Corrections to cpu map event encoding
Ian Rogers
reviewerv2 cover Corrections to cpu map event encoding
A mask encoding of a cpu map is laid out as:
u16 nr
u16 long_size
unsigned long mask[];
However, the mask may be 8-byte aligned meaning there is a 4-byte pad after long_size. This means 32-bit and 64-bit builds see the mask as being at different offsets. On top of this the structure is in the byte data[] encoded as:
u16 type
char data[]
This means the mask's struct isn't the required 4 or 8 byte aligned, but is offset by 2. Consequently the long reads and writes are causing undefined behavior as the alignment is broken.
These changes do minor clean up with const, visibility of functions and using the constant time max function. It then adds 32 and 64-bit mask encoding variants, packed to match current alignment. Taking the address of a packed struct leads to unaligned data, so function arguments are altered to be passed the packed struct. To compact the mask encoding further and drop the padding, the 4-byte variant is preferred. Finally a new range encoding is added, that reduces the size of the common case of a range of CPUs to a single u64.
On a 72 CPU (hyperthread) machine the original encoding of all CPUs is:0x9a98 [0x28]: event: 74
.
. ... raw event: size 40 bytes
. 0000: 4a 00 00 00 00 00 28 00 01 00 02 00 08 00 00 00 J.....(......... . 0010: 00 00 ff ff ff ff ff ff ff ff ff 00 00 00 00 00 ................ . 0020: 00 00 00 00 00 00 00 00 ........
0 0 0x9a98 [0x28]: PERF_RECORD_CPU_MAP
Using the 4-byte encoding it is:
0x9a98@pipe [0x20]: event: 74
.
. ... raw event: size 32 bytes
. 0000: 4a 00 00 00 00 00 20 00 01 00 03 00 04 00 ff ff J..... ......... . 0010: ff ff ff ff ff ff ff 00 00 00 00 00 00 00 00 00 ................
0 0 0x9a98 [0x20]: PERF_RECORD_CPU_MAP
Finally, with the range encoding it is:
0x9ab8@pipe [0x10]: event: 74
.
. ... raw event: size 16 bytes
. 0000: 4a 00 00 00 00 00 10 00 02 00 00 00 00 00 47 00 J.............G.
0 0 0x9ab8 [0x10]: PERF_RECORD_CPU_MAP
v2. Fixes a bug in the size computation of the update header
introduced by the last patch (Add range data encoding) and caught by address sanitizer.
Ian Rogers (6):
perf cpumap: Const map for max
perf cpumap: Synthetic events and const/static perf cpumap: Compute mask size in constant time perf cpumap: Fix alignment for masks in event encoding perf events: Prefer union over variable length array perf cpumap: Add range data encoding
tools/lib/perf/cpumap.c | 2 +- tools/lib/perf/include/perf/cpumap.h | 2 +- tools/lib/perf/include/perf/event.h | 61 ++++++++- tools/perf/tests/cpumap.c | 71 ++++++++--- tools/perf/tests/event_update.c | 14 +-- tools/perf/util/cpumap.c | 111 +++++++++++++--- tools/perf/util/cpumap.h | 4 +- tools/perf/util/event.h | 4 - tools/perf/util/header.c | 24 ++-- tools/perf/util/session.c | 35 +++--- tools/perf/util/synthetic-events.c | 182 +++++++++++++-------------- tools/perf/util/synthetic-events.h | 2 +- 12 files changed, 327 insertions(+), 185 deletions(-)
Signature · 1 line
2.36.1.476.g0c4daa206d-goog
Ian Rogers
reviewerPATCH v2 1/6 perf cpumap: Const map for max
Allows max to be used with const perf_cpu_maps.
--- tools/lib/perf/cpumap.c | 2 +- tools/lib/perf/include/perf/cpumap.h | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-)
tools/lib/perf/cpumap.c +1 -1
@@ -309,7 +309,7 @@ bool perf_cpu_map__has(const struct perf_cpu_map *cpus, struct perf_cpu cpu)return perf_cpu_map__idx (cpus, cpu) !=-1 ; } -struct perf_cpu perf_cpu_map__max (struct perf_cpu_map *map) +struct perf_cpu perf_cpu_map__max (const struct perf_cpu_map *map) {struct perf_cpu result = { .cpu =-1
tools/lib/perf/include/perf/cpumap.h +1 -1
@@ -23,7 +23,7 @@ LIBPERF_API void perf_cpu_map__put(struct perf_cpu_map *map); LIBPERF_API struct perf_cpu perf_cpu_map__cpu(const struct perf_cpu_map *cpus, int idx); LIBPERF_API int perf_cpu_map__nr(const struct perf_cpu_map *cpus); LIBPERF_API bool perf_cpu_map__empty(const struct perf_cpu_map *map); -LIBPERF_API struct perf_cpu perf_cpu_map__max(struct perf_cpu_map *map); +LIBPERF_API struct perf_cpu perf_cpu_map__max(const struct perf_cpu_map *map); LIBPERF_API bool perf_cpu_map__has(const struct perf_cpu_map *map, struct perf_cpu cpu);#define perf_cpu_map__for_each_cpu (cpu, idx, cpus) \
Signature · 1 line
2.36.1.476.g0c4daa206d-goog
Ian Rogers
reviewerPATCH v2 2/6 perf cpumap: Synthetic events and const/static
In reply to Ian RogersMake the cpumap arguments const to make it clearer they are in rather than out arguments. Make two functions static and remove external declarations.
--- tools/perf/util/event.h | 4 ---- tools/perf/util/synthetic-events.c | 20 +++++++++++--------- tools/perf/util/synthetic-events.h | 2 +- 3 files changed, 12 insertions(+), 14 deletions(-)
tools/perf/util/event.h +0 -4
@@ -461,10 +461,6 @@ size_t perf_event__fprintf(union perf_event *event, struct machine *machine, FILint kallsyms__get_function_start (const char *kallsyms_filename,const char *symbol_name,u64 *addr); -void *cpu_map_data__alloc (struct perf_cpu_map *map,size_t *size,u16 *type,int *max); -void cpu_map_data__synthesize (struct perf_record_cpu_map_data *data,struct perf_cpu_map *map, -u16 type,int max); -void event_attr_init (struct perf_event_attr *attr);int perf_event_paranoid (void );
tools/perf/util/synthetic-events.c +11 -9
@@ -1184,7 +1184,7 @@ int perf_event__synthesize_thread_map2(struct perf_tool *tool, }static void synthesize_cpus (struct cpu_map_entries *cpus, -struct perf_cpu_map *map) +const struct perf_cpu_map *map) {int i, map_nr =perf_cpu_map__nr (map);
@@ -1195,7 +1195,7 @@ static void synthesize_cpus(struct cpu_map_entries *cpus, }static void synthesize_mask (struct perf_record_record_cpu_map *mask, -struct perf_cpu_map *map,int max) +const struct perf_cpu_map *map,int max) {int i;
@@ -1206,12 +1206,12 @@ static void synthesize_mask(struct perf_record_record_cpu_map *mask,set_bit (perf_cpu_map__cpu (map, i).cpu, mask->mask); } -static size_t cpus_size (struct perf_cpu_map *map) +static size_t cpus_size (const struct perf_cpu_map *map) {return sizeof (struct cpu_map_entries ) +perf_cpu_map__nr (map) *sizeof (u16); } -static size_t mask_size (struct perf_cpu_map *map,int *max) +static size_t mask_size (const struct perf_cpu_map *map,int *max) {int i;
@@ -1228,7 +1228,8 @@ static size_t mask_size(struct perf_cpu_map *map, int *max)return sizeof (struct perf_record_record_cpu_map ) +BITS_TO_LONGS (*max) *sizeof (long ); } -void *cpu_map_data__alloc (struct perf_cpu_map *map,size_t *size,u16 *type,int *max) +static void *cpu_map_data__alloc (const struct perf_cpu_map *map,size_t *size, +u16 *type,int *max) {size_t size_cpus, size_mask;bool is_dummy =perf_cpu_map__empty (map);
@@ -1262,8 +1263,9 @@ void *cpu_map_data__alloc(struct perf_cpu_map *map, size_t *size, u16 *type, intreturn zalloc (*size); } -void cpu_map_data__synthesize (struct perf_record_cpu_map_data *data,struct perf_cpu_map *map, -u16 type,int max) +static void cpu_map_data__synthesize (struct perf_record_cpu_map_data *data, +const struct perf_cpu_map *map, +u16 type,int max) { data->type = type;
@@ -1278,7 +1280,7 @@ void cpu_map_data__synthesize(struct perf_record_cpu_map_data *data, struct perf } } -static struct perf_record_cpu_map *cpu_map_event__new (struct perf_cpu_map *map) +static struct perf_record_cpu_map *cpu_map_event__new (const struct perf_cpu_map *map) {size_t size =sizeof (struct perf_record_cpu_map );struct perf_record_cpu_map *event;
@@ -1298,7 +1300,7 @@ static struct perf_record_cpu_map *cpu_map_event__new(struct perf_cpu_map *map) }int perf_event__synthesize_cpu_map (struct perf_tool *tool, -struct perf_cpu_map *map, +const struct perf_cpu_map *map,perf_event__handler_t process,struct machine *machine) {
tools/perf/util/synthetic-events.h +1 -1
@@ -46,7 +46,7 @@ typedef int (*perf_event__handler_t)(struct perf_tool *tool, union perf_event *eint perf_event__synthesize_attrs (struct perf_tool *tool,struct evlist *evlist,perf_event__handler_t process);int perf_event__synthesize_attr (struct perf_tool *tool,struct perf_event_attr *attr,u32 ids,u64 *id,perf_event__handler_t process);int perf_event__synthesize_build_id (struct perf_tool *tool,struct dso *pos,u16 misc,perf_event__handler_t process,struct machine *machine); -int perf_event__synthesize_cpu_map (struct perf_tool *tool,struct perf_cpu_map *cpus,perf_event__handler_t process,struct machine *machine); +int perf_event__synthesize_cpu_map (struct perf_tool *tool,const struct perf_cpu_map *cpus,perf_event__handler_t process,struct machine *machine);int perf_event__synthesize_event_update_cpus (struct perf_tool *tool,struct evsel *evsel,perf_event__handler_t process);int perf_event__synthesize_event_update_name (struct perf_tool *tool,struct evsel *evsel,perf_event__handler_t process);int perf_event__synthesize_event_update_scale (struct perf_tool *tool,struct evsel *evsel,perf_event__handler_t process);
Signature · 1 line
2.36.1.476.g0c4daa206d-goog
Ian Rogers
reviewerPATCH v2 3/6 perf cpumap: Compute mask size in constant time
In reply to Ian Rogersperf_cpu_map__max computes the cpumap's maximum value, no need to iterate over all values.
--- tools/perf/util/synthetic-events.c | 13 +------------ 1 file changed, 1 insertion(+), 12 deletions(-)
tools/perf/util/synthetic-events.c +1 -12
@@ -1213,18 +1213,7 @@ static size_t cpus_size(const struct perf_cpu_map *map)static size_t mask_size (const struct perf_cpu_map *map,int *max) { -int i; - - *max =0 ; - -for (i =0 ; i <perf_cpu_map__nr (map); i++) { -/* bit position of the cpu is + 1 */ -int bit =perf_cpu_map__cpu (map, i).cpu +1 ; - -if (bit > *max) - *max = bit; - } - + *max =perf_cpu_map__max (map).cpu;return sizeof (struct perf_record_record_cpu_map ) +BITS_TO_LONGS (*max) *sizeof (long ); }
Signature · 1 line
2.36.1.476.g0c4daa206d-goog
Ian Rogers
reviewerPATCH v2 4/6 perf cpumap: Fix alignment for masks in event encoding
In reply to Ian RogersA mask encoding of a cpu map is laid out as:
u16 nr
u16 long_size
unsigned long mask[];
However, the mask may be 8-byte aligned meaning there is a 4-byte pad after long_size. This means 32-bit and 64-bit builds see the mask as being at different offsets. On top of this the structure is in the byte data[] encoded as:
u16 type
char data[]
This means the mask's struct isn't the required 4 or 8 byte aligned, but is offset by 2. Consequently the long reads and writes are causing undefined behavior as the alignment is broken.
Fix the mask struct by creating explicit 32 and 64-bit variants, use a union to avoid data[] and casts; the struct must be packed so the layout matches the existing perf.data layout. Taking an address of a member of a packed struct breaks alignment so pass the packed perf_record_cpu_map_data to functions, so they can access variables with the right alignment.
As the 64-bit version has 4 bytes of padding, optimizing writing to only write the 32-bit version.
