on a Westmere (or Nehalem) due to the lack of mem-stores.
Indeed, I had definitely overlooked this detail.
Performance Events: PEBS fmt1+, 16-deep LBR, Westmere events, Broken BIOS detected, complain to your hardware vendor.
Thanks,
Davidlohr
PATCH v2 9/19 perf c2c: Dump raw records, decode data_src bits
In reply to Don ZickusFrom: Arnaldo Carvalho de Melo <[email protected]>
Quoted 1 line
From the c2c prototype:
[root@sandy ~]# perf c2c -r report | head -7
T Status Pid Tid CPU Inst Adrs Virt Data Adrs Phys Data Adrs Cycles Source Decoded Source ObJect:Symbol
--------------------------------------------------------------------------------------------------------------------------------------------
raw input 779 779 7 0xffffffff810865dd 0xffff8803f4d75ec8 0 370 0x68080882 [LOAD,LCL_LLC,MISS,SNP NA] [kernel.kallsyms]:try_to_wake_up raw input 779 779 7 0xffffffff8107acb3 0xffff8802a5b73158 0 297 0x6a100142 [LOAD,L1,HIT,SNP NONE,LOCKED] [kernel.kallsyms]:up_read raw input 779 779 7 0x3b7e009814 0x7fff87429ea0 0 925 0x68100142 [LOAD,L1,HIT,SNP NONE] ???:??? raw input 0 0 1 0xffffffff8108bf81 0xffff8803eafebf50 0 172 0x68800842 [LOAD,LCL_LLC,HIT,SNP HITM] [kernel.kallsyms]:update_stats_wait_end raw input 779 779 7 0x3b7e0097cc 0x7fac94b69068 0 228 0x68100242 [LOAD,LFB,HIT,SNP NONE] ???:???
[root@sandy ~]#
The "Phys Data Adrs" column is not available at this point.
--- tools/perf/builtin-c2c.c | 148 +++++++++++++++++++++++++++++++++++++++-------- 1 file changed, 125 insertions(+), 23 deletions(-)
tools/perf/builtin-c2c.c +125 -23
@@ -11,51 +11,148 @@struct perf_c2c {struct perf_tool tool; +bool raw_records; }; -static int perf_sample__fprintf (struct perf_sample *sample, -struct perf_evsel *evsel, -struct addr_location *al,FILE *fp) +enum {OP ,LVL ,SNP ,LCK ,TLB }; + +static int perf_c2c__scnprintf_data_src (char *bf,size_t size,uint64_t val) { -return fprintf (fp,"%25.25s: %5d %5d 0x%016" PRIx64" 0x016%" PRIx64" %5" PRIu64" 0x%06" PRIx64" %s:%s\n" , -perf_evsel__name (evsel), - sample->pid, sample->tid, sample->ip, sample->addr, - sample->weight, sample->data_src, - al->map ? (al->map->dso ? al->map->dso->long_name :"???" ) :"???" , - al->sym ? al->sym->name :"???" ); +#define PREFIX "[" +#define SUFFIX "]" +#define ELLIPSIS "..." +static const struct { +uint64_t bit; +int64_t field; +const char *name; + } decode_bits[] = { + {PERF_MEM_OP_LOAD ,OP ,"LOAD" }, + {PERF_MEM_OP_STORE ,OP ,"STORE" }, + {PERF_MEM_OP_NA ,OP ,"OP_NA" }, + {PERF_MEM_LVL_LFB ,LVL ,"LFB" }, + {PERF_MEM_LVL_L1 ,LVL ,"L1" }, + {PERF_MEM_LVL_L2 ,LVL ,"L2" }, + {PERF_MEM_LVL_L3 ,LVL ,"LCL_LLC" }, + {PERF_MEM_LVL_LOC_RAM ,LVL ,"LCL_RAM" }, + {PERF_MEM_LVL_REM_RAM1 ,LVL ,"RMT_RAM" }, + {PERF_MEM_LVL_REM_RAM2 ,LVL ,"RMT_RAM" }, + {PERF_MEM_LVL_REM_CCE1 ,LVL ,"RMT_LLC" }, + {PERF_MEM_LVL_REM_CCE2 ,LVL ,"RMT_LLC" }, + {PERF_MEM_LVL_IO ,LVL ,"I/O" }, + {PERF_MEM_LVL_UNC ,LVL ,"UNCACHED" }, + {PERF_MEM_LVL_NA ,LVL ,"N" }, + {PERF_MEM_LVL_HIT ,LVL ,"HIT" }, + {PERF_MEM_LVL_MISS ,LVL ,"MISS" }, + {PERF_MEM_SNOOP_NONE ,SNP ,"SNP NONE" }, + {PERF_MEM_SNOOP_HIT ,SNP ,"SNP HIT" }, + {PERF_MEM_SNOOP_MISS ,SNP ,"SNP MISS" }, + {PERF_MEM_SNOOP_HITM ,SNP ,"SNP HITM" }, + {PERF_MEM_SNOOP_NA ,SNP ,"SNP NA" }, + {PERF_MEM_LOCK_LOCKED ,LCK ,"LOCKED" }, + {PERF_MEM_LOCK_NA ,LCK ,"LOCK_NA" }, + }; +union perf_mem_data_src dsrc = { .val = val, }; +int printed =scnprintf (bf, size,PREFIX ); +size_t i; +bool first_present = true; + +for (i =0 ; i <ARRAY_SIZE (decode_bits); i++) { +int bitval; + +switch (decode_bits[i].field) { +case OP : bitval = decode_bits[i].bit & dsrc.mem_op;break ; +case LVL : bitval = decode_bits[i].bit & dsrc.mem_lvl;break ; +case SNP : bitval = decode_bits[i].bit & dsrc.mem_snoop;break ; +case LCK : bitval = decode_bits[i].bit & dsrc.mem_lock;break ; +case TLB : bitval = decode_bits[i].bit & dsrc.mem_dtlb;break ; +default : bitval =0 ;break ; + } + +if (!bitval) +continue ; + +if (strlen (decode_bits[i].name) + !!i > size - printed -sizeof (SUFFIX )) { +sprintf (bf + size -sizeof (SUFFIX ) -sizeof (ELLIPSIS ) +1 ,ELLIPSIS ); + printed = size -sizeof (SUFFIX ); +break ; + } + + printed +=scnprintf (bf + printed, size - printed,"%s%s" , + first_present ?"" :"," , decode_bits[i].name); + first_present = false; + } + + printed +=scnprintf (bf + printed, size - printed,SUFFIX ); +return printed; } -static int perf_c2c__process_load (struct perf_evsel *evsel, -struct perf_sample *sample, -struct addr_location *al) +static int perf_c2c__fprintf_header (FILE *fp) { -perf_sample__fprintf (sample, evsel, al, stdout); -return 0 ; +int printed =fprintf (fp,"%c %-16s %6s %6s %4s %18s %18s %18s %6s %-10s %-60s %s\n" , +'T' , +"Status" , +"Pid" , +"Tid" , +"CPU" , +"Inst Adrs" , +"Virt Data Adrs" , +"Phys Data Adrs" , +"Cycles" , +"Source" , +" Decoded Source" , +"ObJect:Symbol" ); +return printed +fprintf (fp,"%-*.*s\n" , printed, printed, graph_dotted_line); +} + +static int perf_sample__fprintf (struct perf_sample *sample,char tag, +const char *reason,struct addr_location *al,FILE *fp) +{ +char data_src[61 ]; + +perf_c2c__scnprintf_data_src (data_src,sizeof (data_src), sample->data_src); + +return fprintf (fp,"%c %-16s %6d %6d %4d %#18" PRIx64" %#18" PRIx64" %#18" PRIx64" %6" PRIu64" %#10" PRIx64" %-60.60s %s:%s\n" , + tag, + reason ?:"valid record" , + sample->pid, + sample->tid, + sample->cpu, + sample->ip, + sample->addr, +0UL , + sample->weight, + sample->data_src, + data_src, + al->map ? (al->map->dso ? al->map->dso->long_name :"???" ) :"???" , + al->sym ? al->sym->name :"???" ); } -static int perf_c2c__process_store (struct perf_evsel *evsel, -struct perf_sample *sample, -struct addr_location *al) +static int perf_c2c__process_load_store (struct perf_c2c *c2c, +struct perf_sample *sample, +struct addr_location *al) { -perf_sample__fprintf (sample, evsel, al, stdout); +if (c2c->raw_records) +perf_sample__fprintf (sample,' ' ,"raw input" , al, stdout); +return 0 ; }static const struct perf_evsel_str_handler handlers[] = { - {"cpu/mem-loads,ldlat=30/pp" , perf_c2c__process_load, }, - {"cpu/mem-stores/pp" , perf_c2c__process_store, }, + {"cpu/mem-loads,ldlat=30/pp" , perf_c2c__process_load_store, }, + {"cpu/mem-stores/pp" , perf_c2c__process_load_store, }, }; -typedef int (*sample_handler )(struct perf_evsel *evsel, +typedef int (*sample_handler )(struct perf_c2c *c2c,struct perf_sample *sample,struct addr_location *al); -static int perf_c2c__process_sample (struct perf_tool *tool__maybe_unused , +static int perf_c2c__process_sample (struct perf_tool *tool,union perf_event *event,struct perf_sample *sample,struct perf_evsel *evsel,struct machine *machine) { + struct perf_c2c *c2c = container_of(tool, struct perf_c2c, tool);struct addr_location al;int err =0 ;
@@ -67,7 +164,7 @@ static int perf_c2c__process_sample(struct perf_tool *tool __maybe_unused,if (evsel->handler !=NULL ) {sample_handler f = evsel->handler; - err =f (evsel, sample, &al); + err =f (c2c, sample, &al); }return err;
@@ -111,6 +208,10 @@ out:static int perf_c2c__report (struct perf_c2c *c2c) {setup_pager (); + +if (c2c->raw_records) +perf_c2c__fprintf_header (stdout); +return perf_c2c__read_events (c2c); }
@@ -161,6 +262,7 @@ int cmd_c2c(int argc, const char **argv, const char *prefix __maybe_unused) }, };const struct option c2c_options[] = { +OPT_BOOLEAN ('r' ,"raw_records" , &c2c.raw_records,"dump raw events" ),OPT_END () };const char *const c2c_usage[] = {
Signature · 1 line
1.7.11.7
PATCH v2 10/19 perf, c2c: Rework setup code to prepare for features
In reply to Don ZickusA basic patch that re-arranges some of the c2c code and adds a couple of small features to lay the ground work for the rest of the patch series.
Changes include:
o reworking the report path
o replace preprocess_sample with simpler calls
o rework raw output to handle separators
o remove phys id gunk
o add some generic options
There isn't much meat in this patch just a bunch of code movement and cleanups.