--- tools/lib/perf/include/perf/event.h | 36 +++++++++++-- tools/perf/tests/cpumap.c | 19 ++++--- tools/perf/util/cpumap.c | 80 +++++++++++++++++++++++------ tools/perf/util/cpumap.h | 4 +- tools/perf/util/session.c | 30 +++++------ tools/perf/util/synthetic-events.c | 34 +++++++----- 6 files changed, 143 insertions(+), 60 deletions(-)
tools/lib/perf/include/perf/event.h +32 -4
@@ -6,6 +6,7 @@#include <linux/types.h> #include <linux/limits.h> #include <linux/bpf.h> +#include <linux/compiler.h> #include <sys/types.h> /* pid_t */ #define event_contains (obj, mem) ((obj).header.size > offsetof(typeof(obj), mem))
@@ -153,20 +154,47 @@ enum { PERF_CPU_MAP__MASK = 1, }; +/* +* Array encoding of a perf_cpu_map where nr is the number of entries in cpu[] +* and each entry is a value for a CPU in the map. +*/ struct cpu_map_entries {__u16 nr;__u16 cpu[]; }; -struct perf_record_record_cpu_map { +/* Bitmap encoding of a perf_cpu_map where bitmap entries are 32-bit. */ +struct perf_record_mask_cpu_map32 { +/* Number of mask values. */ __u16 nr; +/* Constant 4. */ __u16 long_size; -unsigned long mask[]; +/* Bitmap data. */ +__u32 mask[]; }; -struct perf_record_cpu_map_data { +/* Bitmap encoding of a perf_cpu_map where bitmap entries are 64-bit. */ +struct perf_record_mask_cpu_map64 { +/* Number of mask values. */ +__u16 nr; +/* Constant 8. */ +__u16 long_size; +/* Legacy padding. */ +char __pad[4 ]; +/* Bitmap data. */ +__u64 mask[]; +}; + +struct __packed perf_record_cpu_map_data {__u16 type; -char data[]; +union { +/* Used when type == PERF_CPU_MAP__CPUS. */ +struct cpu_map_entries cpus_data; +/* Used when type == PERF_CPU_MAP__MASK and long_size == 4. */ +struct perf_record_mask_cpu_map32 mask32_data; +/* Used when type == PERF_CPU_MAP__MASK and long_size == 8. */ +struct perf_record_mask_cpu_map64 mask64_data; + }; };struct perf_record_cpu_map {
tools/perf/tests/cpumap.c +9 -10
@@ -17,21 +17,23 @@ static int process_event_mask(struct perf_tool *tool __maybe_unused, struct machine *machine __maybe_unused) {struct perf_record_cpu_map *map_event = &event->cpu_map; -struct perf_record_record_cpu_map *mask;struct perf_record_cpu_map_data *data;struct perf_cpu_map *map;int i; +unsigned int long_size; data = &map_event->data;TEST_ASSERT_VAL ("wrong type" , data->type ==PERF_CPU_MAP__MASK ); - mask = (struct perf_record_record_cpu_map *)data->data; + long_size = data->mask32_data.long_size; -TEST_ASSERT_VAL ("wrong nr" , mask->nr ==1 ); +TEST_ASSERT_VAL ("wrong long_size" , long_size ==4 || long_size ==8 ); + +TEST_ASSERT_VAL ("wrong nr" , data->mask32_data.nr ==1 );for (i =0 ; i <20 ; i++) { -TEST_ASSERT_VAL ("wrong cpu" ,test_bit (i, mask->mask)); +TEST_ASSERT_VAL ("wrong cpu" ,perf_record_cpu_map_data__test_bit (i, data)); } map =cpu_map__new_data (data);
@@ -51,7 +53,6 @@ static int process_event_cpus(struct perf_tool *tool __maybe_unused, struct machine *machine __maybe_unused) {struct perf_record_cpu_map *map_event = &event->cpu_map; -struct cpu_map_entries *cpus;struct perf_record_cpu_map_data *data;struct perf_cpu_map *map;
@@ -59,11 +60,9 @@ static int process_event_cpus(struct perf_tool *tool __maybe_unused,TEST_ASSERT_VAL ("wrong type" , data->type ==PERF_CPU_MAP__CPUS ); - cpus = (struct cpu_map_entries *)data->data; - -TEST_ASSERT_VAL ("wrong nr" , cpus->nr ==2 ); -TEST_ASSERT_VAL ("wrong cpu" , cpus->cpu[0 ] ==1 ); -TEST_ASSERT_VAL ("wrong cpu" , cpus->cpu[1 ] ==256 ); +TEST_ASSERT_VAL ("wrong nr" , data->cpus_data.nr ==2 ); +TEST_ASSERT_VAL ("wrong cpu" , data->cpus_data.cpu[0 ] ==1 ); +TEST_ASSERT_VAL ("wrong cpu" , data->cpus_data.cpu[1 ] ==256 ); map =cpu_map__new_data (data);TEST_ASSERT_VAL ("wrong nr" ,perf_cpu_map__nr (map) ==2 );
tools/perf/util/cpumap.c +64 -16
118 lines not on this page · raw message
tools/perf/util/cpumap.h +3 -1
@@ -37,9 +37,11 @@ struct cpu_aggr_map {struct perf_record_cpu_map_data ; +bool perf_record_cpu_map_data__test_bit (int i,const struct perf_record_cpu_map_data *data); +struct perf_cpu_map *perf_cpu_map__empty_new (int nr); -struct perf_cpu_map *cpu_map__new_data (struct perf_record_cpu_map_data *data); +struct perf_cpu_map *cpu_map__new_data (const struct perf_record_cpu_map_data *data);size_t cpu_map__snprint (struct perf_cpu_map *map,char *buf,size_t size);size_t cpu_map__snprint_mask (struct perf_cpu_map *map,char *buf,size_t size);size_t cpu_map__fprintf (struct perf_cpu_map *map,FILE *fp);
tools/perf/util/session.c +15 -15
@@ -914,30 +914,30 @@ static void perf_event__cpu_map_swap(union perf_event *event, bool sample_id_all __maybe_unused) {struct perf_record_cpu_map_data *data = &event->cpu_map.data; -struct cpu_map_entries *cpus; -struct perf_record_record_cpu_map *mask; -unsigned i; data->type =bswap_16 (data->type);switch (data->type) {case PERF_CPU_MAP__CPUS : - cpus = (struct cpu_map_entries *)data->data; - - cpus->nr =bswap_16 (cpus->nr); + data->cpus_data.nr =bswap_16 (data->cpus_data.nr); -for (i =0 ; i < cpus->nr; i++) - cpus->cpu[i] =bswap_16 (cpus->cpu[i]); +for (unsigned i =0 ; i < data->cpus_data.nr; i++) + data->cpus_data.cpu[i] =bswap_16 (data->cpus_data.cpu[i]);break ;case PERF_CPU_MAP__MASK : - mask = (struct perf_record_record_cpu_map *)data->data; - - mask->nr =bswap_16 (mask->nr); - mask->long_size =bswap_16 (mask->long_size); + data->mask32_data.long_size =bswap_16 (data->mask32_data.long_size); -switch (mask->long_size) { -case 4 :mem_bswap_32 (&mask->mask, mask->nr);break ; -case 8 :mem_bswap_64 (&mask->mask, mask->nr);break ; +switch (data->mask32_data.long_size) { +case 4 : + data->mask32_data.nr =bswap_16 (data->mask32_data.nr); +for (unsigned i =0 ; i < data->mask32_data.nr; i++) + data->mask32_data.mask[i] =bswap_32 (data->mask32_data.mask[i]); +break ; +case 8 : + data->mask64_data.nr =bswap_16 (data->mask64_data.nr); +for (unsigned i =0 ; i < data->mask64_data.nr; i++) + data->mask64_data.mask[i] =bswap_64 (data->mask64_data.mask[i]); +break ;default :pr_err ("cpu_map swap: unsupported long size\n" ); }
tools/perf/util/synthetic-events.c +20 -14
@@ -1183,27 +1183,33 @@ int perf_event__synthesize_thread_map2(struct perf_tool *tool,return err; } -static void synthesize_cpus (struct cpu_map_entries *cpus, +static void synthesize_cpus (struct perf_record_cpu_map_data *data,const struct perf_cpu_map *map) {int i, map_nr =perf_cpu_map__nr (map); - cpus->nr = map_nr; + data->cpus_data.nr = map_nr;for (i =0 ; i < map_nr; i++) - cpus->cpu[i] =perf_cpu_map__cpu (map, i).cpu; + data->cpus_data.cpu[i] =perf_cpu_map__cpu (map, i).cpu; } -static void synthesize_mask (struct perf_record_record_cpu_map *mask, +static void synthesize_mask (struct perf_record_cpu_map_data *data,const struct perf_cpu_map *map,int max) { -int i; +int idx; +struct perf_cpu cpu; + +/* Due to padding, the 4bytes per entry mask variant is always smaller. */ + data->mask32_data.nr =BITS_TO_U32 (max); + data->mask32_data.long_size =4 ; - mask->nr =BITS_TO_LONGS (max); - mask->long_size =sizeof (long ); +perf_cpu_map__for_each_cpu (cpu, idx, map) { +int bit_word = cpu.cpu /32 ; +__u32 bit_mask =1U << (cpu.cpu &31 ); -for (i =0 ; i <perf_cpu_map__nr (map); i++) -set_bit (perf_cpu_map__cpu (map, i).cpu, mask->mask); + data->mask32_data.mask[bit_word] |= bit_mask; + } }static size_t cpus_size (const struct perf_cpu_map *map)
@@ -1214,7 +1220,7 @@ static size_t cpus_size(const struct perf_cpu_map *map)static size_t mask_size (const struct perf_cpu_map *map,int *max) { *max =perf_cpu_map__max (map).cpu; -return sizeof (struct perf_record_record_cpu_map ) +BITS_TO_LONGS (*max) *sizeof (long ); +return sizeof (struct perf_record_mask_cpu_map32 ) +BITS_TO_U32 (*max) *sizeof (__u32); } static void *cpu_map_data__alloc(const struct perf_cpu_map *map, size_t *size,
@@ -1247,7 +1253,7 @@ static void *cpu_map_data__alloc(const struct perf_cpu_map *map, size_t *size, *type =PERF_CPU_MAP__MASK ; } - *size +=sizeof (struct perf_record_cpu_map_data ); + *size +=sizeof (__u16);/* For perf_record_cpu_map_data.type. */ *size =PERF_ALIGN (*size,sizeof (u64));return zalloc (*size); }
@@ -1260,10 +1266,10 @@ static void cpu_map_data__synthesize(struct perf_record_cpu_map_data *data,switch (type) {case PERF_CPU_MAP__CPUS : -synthesize_cpus ((struct cpu_map_entries *) data->data, map); +synthesize_cpus (data, map);break ;case PERF_CPU_MAP__MASK : -synthesize_mask ((struct perf_record_record_cpu_map *)data->data, map, max); +synthesize_mask (data, map, max);default :break ; }
@@ -1271,7 +1277,7 @@ static void cpu_map_data__synthesize(struct perf_record_cpu_map_data *data,static struct perf_record_cpu_map *cpu_map_event__new (const struct perf_cpu_map *map) { -size_t size =sizeof (struct perf_record_cpu_map ); +size_t size =sizeof (struct perf_event_header );struct perf_record_cpu_map *event;int max;u16 type;
Signature · 1 line
2.36.1.476.g0c4daa206d-goog
Namhyung Kim
maintainerRe: [PATCH v2 4/6] perf cpumap: Fix alignment for masks in event encoding
Hi Ian,
Quoted 87 lines · Ian Rogers
A mask encoding of a cpu map is laid out as:
u16 nr
u16long_size
unsigned long mask[];
However, the mask may be 8-byte aligned meaning there is a 4-byte pad afterlong_size. This means 32-bit and 64-bit builds see the mask as being at different offsets. On top of this the structure is in the byte data[] encoded as:
u16 type
char data[]
This means the mask's struct isn't the required 4 or 8 byte aligned, but is offset by 2. Consequently the long reads and writes are causing undefined behavior as the alignment is broken.Fix the mask struct by creating explicit 32 and 64-bit variants, use a union to avoid data[] and casts; the struct must be packed so the layout matches the existing perf.data layout. Taking an address of a member of a packed struct breaks alignment so pass the packed
perf_record_cpu_map_datato functions, so they can access variables with the right alignment.As the 64-bit version has 4 bytes of padding, optimizing writing to only write the 32-bit version.
Signed-off-by Ian Rogers <[email protected]>--- tools/lib/perf/include/perf/event.h | 36 +++++++++++-- tools/perf/tests/cpumap.c | 19 ++++--- tools/perf/util/cpumap.c | 80 +++++++++++++++++++++++------ tools/perf/util/cpumap.h | 4 +- tools/perf/util/session.c | 30 +++++------ tools/perf/util/synthetic-events.c | 34 +++++++----- 6 files changed, 143 insertions(+), 60 deletions(-)tools/lib/perf/include/perf/event.h +24
-3@@ -6,6 +6,7 @@#include <linux/types.h> #include <linux/limits.h> #include <linux/bpf.h> +#include <linux/compiler.h> #include <sys/types.h> /* pid_t */ #define event_contains (obj, mem) ((obj).header.size > offsetof(typeof(obj), mem))@@ -153,20 +154,47 @@ enum { PERF_CPU_MAP__MASK = 1, }; +/* +* Array encoding of a perf_cpu_map where nr is the number of entries in cpu[] +* and each entry is a value for a CPU in the map. +*/ struct cpu_map_entries {__u16 nr;__u16 cpu[]; }; -struct perf_record_record_cpu_map { +/* Bitmap encoding of a perf_cpu_map where bitmap entries are 32-bit. */ +struct perf_record_mask_cpu_map32 { +/* Number of mask values. */ __u16 nr; +/* Constant 4. */ __u16 long_size; -unsigned long mask[]; +/* Bitmap data. */ +__u32 mask[]; }; -struct perf_record_cpu_map_data { +/* Bitmap encoding of a perf_cpu_map where bitmap entries are 64-bit. */ +struct perf_record_mask_cpu_map64 { +/* Number of mask values. */ +__u16 nr; +/* Constant 8. */ +__u16 long_size; +/* Legacy padding. */ +char __pad[4 ]; +/* Bitmap data. */ +__u64 mask[]; +}; + +struct __packed perf_record_cpu_map_data {__u16 type;
tools/lib/perf/include/perf/event.h enum { · line 198
Quoted 10 lines
tools/lib/perf/include/perf/event.h +8
-1-char data[]; +union { +/* Used when type == PERF_CPU_MAP__CPUS. */ +struct cpu_map_entries cpus_data; +/* Used when type == PERF_CPU_MAP__MASK and long_size == 4. */ +struct perf_record_mask_cpu_map32 mask32_data; +/* Used when type == PERF_CPU_MAP__MASK and long_size == 8. */ +struct perf_record_mask_cpu_map64 mask64_data; + }; };
How about moving the 'type' to the union as well?
This way we don't need to pack the entire struct
and can have a common struct for 32 and 64 bit..
struct cpu_map_entries {
__u16 type;
__u16 nr;
__u16 cpu[];
};struct perf_record_mask_cpu_map {
__u16 type;
__u16 nr;
__u16 long_size; // still needed?
__u16 pad;
unsigned long mask[];
};// changed it to union
union perf_record_cpu_map_data {
__u16 type;
struct cpu_map_entries cpus_data;
struct perf_record_mask_cpu_map mask_data;
};Thanks,
Namhyung
Ian Rogers
reviewerRe: [PATCH v2 4/6] perf cpumap: Fix alignment for masks in event encoding
Quoted 120 lines · Namhyung Kim
Hi Ian,
Quoted 87 lines · Ian Rogers
A mask encoding of a cpu map is laid out as:
u16 nr
u16long_size
unsigned long mask[];
However, the mask may be 8-byte aligned meaning there is a 4-byte pad afterlong_size. This means 32-bit and 64-bit builds see the mask as being at different offsets. On top of this the structure is in the byte data[] encoded as:
u16 type
char data[]
This means the mask's struct isn't the required 4 or 8 byte aligned, but is offset by 2. Consequently the long reads and writes are causing undefined behavior as the alignment is broken.Fix the mask struct by creating explicit 32 and 64-bit variants, use a union to avoid data[] and casts; the struct must be packed so the layout matches the existing perf.data layout. Taking an address of a member of a packed struct breaks alignment so pass the packed
perf_record_cpu_map_datato functions, so they can access variables with the right alignment.As the 64-bit version has 4 bytes of padding, optimizing writing to only write the 32-bit version.
Signed-off-by Ian Rogers <[email protected]>--- tools/lib/perf/include/perf/event.h | 36 +++++++++++-- tools/perf/tests/cpumap.c | 19 ++++--- tools/perf/util/cpumap.c | 80 +++++++++++++++++++++++------ tools/perf/util/cpumap.h | 4 +- tools/perf/util/session.c | 30 +++++------ tools/perf/util/synthetic-events.c | 34 +++++++----- 6 files changed, 143 insertions(+), 60 deletions(-)tools/lib/perf/include/perf/event.h +24
-3@@ -6,6 +6,7 @@#include <linux/types.h> #include <linux/limits.h> #include <linux/bpf.h> +#include <linux/compiler.h> #include <sys/types.h> /* pid_t */ #define event_contains (obj, mem) ((obj).header.size > offsetof(typeof(obj), mem))@@ -153,20 +154,47 @@ enum { PERF_CPU_MAP__MASK = 1, }; +/* +* Array encoding of a perf_cpu_map where nr is the number of entries in cpu[] +* and each entry is a value for a CPU in the map. +*/ struct cpu_map_entries {__u16 nr;__u16 cpu[]; }; -struct perf_record_record_cpu_map { +/* Bitmap encoding of a perf_cpu_map where bitmap entries are 32-bit. */ +struct perf_record_mask_cpu_map32 { +/* Number of mask values. */ __u16 nr; +/* Constant 4. */ __u16 long_size; -unsigned long mask[]; +/* Bitmap data. */ +__u32 mask[]; }; -struct perf_record_cpu_map_data { +/* Bitmap encoding of a perf_cpu_map where bitmap entries are 64-bit. */ +struct perf_record_mask_cpu_map64 { +/* Number of mask values. */ +__u16 nr; +/* Constant 8. */ +__u16 long_size; +/* Legacy padding. */ +char __pad[4 ]; +/* Bitmap data. */ +__u64 mask[]; +}; + +struct __packed perf_record_cpu_map_data {__u16 type;tools/lib/perf/include/perf/event.h enum { · line 198
Quoted 10 lines
tools/lib/perf/include/perf/event.h +8
-1-char data[]; +union { +/* Used when type == PERF_CPU_MAP__CPUS. */ +struct cpu_map_entries cpus_data; +/* Used when type == PERF_CPU_MAP__MASK and long_size == 4. */ +struct perf_record_mask_cpu_map32 mask32_data; +/* Used when type == PERF_CPU_MAP__MASK and long_size == 8. */ +struct perf_record_mask_cpu_map64 mask64_data; + }; };How about moving the 'type' to the union as well?
This way we don't need to pack the entire struct
and can have a common struct for 32 and 64 bit..struct cpu_map_entries { __u16 type; __u16 nr; __u16 cpu[]; };struct perf_record_mask_cpu_map { __u16 type; __u16 nr; __u16 long_size; // still needed? __u16 pad; unsigned long mask[]; };
Quoted 6 lines
// changed it to union
union perf_record_cpu_map_data { __u16 type; struct cpu_map_entries cpus_data; struct perf_record_mask_cpu_map mask_data; };
Thanks Namhyung,
Unfortunately this doesn't quite work as I want to make it so that the existing cpu map encodings work with this change - ie, an old perf.data should be readable in a newer perf with this change (the range encoding will require that new perf.data files are read by versions of perf with these changes). For this to work I need the layout to match the existing unaligned code, I either need to make mask bytes and memcpy or use an attribute like packed. Fwiw, this is a little more efficient than the layout above as with long_size == 4 the pad isn't necessary saving 2 bytes. I think with the packed approach we can also add new unpacked variants like above, although I'd be keen not to use a type that varies in size like long. I guess at some future date we could remove the legacy supporting packed versions so that packed or byte copying is unnecessary.
I could use a union as you show above, unfortunately that will need the 'struct perf_record_mask_cpu_map32' and 'struct perf_record_mask_cpu_map64' to be packed or to use bytes. We'd lose one use of packed just to introduce two others. Potentially it is more of a breaking change for users of this code via libperf.
These changes are something of a bug report along with fixes. If there is a consensus that the right way to fix the bug is to break legacy perf.data files then I'm happy to update the code accordingly (as you show above).
Thanks,
Ian
Quoted 2 lines
Thanks,
Namhyung
Re: [PATCH v2 4/6] perf cpumap: Fix alignment for masks in event encoding
In reply to Ian RogersQuoted 25 lines · Ian Rogers
A mask encoding of a cpu map is laid out as:
u16 nr
u16long_size
unsigned long mask[];
However, the mask may be 8-byte aligned meaning there is a 4-byte pad afterlong_size. This means 32-bit and 64-bit builds see the mask as being at different offsets. On top of this the structure is in the byte data[] encoded as:
u16 type
char data[]
This means the mask's struct isn't the required 4 or 8 byte aligned, but is offset by 2. Consequently the long reads and writes are causing undefined behavior as the alignment is broken.Fix the mask struct by creating explicit 32 and 64-bit variants, use a union to avoid data[] and casts; the struct must be packed so the layout matches the existing perf.data layout. Taking an address of a member of a packed struct breaks alignment so pass the packed
perf_record_cpu_map_datato functions, so they can access variables with the right alignment.As the 64-bit version has 4 bytes of padding, optimizing writing to only write the 32-bit version.
Signed-off-by Ian Rogers <[email protected]>
SNIP
Quoted 14 lines
struct
perf_record_cpu_map{tools/perf/tests/cpumap.c +1
-1@@ -17,21 +17,23 @@ static int process_event_mask(struct perf_tool *tool __maybe_unused, struct machine *machine __maybe_unused) {struct perf_record_cpu_map *map_event = &event->cpu_map; -struct perf_record_record_cpu_map *mask;struct perf_record_cpu_map_data *data;struct perf_cpu_map *map;int i; +unsigned int long_size;
tools/perf/tests/cpumap.c static int process_event_mask(struct perf_tool *tool __maybe_unused, · line 31
Quoted 10 lines
tools/perf/tests/cpumap.c +2
-2data = &map_event->data;TEST_ASSERT_VAL ("wrong type" , data->type ==PERF_CPU_MAP__MASK ); - mask = (struct perf_record_record_cpu_map *)data->data; + long_size = data->mask32_data.long_size; -TEST_ASSERT_VAL ("wrong nr" , mask->nr ==1 ); +TEST_ASSERT_VAL ("wrong long_size" , long_size ==4 || long_size ==8 );
should we check here just for long_size == 4 ?
SNIP
Quoted 35 lines
tools/perf/util/cpumap.c +26
-1@@ -22,54 +22,102 @@ static int max_node_num;*/ static int *cpunode_map; -static struct perf_cpu_map *cpu_map__from_entries (struct cpu_map_entries *cpus) +bool perf_record_cpu_map_data__test_bit (int i, +const struct perf_record_cpu_map_data *data) +{ +int bit_word32 = i /32 ; +__u32 bit_mask32 =1U << (i &31 ); +int bit_word64 = i /64 ; +__u64 bit_mask64 = ((__u64 )1 ) << (i &63 ); + +return (data->mask32_data.long_size ==4 ) + ? (bit_word32 < data->mask32_data.nr) && + (data->mask32_data.mask[bit_word32] & bit_mask32) !=0 + : (bit_word64 < data->mask64_data.nr) && + (data->mask64_data.mask[bit_word64] & bit_mask64) !=0 ; +} + +/* Read ith mask value from data into the given 64-bit sized bitmap */ +static void perf_record_cpu_map_data__read_one_mask (const struct perf_record_cpu_map_data *data, +int i,unsigned long *bitmap) +{ +#if __SIZEOF_LONG__ ==8 +if (data->mask32_data.long_size ==4 ) + bitmap[0 ] = data->mask32_data.mask[i]; +else + bitmap[0 ] = data->mask64_data.mask[i]; +#else +if (data->mask32_data.long_size ==4 ) {
tools/perf/util/cpumap.c static int max_node_num; · line 60
Quoted 10 lines
tools/perf/util/cpumap.c +10
-0+ bitmap[0 ] = data->mask32_data.mask[i]; + bitmap[1 ] =0 ; + }else { +#if __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__ + bitmap[0 ] = (unsigned long )(data->mask64_data.mask[i] >>32 ); + bitmap[1 ] = (unsigned long )data->mask64_data.mask[i]; +#else + bitmap[0 ] = (unsigned long )data->mask64_data.mask[i]; + bitmap[1 ] = (unsigned long )(data->mask64_data.mask[i] >>32 ); +#endif
should this be taken care of earlier by perf_event__cpu_map_swap ?
Quoted 127 lines
tools/perf/util/cpumap.c +28
-15+ } +#endif +} +static struct perf_cpu_map *cpu_map__from_entries (const struct perf_record_cpu_map_data *data) {struct perf_cpu_map *map; - map =perf_cpu_map__empty_new (cpus->nr); + map =perf_cpu_map__empty_new (data->cpus_data.nr);if (map) {unsigned i; -for (i =0 ; i < cpus->nr; i++) { +for (i =0 ; i < data->cpus_data.nr; i++) {/* * Special treatment for -1, which is not real cpu number, * and we need to use (int) -1 to initialize map[i], * otherwise it would become 65535. */ -if (cpus->cpu[i] == (u16) -1 ) +if (data->cpus_data.cpu[i] == (u16) -1 ) map->map[i].cpu =-1 ;else - map->map[i].cpu = (int ) cpus->cpu[i]; + map->map[i].cpu = (int ) data->cpus_data.cpu[i]; } }return map; } -static struct perf_cpu_map *cpu_map__from_mask (struct perf_record_record_cpu_map *mask) +static struct perf_cpu_map *cpu_map__from_mask (const struct perf_record_cpu_map_data *data) { +DECLARE_BITMAP (local_copy,64 ); +int weight =0 , mask_nr = data->mask32_data.nr;struct perf_cpu_map *map; -int nr, nbits = mask->nr * mask->long_size *BITS_PER_BYTE ; - nr =bitmap_weight (mask->mask, nbits); +for (int i =0 ; i < mask_nr; i++) { +perf_record_cpu_map_data__read_one_mask (data, i, local_copy); + weight +=bitmap_weight (local_copy,64 ); + } + + map =perf_cpu_map__empty_new (weight); +if (!map) +return NULL ; - map =perf_cpu_map__empty_new (nr); -if (map) { -int cpu, i =0 ; +for (int i =0 , j =0 ; i < mask_nr; i++) { +int cpus_per_i = (i * data->mask32_data.long_size *BITS_PER_BYTE ); +int cpu; -for_each_set_bit (cpu, mask->mask, nbits) - map->map[i++].cpu = cpu; +perf_record_cpu_map_data__read_one_mask (data, i, local_copy); +for_each_set_bit (cpu, local_copy,64 ) + map->map[j++].cpu = cpu + cpus_per_i; }return map; } -struct perf_cpu_map *cpu_map__new_data (struct perf_record_cpu_map_data *data) +struct perf_cpu_map *cpu_map__new_data (const struct perf_record_cpu_map_data *data) {if (data->type ==PERF_CPU_MAP__CPUS ) -return cpu_map__from_entries ((struct cpu_map_entries *)data->data); +return cpu_map__from_entries (data);else -return cpu_map__from_mask ((struct perf_record_record_cpu_map *)data->data); +return cpu_map__from_mask (data); }size_t cpu_map__fprintf (struct perf_cpu_map *map,FILE *fp)tools/perf/util/cpumap.h +3
-1@@ -37,9 +37,11 @@ struct cpu_aggr_map {struct perf_record_cpu_map_data ; +bool perf_record_cpu_map_data__test_bit (int i,const struct perf_record_cpu_map_data *data); +struct perf_cpu_map *perf_cpu_map__empty_new (int nr); -struct perf_cpu_map *cpu_map__new_data (struct perf_record_cpu_map_data *data); +struct perf_cpu_map *cpu_map__new_data (const struct perf_record_cpu_map_data *data);size_t cpu_map__snprint (struct perf_cpu_map *map,char *buf,size_t size);size_t cpu_map__snprint_mask (struct perf_cpu_map *map,char *buf,size_t size);size_t cpu_map__fprintf (struct perf_cpu_map *map,FILE *fp);tools/perf/util/session.c +4
-12@@ -914,30 +914,30 @@ static void perf_event__cpu_map_swap(union perf_event *event, bool sample_id_all __maybe_unused) {struct perf_record_cpu_map_data *data = &event->cpu_map.data; -struct cpu_map_entries *cpus; -struct perf_record_record_cpu_map *mask; -unsigned i; data->type =bswap_16 (data->type);switch (data->type) {case PERF_CPU_MAP__CPUS : - cpus = (struct cpu_map_entries *)data->data; - - cpus->nr =bswap_16 (cpus->nr); + data->cpus_data.nr =bswap_16 (data->cpus_data.nr); -for (i =0 ; i < cpus->nr; i++) - cpus->cpu[i] =bswap_16 (cpus->cpu[i]); +for (unsigned i =0 ; i < data->cpus_data.nr; i++) + data->cpus_data.cpu[i] =bswap_16 (data->cpus_data.cpu[i]);break ;case PERF_CPU_MAP__MASK : - mask = (struct perf_record_record_cpu_map *)data->data; - - mask->nr =bswap_16 (mask->nr); - mask->long_size =bswap_16 (mask->long_size); + data->mask32_data.long_size =bswap_16 (data->mask32_data.long_size);
tools/perf/util/session.c static void perf_event__cpu_map_swap(union perf_event *event, · line 935
Quoted 10 lines
tools/perf/util/session.c +6
-3-switch (mask->long_size) { -case 4 :mem_bswap_32 (&mask->mask, mask->nr);break ; -case 8 :mem_bswap_64 (&mask->mask, mask->nr);break ; +switch (data->mask32_data.long_size) { +case 4 : + data->mask32_data.nr =bswap_16 (data->mask32_data.nr); +for (unsigned i =0 ; i < data->mask32_data.nr; i++) + data->mask32_data.mask[i] =bswap_32 (data->mask32_data.mask[i]); +break ;
why not use the mem_bswap_* functions?