V2: refresh to latest upstream changes
--- tools/perf/builtin-c2c.c | 125 ++++++++++++++++++++++++++++++++++------------- 1 file changed, 92 insertions(+), 33 deletions(-)
tools/perf/builtin-c2c.c +92 -33
@@ -5,6 +5,7 @@#include "util/parse-options.h" #include "util/session.h" #include "util/tool.h" +#include "util/debug.h" #include <linux/compiler.h> #include <linux/kernel.h>
@@ -105,34 +106,55 @@ static int perf_c2c__fprintf_header(FILE *fp) }static int perf_sample__fprintf (struct perf_sample *sample,char tag, -const char *reason,struct addr_location *al,FILE *fp) +const char *reason,struct mem_info *mi,FILE *fp) {char data_src[61 ]; +const char *fmt, *sep; +struct map *map = mi->iaddr.map;perf_c2c__scnprintf_data_src (data_src,sizeof (data_src), sample->data_src); -return fprintf (fp,"%c %-16s %6d %6d %4d %#18" PRIx64" %#18" PRIx64" %#18" PRIx64" %6" PRIu64" %#10" PRIx64" %-60.60s %s:%s\n" , - tag, - reason ?:"valid record" , - sample->pid, - sample->tid, - sample->cpu, - sample->ip, - sample->addr, -0UL , - sample->weight, - sample->data_src, - data_src, - al->map ? (al->map->dso ? al->map->dso->long_name :"???" ) :"???" , - al->sym ? al->sym->name :"???" ); +if (symbol_conf.field_sep) { + fmt ="%c%s%s%s%d%s%d%s%d%s%#" PRIx64"%s%#" PRIx64"%s" +"%" PRIu64"%s%#" PRIx64"%s%s%s%s:%s\n" ; + sep = symbol_conf.field_sep; + }else { + fmt ="%c%s%-16s%s%6d%s%6d%s%4d%s%#18" PRIx64"%s%#18" PRIx64"%s" +"%6" PRIu64"%s%#10" PRIx64"%s%-60.60s%s%s:%s\n" ; + sep =" " ; + } + +return fprintf (fp, fmt, + tag, sep, + reason ?:"valid record" , sep, + sample->pid, sep, + sample->tid, sep, + sample->cpu, sep, + sample->ip, sep, + sample->addr, sep, + sample->weight, sep, + sample->data_src, sep, + data_src, sep, + map ? (map->dso ? map->dso->long_name :"???" ) :"???" , + mi->iaddr.sym ? mi->iaddr.sym->name :"???" ); }static int perf_c2c__process_load_store (struct perf_c2c *c2c, +struct addr_location *al,struct perf_sample *sample, -struct addr_location *al) +struct perf_evsel *evsel) { -if (c2c->raw_records) -perf_sample__fprintf (sample,' ' ,"raw input" , al, stdout); +struct mem_info *mi; + + mi =sample__resolve_mem (sample, al); +if (!mi) +return -ENOMEM ; + +if (c2c->raw_records) { +perf_sample__fprintf (sample,' ' ,"raw input" , mi, stdout); +free (mi); +return 0 ; + }return 0 ; }
@@ -143,8 +165,9 @@ static const struct perf_evsel_str_handler handlers[] = { };typedef int (*sample_handler )(struct perf_c2c *c2c, +struct addr_location *al,struct perf_sample *sample, -struct addr_location *al); +struct perf_evsel *evsel); static int perf_c2c__process_sample(struct perf_tool *tool, union perf_event *event,
@@ -153,20 +176,49 @@ static int perf_c2c__process_sample(struct perf_tool *tool, struct machine *machine) { struct perf_c2c *c2c = container_of(tool, struct perf_c2c, tool); -struct addr_location al; -int err =0 ; +u8 cpumode = event->header.misc &PERF_RECORD_MISC_CPUMODE_MASK ; +struct thread *thread; +sample_handler f; +int err =-1 ; +struct addr_location al = { + .machine = machine, + .cpu = sample->cpu, + .cpumode = cpumode, + }; -if (perf_event__preprocess_sample (event, machine, &al, sample) <0 ) { -pr_err ("problem processing %d event, skipping it.\n" , - event->header.type); -return -1 ; - } +if (evsel->handler ==NULL ) +return 0 ; + + thread =machine__find_thread (machine, sample->tid); +if (thread ==NULL ) + goto err; + + al.thread = thread; -if (evsel->handler !=NULL ) { -sample_handler f = evsel->handler; - err =f (c2c, sample, &al); + f = evsel->handler; + err =f (c2c, &al, sample, evsel); +if (err) + goto err; + +return 0 ; +err: +if (err >0 ) + err =0 ; +return err; +} + +static int perf_c2c__process_events (struct perf_session *session, +struct perf_c2c *c2c) +{ +int err =-1 ; + + err =perf_session__process_events (session, &c2c->tool); +if (err) { +pr_err ("Failed to process count events, error %d\n" , err); + goto err; } +err:return err; }
@@ -197,9 +249,7 @@ static int perf_c2c__read_events(struct perf_c2c *c2c) } } - err =perf_session__process_events (session, &c2c->tool); -if (err) -pr_err ("Failed to process events, error %d" , err); + err =perf_c2c__process_events (session, c2c); out:return err;
@@ -221,7 +271,6 @@ static int perf_c2c__record(int argc, const char **argv)const char **rec_argv;const char *const record_args[] = {"record" , -/* "--phys-addr", */ "-W" ,"-d" ,"-a" ,
@@ -254,6 +303,8 @@ int cmd_c2c(int argc, const char **argv, const char *prefix __maybe_unused)struct perf_c2c c2c = { .tool = { .sample = perf_c2c__process_sample, + .mmap2 = perf_event__process_mmap2, + .mmap = perf_event__process_mmap, .comm = perf_event__process_comm, .exit = perf_event__process_exit, .fork = perf_event__process_fork,
@@ -263,6 +314,14 @@ int cmd_c2c(int argc, const char **argv, const char *prefix __maybe_unused) };const struct option c2c_options[] = {OPT_BOOLEAN ('r' ,"raw_records" , &c2c.raw_records,"dump raw events" ), +OPT_INCR ('v' ,"verbose" , &verbose, +"be more verbose (show counter open errors, etc)" ), +OPT_STRING ('i' ,"input" , &input_name,"file" , +"the input file to process" ), +OPT_STRING ('x' ,"field-separator" , &symbol_conf.field_sep, +"separator" , +"separator for columns, no spaces will be added" +" between columns '.' is reserved." ),OPT_END () };const char *const c2c_usage[] = {
Signature · 1 line
1.7.11.7
PATCH v2 11/19 perf, c2c: Add in sort on physid
In reply to Don ZickusNow that the infrastructure is set, add in the support to use hist_entry to sort on physid.
--- tools/perf/builtin-c2c.c | 63 ++++++++++++++++++++++++++++++++++++++++++++++-- 1 file changed, 61 insertions(+), 2 deletions(-)
tools/perf/builtin-c2c.c +61 -2
@@ -6,6 +6,7 @@#include "util/session.h" #include "util/tool.h" #include "util/debug.h" +#include "util/annotate.h" #include <linux/compiler.h> #include <linux/kernel.h>
@@ -13,6 +14,7 @@struct perf_c2c {struct perf_tool tool;bool raw_records; +struct hists hists; };enum {OP ,LVL ,SNP ,LCK ,TLB };
@@ -144,7 +146,11 @@ static int perf_c2c__process_load_store(struct perf_c2c *c2c,struct perf_sample *sample, struct perf_evsel *evsel) { - struct mem_info *mi; + struct symbol *parent = NULL; +struct hist_entry *he; +struct mem_info *mi, *mx; +uint64_t cost; +int err; mi =sample__resolve_mem (sample, al);if (!mi)
@@ -156,7 +162,42 @@ static int perf_c2c__process_load_store(struct perf_c2c *c2c,return 0 ; } -return 0 ; + cost = sample->weight; +if (!cost) + cost =1 ; + +/* +* must pass period=weight in order to get the correct +* sorting from hists__collapse_resort() which is solely +* based on periods. We want sorting be done on nr_events * weight +* and this is indirectly achieved by passing period=weight here +* and the he_stat__add_period() function. +*/ + he =__hists__add_entry (&c2c->hists, al, parent,NULL , mi, + cost, cost,0 ); +if (!he) { + err = -ENOMEM ; + goto out_mem; + } + + err =hist_entry__inc_addr_samples (he, evsel->idx, al->addr); +if (err) + goto out; + + mx = he->mem_info; + err =addr_map_symbol__inc_samples (&mx->daddr, evsel->idx); +if (err) + goto out; + + c2c->hists.stats.total_period += cost; +hists__inc_nr_events (&c2c->hists,PERF_RECORD_SAMPLE ); +return err; + +out_mem: +/* implicitly freed by __hists__add_entry */ +free (mi); +out: +return err; }static const struct perf_evsel_str_handler handlers[] = {
@@ -255,10 +296,28 @@ out:return err; } +static int perf_c2c__init (struct perf_c2c *c2c) +{ + sort__mode =SORT_MODE__MEMORY ; + sort__wants_unique =1 ; + sort_order ="physid" ; + +if (setup_sorting () <0 ) { +pr_err ("can not setup sorting\n" ); +return -1 ; + } + +hists__init (&c2c->hists); + +return 0 ; +}static int perf_c2c__report (struct perf_c2c *c2c) {setup_pager (); +if (perf_c2c__init (c2c)) +return -1 ; +if (c2c->raw_records)perf_c2c__fprintf_header (stdout);
Signature · 1 line
1.7.11.7
Re: [PATCH 11/19] perf, c2c: Add in sort on physid
tools/perf/builtin-c2c.c static int perf_c2c__process_load_store(struct perf_c2c *c2c, · line 175
Quoted 8 lines · Don Zickus
tools/perf/builtin-c2c.c +8
-0+ +/* +* must pass period=weight in order to get the correct +* sorting from hists__collapse_resort() which is solely +* based on periods. We want sorting be done on nr_events * weight +* and this is indirectly achieved by passing period=weight here +* and the he_stat__add_period() function. +*/
sort.c really must be fixed to support multiple sort keys properly. Lots of other code has workarounds for this too (e.g. it is a bit problem for TSX abort weights)
I would suggest to fix that instead of hacking around it.
-Andi
Signature · 1 line
[email protected] -- Speaking for myself only
Re: [PATCH 11/19] perf, c2c: Add in sort on physid
Quoted 14 lines · Andi Kleen
tools/perf/builtin-c2c.c static int perf_c2c__process_load_store(struct perf_c2c *c2c, · line 175
Quoted 8 lines · Don Zickus
tools/perf/builtin-c2c.c +8
-0+ +/* +* must pass period=weight in order to get the correct +* sorting from hists__collapse_resort() which is solely +* based on periods. We want sorting be done on nr_events * weight +* and this is indirectly achieved by passing period=weight here +* and the he_stat__add_period() function. +*/
sort.creally must be fixed to support multiple sort keys properly. Lots of other code has workarounds for this too (e.g. it is a bit problem for TSX abort weights)
Quoted 1 line
I would suggest to fix that instead of hacking around it.
I don't think I understand the problem enough to know what to fix. I just copied this piece of code from builtin-report.c and things seemed to work.
Mind giving me some details and I can look at fixing it. :-)
Cheers,
Don
Re: [PATCH 11/19] perf, c2c: Add in sort on physid
Quoted 4 lines
I don't think I understand the problem enough to know what to fix. I just copied this piece of code from
builtin-report.cand things seemed to work.Mind giving me some details and I can look at fixing it. :-)
sort.c even though has all these sort keys only sorts by period. It should instead sort by all the specified keys in order instead.
Namhyung looked at it at some point.
-Andi
Signature · 1 line
[email protected] -- Speaking for myself only.
Namhyung Kim
maintainerRe: [PATCH 11/19] perf, c2c: Add in sort on physid
Hi Andi,
Quoted 9 lines · Andi Kleen
Quoted 4 lines
I don't think I understand the problem enough to know what to fix. I just copied this piece of code from
builtin-report.cand things seemed to work.Mind giving me some details and I can look at fixing it. :-)
sort.ceven though has all these sort keys only sorts by period. It should instead sort by all the specified keys in order instead.Namhyung looked at it at some point.
Yes, I'm working on it now, but I only have a little time. Hopefully I can send a rfc version next week..
Thanks,
Namhyung
PATCH v2 12/19 perf, c2c: Add stats to track data source bits and cpu to node maps
In reply to Don ZickusThis patch adds a bunch of stats that will be used later in post-processing to determine where and with what frequency the HITMs are coming from.
Most of the stats are decoded from the data source response. Another piece of the stats is tracking which cpu the record came in on.
Credit to Dick Fowles for determining which bits are important and how to properly track them. Ported to perf by me.