looks like we never swap it completely, because we passed mask->nr where should be the size
Quoted 31 lines
tools/perf/util/session.c +5
-0+case 8 : + data->mask64_data.nr =bswap_16 (data->mask64_data.nr); +for (unsigned i =0 ; i < data->mask64_data.nr; i++) + data->mask64_data.mask[i] =bswap_64 (data->mask64_data.mask[i]); +break ;default :pr_err ("cpu_map swap: unsupported long size\n" ); }tools/perf/util/synthetic-events.c +3
-4@@ -1183,27 +1183,33 @@ int perf_event__synthesize_thread_map2(struct perf_tool *tool,return err; } -static void synthesize_cpus (struct cpu_map_entries *cpus, +static void synthesize_cpus (struct perf_record_cpu_map_data *data,const struct perf_cpu_map *map) {int i, map_nr =perf_cpu_map__nr (map); - cpus->nr = map_nr; + data->cpus_data.nr = map_nr;for (i =0 ; i < map_nr; i++) - cpus->cpu[i] =perf_cpu_map__cpu (map, i).cpu; + data->cpus_data.cpu[i] =perf_cpu_map__cpu (map, i).cpu; } -static void synthesize_mask(struct perf_record_record_cpu_map *mask,
tools/perf/util/synthetic-events.c int perf_event__synthesize_thread_map2(struct perf_tool *tool, · line 1205
Quoted 10 lines
tools/perf/util/synthetic-events.c +7
-1+static void synthesize_mask (struct perf_record_cpu_map_data *data,const struct perf_cpu_map *map,int max) { -int i; +int idx; +struct perf_cpu cpu; + +/* Due to padding, the 4bytes per entry mask variant is always smaller. */ + data->mask32_data.nr =BITS_TO_U32 (max); + data->mask32_data.long_size =4 ;
ok, so we go always with 32 bit version
tools/perf/util/synthetic-events.c int perf_event__synthesize_thread_map2(struct perf_tool *tool, · line 1209
Quoted 6 lines
tools/perf/util/synthetic-events.c +3
-2- mask->nr =BITS_TO_LONGS (max); - mask->long_size =sizeof (long ); +perf_cpu_map__for_each_cpu (cpu, idx, map) { +int bit_word = cpu.cpu /32 ; +__u32 bit_mask =1U << (cpu.cpu &31 );
set_bit uses (nr % 32), but I guess it does not matter
jirka
Quoted 51 lines
tools/perf/util/synthetic-events.c +7
-7-for (i =0 ; i <perf_cpu_map__nr (map); i++) -set_bit (perf_cpu_map__cpu (map, i).cpu, mask->mask); + data->mask32_data.mask[bit_word] |= bit_mask; + } }static size_t cpus_size (const struct perf_cpu_map *map)@@ -1214,7 +1220,7 @@ static size_t cpus_size(const struct perf_cpu_map *map)static size_t mask_size (const struct perf_cpu_map *map,int *max) { *max =perf_cpu_map__max (map).cpu; -return sizeof (struct perf_record_record_cpu_map ) +BITS_TO_LONGS (*max) *sizeof (long ); +return sizeof (struct perf_record_mask_cpu_map32 ) +BITS_TO_U32 (*max) *sizeof (__u32); } static void *cpu_map_data__alloc(const struct perf_cpu_map *map, size_t *size,@@ -1247,7 +1253,7 @@ static void *cpu_map_data__alloc(const struct perf_cpu_map *map, size_t *size, *type =PERF_CPU_MAP__MASK ; } - *size +=sizeof (struct perf_record_cpu_map_data ); + *size +=sizeof (__u16);/* For perf_record_cpu_map_data.type. */ *size =PERF_ALIGN (*size,sizeof (u64));return zalloc (*size); }@@ -1260,10 +1266,10 @@ static void cpu_map_data__synthesize(struct perf_record_cpu_map_data *data,switch (type) {case PERF_CPU_MAP__CPUS : -synthesize_cpus ((struct cpu_map_entries *) data->data, map); +synthesize_cpus (data, map);break ;case PERF_CPU_MAP__MASK : -synthesize_mask ((struct perf_record_record_cpu_map *)data->data, map, max); +synthesize_mask (data, map, max);default :break ; }@@ -1271,7 +1277,7 @@ static void cpu_map_data__synthesize(struct perf_record_cpu_map_data *data,static struct perf_record_cpu_map *cpu_map_event__new (const struct perf_cpu_map *map) { -size_t size =sizeof (struct perf_record_cpu_map ); +size_t size =sizeof (struct perf_event_header );struct perf_record_cpu_map *event;int max;u16 type;Signature · 1 line
2.36.1.476.g0c4daa206d-goog
Ian Rogers
reviewerRe: [PATCH v2 4/6] perf cpumap: Fix alignment for masks in event encoding
Quoted 49 lines · Jiri Olsa
Quoted 25 lines · Ian Rogers
A mask encoding of a cpu map is laid out as:
u16 nr
u16long_size
unsigned long mask[];
However, the mask may be 8-byte aligned meaning there is a 4-byte pad afterlong_size. This means 32-bit and 64-bit builds see the mask as being at different offsets. On top of this the structure is in the byte data[] encoded as:
u16 type
char data[]
This means the mask's struct isn't the required 4 or 8 byte aligned, but is offset by 2. Consequently the long reads and writes are causing undefined behavior as the alignment is broken.Fix the mask struct by creating explicit 32 and 64-bit variants, use a union to avoid data[] and casts; the struct must be packed so the layout matches the existing perf.data layout. Taking an address of a member of a packed struct breaks alignment so pass the packed
perf_record_cpu_map_datato functions, so they can access variables with the right alignment.As the 64-bit version has 4 bytes of padding, optimizing writing to only write the 32-bit version.
Signed-off-by Ian Rogers <[email protected]>SNIP
Quoted 19 lines
struct
perf_record_cpu_map{tools/perf/tests/cpumap.c +1
-1@@ -17,21 +17,23 @@ static int process_event_mask(struct perf_tool *tool __maybe_unused, struct machine *machine __maybe_unused) {struct perf_record_cpu_map *map_event = &event->cpu_map; -struct perf_record_record_cpu_map *mask;struct perf_record_cpu_map_data *data;struct perf_cpu_map *map;int i; +unsigned int long_size; data = &map_event->data;TEST_ASSERT_VAL ("wrong type" , data->type ==PERF_CPU_MAP__MASK );
Quoted 7 lines
tools/perf/tests/cpumap.c static int process_event_mask(struct perf_tool *tool __maybe_unused, · line 31
Quoted 5 lines
tools/perf/tests/cpumap.c +2
-2- mask = (struct perf_record_record_cpu_map *)data->data; + long_size = data->mask32_data.long_size; -TEST_ASSERT_VAL ("wrong nr" , mask->nr ==1 ); +TEST_ASSERT_VAL ("wrong long_size" , long_size ==4 || long_size ==8 );should we check here just for
long_size== 4 ?
We could given that we expect 4 byte versions only to be generated after this change. Unit tests act like documentation, so I am hesitant to remove the == 8 given it is a valid option.
Quoted 41 lines
SNIP
Quoted 39 lines
tools/perf/util/cpumap.c +30
-1@@ -22,54 +22,102 @@ static int max_node_num;*/ static int *cpunode_map; -static struct perf_cpu_map *cpu_map__from_entries (struct cpu_map_entries *cpus) +bool perf_record_cpu_map_data__test_bit (int i, +const struct perf_record_cpu_map_data *data) +{ +int bit_word32 = i /32 ; +__u32 bit_mask32 =1U << (i &31 ); +int bit_word64 = i /64 ; +__u64 bit_mask64 = ((__u64 )1 ) << (i &63 ); + +return (data->mask32_data.long_size ==4 ) + ? (bit_word32 < data->mask32_data.nr) && + (data->mask32_data.mask[bit_word32] & bit_mask32) !=0 + : (bit_word64 < data->mask64_data.nr) && + (data->mask64_data.mask[bit_word64] & bit_mask64) !=0 ; +} + +/* Read ith mask value from data into the given 64-bit sized bitmap */ +static void perf_record_cpu_map_data__read_one_mask (const struct perf_record_cpu_map_data *data, +int i,unsigned long *bitmap) +{ +#if __SIZEOF_LONG__ ==8 +if (data->mask32_data.long_size ==4 ) + bitmap[0 ] = data->mask32_data.mask[i]; +else + bitmap[0 ] = data->mask64_data.mask[i]; +#else +if (data->mask32_data.long_size ==4 ) { + bitmap[0 ] = data->mask32_data.mask[i]; + bitmap[1 ] =0 ; + }else { +#if __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__
Quoted 8 lines
tools/perf/util/cpumap.c static int max_node_num; · line 60
Quoted 6 lines
tools/perf/util/cpumap.c +6
-0+ bitmap[0 ] = (unsigned long )(data->mask64_data.mask[i] >>32 ); + bitmap[1 ] = (unsigned long )data->mask64_data.mask[i]; +#else + bitmap[0 ] = (unsigned long )data->mask64_data.mask[i]; + bitmap[1 ] = (unsigned long )(data->mask64_data.mask[i] >>32 ); +#endif should this be taken care of earlier by
perf_event__cpu_map_swap?
By the #ifdefs we're either in big or little endian 32-bit code here, so long is 32-bit but the bitmap is 64-bit. For little endian we pretty much just copy the bytes over, but for big endian the values are swapped so the littler value occurs in the low-bits. The bytes are still swapped from perf_event__cpu_map_swap but I was trying to make sure that the order of the words was correct.
Quoted 128 lines
Quoted 128 lines
tools/perf/util/cpumap.c +28
-15+ } +#endif +} +static struct perf_cpu_map *cpu_map__from_entries (const struct perf_record_cpu_map_data *data) {struct perf_cpu_map *map; - map =perf_cpu_map__empty_new (cpus->nr); + map =perf_cpu_map__empty_new (data->cpus_data.nr);if (map) {unsigned i; -for (i =0 ; i < cpus->nr; i++) { +for (i =0 ; i < data->cpus_data.nr; i++) {/* * Special treatment for -1, which is not real cpu number, * and we need to use (int) -1 to initialize map[i], * otherwise it would become 65535. */ -if (cpus->cpu[i] == (u16) -1 ) +if (data->cpus_data.cpu[i] == (u16) -1 ) map->map[i].cpu =-1 ;else - map->map[i].cpu = (int ) cpus->cpu[i]; + map->map[i].cpu = (int ) data->cpus_data.cpu[i]; } }return map; } -static struct perf_cpu_map *cpu_map__from_mask (struct perf_record_record_cpu_map *mask) +static struct perf_cpu_map *cpu_map__from_mask (const struct perf_record_cpu_map_data *data) { +DECLARE_BITMAP (local_copy,64 ); +int weight =0 , mask_nr = data->mask32_data.nr;struct perf_cpu_map *map; -int nr, nbits = mask->nr * mask->long_size *BITS_PER_BYTE ; - nr =bitmap_weight (mask->mask, nbits); +for (int i =0 ; i < mask_nr; i++) { +perf_record_cpu_map_data__read_one_mask (data, i, local_copy); + weight +=bitmap_weight (local_copy,64 ); + } + + map =perf_cpu_map__empty_new (weight); +if (!map) +return NULL ; - map =perf_cpu_map__empty_new (nr); -if (map) { -int cpu, i =0 ; +for (int i =0 , j =0 ; i < mask_nr; i++) { +int cpus_per_i = (i * data->mask32_data.long_size *BITS_PER_BYTE ); +int cpu; -for_each_set_bit (cpu, mask->mask, nbits) - map->map[i++].cpu = cpu; +perf_record_cpu_map_data__read_one_mask (data, i, local_copy); +for_each_set_bit (cpu, local_copy,64 ) + map->map[j++].cpu = cpu + cpus_per_i; }return map; } -struct perf_cpu_map *cpu_map__new_data (struct perf_record_cpu_map_data *data) +struct perf_cpu_map *cpu_map__new_data (const struct perf_record_cpu_map_data *data) {if (data->type ==PERF_CPU_MAP__CPUS ) -return cpu_map__from_entries ((struct cpu_map_entries *)data->data); +return cpu_map__from_entries (data);else -return cpu_map__from_mask ((struct perf_record_record_cpu_map *)data->data); +return cpu_map__from_mask (data); }size_t cpu_map__fprintf (struct perf_cpu_map *map,FILE *fp)tools/perf/util/cpumap.h +3
-1@@ -37,9 +37,11 @@ struct cpu_aggr_map {struct perf_record_cpu_map_data ; +bool perf_record_cpu_map_data__test_bit (int i,const struct perf_record_cpu_map_data *data); +struct perf_cpu_map *perf_cpu_map__empty_new (int nr); -struct perf_cpu_map *cpu_map__new_data (struct perf_record_cpu_map_data *data); +struct perf_cpu_map *cpu_map__new_data (const struct perf_record_cpu_map_data *data);size_t cpu_map__snprint (struct perf_cpu_map *map,char *buf,size_t size);size_t cpu_map__snprint_mask (struct perf_cpu_map *map,char *buf,size_t size);size_t cpu_map__fprintf (struct perf_cpu_map *map,FILE *fp);tools/perf/util/session.c +4
-12@@ -914,30 +914,30 @@ static void perf_event__cpu_map_swap(union perf_event *event, bool sample_id_all __maybe_unused) {struct perf_record_cpu_map_data *data = &event->cpu_map.data; -struct cpu_map_entries *cpus; -struct perf_record_record_cpu_map *mask; -unsigned i; data->type =bswap_16 (data->type);switch (data->type) {case PERF_CPU_MAP__CPUS : - cpus = (struct cpu_map_entries *)data->data; - - cpus->nr =bswap_16 (cpus->nr); + data->cpus_data.nr =bswap_16 (data->cpus_data.nr); -for (i =0 ; i < cpus->nr; i++) - cpus->cpu[i] =bswap_16 (cpus->cpu[i]); +for (unsigned i =0 ; i < data->cpus_data.nr; i++) + data->cpus_data.cpu[i] =bswap_16 (data->cpus_data.cpu[i]);break ;case PERF_CPU_MAP__MASK : - mask = (struct perf_record_record_cpu_map *)data->data; - - mask->nr =bswap_16 (mask->nr); - mask->long_size =bswap_16 (mask->long_size); + data->mask32_data.long_size =bswap_16 (data->mask32_data.long_size);
Quoted 11 lines
tools/perf/util/session.c static void perf_event__cpu_map_swap(union perf_event *event, · line 935
Quoted 9 lines
tools/perf/util/session.c +6
-3-switch (mask->long_size) { -case 4 :mem_bswap_32 (&mask->mask, mask->nr);break ; -case 8 :mem_bswap_64 (&mask->mask, mask->nr);break ; +switch (data->mask32_data.long_size) { +case 4 : + data->mask32_data.nr =bswap_16 (data->mask32_data.nr); +for (unsigned i =0 ; i < data->mask32_data.nr; i++) + data->mask32_data.mask[i] =bswap_32 (data->mask32_data.mask[i]); +break ;why not use the mem_bswap_* functions?
The mem_bswap_* assume the arrays of 32 or 64-bit values are aligned, which isn't true here because of the packing. Passing the values results in undefined behavior.