V2: refresh with hist_entry
--- tools/perf/builtin-c2c.c | 184 +++++++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 184 insertions(+)
tools/perf/builtin-c2c.c +184 -0
@@ -5,20 +5,84 @@#include "util/parse-options.h" #include "util/session.h" #include "util/tool.h" +#include "util/stat.h" +#include "util/cpumap.h" #include "util/debug.h" #include "util/annotate.h" #include <linux/compiler.h> #include <linux/kernel.h> +#include <sched.h> + +typedef struct { +int locks;/* count of 'lock' transactions */ +int store;/* count of all stores in trace */ +int st_uncache;/* stores to uncacheable address */ +int st_noadrs;/* cacheable store with no address */ +int st_l1hit;/* count of stores that hit L1D */ +int st_l1miss;/* count of stores that miss L1D */ +int load;/* count of all loads in trace */ +int ld_excl;/* exclusive loads, rmt/lcl DRAM - snp none/miss */ +int ld_shared;/* shared loads, rmt/lcl DRAM - snp hit */ +int ld_uncache;/* loads to uncacheable address */ +int ld_io;/* loads to io address */ +int ld_miss;/* loads miss */ +int ld_noadrs;/* cacheable load with no address */ +int ld_fbhit;/* count of loads hitting Fill Buffer */ +int ld_l1hit;/* count of loads that hit L1D */ +int ld_l2hit;/* count of loads that hit L2D */ +int ld_llchit;/* count of loads that hit LLC */ +int lcl_hitm;/* count of loads with local HITM */ +int rmt_hitm;/* count of loads with remote HITM */ +int rmt_hit;/* count of loads with remote hit clean; */ +int lcl_dram;/* count of loads miss to local DRAM */ +int rmt_dram;/* count of loads miss to remote DRAM */ +int nomap;/* count of load/stores with no phys adrs */ +int noparse;/* count of unparsable data sources */ +}trinfo_t ; + +struct c2c_stats { +cpu_set_t cpuset; +int nr_entries; +u64 total_period; +trinfo_t t; +struct stats stats; +};struct perf_c2c {struct perf_tool tool;bool raw_records;struct hists hists; + +/* stats */ +struct c2c_stats stats; };enum {OP ,LVL ,SNP ,LCK ,TLB }; +#define RMT_RAM (PERF_MEM_LVL_REM_RAM1 | PERF_MEM_LVL_REM_RAM2) +#define RMT_LLC (PERF_MEM_LVL_REM_CCE1 | PERF_MEM_LVL_REM_CCE2) + +#define L1CACHE_HIT (a) (((a) & PERF_MEM_LVL_L1 ) && ((a) & PERF_MEM_LVL_HIT)) +#define FILLBUF_HIT (a) (((a) & PERF_MEM_LVL_LFB) && ((a) & PERF_MEM_LVL_HIT)) +#define L2CACHE_HIT (a) (((a) & PERF_MEM_LVL_L2 ) && ((a) & PERF_MEM_LVL_HIT)) +#define L3CACHE_HIT (a) (((a) & PERF_MEM_LVL_L3 ) && ((a) & PERF_MEM_LVL_HIT)) + +#define L1CACHE_MISS (a) (((a) & PERF_MEM_LVL_L1 ) && ((a) & PERF_MEM_LVL_MISS)) +#define L3CACHE_MISS (a) (((a) & PERF_MEM_LVL_L3 ) && ((a) & PERF_MEM_LVL_MISS)) + +#define LD_UNCACHED (a) (((a) & PERF_MEM_LVL_UNC) && ((a) & PERF_MEM_LVL_HIT)) +#define ST_UNCACHED (a) (((a) & PERF_MEM_LVL_UNC) && ((a) & PERF_MEM_LVL_HIT)) + +#define RMT_LLCHIT (a) (((a) & RMT_LLC) && ((a) & PERF_MEM_LVL_HIT)) +#define RMT_HIT (a,b) (((a) & RMT_LLC) && ((b) & PERF_MEM_SNOOP_HIT)) +#define RMT_HITM (a,b) (((a) & RMT_LLC) && ((b) & PERF_MEM_SNOOP_HITM)) +#define RMT_MEM (a) (((a) & RMT_RAM) && ((a) & PERF_MEM_LVL_HIT)) + +#define LCL_HIT (a,b) (L3CACHE_HIT(a) && ((b) & PERF_MEM_SNOOP_HIT)) +#define LCL_HITM (a,b) (L3CACHE_HIT(a) && ((b) & PERF_MEM_SNOOP_HITM)) +#define LCL_MEM (a) (((a) & PERF_MEM_LVL_LOC_RAM) && ((a) & PERF_MEM_LVL_HIT)) +static int perf_c2c__scnprintf_data_src (char *bf,size_t size,uint64_t val) {#define PREFIX "["
@@ -141,6 +205,106 @@ static int perf_sample__fprintf(struct perf_sample *sample, char tag, mi->iaddr.sym ? mi->iaddr.sym->name : "???"); } +static int c2c_decode_stats (struct c2c_stats *stats,struct hist_entry *entry) +{ +union perf_mem_data_src *data_src = &entry->mem_info->data_src; +u64 daddr = entry->mem_info->daddr.addr; +u64 weight = entry->stat.weight; +int err =0 ; + +u64 op = data_src->mem_op; +u64 lvl = data_src->mem_lvl; +u64 snoop = data_src->mem_snoop; +u64 lock = data_src->mem_lock; + +#define P (a,b) PERF_MEM_##a##_##b + + stats->nr_entries++; + stats->total_period += entry->stat.period; + +if (lock &P (LOCK ,LOCKED )) stats->t.locks++; + +if (op &P (OP ,LOAD )) { + stats->t.load++; + +if (!daddr) { + stats->t.ld_noadrs++; +return -1 ; + } + +if (lvl &P (LVL ,HIT )) { +if (lvl &P (LVL ,UNC )) stats->t.ld_uncache++; +if (lvl &P (LVL ,IO )) stats->t.ld_io++; +if (lvl &P (LVL ,LFB )) stats->t.ld_fbhit++; +if (lvl &P (LVL ,L1 )) stats->t.ld_l1hit++; +if (lvl &P (LVL ,L2 )) stats->t.ld_l2hit++; +if (lvl &P (LVL ,L3 )) { +if (snoop &P (SNOOP ,HITM )) + stats->t.lcl_hitm++; +else + stats->t.ld_llchit++; + } + +if (lvl &P (LVL ,LOC_RAM )) { + stats->t.lcl_dram++; +if (snoop &P (SNOOP ,HIT )) + stats->t.ld_shared++; +else + stats->t.ld_excl++; + } + +if ((lvl &P (LVL ,REM_RAM1 )) || + (lvl &P (LVL ,REM_RAM2 ))) { + stats->t.rmt_dram++; +if (snoop &P (SNOOP ,HIT )) + stats->t.ld_shared++; +else + stats->t.ld_excl++; + } + } + +if ((lvl &P (LVL ,REM_CCE1 )) || + (lvl &P (LVL ,REM_CCE2 ))) { +if (snoop &P (SNOOP ,HIT )) + stats->t.rmt_hit++; +else if (snoop &P (SNOOP ,HITM )) { + stats->t.rmt_hitm++; +update_stats (&stats->stats, weight); + } + } + +if ((lvl &P (LVL ,MISS ))) + stats->t.ld_miss++; + + }else if (op &P (OP ,STORE )) { +/* store */ + stats->t.store++; + +if (!daddr) { + stats->t.st_noadrs++; +return -1 ; + } + +if (lvl &P (LVL ,HIT )) { +if (lvl &P (LVL ,UNC )) stats->t.st_uncache++; +if (lvl &P (LVL ,L1 )) stats->t.st_l1hit++; + } +if (lvl &P (LVL ,MISS )) +if (lvl &P (LVL ,L1 )) stats->t.st_l1miss++; + }else { +/* unparsable data_src? */ + stats->t.noparse++; +return -1 ; + } + +if (!entry->mem_info->daddr.map || !entry->mem_info->iaddr.map) { + stats->t.nomap++; +return -1 ; + } + +return err; +} + static int perf_c2c__process_load_store(struct perf_c2c *c2c, struct addr_location *al, struct perf_sample *sample,
@@ -180,6 +344,14 @@ static int perf_c2c__process_load_store(struct perf_c2c *c2c, goto out_mem; } + err =c2c_decode_stats (&c2c->stats, he); +if (err <0 ) { + err =0 ; +rb_erase (&he->rb_node_in, c2c->hists.entries_in); +free (he); + goto out; + } + err =hist_entry__inc_addr_samples (he, evsel->idx, al->addr);if (err) goto out;
@@ -279,6 +451,9 @@ static int perf_c2c__read_events(struct perf_c2c *c2c) goto out; } +if (symbol__init () <0 ) + goto out_delete; +/* setup the evsel handlers for each event type */ evlist__for_each (session->evlist, evsel) {const char *name =perf_evsel__name (evsel);
@@ -292,12 +467,20 @@ static int perf_c2c__read_events(struct perf_c2c *c2c) err =perf_c2c__process_events (session, c2c); +out_delete: +perf_session__delete (session); out:return err; }static int perf_c2c__init (struct perf_c2c *c2c) { +/* setup cpu map */ +if (cpu_map__setup_cpunode_map () <0 ) { +pr_err ("can not setup cpu map\n" ); +return -1 ; + } + sort__mode =SORT_MODE__MEMORY ; sort__wants_unique =1 ; sort_order ="physid" ;
@@ -308,6 +491,7 @@ static int perf_c2c__init(struct perf_c2c *c2c) }hists__init (&c2c->hists); +CPU_ZERO (&c2c->stats.cpuset);return 0 ; }
Signature · 1 line
1.7.11.7
PATCH v2 13/19 perf, c2c: Sort based on hottest cache line
In reply to Don ZickusNow that we have all the events sort on a unique address, we can walk the rbtree sequential and count up all the HITMs for each cacheline fairly easily.
Once we encounter a new event on a different cacheline, process the previous cacheline. That includes determining if any HITMs were present on that cacheline and if so, add it to another rbtree sorted on the number of HITMs.
This second rbtree sorted on number of HITMs will be the interesting data we want to report and will be displayed in a follow up patch.
For now, organize the data properly.
V2: re-work using hist_entries
--- tools/perf/builtin-c2c.c | 201 +++++++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 201 insertions(+)
tools/perf/builtin-c2c.c +201 -0
@@ -58,6 +58,25 @@ struct perf_c2c {struct c2c_stats stats; }; +#define CACHE_LINESIZE 64 +#define CLINE_OFFSET_MSK (CACHE_LINESIZE - 1) +#define CLADRS (a) ((a) & ~(CLINE_OFFSET_MSK)) +#define CLOFFSET (a) (int)((a) & (CLINE_OFFSET_MSK)) + +struct c2c_hit { +struct rb_node rb_node; +struct rb_root tree; +struct list_head list; +u64 cacheline; +int color; +struct c2c_stats stats; +pid_t pid; +pid_t tid; +u64 daddr; +u64 iaddr; +struct mem_info *mi; +}; +enum {OP ,LVL ,SNP ,LCK ,TLB };#define RMT_RAM (PERF_MEM_LVL_REM_RAM1 | PERF_MEM_LVL_REM_RAM2)
@@ -153,6 +172,44 @@ static int perf_c2c__scnprintf_data_src(char *bf, size_t size, uint64_t val)return printed; } +static int c2c_hitm__add_to_list (struct rb_root *root,struct c2c_hit *h) +{ +struct rb_node **p; +struct rb_node *parent =NULL ; +struct c2c_hit *he; +int64_t cmp; +u64 l_hitms, r_hitms; + + p = &root->rb_node; + +while (*p !=NULL ) { + parent = *p; + he = rb_entry(parent, struct c2c_hit, rb_node); + +/* sort on remote hitms first */ + l_hitms = he->stats.t.rmt_hitm; + r_hitms = h->stats.t.rmt_hitm; + cmp = r_hitms - l_hitms; + +if (!cmp) { +/* sort on local hitms */ + l_hitms = he->stats.t.lcl_hitm; + r_hitms = h->stats.t.lcl_hitm; + cmp = r_hitms - l_hitms; + } + +if (cmp >0 ) + p = &(*p)->rb_left; +else + p = &(*p)->rb_right; + } + +rb_link_node (&h->rb_node, parent, p); +rb_insert_color (&h->rb_node, root); + +return 0 ; +} +static int perf_c2c__fprintf_header (FILE *fp) { int printed = fprintf(fp, "%c %-16s %6s %6s %4s %18s %18s %18s %6s %-10s %-60s %s\n",
@@ -305,6 +362,50 @@ static int c2c_decode_stats(struct c2c_stats *stats, struct hist_entry *entry)return err; } +static struct c2c_hit *c2c_hit__new (u64 cacheline,struct hist_entry *entry) +{ +struct c2c_hit *h =zalloc (sizeof (struct c2c_hit )); + +if (!h) { +pr_err ("Could not allocate c2c_hit memory\n" ); +return NULL ; + } + +CPU_ZERO (&h->stats.cpuset); +INIT_LIST_HEAD (&h->list); +init_stats (&h->stats.stats); + h->tree =RB_ROOT ; + h->cacheline = cacheline; + h->pid = entry->thread->pid_; + h->tid = entry->thread->tid; + +/* use original addresses here, not adjusted al_addr */ + h->iaddr = entry->mem_info->iaddr.addr; + h->daddr = entry->mem_info->daddr.addr; + + h->mi = entry->mem_info; +return h; +} + +static void c2c_hit__update_strings (struct c2c_hit *h, +struct hist_entry *n) +{ +if (h->pid != n->thread->pid_) + h->pid =-1 ; + +if (h->tid != n->thread->tid) + h->tid =-1 ; + +/* use original addresses here, not adjusted al_addr */ +if (h->iaddr != n->mem_info->iaddr.addr) + h->iaddr =-1 ; + +if (CLADRS (h->daddr) !=CLADRS (n->mem_info->daddr.addr)) + h->daddr =-1 ; + +CPU_SET (n->cpu, &h->stats.cpuset); +} + static int perf_c2c__process_load_store(struct perf_c2c *c2c, struct addr_location *al, struct perf_sample *sample,
@@ -420,6 +521,104 @@ err:return err; } +#define HAS_HITMS (h) (h->stats.t.lcl_hitm || h->stats.t.rmt_hitm) + +static void c2c_hit__update_stats (struct c2c_stats *new, +struct c2c_stats *old) +{ + new->t.load += old->t.load; + new->t.ld_fbhit += old->t.ld_fbhit; + new->t.ld_l1hit += old->t.ld_l1hit; + new->t.ld_l2hit += old->t.ld_l2hit; + new->t.ld_llchit += old->t.ld_llchit; + new->t.locks += old->t.locks; + new->t.lcl_dram += old->t.lcl_dram; + new->t.rmt_dram += old->t.rmt_dram; + new->t.lcl_hitm += old->t.lcl_hitm; + new->t.rmt_hitm += old->t.rmt_hitm; + new->t.rmt_hit += old->t.rmt_hit; + new->t.store += old->t.store; + new->t.st_l1hit += old->t.st_l1hit; + + new->total_period += old->total_period; +} + +static inline int valid_hitm_or_store (union perf_mem_data_src *dsrc) +{ +return ((dsrc->mem_snoop &P (SNOOP ,HITM )) || + (dsrc->mem_op &P (OP ,STORE ))); +} + +static void c2c_analyze_hitms (struct perf_c2c *c2c) +{ + +struct rb_node *next =rb_first (c2c->hists.entries_in); +struct hist_entry *he; +struct c2c_hit *h =NULL ; +struct c2c_stats hitm_stats; +struct rb_root hitm_tree =RB_ROOT ; +int shared_clines =0 ; +u64 cl =0 ; + +memset (&hitm_stats,0 ,sizeof (struct c2c_stats )); + +/* find HITMs */ +while (next) { + he = rb_entry(next, struct hist_entry, rb_node_in); + next =rb_next (&he->rb_node_in); + + cl = he->mem_info->daddr.al_addr; + +/* switch cache line objects */ +/* 'color' forces a boundary change based on the original sort */ +if (!h || !he->color || (CLADRS (cl) != h->cacheline)) { +if (h &&HAS_HITMS (h)) { +c2c_hit__update_stats (&hitm_stats, &h->stats); + +/* sort based on hottest cacheline */ +c2c_hitm__add_to_list (&hitm_tree, h); + shared_clines++; + }else { +/* stores-only are un-interesting */ +free (h); + } + h =c2c_hit__new (CLADRS (cl), he); +if (!h) + goto cleanup; + } + + +c2c_decode_stats (&h->stats, he); + +/* filter out non-hitms as un-interesting noise */ +if (valid_hitm_or_store (&he->mem_info->data_src)) { +/* save the entry for later processing */ +list_add_tail (&he->pairs.node, &h->list); + +c2c_hit__update_strings (h, he); + } + } + +/* last chunk */ +if (HAS_HITMS (h)) { +c2c_hit__update_stats (&hitm_stats, &h->stats); +c2c_hitm__add_to_list (&hitm_tree, h); + shared_clines++; + }else +free (h); + +cleanup: + next =rb_first (&hitm_tree); +while (next) { + h = rb_entry(next, struct c2c_hit, rb_node); + next =rb_next (&h->rb_node); +rb_erase (&h->rb_node, &hitm_tree); + +free (h); + } +return ; +} +static int perf_c2c__process_events (struct perf_session *session,struct perf_c2c *c2c) {
@@ -431,6 +630,8 @@ static int perf_c2c__process_events(struct perf_session *session, goto err; } +c2c_analyze_hitms (c2c); + err:return err; }
Signature · 1 line
1.7.11.7
PATCH v2 14/19 perf, c2c: Display cacheline HITM analysis to stdout
In reply to Don ZickusThis patch mainly focuses on processing and displaying the collected HITMs to stdout. Most of it is just printing data in a pretty way.
There is one trick used when walking the cacheline. When we get this far we have two rbtrees. One rbtree holds every record sorted on a unique id (using the mmap2 decoder), the other rbtree holds every cacheline with at least one HITM sorted on number of HITMs.
To display the output, the tool walks the second rbtree to display the hottest cachelines. Inside each hot cacheline, the tool displays the offsets and the loads/stores it generates. To determine the cacheline offset, it uses linked list inside the cacheline elment to walk the first rbtree elements for that particular cacheline.
The first rbtree elements are already sorted correctly in offset order, so processing the offsets is fairly trivial and is done sequentially.
This is why you will see two while loops in the print_c2c_hitm_report(), the outer one walks the cachelines, the inner one walks the offsets.
A knob has been added to display node information, which is useful to see which cpus are involved in the contention and their nodes.
Another knob has been added to change the coalescing levels. You can coalesce the output based on pid, tid, ip, and/or symbol.
Original output and statistics done by Dick Fowles, backported by me.
Sample output:
=================================================
Global Shared Cache Line Event Information
=================================================
Total Shared Cache Lines : 1327 Load HITs on shared lines : 167131 Fill Buffer Hits on shared lines : 43469 L1D hits on shared lines : 50497 L2D hits on shared lines : 960 LLC hits on shared lines : 38467 Locked Access on shared lines : 100032 Store HITs on shared lines : 118659 Store L1D hits on shared lines : 113783 Total Merged records : 160807
===========================================================================================================================================================
Shared Cache Line Distribution Pareto
---- All ---- -- Shared -- ---- HITM ---- Load Inst Execute Latency
Data Misses Data Misses Remote Local -- Store Refs --
---- cycles ---- cpu
Num %dist %cumm %dist %cumm LLCmiss LLChit L1 hit L1 Miss Data Address Pid Tid Inst Address median mean CV cnt
==========================================================================================================================================================
-----------------------------------------------------------------------------------------------
0 17.0% 17.0% 23.3% 23.3% 6238 3288 28098 813 0xffff881fa55b0140 ***
-----------------------------------------------------------------------------------------------
0.0% 0.0% 0.0% 0.0% 0x00 375 375 0xffffffffa018ff5b n/a n/a n/a 1
0.0% 0.0% 0.0% 0.0% 0x08 18156 18156 0xffffffffa018b7f9 -1 384 0.0% 1
0.2% 0.0% 0.0% 0.0% 0x10 18156 18156 0xffffffff811ca1aa -1 387 10.7% 7
0.0% 0.0% 23.2% 0.0% 0x18 18156 18156 0xffffffff815c1615 -1 684 0.0% 50-----------------------------------------------------------------------------------------------
1 5.3% 22.3% 7.3% 30.6% 1944 1143 7916 0 0xffff881fba47f000 18156
-----------------------------------------------------------------------------------------------
100.0% 100.0% 0.0% 0.0% 0x00 18156 18156 0xffffffffa01b410e -1 401 13.5% 50
0.0% 0.0% 10.1% 0.0% 0x28 18156 18156 0xffffffffa0167409 n/a n/a n/a 50
0.0% 0.0% 89.9% 0.0% 0x28 18156 18156 0xffffffff815c4be9 n/a n/a n/a 50--- tools/perf/builtin-c2c.c | 519 +++++++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 519 insertions(+)
tools/perf/builtin-c2c.c +519 -0
@@ -58,10 +58,13 @@ struct perf_c2c {struct c2c_stats stats; }; +#define DISPLAY_LINE_LIMIT 0.0015#define CACHE_LINESIZE 64#define CLINE_OFFSET_MSK (CACHE_LINESIZE - 1)#define CLADRS (a) ((a) & ~(CLINE_OFFSET_MSK))#define CLOFFSET (a) (int)((a) & (CLINE_OFFSET_MSK)) +#define MAXTITLE_SZ 400 +#define MAXLBL_SZ 256struct c2c_hit {struct rb_node rb_node;
@@ -102,6 +105,11 @@ enum { OP, LVL, SNP, LCK, TLB };#define LCL_HITM (a,b) (L3CACHE_HIT(a) && ((b) & PERF_MEM_SNOOP_HITM))#define LCL_MEM (a) (((a) & PERF_MEM_LVL_LOC_RAM) && ((a) & PERF_MEM_LVL_HIT)) +enum {LVL0 ,LVL1 ,LVL2 ,LVL3 ,LVL4 ,MAX_LVL }; +static int cloffset =LVL1 ; +static int node_info =0 ; +static int coalesce_level =LVL1 ; +static int perf_c2c__scnprintf_data_src (char *bf,size_t size,uint64_t val) {#define PREFIX "["
@@ -406,6 +414,80 @@ static void c2c_hit__update_strings(struct c2c_hit *h,CPU_SET (n->cpu, &h->stats.cpuset); } +static inline bool matching_coalescing (struct c2c_hit *h, +struct hist_entry *e) +{ +bool value = false; +struct mem_info *mi = e->mem_info; + +if (coalesce_level >MAX_LVL ) +printf ("DON: bad coalesce level %d\n" , coalesce_level); + +if (e->cpumode !=PERF_RECORD_MISC_KERNEL ) { + +switch (coalesce_level) { + +case LVL0 : +case LVL1 : + value = ((h->daddr == mi->daddr.addr) && + (h->pid == e->thread->pid_) && + (h->tid == e->thread->tid) && + (h->iaddr == mi->iaddr.addr)); +break ; + +case LVL2 : + value = ((h->daddr == mi->daddr.addr) && + (h->pid == e->thread->pid_) && + (h->iaddr == mi->iaddr.addr)); +break ; + +case LVL3 : + value = ((h->daddr == mi->daddr.addr) && + (h->iaddr == mi->iaddr.addr)); +break ; + +case LVL4 : + value = ((h->daddr == mi->daddr.addr) && + (h->mi->iaddr.sym == mi->iaddr.sym)); +break ; + +default : +break ; + + } + + }else { + +switch (coalesce_level) { + +case LVL0 : + value = ((h->daddr == mi->daddr.addr) && + (h->pid == e->thread->pid_) && + (h->tid == e->thread->tid) && + (h->iaddr == mi->iaddr.addr)); +break ; + +case LVL1 : +case LVL2 : +case LVL3 : + value = ((h->daddr == mi->daddr.addr) && + (h->iaddr == mi->iaddr.addr)); +break ; + +case LVL4 : + value = ((h->daddr == mi->daddr.addr) && + (h->mi->iaddr.sym == mi->iaddr.sym)); +break ; + +default : +break ; + + } + } + +return value; +} + static int perf_c2c__process_load_store(struct perf_c2c *c2c, struct addr_location *al, struct perf_sample *sample,