Quoted 33 lines
looks like we never swap it completely, because we passed
mask->nrwhere should be the sizeQuoted 30 lines
tools/perf/util/session.c +5
-0+case 8 : + data->mask64_data.nr =bswap_16 (data->mask64_data.nr); +for (unsigned i =0 ; i < data->mask64_data.nr; i++) + data->mask64_data.mask[i] =bswap_64 (data->mask64_data.mask[i]); +break ;default :pr_err ("cpu_map swap: unsupported long size\n" ); }tools/perf/util/synthetic-events.c +3
-3@@ -1183,27 +1183,33 @@ int perf_event__synthesize_thread_map2(struct perf_tool *tool,return err; } -static void synthesize_cpus (struct cpu_map_entries *cpus, +static void synthesize_cpus (struct perf_record_cpu_map_data *data,const struct perf_cpu_map *map) {int i, map_nr =perf_cpu_map__nr (map); - cpus->nr = map_nr; + data->cpus_data.nr = map_nr;for (i =0 ; i < map_nr; i++) - cpus->cpu[i] =perf_cpu_map__cpu (map, i).cpu; + data->cpus_data.cpu[i] =perf_cpu_map__cpu (map, i).cpu; }
Quoted 13 lines
tools/perf/util/synthetic-events.c int perf_event__synthesize_thread_map2(struct perf_tool *tool, · line 1205
Quoted 11 lines
tools/perf/util/synthetic-events.c +7
-2-static void synthesize_mask (struct perf_record_record_cpu_map *mask, +static void synthesize_mask (struct perf_record_cpu_map_data *data,const struct perf_cpu_map *map,int max) { -int i; +int idx; +struct perf_cpu cpu; + +/* Due to padding, the 4bytes per entry mask variant is always smaller. */ + data->mask32_data.nr =BITS_TO_U32 (max); + data->mask32_data.long_size =4 ;ok, so we go always with 32 bit version
I couldn't come up with a condition that made the 64-bit version sensible :-)
Quoted 8 lines
tools/perf/util/synthetic-events.c int perf_event__synthesize_thread_map2(struct perf_tool *tool, · line 1209
Quoted 6 lines
tools/perf/util/synthetic-events.c +3
-2- mask->nr =BITS_TO_LONGS (max); - mask->long_size =sizeof (long ); +perf_cpu_map__for_each_cpu (cpu, idx, map) { +int bit_word = cpu.cpu /32 ; +__u32 bit_mask =1U << (cpu.cpu &31 );
set_bituses (nr % 32), but I guess it does not matter
Perhaps if nr were negative, better not thought about :-)
Thanks,
Ian
Quoted 54 lines
jirka
Quoted 51 lines
tools/perf/util/synthetic-events.c +7
-8-for (i =0 ; i <perf_cpu_map__nr (map); i++) -set_bit (perf_cpu_map__cpu (map, i).cpu, mask->mask); + data->mask32_data.mask[bit_word] |= bit_mask; + } }static size_t cpus_size (const struct perf_cpu_map *map)@@ -1214,7 +1220,7 @@ static size_t cpus_size(const struct perf_cpu_map *map)static size_t mask_size (const struct perf_cpu_map *map,int *max) { *max =perf_cpu_map__max (map).cpu; -return sizeof (struct perf_record_record_cpu_map ) +BITS_TO_LONGS (*max) *sizeof (long ); +return sizeof (struct perf_record_mask_cpu_map32 ) +BITS_TO_U32 (*max) *sizeof (__u32); } static void *cpu_map_data__alloc(const struct perf_cpu_map *map, size_t *size,@@ -1247,7 +1253,7 @@ static void *cpu_map_data__alloc(const struct perf_cpu_map *map, size_t *size, *type =PERF_CPU_MAP__MASK ; } - *size +=sizeof (struct perf_record_cpu_map_data ); + *size +=sizeof (__u16);/* For perf_record_cpu_map_data.type. */ *size =PERF_ALIGN (*size,sizeof (u64));return zalloc (*size); }@@ -1260,10 +1266,10 @@ static void cpu_map_data__synthesize(struct perf_record_cpu_map_data *data,switch (type) {case PERF_CPU_MAP__CPUS : -synthesize_cpus ((struct cpu_map_entries *) data->data, map); +synthesize_cpus (data, map);break ;case PERF_CPU_MAP__MASK : -synthesize_mask ((struct perf_record_record_cpu_map *)data->data, map, max); +synthesize_mask (data, map, max);default :break ; }@@ -1271,7 +1277,7 @@ static void cpu_map_data__synthesize(struct perf_record_cpu_map_data *data,static struct perf_record_cpu_map *cpu_map_event__new (const struct perf_cpu_map *map) { -size_t size =sizeof (struct perf_record_cpu_map ); +size_t size =sizeof (struct perf_event_header );struct perf_record_cpu_map *event;int max;u16 type; --2.36.1.476.g0c4daa206d-goog
Arnaldo Carvalho de Melo
maintainerRe: [PATCH v2 4/6] perf cpumap: Fix alignment for masks in event encoding
In reply to Ian RogersEm Tue, Jun 14, 2022 at 07:33:51AM -0700, Ian Rogers escreveu:
Quoted 75 lines
A mask encoding of a cpu map is laid out as:
u16 nr
u16long_size
unsigned long mask[];
However, the mask may be 8-byte aligned meaning there is a 4-byte pad afterlong_size. This means 32-bit and 64-bit builds see the mask as being at different offsets. On top of this the structure is in the byte data[] encoded as:
u16 type
char data[]
This means the mask's struct isn't the required 4 or 8 byte aligned, but is offset by 2. Consequently the long reads and writes are causing undefined behavior as the alignment is broken.Fix the mask struct by creating explicit 32 and 64-bit variants, use a union to avoid data[] and casts; the struct must be packed so the layout matches the existing perf.data layout. Taking an address of a member of a packed struct breaks alignment so pass the packed
perf_record_cpu_map_datato functions, so they can access variables with the right alignment.As the 64-bit version has 4 bytes of padding, optimizing writing to only write the 32-bit version.
Signed-off-by Ian Rogers <[email protected]>--- tools/lib/perf/include/perf/event.h | 36 +++++++++++-- tools/perf/tests/cpumap.c | 19 ++++--- tools/perf/util/cpumap.c | 80 +++++++++++++++++++++++------ tools/perf/util/cpumap.h | 4 +- tools/perf/util/session.c | 30 +++++------ tools/perf/util/synthetic-events.c | 34 +++++++----- 6 files changed, 143 insertions(+), 60 deletions(-)tools/lib/perf/include/perf/event.h +14
-3@@ -6,6 +6,7 @@#include <linux/types.h> #include <linux/limits.h> #include <linux/bpf.h> +#include <linux/compiler.h> #include <sys/types.h> /* pid_t */ #define event_contains (obj, mem) ((obj).header.size > offsetof(typeof(obj), mem))@@ -153,20 +154,47 @@ enum { PERF_CPU_MAP__MASK = 1, }; +/* +* Array encoding of a perf_cpu_map where nr is the number of entries in cpu[] +* and each entry is a value for a CPU in the map. +*/ struct cpu_map_entries {__u16 nr;__u16 cpu[]; }; -struct perf_record_record_cpu_map { +/* Bitmap encoding of a perf_cpu_map where bitmap entries are 32-bit. */ +struct perf_record_mask_cpu_map32 { +/* Number of mask values. */ __u16 nr; +/* Constant 4. */ __u16 long_size; -unsigned long mask[]; +/* Bitmap data. */ +__u32 mask[]; }; -struct perf_record_cpu_map_data { +/* Bitmap encoding of a perf_cpu_map where bitmap entries are 64-bit. */ +struct perf_record_mask_cpu_map64 { +/* Number of mask values. */
tools/lib/perf/include/perf/event.h enum { · line 188
Quoted 10 lines
tools/lib/perf/include/perf/event.h +10
-0+__u16 nr; +/* Constant 8. */ +__u16 long_size; +/* Legacy padding. */ +char __pad[4 ]; +/* Bitmap data. */ +__u64 mask[]; +}; + +struct __packed perf_record_cpu_map_data {
In various places I'm getting this:
[perfbuilder@five x-riscv]$ export BUILD_TARBALL=http://192.168.86.14/perf/perf-6.0.0-rc1.tar.xz
[perfbuilder@five x-riscv]$ time dm .
1 5.47 ubuntu:22.04-x-riscv64 : FAIL gcc version 11.2.0 (Ubuntu 11.2.0-16ubuntu1)
In file included from mmap.c:10:
/git/perf-6.0.0-rc1/tools/lib/perf/include/perf/event.h:190:34: error: packed attribute causes inefficient alignment for 'type' [-Werror=attributes]
190 | __u16 type;
| ^~~~
cc1: all warnings being treated as errors
In file included from util/event.h:12,
from builtin-diff.c:12:
/git/perf-6.0.0-rc1/tools/lib/perf/include/perf/event.h:190:34: error: packed attribute causes inefficient alignment for 'type' [-Werror=attributes]
190 | __u16 type;
| ^~~~
In file included from util/events_stats.h:6,
from util/evlist.h:12,
from builtin-evlist.c:11:
/git/perf-6.0.0-rc1/tools/lib/perf/include/perf/event.h:190:34: error: packed attribute causes inefficient alignment for 'type' [-Werror=attributes]
190 | __u16 type;
| ^~~~
So probably we need to disable this -Werror=attributes in some architectures?
- Arnaldo
Quoted 347 lines
tools/lib/perf/include/perf/event.h +8
-1__u16 type; -char data[]; +union { +/* Used when type == PERF_CPU_MAP__CPUS. */ +struct cpu_map_entries cpus_data; +/* Used when type == PERF_CPU_MAP__MASK and long_size == 4. */ +struct perf_record_mask_cpu_map32 mask32_data; +/* Used when type == PERF_CPU_MAP__MASK and long_size == 8. */ +struct perf_record_mask_cpu_map64 mask64_data; + }; };struct perf_record_cpu_map {tools/perf/tests/cpumap.c +9
-10@@ -17,21 +17,23 @@ static int process_event_mask(struct perf_tool *tool __maybe_unused, struct machine *machine __maybe_unused) {struct perf_record_cpu_map *map_event = &event->cpu_map; -struct perf_record_record_cpu_map *mask;struct perf_record_cpu_map_data *data;struct perf_cpu_map *map;int i; +unsigned int long_size; data = &map_event->data;TEST_ASSERT_VAL ("wrong type" , data->type ==PERF_CPU_MAP__MASK ); - mask = (struct perf_record_record_cpu_map *)data->data; + long_size = data->mask32_data.long_size; -TEST_ASSERT_VAL ("wrong nr" , mask->nr ==1 ); +TEST_ASSERT_VAL ("wrong long_size" , long_size ==4 || long_size ==8 ); + +TEST_ASSERT_VAL ("wrong nr" , data->mask32_data.nr ==1 );for (i =0 ; i <20 ; i++) { -TEST_ASSERT_VAL ("wrong cpu" ,test_bit (i, mask->mask)); +TEST_ASSERT_VAL ("wrong cpu" ,perf_record_cpu_map_data__test_bit (i, data)); } map =cpu_map__new_data (data);@@ -51,7 +53,6 @@ static int process_event_cpus(struct perf_tool *tool __maybe_unused, struct machine *machine __maybe_unused) {struct perf_record_cpu_map *map_event = &event->cpu_map; -struct cpu_map_entries *cpus;struct perf_record_cpu_map_data *data;struct perf_cpu_map *map;@@ -59,11 +60,9 @@ static int process_event_cpus(struct perf_tool *tool __maybe_unused,TEST_ASSERT_VAL ("wrong type" , data->type ==PERF_CPU_MAP__CPUS ); - cpus = (struct cpu_map_entries *)data->data; - -TEST_ASSERT_VAL ("wrong nr" , cpus->nr ==2 ); -TEST_ASSERT_VAL ("wrong cpu" , cpus->cpu[0 ] ==1 ); -TEST_ASSERT_VAL ("wrong cpu" , cpus->cpu[1 ] ==256 ); +TEST_ASSERT_VAL ("wrong nr" , data->cpus_data.nr ==2 ); +TEST_ASSERT_VAL ("wrong cpu" , data->cpus_data.cpu[0 ] ==1 ); +TEST_ASSERT_VAL ("wrong cpu" , data->cpus_data.cpu[1 ] ==256 ); map =cpu_map__new_data (data);TEST_ASSERT_VAL ("wrong nr" ,perf_cpu_map__nr (map) ==2 );tools/perf/util/cpumap.c +64
-16@@ -22,54 +22,102 @@ static int max_node_num;*/ static int *cpunode_map; -static struct perf_cpu_map *cpu_map__from_entries (struct cpu_map_entries *cpus) +bool perf_record_cpu_map_data__test_bit (int i, +const struct perf_record_cpu_map_data *data) +{ +int bit_word32 = i /32 ; +__u32 bit_mask32 =1U << (i &31 ); +int bit_word64 = i /64 ; +__u64 bit_mask64 = ((__u64 )1 ) << (i &63 ); + +return (data->mask32_data.long_size ==4 ) + ? (bit_word32 < data->mask32_data.nr) && + (data->mask32_data.mask[bit_word32] & bit_mask32) !=0 + : (bit_word64 < data->mask64_data.nr) && + (data->mask64_data.mask[bit_word64] & bit_mask64) !=0 ; +} + +/* Read ith mask value from data into the given 64-bit sized bitmap */ +static void perf_record_cpu_map_data__read_one_mask (const struct perf_record_cpu_map_data *data, +int i,unsigned long *bitmap) +{ +#if __SIZEOF_LONG__ ==8 +if (data->mask32_data.long_size ==4 ) + bitmap[0 ] = data->mask32_data.mask[i]; +else + bitmap[0 ] = data->mask64_data.mask[i]; +#else +if (data->mask32_data.long_size ==4 ) { + bitmap[0 ] = data->mask32_data.mask[i]; + bitmap[1 ] =0 ; + }else { +#if __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__ + bitmap[0 ] = (unsigned long )(data->mask64_data.mask[i] >>32 ); + bitmap[1 ] = (unsigned long )data->mask64_data.mask[i]; +#else + bitmap[0 ] = (unsigned long )data->mask64_data.mask[i]; + bitmap[1 ] = (unsigned long )(data->mask64_data.mask[i] >>32 ); +#endif + } +#endif +} +static struct perf_cpu_map *cpu_map__from_entries (const struct perf_record_cpu_map_data *data) {struct perf_cpu_map *map; - map =perf_cpu_map__empty_new (cpus->nr); + map =perf_cpu_map__empty_new (data->cpus_data.nr);if (map) {unsigned i; -for (i =0 ; i < cpus->nr; i++) { +for (i =0 ; i < data->cpus_data.nr; i++) {/* * Special treatment for -1, which is not real cpu number, * and we need to use (int) -1 to initialize map[i], * otherwise it would become 65535. */ -if (cpus->cpu[i] == (u16) -1 ) +if (data->cpus_data.cpu[i] == (u16) -1 ) map->map[i].cpu =-1 ;else - map->map[i].cpu = (int ) cpus->cpu[i]; + map->map[i].cpu = (int ) data->cpus_data.cpu[i]; } }return map; } -static struct perf_cpu_map *cpu_map__from_mask (struct perf_record_record_cpu_map *mask) +static struct perf_cpu_map *cpu_map__from_mask (const struct perf_record_cpu_map_data *data) { +DECLARE_BITMAP (local_copy,64 ); +int weight =0 , mask_nr = data->mask32_data.nr;struct perf_cpu_map *map; -int nr, nbits = mask->nr * mask->long_size *BITS_PER_BYTE ; - nr =bitmap_weight (mask->mask, nbits); +for (int i =0 ; i < mask_nr; i++) { +perf_record_cpu_map_data__read_one_mask (data, i, local_copy); + weight +=bitmap_weight (local_copy,64 ); + } + + map =perf_cpu_map__empty_new (weight); +if (!map) +return NULL ; - map =perf_cpu_map__empty_new (nr); -if (map) { -int cpu, i =0 ; +for (int i =0 , j =0 ; i < mask_nr; i++) { +int cpus_per_i = (i * data->mask32_data.long_size *BITS_PER_BYTE ); +int cpu; -for_each_set_bit (cpu, mask->mask, nbits) - map->map[i++].cpu = cpu; +perf_record_cpu_map_data__read_one_mask (data, i, local_copy); +for_each_set_bit (cpu, local_copy,64 ) + map->map[j++].cpu = cpu + cpus_per_i; }return map; } -struct perf_cpu_map *cpu_map__new_data (struct perf_record_cpu_map_data *data) +struct perf_cpu_map *cpu_map__new_data (const struct perf_record_cpu_map_data *data) {if (data->type ==PERF_CPU_MAP__CPUS ) -return cpu_map__from_entries ((struct cpu_map_entries *)data->data); +return cpu_map__from_entries (data);else -return cpu_map__from_mask ((struct perf_record_record_cpu_map *)data->data); +return cpu_map__from_mask (data); }size_t cpu_map__fprintf (struct perf_cpu_map *map,FILE *fp)tools/perf/util/cpumap.h +3
-1@@ -37,9 +37,11 @@ struct cpu_aggr_map {struct perf_record_cpu_map_data ; +bool perf_record_cpu_map_data__test_bit (int i,const struct perf_record_cpu_map_data *data); +struct perf_cpu_map *perf_cpu_map__empty_new (int nr); -struct perf_cpu_map *cpu_map__new_data (struct perf_record_cpu_map_data *data); +struct perf_cpu_map *cpu_map__new_data (const struct perf_record_cpu_map_data *data);size_t cpu_map__snprint (struct perf_cpu_map *map,char *buf,size_t size);size_t cpu_map__snprint_mask (struct perf_cpu_map *map,char *buf,size_t size);size_t cpu_map__fprintf (struct perf_cpu_map *map,FILE *fp);tools/perf/util/session.c +15
-15@@ -914,30 +914,30 @@ static void perf_event__cpu_map_swap(union perf_event *event, bool sample_id_all __maybe_unused) {struct perf_record_cpu_map_data *data = &event->cpu_map.data; -struct cpu_map_entries *cpus; -struct perf_record_record_cpu_map *mask; -unsigned i; data->type =bswap_16 (data->type);switch (data->type) {case PERF_CPU_MAP__CPUS : - cpus = (struct cpu_map_entries *)data->data; - - cpus->nr =bswap_16 (cpus->nr); + data->cpus_data.nr =bswap_16 (data->cpus_data.nr); -for (i =0 ; i < cpus->nr; i++) - cpus->cpu[i] =bswap_16 (cpus->cpu[i]); +for (unsigned i =0 ; i < data->cpus_data.nr; i++) + data->cpus_data.cpu[i] =bswap_16 (data->cpus_data.cpu[i]);break ;case PERF_CPU_MAP__MASK : - mask = (struct perf_record_record_cpu_map *)data->data; - - mask->nr =bswap_16 (mask->nr); - mask->long_size =bswap_16 (mask->long_size); + data->mask32_data.long_size =bswap_16 (data->mask32_data.long_size); -switch (mask->long_size) { -case 4 :mem_bswap_32 (&mask->mask, mask->nr);break ; -case 8 :mem_bswap_64 (&mask->mask, mask->nr);break ; +switch (data->mask32_data.long_size) { +case 4 : + data->mask32_data.nr =bswap_16 (data->mask32_data.nr); +for (unsigned i =0 ; i < data->mask32_data.nr; i++) + data->mask32_data.mask[i] =bswap_32 (data->mask32_data.mask[i]); +break ; +case 8 : + data->mask64_data.nr =bswap_16 (data->mask64_data.nr); +for (unsigned i =0 ; i < data->mask64_data.nr; i++) + data->mask64_data.mask[i] =bswap_64 (data->mask64_data.mask[i]); +break ;default :pr_err ("cpu_map swap: unsupported long size\n" ); }tools/perf/util/synthetic-events.c +20
-14@@ -1183,27 +1183,33 @@ int perf_event__synthesize_thread_map2(struct perf_tool *tool,return err; } -static void synthesize_cpus (struct cpu_map_entries *cpus, +static void synthesize_cpus (struct perf_record_cpu_map_data *data,const struct perf_cpu_map *map) {int i, map_nr =perf_cpu_map__nr (map); - cpus->nr = map_nr; + data->cpus_data.nr = map_nr;for (i =0 ; i < map_nr; i++) - cpus->cpu[i] =perf_cpu_map__cpu (map, i).cpu; + data->cpus_data.cpu[i] =perf_cpu_map__cpu (map, i).cpu; } -static void synthesize_mask (struct perf_record_record_cpu_map *mask, +static void synthesize_mask (struct perf_record_cpu_map_data *data,const struct perf_cpu_map *map,int max) { -int i; +int idx; +struct perf_cpu cpu; + +/* Due to padding, the 4bytes per entry mask variant is always smaller. */ + data->mask32_data.nr =BITS_TO_U32 (max); + data->mask32_data.long_size =4 ; - mask->nr =BITS_TO_LONGS (max); - mask->long_size =sizeof (long ); +perf_cpu_map__for_each_cpu (cpu, idx, map) { +int bit_word = cpu.cpu /32 ; +__u32 bit_mask =1U << (cpu.cpu &31 ); -for (i =0 ; i <perf_cpu_map__nr (map); i++) -set_bit (perf_cpu_map__cpu (map, i).cpu, mask->mask); + data->mask32_data.mask[bit_word] |= bit_mask; + } }static size_t cpus_size (const struct perf_cpu_map *map)@@ -1214,7 +1220,7 @@ static size_t cpus_size(const struct perf_cpu_map *map)static size_t mask_size (const struct perf_cpu_map *map,int *max) { *max =perf_cpu_map__max (map).cpu; -return sizeof (struct perf_record_record_cpu_map ) +BITS_TO_LONGS (*max) *sizeof (long ); +return sizeof (struct perf_record_mask_cpu_map32 ) +BITS_TO_U32 (*max) *sizeof (__u32); } static void *cpu_map_data__alloc(const struct perf_cpu_map *map, size_t *size,@@ -1247,7 +1253,7 @@ static void *cpu_map_data__alloc(const struct perf_cpu_map *map, size_t *size, *type =PERF_CPU_MAP__MASK ; } - *size +=sizeof (struct perf_record_cpu_map_data ); + *size +=sizeof (__u16);/* For perf_record_cpu_map_data.type. */ *size =PERF_ALIGN (*size,sizeof (u64));return zalloc (*size); }@@ -1260,10 +1266,10 @@ static void cpu_map_data__synthesize(struct perf_record_cpu_map_data *data,switch (type) {case PERF_CPU_MAP__CPUS : -synthesize_cpus ((struct cpu_map_entries *) data->data, map); +synthesize_cpus (data, map);break ;case PERF_CPU_MAP__MASK : -synthesize_mask ((struct perf_record_record_cpu_map *)data->data, map, max); +synthesize_mask (data, map, max);default :break ; }@@ -1271,7 +1277,7 @@ static void cpu_map_data__synthesize(struct perf_record_cpu_map_data *data,static struct perf_record_cpu_map *cpu_map_event__new (const struct perf_cpu_map *map) { -size_t size =sizeof (struct perf_record_cpu_map ); +size_t size =sizeof (struct perf_event_header );struct perf_record_cpu_map *event;int max;u16 type;Signature · 1 line
2.36.1.476.g0c4daa206d-goog
Signature · 1 line
- Arnaldo
Ian Rogers
reviewerRe: [PATCH v2 4/6] perf cpumap: Fix alignment for masks in event encoding
Quoted 110 lines · Arnaldo Carvalho de Melo
Em Tue, Jun 14, 2022 at 07:33:51AM -0700, Ian Rogers escreveu:
Quoted 75 lines
A mask encoding of a cpu map is laid out as:
u16 nr
u16long_size
unsigned long mask[];
However, the mask may be 8-byte aligned meaning there is a 4-byte pad afterlong_size. This means 32-bit and 64-bit builds see the mask as being at different offsets. On top of this the structure is in the byte data[] encoded as:
u16 type
char data[]
This means the mask's struct isn't the required 4 or 8 byte aligned, but is offset by 2. Consequently the long reads and writes are causing undefined behavior as the alignment is broken.Fix the mask struct by creating explicit 32 and 64-bit variants, use a union to avoid data[] and casts; the struct must be packed so the layout matches the existing perf.data layout. Taking an address of a member of a packed struct breaks alignment so pass the packed
perf_record_cpu_map_datato functions, so they can access variables with the right alignment.As the 64-bit version has 4 bytes of padding, optimizing writing to only write the 32-bit version.
Signed-off-by Ian Rogers <[email protected]>--- tools/lib/perf/include/perf/event.h | 36 +++++++++++-- tools/perf/tests/cpumap.c | 19 ++++--- tools/perf/util/cpumap.c | 80 +++++++++++++++++++++++------ tools/perf/util/cpumap.h | 4 +- tools/perf/util/session.c | 30 +++++------ tools/perf/util/synthetic-events.c | 34 +++++++----- 6 files changed, 143 insertions(+), 60 deletions(-)tools/lib/perf/include/perf/event.h +14
-3@@ -6,6 +6,7 @@#include <linux/types.h> #include <linux/limits.h> #include <linux/bpf.h> +#include <linux/compiler.h> #include <sys/types.h> /* pid_t */ #define event_contains (obj, mem) ((obj).header.size > offsetof(typeof(obj), mem))@@ -153,20 +154,47 @@ enum { PERF_CPU_MAP__MASK = 1, }; +/* +* Array encoding of a perf_cpu_map where nr is the number of entries in cpu[] +* and each entry is a value for a CPU in the map. +*/ struct cpu_map_entries {__u16 nr;__u16 cpu[]; }; -struct perf_record_record_cpu_map { +/* Bitmap encoding of a perf_cpu_map where bitmap entries are 32-bit. */ +struct perf_record_mask_cpu_map32 { +/* Number of mask values. */ __u16 nr; +/* Constant 4. */ __u16 long_size; -unsigned long mask[]; +/* Bitmap data. */ +__u32 mask[]; }; -struct perf_record_cpu_map_data { +/* Bitmap encoding of a perf_cpu_map where bitmap entries are 64-bit. */ +struct perf_record_mask_cpu_map64 { +/* Number of mask values. */ tools/lib/perf/include/perf/event.h enum { · line 188
Quoted 10 lines
tools/lib/perf/include/perf/event.h +10
-0+__u16 nr; +/* Constant 8. */ +__u16 long_size; +/* Legacy padding. */ +char __pad[4 ]; +/* Bitmap data. */ +__u64 mask[]; +}; + +struct __packed perf_record_cpu_map_data {In various places I'm getting this:
[perfbuilder@five x-riscv]$ export BUILD_TARBALL=http://192.168.86.14/perf/perf-6.0.0-rc1.tar.xz
[perfbuilder@five x-riscv]$ time dm .
1 5.47 ubuntu:22.04-x-riscv64 : FAIL gcc version 11.2.0 (Ubuntu 11.2.0-16ubuntu1)
In file included from mmap.c:10:
/git/perf-6.0.0-rc1/tools/lib/perf/include/perf/event.h:190:34: error: packed attribute causes inefficient alignment for 'type' [-Werror=attributes]
190 |__u16type;
| ^~~~
cc1: all warnings being treated as errors
In file included fromutil/event.h:12,
from builtin-diff.c:12:
/git/perf-6.0.0-rc1/tools/lib/perf/include/perf/event.h:190:34: error: packed attribute causes inefficient alignment for 'type' [-Werror=attributes]
190 |__u16type;
| ^~~~
In file included fromutil/events_stats.h:6,
fromutil/evlist.h:12,
from builtin-evlist.c:11:
/git/perf-6.0.0-rc1/tools/lib/perf/include/perf/event.h:190:34: error: packed attribute causes inefficient alignment for 'type' [-Werror=attributes]
190 |__u16type;
| ^~~~
Quoted 2 lines
So probably we need to disable this -Werror=attributes in some architectures?
Looks like it. An inefficient load will be better than a SIGBUS.