@@ -543,12 +625,442 @@ static void c2c_hit__update_stats(struct c2c_stats *new, new->total_period += old->total_period; } +static void print_hitm_cacheline_header (void ) +{ +#define SHARING_REPORT_TITLE "Shared Cache Line Distribution Pareto" +#define PARTICIPANTS1 "Node{cpus %hitms %stores} Node{cpus %hitms %stores} ..." +#define PARTICIPANTS2 "Node{cpu list}; Node{cpu list}; Node{cpu list}; ..." + +int i; +const char *docptr; +static char delimit[MAXTITLE_SZ ]; +static char title2[MAXTITLE_SZ ]; +int pad; + + docptr =" " ; +if (node_info ==1 ) + docptr =PARTICIPANTS1 ; +if (node_info ==2 ) + docptr =PARTICIPANTS2 ; + +sprintf (title2,"%4s %6s %6s %6s %6s %8s %8s %8s %8s %18s %6s %6s %18s %8s %8s %8s %6s %-30s %-20s %s" , +"Num" , +"%dist" , +"%cumm" , +"%dist" , +"%cumm" , +"LLCmiss" , +"LLChit" , +"L1 hit" , +"L1 Miss" , +"Data Address" , +"Pid" , +"Tid" , +"Inst Address" , +"median" , +"mean" , +"CV " , +"cnt" , +"Symbol" , +"Object" , + docptr); + +for (i =0 ; i < (int )strlen (title2); i++)strcat (delimit,"=" ); + + +printf ("\n\n" ); +printf ("%s\n" , delimit); +printf ("\n" ); + + pad = (strlen (title2)/2 ) - (strlen (SHARING_REPORT_TITLE )/2 ); +for (i =0 ; i < pad; i++)printf (" " ); +printf ("%s\n" ,SHARING_REPORT_TITLE ); + +printf ("\n" ); +printf ("%4s %13s %13s %17s %8s %8s %18s %6s %6s %18s %26s %6s %30s %20s %s\n" , +" " , +"---- All ----" , +"-- Shared --" , +"---- HITM ----" , +" " , +" " , +" " , +" " , +" " , +" " , +"Load Inst Execute Latency" , +" " , +" " , +" " , + node_info ?"Shared Data Participants" :" " ); + + +printf ("%4s %13s %13s %8s %8s %17s %18s %6s %6s %17s %18s\n" , +" " , +" Data Misses" , +" Data Misses" , +"Remote" , +"Local" , +"-- Store Refs --" , +" " , +" " , +" " , +" " , +" " ); + +printf ("%4s %13s %13s %8s %8s %8s %8s %18s %6s %6s %17s %18s %8s %6s\n" , +" " , +" " , +" " , +" " , +" " , +" " , +" " , +" " , +" " , +" " , +" " , +"---- cycles ----" , +" " , +"cpu" ); + +printf ("%s\n" , title2); +printf ("%s\n" , delimit); +} + +static void print_hitm_cacheline (struct c2c_hit *h, +int record, +double tot_cumm, +double ld_cumm, +double tot_dist, +double ld_dist) +{ +char pidstr[7 ]; +char addrstr[20 ]; +static char summary[MAXLBL_SZ ]; +int j; + +if (h->pid >0 ) +sprintf (pidstr,"%6d" , h->pid); +else +sprintf (pidstr,"***" ); +/* +* It is possible to have none distinct virtual addresses +* pointing to a distinct SYstem V shared memory region. +* if there are multple virtual addresses the address +* field will be astericks. It would be possible to subsitute +* the physical address but this count be confusing as some +* times the field is a virtual address while or times it +* may be a physical address which may lead to confusion. +*/ +if (h->daddr != ~0UL ) +sprintf (addrstr,"%#18lx" ,CLADRS (h->daddr)); +else +sprintf (addrstr,"****************" ); + + +sprintf (summary,"%4d %5.1f%% %5.1f%% %5.1f%% %5.1f%% %8d %8d %8d %8d %18s %6s\n" , + record, + tot_dist *100. , + tot_cumm *100. , + ld_dist *100. , + ld_cumm *100. , + h->stats.t.rmt_hitm, + h->stats.t.lcl_hitm, + h->stats.t.st_l1hit, + h->stats.t.st_l1miss, + addrstr, + pidstr); + +for (j =0 ; j < (int )strlen (summary); j++)printf ("-" ); +printf ("\n" ); +printf ("%s" , summary); +for (j =0 ; j < (int )strlen (summary); j++)printf ("-" ); +printf ("\n" ); +} + +static void print_socket_stats_str (struct c2c_hit *clo, +struct c2c_stats *node_stats) +{ +int i, j; + +if (!node_stats) +return ; + +for (i =0 ; i < max_node_num; i++) { +struct c2c_stats *stats = &node_stats[i]; +int num =CPU_COUNT (&stats->cpuset); + +if (!num) { +/* pad align socket info */ +for (j =0 ; j <21 ; j++) +printf (" " ); +continue ; + } + +printf ("%2d{%2d " , i, num); + +if (clo->stats.t.rmt_hitm >0 ) +printf ("%5.1f%% " ,100. * ((double )stats->t.rmt_hitm / (double ) clo->stats.t.rmt_hitm)); +else +printf ("%6s " ,"n/a" ); + +if (clo->stats.t.store >0 ) +printf ("%5.1f%%} " ,100. * ((double )stats->t.store / (double )clo->stats.t.store)); +else +printf ("%6s} " ,"n/a" ); + } +} + +static void print_socket_shared_str (struct c2c_stats *node_stats) +{ +int i, j; + +if (!node_stats) +return ; + +for (i =0 ; i < max_node_num; i++) { +struct c2c_stats *stats = &node_stats[i]; +int num =CPU_COUNT (&stats->cpuset); +int start =-1 ; +bool first = true; + +if (!num) +continue ; + +printf ("%d{" , i); + +for (j =0 ; j < max_cpu_num; j++) { +if (!CPU_ISSET (j, &stats->cpuset)) { +if (start !=-1 ) { +if ((j-1 ) - start) +/* print the range */ +printf ("%s%d-%d" , (first ?"" :"," ), start, j-1 ); +else +/* standalone */ +printf ("%s%d" , (first ?"" :"," ) , start); + start =-1 ; + first = false; + } +continue ; + } + +if (start ==-1 ) + start = j; + } +/* last chunk */ +if (start !=-1 ) { +if ((j-1 ) - start) +/* print the range */ +printf ("%s%d-%d" , (first ?"" :"," ), start, j-1 ); +else +/* standalone */ +printf ("%s%d" , (first ?"" :"," ) , start); + } + +printf ("}; " ); + } +} + +static void print_hitm_cacheline_offset (struct c2c_hit *clo, +struct c2c_hit *h, +struct c2c_stats *node_stats) +{ +#define SHORT_STR_LEN 7 +#define LONG_STR_LEN 30 + +char pidstr[SHORT_STR_LEN ]; +char tidstr[SHORT_STR_LEN ]; +char addrstr[LONG_STR_LEN ]; +char latstr[LONG_STR_LEN ]; +char objptr[LONG_STR_LEN ]; +char symptr[LONG_STR_LEN ]; +struct c2c_stats *stats = &clo->stats; +struct addr_map_symbol *ams; + + ams = &clo->mi->iaddr; + +if (clo->pid >=0 ) +snprintf (pidstr,SHORT_STR_LEN ,"%6d" , clo->pid); +else +sprintf (pidstr,"***" ); + +if (clo->tid >=0 ) +snprintf (tidstr,SHORT_STR_LEN ,"%6d" , clo->tid); +else +sprintf (tidstr,"***" ); + +if (clo->iaddr != ~0UL ) +snprintf (addrstr,LONG_STR_LEN ,"%#18lx" , clo->iaddr); +else +sprintf (addrstr,"****************" ); +snprintf (objptr,LONG_STR_LEN ,"%-18s" , ams->map->dso->short_name); +snprintf (symptr,LONG_STR_LEN ,"%-18s" , (ams->sym ? ams->sym->name :"?????" )); + +if (stats->t.rmt_hitm >0 ) { +double mean =avg_stats (&stats->stats); +double std =stddev_stats (&stats->stats); + +sprintf (latstr,"%8.0f %8.0f %7.1f%%" , +-1.0 ,/* FIXME */ + mean, +rel_stddev_stats (std, mean)); + }else { +sprintf (latstr,"%8s %8s %8s" , +"n/a" , +"n/a" , +"n/a" ); + + } + +/* +* implicit assumption that we are not coalescing over IPs +*/ +printf ("%4s %6s %6s %6s %6s %7.1f%% %7.1f%% %7.1f%% %7.1f%% %14s0x%02lx %6s %6s %18s %8s %6d %-30s %-20s" , +" " , +" " , +" " , +" " , +" " , + (stats->t.rmt_hitm >0 ) ? (100. * ((double )stats->t.rmt_hitm / (double )h->stats.t.rmt_hitm)) :0.0 , + (stats->t.lcl_hitm >0 ) ? (100. * ((double )stats->t.lcl_hitm / (double )h->stats.t.lcl_hitm)) :0.0 , + (stats->t.st_l1hit >0 ) ? (100. * ((double )stats->t.st_l1hit / (double )h->stats.t.st_l1hit)) :0.0 , + (stats->t.st_l1miss >0 ) ? (100. * ((double )stats->t.st_l1miss / (double )h->stats.t.st_l1miss)) :0.0 , +" " , + (cloffset ==LVL2 ) ? (clo->daddr &0xff ) :CLOFFSET (clo->daddr), + pidstr, + tidstr, + addrstr, + latstr, +CPU_COUNT (&clo->stats.cpuset), + symptr, + objptr); + +if (node_info ==0 ) +printf (" " ); +else if (node_info ==1 ) +print_socket_stats_str (clo, node_stats); +else if (node_info ==2 ) +print_socket_shared_str (node_stats); + +printf ("\n" ); +} + +static void print_c2c_hitm_report (struct rb_root *hitm_tree, +struct c2c_stats *hitm_stats__maybe_unused , +struct c2c_stats *c2c_stats) +{ +struct rb_node *next =rb_first (hitm_tree); +struct c2c_hit *h, *clo =NULL ; +u64 addr; +double tot_dist, tot_cumm; +double ld_dist, ld_cumm; +int llc_misses; +int record =0 ; +struct c2c_stats *node_stats =NULL ; + +if (node_info) { + node_stats =zalloc (sizeof (struct c2c_stats ) *cpu_map__max_node ()); +if (!node_stats) { +printf ("Can not allocate stats for node output\n" ); +return ; + } + } + +print_hitm_cacheline_header (); + + llc_misses = c2c_stats->t.lcl_dram + + c2c_stats->t.rmt_dram + + c2c_stats->t.rmt_hit + + c2c_stats->t.rmt_hitm; + +/* +* generate distinct cache line report +*/ + tot_cumm =0.0 ; + ld_cumm =0.0 ; + +while (next) { +struct hist_entry *entry; + + h = rb_entry(next, struct c2c_hit, rb_node); + next =rb_next (&h->rb_node); + + tot_dist = ((double )h->stats.t.rmt_hitm / llc_misses); + tot_cumm += tot_dist; + + ld_dist = ((double )h->stats.t.rmt_hitm / c2c_stats->t.rmt_hitm); + ld_cumm += ld_dist; + +/* +* don't display lines with insignificant sharing contribution +*/ +if (ld_dist <DISPLAY_LINE_LIMIT ) +break ; + +print_hitm_cacheline (h, record, tot_cumm, ld_cumm, tot_dist, ld_dist); + +list_for_each_entry (entry, &h->list, pairs.node) { + +if (!clo || !matching_coalescing (clo, entry)) { +if (clo) +print_hitm_cacheline_offset (clo, h, node_stats); + +free (clo); + addr = entry->mem_info->iaddr.al_addr; + clo =c2c_hit__new (addr, entry); +if (node_info) +memset (node_stats,0 ,sizeof (struct c2c_stats ) *cpu_map__max_node ()); + } +c2c_decode_stats (&clo->stats, entry); +c2c_hit__update_strings (clo, entry); + +if (node_info) { +int node =cpu_map__get_node (entry->cpu); +c2c_decode_stats (&node_stats[node], entry); +CPU_SET (entry->cpu, &(node_stats[node].cpuset)); + } + + } +if (clo) { +print_hitm_cacheline_offset (clo, h, node_stats); +free (clo); + clo =NULL ; + } + +if (node_info) +memset (node_stats,0 ,sizeof (struct c2c_stats ) *cpu_map__max_node ()); + +printf ("\n" ); + record++; + } +} +static inline int valid_hitm_or_store (union perf_mem_data_src *dsrc) {return ((dsrc->mem_snoop &P (SNOOP ,HITM )) || (dsrc->mem_op &P (OP ,STORE ))); } +static void print_shared_cacheline_info (struct c2c_stats *stats,int cline_cnt) +{ +int hitm_cnt = stats->t.lcl_hitm + stats->t.rmt_hitm; + +printf ("=================================================\n" ); +printf (" Global Shared Cache Line Event Information \n" ); +printf ("=================================================\n" ); +printf (" Total Shared Cache Lines : %10d\n" , cline_cnt); +printf (" Load HITs on shared lines : %10d\n" , stats->t.load); +printf (" Fill Buffer Hits on shared lines : %10d\n" , stats->t.ld_fbhit); +printf (" L1D hits on shared lines : %10d\n" , stats->t.ld_l1hit); +printf (" L2D hits on shared lines : %10d\n" , stats->t.ld_l2hit); +printf (" LLC hits on shared lines : %10d\n" , stats->t.ld_llchit + stats->t.lcl_hitm); +printf (" Locked Access on shared lines : %10d\n" , stats->t.locks); +printf (" Store HITs on shared lines : %10d\n" , stats->t.store); +printf (" Store L1D hits on shared lines : %10d\n" , stats->t.st_l1hit); +printf (" Total Merged records : %10d\n" , hitm_cnt + stats->t.store); +} +static void c2c_analyze_hitms (struct perf_c2c *c2c) {
@@ -607,6 +1119,9 @@ static void c2c_analyze_hitms(struct perf_c2c *c2c) }else free (h); +print_shared_cacheline_info (&hitm_stats, shared_clines); +print_c2c_hitm_report (&hitm_tree, &hitm_stats, &c2c->stats); + cleanup: next =rb_first (&hitm_tree);while (next) {
@@ -758,6 +1273,10 @@ int cmd_c2c(int argc, const char **argv, const char *prefix __maybe_unused) };const struct option c2c_options[] = {OPT_BOOLEAN ('r' ,"raw_records" , &c2c.raw_records,"dump raw events" ), +OPT_INCR ('N' ,"node-info" , &node_info, +"show extra node info in report (repeat for more info)" ), +OPT_INTEGER ('c' ,"coalesce-level" , &coalesce_level, +"how much coalescing for tid, pid, and ip is done (repeat for more coalescing)" ),OPT_INCR ('v' ,"verbose" , &verbose,"be more verbose (show counter open errors, etc)" ), OPT_STRING('i', "input", &input_name, "file",
Signature · 1 line
1.7.11.7
PATCH v2 15/19 perf, c2c: Add callchain support
In reply to Don ZickusSeeing cacheline statistics is useful by itself. Seeing the callchain for these cache contentions saves time tracking things down.