Thanks,
Ian
Quoted 353 lines
- Arnaldo
Quoted 346 lines
tools/lib/perf/include/perf/event.h +8
-1__u16 type; -char data[]; +union { +/* Used when type == PERF_CPU_MAP__CPUS. */ +struct cpu_map_entries cpus_data; +/* Used when type == PERF_CPU_MAP__MASK and long_size == 4. */ +struct perf_record_mask_cpu_map32 mask32_data; +/* Used when type == PERF_CPU_MAP__MASK and long_size == 8. */ +struct perf_record_mask_cpu_map64 mask64_data; + }; };struct perf_record_cpu_map {tools/perf/tests/cpumap.c +9
-10@@ -17,21 +17,23 @@ static int process_event_mask(struct perf_tool *tool __maybe_unused, struct machine *machine __maybe_unused) {struct perf_record_cpu_map *map_event = &event->cpu_map; -struct perf_record_record_cpu_map *mask;struct perf_record_cpu_map_data *data;struct perf_cpu_map *map;int i; +unsigned int long_size; data = &map_event->data;TEST_ASSERT_VAL ("wrong type" , data->type ==PERF_CPU_MAP__MASK ); - mask = (struct perf_record_record_cpu_map *)data->data; + long_size = data->mask32_data.long_size; -TEST_ASSERT_VAL ("wrong nr" , mask->nr ==1 ); +TEST_ASSERT_VAL ("wrong long_size" , long_size ==4 || long_size ==8 ); + +TEST_ASSERT_VAL ("wrong nr" , data->mask32_data.nr ==1 );for (i =0 ; i <20 ; i++) { -TEST_ASSERT_VAL ("wrong cpu" ,test_bit (i, mask->mask)); +TEST_ASSERT_VAL ("wrong cpu" ,perf_record_cpu_map_data__test_bit (i, data)); } map =cpu_map__new_data (data);@@ -51,7 +53,6 @@ static int process_event_cpus(struct perf_tool *tool __maybe_unused, struct machine *machine __maybe_unused) {struct perf_record_cpu_map *map_event = &event->cpu_map; -struct cpu_map_entries *cpus;struct perf_record_cpu_map_data *data;struct perf_cpu_map *map;@@ -59,11 +60,9 @@ static int process_event_cpus(struct perf_tool *tool __maybe_unused,TEST_ASSERT_VAL ("wrong type" , data->type ==PERF_CPU_MAP__CPUS ); - cpus = (struct cpu_map_entries *)data->data; - -TEST_ASSERT_VAL ("wrong nr" , cpus->nr ==2 ); -TEST_ASSERT_VAL ("wrong cpu" , cpus->cpu[0 ] ==1 ); -TEST_ASSERT_VAL ("wrong cpu" , cpus->cpu[1 ] ==256 ); +TEST_ASSERT_VAL ("wrong nr" , data->cpus_data.nr ==2 ); +TEST_ASSERT_VAL ("wrong cpu" , data->cpus_data.cpu[0 ] ==1 ); +TEST_ASSERT_VAL ("wrong cpu" , data->cpus_data.cpu[1 ] ==256 ); map =cpu_map__new_data (data);TEST_ASSERT_VAL ("wrong nr" ,perf_cpu_map__nr (map) ==2 );tools/perf/util/cpumap.c +64
-16@@ -22,54 +22,102 @@ static int max_node_num;*/ static int *cpunode_map; -static struct perf_cpu_map *cpu_map__from_entries (struct cpu_map_entries *cpus) +bool perf_record_cpu_map_data__test_bit (int i, +const struct perf_record_cpu_map_data *data) +{ +int bit_word32 = i /32 ; +__u32 bit_mask32 =1U << (i &31 ); +int bit_word64 = i /64 ; +__u64 bit_mask64 = ((__u64 )1 ) << (i &63 ); + +return (data->mask32_data.long_size ==4 ) + ? (bit_word32 < data->mask32_data.nr) && + (data->mask32_data.mask[bit_word32] & bit_mask32) !=0 + : (bit_word64 < data->mask64_data.nr) && + (data->mask64_data.mask[bit_word64] & bit_mask64) !=0 ; +} + +/* Read ith mask value from data into the given 64-bit sized bitmap */ +static void perf_record_cpu_map_data__read_one_mask (const struct perf_record_cpu_map_data *data, +int i,unsigned long *bitmap) +{ +#if __SIZEOF_LONG__ ==8 +if (data->mask32_data.long_size ==4 ) + bitmap[0 ] = data->mask32_data.mask[i]; +else + bitmap[0 ] = data->mask64_data.mask[i]; +#else +if (data->mask32_data.long_size ==4 ) { + bitmap[0 ] = data->mask32_data.mask[i]; + bitmap[1 ] =0 ; + }else { +#if __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__ + bitmap[0 ] = (unsigned long )(data->mask64_data.mask[i] >>32 ); + bitmap[1 ] = (unsigned long )data->mask64_data.mask[i]; +#else + bitmap[0 ] = (unsigned long )data->mask64_data.mask[i]; + bitmap[1 ] = (unsigned long )(data->mask64_data.mask[i] >>32 ); +#endif + } +#endif +} +static struct perf_cpu_map *cpu_map__from_entries (const struct perf_record_cpu_map_data *data) {struct perf_cpu_map *map; - map =perf_cpu_map__empty_new (cpus->nr); + map =perf_cpu_map__empty_new (data->cpus_data.nr);if (map) {unsigned i; -for (i =0 ; i < cpus->nr; i++) { +for (i =0 ; i < data->cpus_data.nr; i++) {/* * Special treatment for -1, which is not real cpu number, * and we need to use (int) -1 to initialize map[i], * otherwise it would become 65535. */ -if (cpus->cpu[i] == (u16) -1 ) +if (data->cpus_data.cpu[i] == (u16) -1 ) map->map[i].cpu =-1 ;else - map->map[i].cpu = (int ) cpus->cpu[i]; + map->map[i].cpu = (int ) data->cpus_data.cpu[i]; } }return map; } -static struct perf_cpu_map *cpu_map__from_mask (struct perf_record_record_cpu_map *mask) +static struct perf_cpu_map *cpu_map__from_mask (const struct perf_record_cpu_map_data *data) { +DECLARE_BITMAP (local_copy,64 ); +int weight =0 , mask_nr = data->mask32_data.nr;struct perf_cpu_map *map; -int nr, nbits = mask->nr * mask->long_size *BITS_PER_BYTE ; - nr =bitmap_weight (mask->mask, nbits); +for (int i =0 ; i < mask_nr; i++) { +perf_record_cpu_map_data__read_one_mask (data, i, local_copy); + weight +=bitmap_weight (local_copy,64 ); + } + + map =perf_cpu_map__empty_new (weight); +if (!map) +return NULL ; - map =perf_cpu_map__empty_new (nr); -if (map) { -int cpu, i =0 ; +for (int i =0 , j =0 ; i < mask_nr; i++) { +int cpus_per_i = (i * data->mask32_data.long_size *BITS_PER_BYTE ); +int cpu; -for_each_set_bit (cpu, mask->mask, nbits) - map->map[i++].cpu = cpu; +perf_record_cpu_map_data__read_one_mask (data, i, local_copy); +for_each_set_bit (cpu, local_copy,64 ) + map->map[j++].cpu = cpu + cpus_per_i; }return map; } -struct perf_cpu_map *cpu_map__new_data (struct perf_record_cpu_map_data *data) +struct perf_cpu_map *cpu_map__new_data (const struct perf_record_cpu_map_data *data) {if (data->type ==PERF_CPU_MAP__CPUS ) -return cpu_map__from_entries ((struct cpu_map_entries *)data->data); +return cpu_map__from_entries (data);else -return cpu_map__from_mask ((struct perf_record_record_cpu_map *)data->data); +return cpu_map__from_mask (data); }size_t cpu_map__fprintf (struct perf_cpu_map *map,FILE *fp)tools/perf/util/cpumap.h +3
-1@@ -37,9 +37,11 @@ struct cpu_aggr_map {struct perf_record_cpu_map_data ; +bool perf_record_cpu_map_data__test_bit (int i,const struct perf_record_cpu_map_data *data); +struct perf_cpu_map *perf_cpu_map__empty_new (int nr); -struct perf_cpu_map *cpu_map__new_data (struct perf_record_cpu_map_data *data); +struct perf_cpu_map *cpu_map__new_data (const struct perf_record_cpu_map_data *data);size_t cpu_map__snprint (struct perf_cpu_map *map,char *buf,size_t size);size_t cpu_map__snprint_mask (struct perf_cpu_map *map,char *buf,size_t size);size_t cpu_map__fprintf (struct perf_cpu_map *map,FILE *fp);tools/perf/util/session.c +15
-15@@ -914,30 +914,30 @@ static void perf_event__cpu_map_swap(union perf_event *event, bool sample_id_all __maybe_unused) {struct perf_record_cpu_map_data *data = &event->cpu_map.data; -struct cpu_map_entries *cpus; -struct perf_record_record_cpu_map *mask; -unsigned i; data->type =bswap_16 (data->type);switch (data->type) {case PERF_CPU_MAP__CPUS : - cpus = (struct cpu_map_entries *)data->data; - - cpus->nr =bswap_16 (cpus->nr); + data->cpus_data.nr =bswap_16 (data->cpus_data.nr); -for (i =0 ; i < cpus->nr; i++) - cpus->cpu[i] =bswap_16 (cpus->cpu[i]); +for (unsigned i =0 ; i < data->cpus_data.nr; i++) + data->cpus_data.cpu[i] =bswap_16 (data->cpus_data.cpu[i]);break ;case PERF_CPU_MAP__MASK : - mask = (struct perf_record_record_cpu_map *)data->data; - - mask->nr =bswap_16 (mask->nr); - mask->long_size =bswap_16 (mask->long_size); + data->mask32_data.long_size =bswap_16 (data->mask32_data.long_size); -switch (mask->long_size) { -case 4 :mem_bswap_32 (&mask->mask, mask->nr);break ; -case 8 :mem_bswap_64 (&mask->mask, mask->nr);break ; +switch (data->mask32_data.long_size) { +case 4 : + data->mask32_data.nr =bswap_16 (data->mask32_data.nr); +for (unsigned i =0 ; i < data->mask32_data.nr; i++) + data->mask32_data.mask[i] =bswap_32 (data->mask32_data.mask[i]); +break ; +case 8 : + data->mask64_data.nr =bswap_16 (data->mask64_data.nr); +for (unsigned i =0 ; i < data->mask64_data.nr; i++) + data->mask64_data.mask[i] =bswap_64 (data->mask64_data.mask[i]); +break ;default :pr_err ("cpu_map swap: unsupported long size\n" ); }tools/perf/util/synthetic-events.c +20
-15@@ -1183,27 +1183,33 @@ int perf_event__synthesize_thread_map2(struct perf_tool *tool,return err; } -static void synthesize_cpus (struct cpu_map_entries *cpus, +static void synthesize_cpus (struct perf_record_cpu_map_data *data,const struct perf_cpu_map *map) {int i, map_nr =perf_cpu_map__nr (map); - cpus->nr = map_nr; + data->cpus_data.nr = map_nr;for (i =0 ; i < map_nr; i++) - cpus->cpu[i] =perf_cpu_map__cpu (map, i).cpu; + data->cpus_data.cpu[i] =perf_cpu_map__cpu (map, i).cpu; } -static void synthesize_mask (struct perf_record_record_cpu_map *mask, +static void synthesize_mask (struct perf_record_cpu_map_data *data,const struct perf_cpu_map *map,int max) { -int i; +int idx; +struct perf_cpu cpu; + +/* Due to padding, the 4bytes per entry mask variant is always smaller. */ + data->mask32_data.nr =BITS_TO_U32 (max); + data->mask32_data.long_size =4 ; - mask->nr =BITS_TO_LONGS (max); - mask->long_size =sizeof (long ); +perf_cpu_map__for_each_cpu (cpu, idx, map) { +int bit_word = cpu.cpu /32 ; +__u32 bit_mask =1U << (cpu.cpu &31 ); -for (i =0 ; i <perf_cpu_map__nr (map); i++) -set_bit (perf_cpu_map__cpu (map, i).cpu, mask->mask); + data->mask32_data.mask[bit_word] |= bit_mask; + } }static size_t cpus_size (const struct perf_cpu_map *map)@@ -1214,7 +1220,7 @@ static size_t cpus_size(const struct perf_cpu_map *map)static size_t mask_size (const struct perf_cpu_map *map,int *max) { *max =perf_cpu_map__max (map).cpu; -return sizeof (struct perf_record_record_cpu_map ) +BITS_TO_LONGS (*max) *sizeof (long ); +return sizeof (struct perf_record_mask_cpu_map32 ) +BITS_TO_U32 (*max) *sizeof (__u32); } static void *cpu_map_data__alloc(const struct perf_cpu_map *map, size_t *size,@@ -1247,7 +1253,7 @@ static void *cpu_map_data__alloc(const struct perf_cpu_map *map, size_t *size, *type =PERF_CPU_MAP__MASK ; } - *size +=sizeof (struct perf_record_cpu_map_data ); + *size +=sizeof (__u16);/* For perf_record_cpu_map_data.type. */ *size =PERF_ALIGN (*size,sizeof (u64));return zalloc (*size); }@@ -1260,10 +1266,10 @@ static void cpu_map_data__synthesize(struct perf_record_cpu_map_data *data,switch (type) {case PERF_CPU_MAP__CPUS : -synthesize_cpus ((struct cpu_map_entries *) data->data, map); +synthesize_cpus (data, map);break ;case PERF_CPU_MAP__MASK : -synthesize_mask ((struct perf_record_record_cpu_map *)data->data, map, max); +synthesize_mask (data, map, max);default :break ; }@@ -1271,7 +1277,7 @@ static void cpu_map_data__synthesize(struct perf_record_cpu_map_data *data,static struct perf_record_cpu_map *cpu_map_event__new (const struct perf_cpu_map *map) { -size_t size =sizeof (struct perf_record_cpu_map ); +size_t size =sizeof (struct perf_event_header );struct perf_record_cpu_map *event;int max;u16 type; --Quoted 1 line
2.36.1.476.g0c4daa206d-goog
--
- Arnaldo
Arnaldo Carvalho de Melo
maintainerRe: [PATCH v2 4/6] perf cpumap: Fix alignment for masks in event encoding
In reply to Arnaldo Carvalho de MeloEm Thu, Aug 18, 2022 at 06:50:00PM -0300, Arnaldo Carvalho de Melo escreveu:
Quoted 109 lines
Em Tue, Jun 14, 2022 at 07:33:51AM -0700, Ian Rogers escreveu:
Quoted 75 lines
A mask encoding of a cpu map is laid out as:
u16 nr
u16long_size
unsigned long mask[];
However, the mask may be 8-byte aligned meaning there is a 4-byte pad afterlong_size. This means 32-bit and 64-bit builds see the mask as being at different offsets. On top of this the structure is in the byte data[] encoded as:
u16 type
char data[]
This means the mask's struct isn't the required 4 or 8 byte aligned, but is offset by 2. Consequently the long reads and writes are causing undefined behavior as the alignment is broken.Fix the mask struct by creating explicit 32 and 64-bit variants, use a union to avoid data[] and casts; the struct must be packed so the layout matches the existing perf.data layout. Taking an address of a member of a packed struct breaks alignment so pass the packed
perf_record_cpu_map_datato functions, so they can access variables with the right alignment.As the 64-bit version has 4 bytes of padding, optimizing writing to only write the 32-bit version.
Signed-off-by Ian Rogers <[email protected]>--- tools/lib/perf/include/perf/event.h | 36 +++++++++++-- tools/perf/tests/cpumap.c | 19 ++++--- tools/perf/util/cpumap.c | 80 +++++++++++++++++++++++------ tools/perf/util/cpumap.h | 4 +- tools/perf/util/session.c | 30 +++++------ tools/perf/util/synthetic-events.c | 34 +++++++----- 6 files changed, 143 insertions(+), 60 deletions(-)tools/lib/perf/include/perf/event.h +14
-3@@ -6,6 +6,7 @@#include <linux/types.h> #include <linux/limits.h> #include <linux/bpf.h> +#include <linux/compiler.h> #include <sys/types.h> /* pid_t */ #define event_contains (obj, mem) ((obj).header.size > offsetof(typeof(obj), mem))@@ -153,20 +154,47 @@ enum { PERF_CPU_MAP__MASK = 1, }; +/* +* Array encoding of a perf_cpu_map where nr is the number of entries in cpu[] +* and each entry is a value for a CPU in the map. +*/ struct cpu_map_entries {__u16 nr;__u16 cpu[]; }; -struct perf_record_record_cpu_map { +/* Bitmap encoding of a perf_cpu_map where bitmap entries are 32-bit. */ +struct perf_record_mask_cpu_map32 { +/* Number of mask values. */ __u16 nr; +/* Constant 4. */ __u16 long_size; -unsigned long mask[]; +/* Bitmap data. */ +__u32 mask[]; }; -struct perf_record_cpu_map_data { +/* Bitmap encoding of a perf_cpu_map where bitmap entries are 64-bit. */ +struct perf_record_mask_cpu_map64 { +/* Number of mask values. */ tools/lib/perf/include/perf/event.h enum { · line 188
Quoted 10 lines
tools/lib/perf/include/perf/event.h +10
-0+__u16 nr; +/* Constant 8. */ +__u16 long_size; +/* Legacy padding. */ +char __pad[4 ]; +/* Bitmap data. */ +__u64 mask[]; +}; + +struct __packed perf_record_cpu_map_data {In various places I'm getting this:
[perfbuilder@five x-riscv]$ export BUILD_TARBALL=http://192.168.86.14/perf/perf-6.0.0-rc1.tar.xz
[perfbuilder@five x-riscv]$ time dm .