This patch tries to add callchain support. I had to use the generic interface from a previous patch to output things to stdout easily.
Other than the displaying the results, collecting the callchain and merging it was fairly straightforward.
I used a lot of copying-n-pasting from other builtin tools to get the intial parameter setup correctly and the automatic reading of 'symbol_conf.use_callchain' from the data file.
Hopefully this is all correct. The amount of memory corruption (from the callchain dynamic array) seems to have dwindled done to nothing. :-)
V2: update to latest api
--- tools/perf/builtin-c2c.c | 153 ++++++++++++++++++++++++++++++++++++++++++++++- 1 file changed, 150 insertions(+), 3 deletions(-)
tools/perf/builtin-c2c.c +150 -3
@@ -52,6 +52,7 @@ struct c2c_stats {struct perf_c2c {struct perf_tool tool;bool raw_records; +bool call_graph;struct hists hists;/* stats */
@@ -78,6 +79,8 @@ struct c2c_hit {u64 daddr;u64 iaddr;struct mem_info *mi; + +struct callchain_root callchain[0 ];/* must be last member */ };enum {OP ,LVL ,SNP ,LCK ,TLB };
@@ -372,7 +375,8 @@ static int c2c_decode_stats(struct c2c_stats *stats, struct hist_entry *entry)static struct c2c_hit *c2c_hit__new (u64 cacheline,struct hist_entry *entry) { -struct c2c_hit *h =zalloc (sizeof (struct c2c_hit )); +size_t callchain_size = symbol_conf.use_callchain ?sizeof (struct callchain_root ) :0 ; +struct c2c_hit *h =zalloc (sizeof (struct c2c_hit ) + callchain_size);if (!h) {pr_err ("Could not allocate c2c_hit memory\n" );
@@ -386,6 +390,8 @@ static struct c2c_hit *c2c_hit__new(u64 cacheline, struct hist_entry *entry) h->cacheline = cacheline; h->pid = entry->thread->pid_; h->tid = entry->thread->tid; +if (symbol_conf.use_callchain) +callchain_init (h->callchain);/* use original addresses here, not adjusted al_addr */ h->iaddr = entry->mem_info->iaddr.addr;
@@ -509,6 +515,10 @@ static int perf_c2c__process_load_store(struct perf_c2c *c2c,return 0 ; } + err =sample__resolve_callchain (sample, &parent, evsel, al,PERF_MAX_STACK_DEPTH ); +if (err) +return err; + cost = sample->weight;if (!cost) cost =1 ;
@@ -544,8 +554,9 @@ static int perf_c2c__process_load_store(struct perf_c2c *c2c,if (err) goto out; - c2c->hists.stats.total_period += cost; -hists__inc_nr_events (&c2c->hists,PERF_RECORD_SAMPLE ); + c2c->hists.stats.total_period += cost; +hists__inc_nr_events (&c2c->hists,PERF_RECORD_SAMPLE ); + err =hist_entry__append_callchain (he, sample);return err; out_mem:
@@ -944,6 +955,13 @@ static void print_hitm_cacheline_offset(struct c2c_hit *clo,print_socket_shared_str (node_stats);printf ("\n" ); + +if (symbol_conf.use_callchain) { +generic_entry_callchain__fprintf (clo->callchain, + h->stats.total_period, + clo->stats.total_period, +23 , stdout); + } } static void print_c2c_hitm_report(struct rb_root *hitm_tree,
@@ -1020,6 +1038,12 @@ static void print_c2c_hitm_report(struct rb_root *hitm_tree,c2c_decode_stats (&node_stats[node], entry);CPU_SET (entry->cpu, &(node_stats[node].cpuset)); } +if (symbol_conf.use_callchain) { +callchain_cursor_reset (&callchain_cursor); +callchain_merge (&callchain_cursor, + clo->callchain, + entry->callchain); + } }if (clo) {
@@ -1151,6 +1175,30 @@ err:return err; } +static int perf_c2c__setup_sample_type (struct perf_c2c *c2c, +struct perf_session *session) +{ +u64 sample_type =perf_evlist__combined_sample_type (session->evlist); + +if (!(sample_type &PERF_SAMPLE_CALLCHAIN )) { +if (symbol_conf.use_callchain) { +printf ("Selected -g but no callchain data. Did " +"you call 'perf c2c record' without -g?\n" ); +return -1 ; + } + }else if (callchain_param.mode !=CHAIN_NONE && + !symbol_conf.use_callchain) { + symbol_conf.use_callchain = true; + c2c->call_graph = true; +if (callchain_register_param (&callchain_param) <0 ) { +printf ("Can't register callchain params.\n" ); +return -EINVAL ; + } + } + +return 0 ; +} +static int perf_c2c__read_events (struct perf_c2c *c2c) {int err =-1 ;
@@ -1170,6 +1218,9 @@ static int perf_c2c__read_events(struct perf_c2c *c2c)if (symbol__init () <0 ) goto out_delete; +if (perf_c2c__setup_sample_type (c2c, session) <0 ) + goto out_delete; +/* setup the evsel handlers for each event type */ evlist__for_each (session->evlist, evsel) {const char *name =perf_evsel__name (evsel);
@@ -1257,8 +1308,101 @@ static int perf_c2c__record(int argc, const char **argv)return cmd_record (i, rec_argv,NULL ); } +static int +opt_callchain_cb (const struct option *opt,const char *arg,int unset) +{ +struct perf_c2c *c2c = (struct perf_c2c *)opt->value; +char *tok, *tok2; +char *endptr; + +/* +* --no-call-graph +*/ +if (unset) { + c2c->call_graph = false; +return 0 ; + } + + symbol_conf.use_callchain = true; + c2c->call_graph = true; + +if (!arg) +return 0 ; + + tok =strtok ((char *)arg,"," ); +if (!tok) +return -1 ; + +/* get the output mode */ +if (!strncmp (tok,"graph" ,strlen (arg))) + callchain_param.mode =CHAIN_GRAPH_ABS ; + +else if (!strncmp (tok,"flat" ,strlen (arg))) + callchain_param.mode =CHAIN_FLAT ; + +else if (!strncmp (tok,"fractal" ,strlen (arg))) + callchain_param.mode =CHAIN_GRAPH_REL ; + +else if (!strncmp (tok,"none" ,strlen (arg))) { + callchain_param.mode =CHAIN_NONE ; + symbol_conf.use_callchain = false; + +return 0 ; + } + +else +return -1 ; + +/* get the min percentage */ + tok =strtok (NULL ,"," ); +if (!tok) + goto setup; + + callchain_param.min_percent =strtod (tok, &endptr); +if (tok == endptr) +return -1 ; + +/* get the print limit */ + tok2 =strtok (NULL ,"," ); +if (!tok2) + goto setup; + +if (tok2[0 ] !='c' ) { + callchain_param.print_limit =strtoul (tok2, &endptr,0 ); + tok2 =strtok (NULL ,"," ); +if (!tok2) + goto setup; + } + +/* get the call chain order */ +if (!strncmp (tok2,"caller" ,strlen ("caller" ))) + callchain_param.order =ORDER_CALLER ; +else if (!strncmp (tok2,"callee" ,strlen ("callee" ))) + callchain_param.order =ORDER_CALLEE ; +else +return -1 ; + +/* Get the sort key */ + tok2 =strtok (NULL ,"," ); +if (!tok2) + goto setup; +if (!strncmp (tok2,"function" ,strlen ("function" ))) + callchain_param.key =CCKEY_FUNCTION ; +else if (!strncmp (tok2,"address" ,strlen ("address" ))) + callchain_param.key =CCKEY_ADDRESS ; +else +return -1 ; +setup: +if (callchain_register_param (&callchain_param) <0 ) { +fprintf (stderr,"Can't register callchain params\n" ); +return -1 ; + } +return 0 ; +} +int cmd_c2c (int argc,const char **argv,const char *prefix__maybe_unused ) { +char callchain_default_opt[] ="fractal,0.05,callee" ;struct perf_c2c c2c = { .tool = { .sample = perf_c2c__process_sample,
@@ -1285,6 +1429,9 @@ int cmd_c2c(int argc, const char **argv, const char *prefix __maybe_unused) "separator", "separator for columns, no spaces will be added" " between columns '.' is reserved."), +OPT_CALLBACK_DEFAULT ('g' ,"call-graph" , &c2c,"output_type,min_percent[,print_limit],call_order" , +"Display callchains using output_type (graph, flat, fractal, or none) , min percent threshold, optional print limit, callchain order, key (function or address). " +"Default: fractal,0.5,callee,function" , &opt_callchain_cb, callchain_default_opt),OPT_END () }; const char * const c2c_usage[] = {
Signature · 1 line
1.7.11.7
Re: [PATCH 15/19] perf, c2c: Add callchain support
On Fri, Feb 28, 2014 at 12:43:04PM -0500, Don Zickus wrote:
SNIP
tools/perf/builtin-c2c.c static int perf_c2c__record(int argc, const char **argv) · line 1316
Quoted 7 lines
tools/perf/builtin-c2c.c +6
-0+static int +opt_callchain_cb (const struct option *opt,const char *arg,int unset) +{ +struct perf_c2c *c2c = (struct perf_c2c *)opt->value; +char *tok, *tok2; +char *endptr;
this still needs to be shared with report code, but I'm guessing V2 was about moving to hist_entry-ies
jirka
Re: [PATCH 15/19] perf, c2c: Add callchain support
Quoted 5 lines · Jiri Olsa
On Fri, Feb 28, 2014 at 12:43:04PM -0500, Don Zickus wrote:
SNIP
Quoted 1 line
Quoted 10 lines
tools/perf/builtin-c2c.c static int perf_c2c__record(int argc, const char **argv) · line 1316
Quoted 6 lines
tools/perf/builtin-c2c.c +6
-0+static int +opt_callchain_cb (const struct option *opt,const char *arg,int unset) +{ +struct perf_c2c *c2c = (struct perf_c2c *)opt->value; +char *tok, *tok2; +char *endptr;this still needs to be shared with report code, but I'm guessing V2 was about moving to hist_entry-ies
I guess I don't understand this? I don't use the report code. Is there another way I can get this?
Cheers,
Don
Re: [PATCH 15/19] perf, c2c: Add callchain support
Quoted 17 lines · Don Zickus
Quoted 5 lines · Jiri Olsa
On Fri, Feb 28, 2014 at 12:43:04PM -0500, Don Zickus wrote:
SNIP
Quoted 1 line
Quoted 10 lines
tools/perf/builtin-c2c.c static int perf_c2c__record(int argc, const char **argv) · line 1316
Quoted 6 lines
tools/perf/builtin-c2c.c +6
-0+static int +opt_callchain_cb (const struct option *opt,const char *arg,int unset) +{ +struct perf_c2c *c2c = (struct perf_c2c *)opt->value; +char *tok, *tok2; +char *endptr;this still needs to be shared with report code, but I'm guessing V2 was about moving to hist_entry-ies
Quoted 2 lines
I guess I don't understand this? I don't use the report code. Is there another way I can get this?
opt_callchain_cb is same as parse_callchain_opt function, the argument parsing code could be shared
jirka
PATCH v2 16/19 perf, c2c: Output summary stats
In reply to Don ZickusOutput some summary stats based on the processed records. Mainly diagnostic uses.
Stats done by Dick Fowles, backported by me.
Sample output:
=================================================
Trace Event Information
=================================================
Total records : 1322047 Locked Load/Store Operations : 206317 Load Operations : 355701 Loads - uncacheable : 590 Loads - IO : 0 Loads - Miss : 440 Loads - no mapping : 207 Load Fill Buffer Hit : 100214 Load L1D hit : 148454 Load L2D hit : 15170 Load LLC hit : 53872 Load Local HITM : 15388 Load Remote HITM : 26760 Load Remote HIT : 3910 Load Local DRAM : 2436 Load Remote DRAM : 3648 Load MESI State Exclusive : 2883 Load MESI State Shared : 3201 Load LLC Misses : 36754 LLC Misses to Local DRAM : 6.6% LLC Misses to Remote DRAM : 9.9% LLC Misses to Remote cache (HIT) : 10.6% LLC Misses to Remote cache (HITM) : 72.8% Store Operations : 966322 Store - uncacheable : 0 Store - no mapping : 42931 Store L1D Hit : 915696 Store L1D Miss : 7695 No Page Map Rejects : 1193 Unable to parse data source : 24
V2: refresh to hist_entry
--- tools/perf/builtin-c2c.c | 47 ++++++++++++++++++++++++++++++++++++++++++++++- 1 file changed, 46 insertions(+), 1 deletion(-)
tools/perf/builtin-c2c.c +46 -1
@@ -963,7 +963,6 @@ static void print_hitm_cacheline_offset(struct c2c_hit *clo, 23, stdout); } } -static void print_c2c_hitm_report (struct rb_root *hitm_tree,struct c2c_stats *hitm_stats__maybe_unused ,struct c2c_stats *c2c_stats)
@@ -1158,6 +1157,51 @@ cleanup:return ; } +static void print_c2c_trace_report (struct perf_c2c *c2c) +{ +int llc_misses; +struct c2c_stats *stats = &c2c->stats; + + llc_misses = stats->t.lcl_dram + + stats->t.rmt_dram + + stats->t.rmt_hit + + stats->t.rmt_hitm; + +printf ("=================================================\n" ); +printf (" Trace Event Information \n" ); +printf ("=================================================\n" ); +printf (" Total records : %10d\n" , c2c->stats.nr_entries); +printf (" Locked Load/Store Operations : %10d\n" , stats->t.locks); +printf (" Load Operations : %10d\n" , stats->t.load); +printf (" Loads - uncacheable : %10d\n" , stats->t.ld_uncache); +printf (" Loads - IO : %10d\n" , stats->t.ld_io); +printf (" Loads - Miss : %10d\n" , stats->t.ld_miss); +printf (" Loads - no mapping : %10d\n" , stats->t.ld_noadrs); +printf (" Load Fill Buffer Hit : %10d\n" , stats->t.ld_fbhit); +printf (" Load L1D hit : %10d\n" , stats->t.ld_l1hit); +printf (" Load L2D hit : %10d\n" , stats->t.ld_l2hit); +printf (" Load LLC hit : %10d\n" , stats->t.ld_llchit + stats->t.lcl_hitm); +printf (" Load Local HITM : %10d\n" , stats->t.lcl_hitm); +printf (" Load Remote HITM : %10d\n" , stats->t.rmt_hitm); +printf (" Load Remote HIT : %10d\n" , stats->t.rmt_hit); +printf (" Load Local DRAM : %10d\n" , stats->t.lcl_dram); +printf (" Load Remote DRAM : %10d\n" , stats->t.rmt_dram); +printf (" Load MESI State Exclusive : %10d\n" , stats->t.ld_excl); +printf (" Load MESI State Shared : %10d\n" , stats->t.ld_shared); +printf (" Load LLC Misses : %10d\n" , llc_misses); +printf (" LLC Misses to Local DRAM : %10.1f%%\n" , ((double )stats->t.lcl_dram/(double )llc_misses) *100. ); +printf (" LLC Misses to Remote DRAM : %10.1f%%\n" , ((double )stats->t.rmt_dram/(double )llc_misses) *100. ); +printf (" LLC Misses to Remote cache (HIT) : %10.1f%%\n" , ((double )stats->t.rmt_hit /(double )llc_misses) *100. ); +printf (" LLC Misses to Remote cache (HITM) : %10.1f%%\n" , ((double )stats->t.rmt_hitm/(double )llc_misses) *100. ); +printf (" Store Operations : %10d\n" , stats->t.store); +printf (" Store - uncacheable : %10d\n" , stats->t.st_uncache); +printf (" Store - no mapping : %10d\n" , stats->t.st_noadrs); +printf (" Store L1D Hit : %10d\n" , stats->t.st_l1hit); +printf (" Store L1D Miss : %10d\n" , stats->t.st_l1miss); +printf (" No Page Map Rejects : %10d\n" , stats->t.nomap); +printf (" Unable to parse data source : %10d\n" , stats->t.noparse); +} +static int perf_c2c__process_events (struct perf_session *session,struct perf_c2c *c2c) {
@@ -1169,6 +1213,7 @@ static int perf_c2c__process_events(struct perf_session *session, goto err; } +print_c2c_trace_report (c2c);c2c_analyze_hitms (c2c); err:
Signature · 1 line
1.7.11.7
PATCH v2 17/19 perf, c2c: Dump rbtree for debugging
In reply to Don ZickusSometimes you want to verify the rbtree sorting on a unique id is working correctly. This allows you to dump it.
Sample output:
Idx Hit Maj Min Ino InoGen Pid Daddr Iaddr Data Src (string) cpumode
0 0 0 0 0 0 22 ffffffff813044cf 48080184 [STORE,L1,MISS,SNP NA] 1
1 0 0 0 0 2332 ca3 ffffffffa0226032 48080184 [STORE,L1,MISS,SNP NA] 1
2 0 0 0 0 2332 ca3 ffffffffa0226032 48080184 [STORE,L1,MISS,SNP NA] 1
3 0 0 0 0 2332 ca3 ffffffffa0226032 48080184 [STORE,L1,MISS,SNP NA] 1
4 0 0 0 0 2332 ca3 ffffffffa0226032 48080184 [STORE,L1,MISS,SNP NA] 1
5 0 0 0 0 2332 ca3 ffffffffa0226032 48080184 [STORE,L1,MISS,SNP NA] 1
6 0 0 0 0 2332 ca3 ffffffffa0226032 48080184 [STORE,L1,MISS,SNP NA] 1
7 0 0 0 0 18179 135f860 ffffffff812ad509 68100242 [LOAD,LFB,HIT,SNP NONE] 1
8 0 0 0 0 18179 7ff9d7fbaf98 ffffffff812ad509 68100242 [LOAD,LFB,HIT,SNP NONE] 1
V2: refresh with hist_entry
--- tools/perf/builtin-c2c.c | 51 ++++++++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 51 insertions(+)
tools/perf/builtin-c2c.c +51 -0
@@ -616,6 +616,55 @@ err:#define HAS_HITMS (h) (h->stats.t.lcl_hitm || h->stats.t.rmt_hitm) +static void dump_rb_tree (struct rb_root *tree, +struct perf_c2c *c2c__maybe_unused ) +{ +struct rb_node *next =rb_first (tree); +struct hist_entry *he; +u64 cl =0 ; +int idx =0 ; + +printf ("# Summary: Total entries - %d\n" , c2c->stats.nr_entries); +printf ("# HITMs: Local - %d Remote - %d Total - %d\n" , + c2c->stats.t.lcl_hitm, c2c->stats.t.rmt_hitm, + (c2c->stats.t.lcl_hitm + c2c->stats.t.rmt_hitm)); + +printf ("%6s %3s %3s %3s %8s %16s %6s %16s %16s %16s %32s %8s\n" , +"Idx" ,"Hit" ,"Maj" ,"Min" ,"Ino" ,"InoGen" ,"Pid" , +"Daddr" ,"Iaddr" ,"Data Src" ,"(string)" ,"cpumode" ); +while (next) { +char data_src[32 ]; +u64 val; + + he = rb_entry(next, struct hist_entry, rb_node_in); + next =rb_next (&he->rb_node_in); + +if ((!he->color) || (cl !=CLADRS (he->mem_info->daddr.al_addr))) { +printf ("\n" ); + cl =CLADRS (he->mem_info->daddr.al_addr); + } + + val = he->mem_info->data_src.val; +perf_c2c__scnprintf_data_src (data_src,sizeof (data_src), val); + +printf ("%6d %3s %3x %3x %8lx %16lx %6d %16lx %16lx %16lx %32s %8x\n" , + idx, + (PERF_MEM_S (SNOOP ,HITM ) & val) ?" * " :" " , + he->mem_info->daddr.map->maj, + he->mem_info->daddr.map->min, + he->mem_info->daddr.map->ino, + he->mem_info->daddr.map->ino_generation, + he->thread->pid_, + he->mem_info->daddr.addr, + he->mem_info->iaddr.addr, + val, + data_src, + he->cpumode); + + idx++; + } +} +static void c2c_hit__update_stats (struct c2c_stats *new,struct c2c_stats *old) {
@@ -1213,6 +1262,8 @@ static int perf_c2c__process_events(struct perf_session *session, goto err; } +if (verbose >2 ) +dump_rb_tree (c2c->hists.entries_in, c2c);print_c2c_trace_report (c2c);c2c_analyze_hitms (c2c);
Signature · 1 line
1.7.11.7
PATCH v2 18/19 perf, c2c: Add symbol count table
In reply to Don ZickusJust another table that displays the referenced symbols in the analysis report. The table lists the most frequently used symbols first.
It is just another way to look at similar data to figure out who is causing the most contention (based on the workload used).
Original done by Dick Fowles, backported by me.
Sample output:
=======================================================================================================
Object Name, Path & Reference Counts
Index Records Object Name Object Path
=======================================================================================================
0 931379 [kernel.kallsyms] [kernel.kallsyms]
1 192258 fio /home/joe/old_fio-2.0.15/fio
2 80302 [jbd2] /lib/modules/3.10.0c2c_all+/kernel/fs/jbd2/jbd2.ko
3 65392 [ext4] /lib/modules/3.10.0c2c_all+/kernel/fs/ext4/ext4.koV2: refresh to latest upstream changes and hist_entry
--- tools/perf/builtin-c2c.c | 99 ++++++++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 99 insertions(+)
tools/perf/builtin-c2c.c +99 -0
@@ -685,6 +685,104 @@ static void c2c_hit__update_stats(struct c2c_stats *new, new->total_period += old->total_period; } +LIST_HEAD (ref_tree); +LIST_HEAD (ref_tree_sorted); +struct refs { +struct list_head list; +int nr; +const char *name; +const char *long_name; +}; + +static int update_ref_list (struct hist_entry *entry) +{ +struct refs *p; +struct dso *dso = entry->mem_info->iaddr.map->dso; +const char *name = dso->short_name; + +list_for_each_entry (p, &ref_tree, list) { +if (!strcmp (p->name, name)) + goto found; + } + + p =zalloc (sizeof (struct refs )); +if (!p) +return -1 ; + p->name = name; + p->long_name = dso->long_name; +list_add_tail (&p->list, &ref_tree); + +found: + p->nr++; +return 0 ; +} + +static void print_symbol_record_count (struct rb_root *tree) +{ +struct rb_node *next =rb_first (tree); +struct hist_entry *he; +struct refs *p, *q, *pn; +char string[256 ]; +char delimit[256 ]; +int i; +int idx =0 ; + +/* gather symbol references */ +while (next) { + he = rb_entry(next, struct hist_entry, rb_node_in); + next =rb_next (&he->rb_node_in); + +if (update_ref_list (he)) { +pr_err ("Could not update reference tree\n" ); + goto cleanup; + } + } + +/* sort on number of references per symbol */ +list_for_each_entry_safe (p, pn, &ref_tree, list) { +list_del_init (&p->list); +list_for_each_entry (q, &ref_tree_sorted, list) { +if (p->nr > q->nr) { +list_add_tail (&p->list, &q->list); +break ; + } + } +if (list_empty (&p->list)) +list_add_tail (&p->list, &ref_tree_sorted); + } + +/* print header info */ +sprintf (string,"%5s %8s %-32s %-80s" , +"Index" , +"Records" , +"Object Name" , +"Object Path" ); + + delimit[0 ] ='\0' ; +for (i =0 ; i < (int )strlen (string); i++)strcat (delimit,"=" ); + +printf ("\n\n" ); +printf ("%s\n" , delimit); +printf ("%50s %s\n" ," " ,"Object Name, Path & Reference Counts" ); +printf ("\n" ); +printf ("%s\n" , string); +printf ("%s\n" , delimit); + +/* print out table */ +list_for_each_entry (p, &ref_tree_sorted, list) { +printf ("%5d %8d %-32s %-80s\n" , + idx, p->nr, p->name, p->long_name); + idx++; + } +printf ("\n" ); + +cleanup: +list_for_each_entry_safe (p, pn, &ref_tree_sorted, list) { +list_del (&p->list); +free (p); + } +} +static void print_hitm_cacheline_header (void ) {#define SHARING_REPORT_TITLE "Shared Cache Line Distribution Pareto"
@@ -1266,6 +1364,7 @@ static int perf_c2c__process_events(struct perf_session *session,dump_rb_tree (c2c->hists.entries_in, c2c);print_c2c_trace_report (c2c);c2c_analyze_hitms (c2c); +print_symbol_record_count (c2c->hists.entries_in); err:return err;
Signature · 1 line
1.7.11.7
PATCH v2 19/19 perf, c2c: Add shared cachline summary table
In reply to Don ZickusThis adds a quick summary of the hottest cache contention lines based on the input data. This summarizes what the broken table shows you, so you can see at a quick glance which cachelines are interesting.
Originally done by Dick Fowles, backported by me.
Sample output (width trimmed):
===================================================================================================================================================
Shared Data Cache Line Table
Total %All Total ---- Core Load Hit ---- -- LLC Load Hit -- ----- LLC Load Hitm -----
Index Phys Adrs Records Ld Miss %hitm Loads FB L1D L2D Lcl Rmt Total Lcl Rmt
====================================================================================================================================================
0 0xffff881fa55b0140 72006 16.97% 23.31% 43095 13591 16860 45 2651 25 9526 3288 6238
1 0xffff881fba47f000 21854 5.29% 7.26% 13938 3887 6941 15 1 7 3087 1143 1944
2 0xffff881fc21b9cc0 2153 1.61% 2.21% 862 32 70 0 15 1 740 148 592
3 0xffff881fc7d91cc0 1957 1.40% 1.92% 866 34 94 0 14 3 720 207 513
4 0xffff881fba539cc0 1813 1.35% 1.85% 808 33 84 3 14 1 665 170 495--- tools/perf/builtin-c2c.c | 136 +++++++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 136 insertions(+)
tools/perf/builtin-c2c.c +136 -0
@@ -783,6 +783,141 @@ cleanup: } } +static void print_c2c_shared_cacheline_report (struct rb_root *hitm_tree, +struct c2c_stats *shared_stats__maybe_unused , +struct c2c_stats *c2c_stats__maybe_unused ) +{ +#define SHM_TITLE "Shared Data Cache Line Table" + +struct rb_node *next =rb_first (hitm_tree); +struct c2c_hit *h; +char header[256 ]; +char delimit[256 ]; +u32 crecords; +u32 lclmiss; +u32 ldcnt; +double p_hitm; +double p_all; +int totmiss; +int rmt_hitm; +int len; +int pad; +int i; + +sprintf (header,"%28s %8s %8s %8s %8s %28s %18s %28s %18s %8s %28s" , +" " , +"Total" , +"%All " , +" " , +"Total" , +"---- Core Load Hit ----" , +"-- LLC Load Hit --" , +"----- LLC Load Hitm -----" , +"-- Load Dram --" , +"LLC " , +"---- Store Reference ----" ); + + len =strlen (header); + delimit[0 ] ='\0' ; + +for (i =0 ; i < len; i++) +strcat (delimit,"=" ); + +printf ("\n\n" ); +printf ("%s\n" , delimit); +printf ("\n" ); + pad = (strlen (header)/2 ) - (strlen (SHM_TITLE )/2 ); +for (i =0 ; i < pad; i++) +printf (" " ); +printf ("%s\n" ,SHM_TITLE ); +printf ("\n" ); +printf ("%s\n" , header); + +sprintf (header,"%8s %18s %8s %8s %8s %8s %8s %8s %8s %8s %8s %8s %8s %8s %8s %8s %8s %8s %8s %8s" , +"Index" , +"Phys Adrs" , +"Records" , +"Ld Miss" , +"%hitm" , +"Loads" , +"FB" , +"L1D" , +"L2D" , +"Lcl" , +"Rmt" , +"Total" , +"Lcl" , +"Rmt" , +"Lcl" , +"Rmt" , +"Ld Miss" , +"Total" , +"L1Hit" , +"L1Miss" ); + +printf ("%s\n" , header); +printf ("%s\n" , delimit); + + rmt_hitm = c2c_stats->t.rmt_hitm; + totmiss = c2c_stats->t.lcl_dram + + c2c_stats->t.rmt_dram + + c2c_stats->t.rmt_hit + + c2c_stats->t.rmt_hitm; + + i =0 ; +while (next) { + h = rb_entry(next, struct c2c_hit, rb_node); + next =rb_next (&h->rb_node); + + lclmiss = h->stats.t.lcl_dram + + h->stats.t.rmt_dram + + h->stats.t.rmt_hitm + + h->stats.t.rmt_hit; + + ldcnt = lclmiss + + h->stats.t.ld_fbhit + + h->stats.t.ld_l1hit + + h->stats.t.ld_l2hit + + h->stats.t.ld_llchit + + h->stats.t.lcl_hitm; + + crecords = ldcnt + + h->stats.t.st_l1hit + + h->stats.t.st_l1miss; + + p_hitm = (double )h->stats.t.rmt_hitm / (double )rmt_hitm; + p_all = (double )h->stats.t.rmt_hitm / (double )totmiss; + +/* stop when the percentage gets to low */ +if (p_hitm <DISPLAY_LINE_LIMIT ) +break ; + +printf ("%8d %#18lx %8u %7.2f%% %7.2f%% %8u %8u %8u %8u %8u %8u %8u %8u %8u %8u %8u %8u %8u %8u %8u\n" , + i, + h->cacheline, + crecords, +100. * p_all, +100. * p_hitm, + ldcnt, + h->stats.t.ld_fbhit, + h->stats.t.ld_l1hit, + h->stats.t.ld_l2hit, + h->stats.t.ld_llchit, + h->stats.t.rmt_hit, + h->stats.t.lcl_hitm + h->stats.t.rmt_hitm, + h->stats.t.lcl_hitm, + h->stats.t.rmt_hitm, + h->stats.t.lcl_dram, + h->stats.t.rmt_dram, + lclmiss, + h->stats.t.store, + h->stats.t.st_l1hit, + h->stats.t.st_l1miss); + + i++; + } +} +static void print_hitm_cacheline_header (void ) {#define SHARING_REPORT_TITLE "Shared Cache Line Distribution Pareto"
@@ -1290,6 +1425,7 @@ static void c2c_analyze_hitms(struct perf_c2c *c2c)free (h);print_shared_cacheline_info (&hitm_stats, shared_clines); +print_c2c_shared_cacheline_report (&hitm_tree, &hitm_stats, &c2c->stats);print_c2c_hitm_report (&hitm_tree, &hitm_stats, &c2c->stats); cleanup:
Signature · 1 line
1.7.11.7
Quoted 3 lines · Don Zickus
A handful of patches include re-enabling MMAP2 support and some fixes to perf itself.
I would suggest to pursue the lone kernel patch separately. Hopefully that can be merged soon, once the remainin problems with that are addressed.
Quoted 2 lines
Comemnts, feedback, anything else welcomed.
As a high level comment, can you add support for CSV mode? (-x, in other tools)
I assume most people would data mine the output in some form, and that's much easier with a more machine oriented format.
Also you should probably have the standard perf conventions for output fds (--log-fd, --output, default to stderr). otherwise the output may often be lost.
-Andi
Signature · 1 line
[email protected] -- Speaking for myself only
Quoted 8 lines · Andi Kleen
Quoted 3 lines · Don Zickus
A handful of patches include re-enabling MMAP2 support and some fixes to perf itself.
I would suggest to pursue the lone kernel patch separately. Hopefully that can be merged soon, once the remainin problems with that are addressed.
Good point.
Quoted 6 lines
Quoted 3 lines
Comemnts, feedback, anything else welcomed.
As a high level comment, can you add support for CSV mode? (-x, in other tools)
It's there. :-)
perf c2c -r -x, report
will spit out raw records in CSV format.
Quoted 7 lines
I assume most people would data mine the output in some form, and that's much easier with a more machine oriented format.
Also you should probably have the standard perf conventions for output fds (--log-fd, --output, default to stderr). otherwise the output may often be lost.
Ok, I'll look into what that entails.
Thanks!
Cheers,
Don
Quoted 5 lines
It's there. :-)
perf c2c -r -x, report
will spit out raw records in CSV format.
But most of the cooked data seems not? I saw
lots of printfs without separator support.
Sorry I meant data mining the cooked data,
e.g. plotting it.
-Andi
Quoted 11 lines · Andi Kleen
Quoted 5 lines
It's there. :-)
perf c2c -r -x, report
will spit out raw records in CSV format.
But most of the cooked data seems not? I saw
lots of printfs without separator support.Sorry I meant data mining the cooked data,
e.g. plotting it.
Ah. Ok. I will look into that.
Cheers,
Don