1 5.47 ubuntu:22.04-x-riscv64 : FAIL gcc version 11.2.0 (Ubuntu 11.2.0-16ubuntu1)
In file included from mmap.c:10:
/git/perf-6.0.0-rc1/tools/lib/perf/include/perf/event.h:190:34: error: packed attribute causes inefficient alignment for 'type' [-Werror=attributes]
190 |__u16type;
| ^~~~
cc1: all warnings being treated as errors
In file included fromutil/event.h:12,
from builtin-diff.c:12:
/git/perf-6.0.0-rc1/tools/lib/perf/include/perf/event.h:190:34: error: packed attribute causes inefficient alignment for 'type' [-Werror=attributes]
190 |__u16type;
| ^~~~
In file included fromutil/events_stats.h:6,
fromutil/evlist.h:12,
from builtin-evlist.c:11:
/git/perf-6.0.0-rc1/tools/lib/perf/include/perf/event.h:190:34: error: packed attribute causes inefficient alignment for 'type' [-Werror=attributes]
190 |__u16type;
| ^~~~
Quoted 2 lines
So probably we need to disable this -Werror=attributes in some architectures?
Slapped this there:
#pragma GCC diagnostic push
#pragma GCC diagnostic ignored "-Wpacked"
#pragma GCC diagnostic ignored "-Wattributes"
struct __packed perf_record_cpu_map_data {
__u16 type;
union {
/* Used when type == PERF_CPU_MAP__CPUS. */
struct cpu_map_entries cpus_data;
/* Used when type == PERF_CPU_MAP__MASK and long_size == 4. */
struct perf_record_mask_cpu_map32 mask32_data;
/* Used when type == PERF_CPU_MAP__MASK and long_size == 8. */
struct perf_record_mask_cpu_map64 mask64_data;
};
};#pragma GCC diagnostic pop
Ian Rogers
reviewerRe: [PATCH v2 4/6] perf cpumap: Fix alignment for masks in event encoding
Quoted 132 lines · Arnaldo Carvalho de Melo
Em Thu, Aug 18, 2022 at 06:50:00PM -0300, Arnaldo Carvalho de Melo escreveu:
Quoted 109 lines
Em Tue, Jun 14, 2022 at 07:33:51AM -0700, Ian Rogers escreveu:
Quoted 75 lines
A mask encoding of a cpu map is laid out as:
u16 nr
u16long_size
unsigned long mask[];
However, the mask may be 8-byte aligned meaning there is a 4-byte pad afterlong_size. This means 32-bit and 64-bit builds see the mask as being at different offsets. On top of this the structure is in the byte data[] encoded as:
u16 type
char data[]
This means the mask's struct isn't the required 4 or 8 byte aligned, but is offset by 2. Consequently the long reads and writes are causing undefined behavior as the alignment is broken.Fix the mask struct by creating explicit 32 and 64-bit variants, use a union to avoid data[] and casts; the struct must be packed so the layout matches the existing perf.data layout. Taking an address of a member of a packed struct breaks alignment so pass the packed
perf_record_cpu_map_datato functions, so they can access variables with the right alignment.As the 64-bit version has 4 bytes of padding, optimizing writing to only write the 32-bit version.
Signed-off-by Ian Rogers <[email protected]>--- tools/lib/perf/include/perf/event.h | 36 +++++++++++-- tools/perf/tests/cpumap.c | 19 ++++--- tools/perf/util/cpumap.c | 80 +++++++++++++++++++++++------ tools/perf/util/cpumap.h | 4 +- tools/perf/util/session.c | 30 +++++------ tools/perf/util/synthetic-events.c | 34 +++++++----- 6 files changed, 143 insertions(+), 60 deletions(-)tools/lib/perf/include/perf/event.h +14
-3@@ -6,6 +6,7 @@#include <linux/types.h> #include <linux/limits.h> #include <linux/bpf.h> +#include <linux/compiler.h> #include <sys/types.h> /* pid_t */ #define event_contains (obj, mem) ((obj).header.size > offsetof(typeof(obj), mem))@@ -153,20 +154,47 @@ enum { PERF_CPU_MAP__MASK = 1, }; +/* +* Array encoding of a perf_cpu_map where nr is the number of entries in cpu[] +* and each entry is a value for a CPU in the map. +*/ struct cpu_map_entries {__u16 nr;__u16 cpu[]; }; -struct perf_record_record_cpu_map { +/* Bitmap encoding of a perf_cpu_map where bitmap entries are 32-bit. */ +struct perf_record_mask_cpu_map32 { +/* Number of mask values. */ __u16 nr; +/* Constant 4. */ __u16 long_size; -unsigned long mask[]; +/* Bitmap data. */ +__u32 mask[]; }; -struct perf_record_cpu_map_data { +/* Bitmap encoding of a perf_cpu_map where bitmap entries are 64-bit. */ +struct perf_record_mask_cpu_map64 { +/* Number of mask values. */ tools/lib/perf/include/perf/event.h enum { · line 188
Quoted 10 lines
tools/lib/perf/include/perf/event.h +10
-0+__u16 nr; +/* Constant 8. */ +__u16 long_size; +/* Legacy padding. */ +char __pad[4 ]; +/* Bitmap data. */ +__u64 mask[]; +}; + +struct __packed perf_record_cpu_map_data {In various places I'm getting this:
[perfbuilder@five x-riscv]$ export BUILD_TARBALL=http://192.168.86.14/perf/perf-6.0.0-rc1.tar.xz
[perfbuilder@five x-riscv]$ time dm .
1 5.47 ubuntu:22.04-x-riscv64 : FAIL gcc version 11.2.0 (Ubuntu 11.2.0-16ubuntu1)
In file included from mmap.c:10:
/git/perf-6.0.0-rc1/tools/lib/perf/include/perf/event.h:190:34: error: packed attribute causes inefficient alignment for 'type' [-Werror=attributes]
190 |__u16type;
| ^~~~
cc1: all warnings being treated as errors
In file included fromutil/event.h:12,
from builtin-diff.c:12:
/git/perf-6.0.0-rc1/tools/lib/perf/include/perf/event.h:190:34: error: packed attribute causes inefficient alignment for 'type' [-Werror=attributes]
190 |__u16type;
| ^~~~
In file included fromutil/events_stats.h:6,
fromutil/evlist.h:12,
from builtin-evlist.c:11:
/git/perf-6.0.0-rc1/tools/lib/perf/include/perf/event.h:190:34: error: packed attribute causes inefficient alignment for 'type' [-Werror=attributes]
190 |__u16type;
| ^~~~Quoted 2 lines
So probably we need to disable this -Werror=attributes in some architectures?
Slapped this there:
#pragma GCC diagnostic push
#pragma GCC diagnostic ignored "-Wpacked"
#pragma GCC diagnostic ignored "-Wattributes"struct __packed perf_record_cpu_map_data { __u16 type; union { /* Used when type == PERF_CPU_MAP__CPUS. */ struct cpu_map_entries cpus_data; /* Used when type == PERF_CPU_MAP__MASK and long_size == 4. */ struct perf_record_mask_cpu_map32 mask32_data; /* Used when type == PERF_CPU_MAP__MASK and long_size == 8. */ struct perf_record_mask_cpu_map64 mask64_data; }; };
Quoted 1 line
#pragma GCC diagnostic pop
Perhaps we should also carry a comment like:perf_record_cpu_map_data is packed as unfortunately an earlier version had unaligned data and we wish to retain file format compatibility.
Thanks,
Ian
Re: [PATCH v2 4/6] perf cpumap: Fix alignment for masks in event encoding
Quoted 134 lines · Ian Rogers
Quoted 132 lines · Arnaldo Carvalho de Melo
Em Thu, Aug 18, 2022 at 06:50:00PM -0300, Arnaldo Carvalho de Melo escreveu:
Quoted 109 lines
Em Tue, Jun 14, 2022 at 07:33:51AM -0700, Ian Rogers escreveu:
Quoted 75 lines
A mask encoding of a cpu map is laid out as:
u16 nr
u16long_size
unsigned long mask[];
However, the mask may be 8-byte aligned meaning there is a 4-byte pad afterlong_size. This means 32-bit and 64-bit builds see the mask as being at different offsets. On top of this the structure is in the byte data[] encoded as:
u16 type
char data[]
This means the mask's struct isn't the required 4 or 8 byte aligned, but is offset by 2. Consequently the long reads and writes are causing undefined behavior as the alignment is broken.Fix the mask struct by creating explicit 32 and 64-bit variants, use a union to avoid data[] and casts; the struct must be packed so the layout matches the existing perf.data layout. Taking an address of a member of a packed struct breaks alignment so pass the packed
perf_record_cpu_map_datato functions, so they can access variables with the right alignment.As the 64-bit version has 4 bytes of padding, optimizing writing to only write the 32-bit version.
Signed-off-by Ian Rogers <[email protected]>--- tools/lib/perf/include/perf/event.h | 36 +++++++++++-- tools/perf/tests/cpumap.c | 19 ++++--- tools/perf/util/cpumap.c | 80 +++++++++++++++++++++++------ tools/perf/util/cpumap.h | 4 +- tools/perf/util/session.c | 30 +++++------ tools/perf/util/synthetic-events.c | 34 +++++++----- 6 files changed, 143 insertions(+), 60 deletions(-)tools/lib/perf/include/perf/event.h +14
-3@@ -6,6 +6,7 @@#include <linux/types.h> #include <linux/limits.h> #include <linux/bpf.h> +#include <linux/compiler.h> #include <sys/types.h> /* pid_t */ #define event_contains (obj, mem) ((obj).header.size > offsetof(typeof(obj), mem))@@ -153,20 +154,47 @@ enum { PERF_CPU_MAP__MASK = 1, }; +/* +* Array encoding of a perf_cpu_map where nr is the number of entries in cpu[] +* and each entry is a value for a CPU in the map. +*/ struct cpu_map_entries {__u16 nr;__u16 cpu[]; }; -struct perf_record_record_cpu_map { +/* Bitmap encoding of a perf_cpu_map where bitmap entries are 32-bit. */ +struct perf_record_mask_cpu_map32 { +/* Number of mask values. */ __u16 nr; +/* Constant 4. */ __u16 long_size; -unsigned long mask[]; +/* Bitmap data. */ +__u32 mask[]; }; -struct perf_record_cpu_map_data { +/* Bitmap encoding of a perf_cpu_map where bitmap entries are 64-bit. */ +struct perf_record_mask_cpu_map64 { +/* Number of mask values. */ tools/lib/perf/include/perf/event.h enum { · line 188
Quoted 10 lines
tools/lib/perf/include/perf/event.h +10
-0+__u16 nr; +/* Constant 8. */ +__u16 long_size; +/* Legacy padding. */ +char __pad[4 ]; +/* Bitmap data. */ +__u64 mask[]; +}; + +struct __packed perf_record_cpu_map_data {In various places I'm getting this:
[perfbuilder@five x-riscv]$ export BUILD_TARBALL=http://192.168.86.14/perf/perf-6.0.0-rc1.tar.xz
[perfbuilder@five x-riscv]$ time dm .
1 5.47 ubuntu:22.04-x-riscv64 : FAIL gcc version 11.2.0 (Ubuntu 11.2.0-16ubuntu1)
In file included from mmap.c:10:
/git/perf-6.0.0-rc1/tools/lib/perf/include/perf/event.h:190:34: error: packed attribute causes inefficient alignment for 'type' [-Werror=attributes]
190 |__u16type;
| ^~~~
cc1: all warnings being treated as errors
In file included fromutil/event.h:12,
from builtin-diff.c:12:
/git/perf-6.0.0-rc1/tools/lib/perf/include/perf/event.h:190:34: error: packed attribute causes inefficient alignment for 'type' [-Werror=attributes]
190 |__u16type;
| ^~~~
In file included fromutil/events_stats.h:6,
fromutil/evlist.h:12,
from builtin-evlist.c:11:
/git/perf-6.0.0-rc1/tools/lib/perf/include/perf/event.h:190:34: error: packed attribute causes inefficient alignment for 'type' [-Werror=attributes]
190 |__u16type;
| ^~~~Quoted 2 lines
So probably we need to disable this -Werror=attributes in some architectures?
Slapped this there:
#pragma GCC diagnostic push
#pragma GCC diagnostic ignored "-Wpacked"
#pragma GCC diagnostic ignored "-Wattributes"struct __packed perf_record_cpu_map_data { __u16 type; union { /* Used when type == PERF_CPU_MAP__CPUS. */ struct cpu_map_entries cpus_data; /* Used when type == PERF_CPU_MAP__MASK and long_size == 4. */ struct perf_record_mask_cpu_map32 mask32_data; /* Used when type == PERF_CPU_MAP__MASK and long_size == 8. */ struct perf_record_mask_cpu_map64 mask64_data; }; };
Quoted 5 lines
Quoted 1 line
#pragma GCC diagnostic pop
Perhaps we should also carry a comment like:
perf_record_cpu_map_datais packed as unfortunately an earlier version had unaligned data and we wish to retain file format compatibility.
Thanks, I'll add it.
- Arnaldo
Re: [PATCH v2 4/6] perf cpumap: Fix alignment for masks in event encoding
In reply to Ian RogersOn Tue, Jun 14, 2022 at 07:33:51AM -0700, Ian Rogers wrote:
Hi Ian,
Although it is committed, still have a question.
tools/lib/perf/include/perf/event.h line 9
Quoted 8 lines
index
e7758707cadd..d2d32589758a100644tools/lib/perf/include/perf/event.h +1
-0@@ -6,6 +6,7 @@#include <linux/types.h> #include <linux/limits.h> #include <linux/bpf.h> +#include <linux/compiler.h>
Is it correct approach to include it into user-exposed headers? AFAICT headers_install.sh strips #include <linux/compiler.h> and compiler*.h itself do not get installed with make headers_install.
[...]
tools/lib/perf/include/perf/event.h enum { · line 188
Quoted 1 line
tools/lib/perf/include/perf/event.h +1
-0+struct __packed perf_record_cpu_map_data {
And it is only needed to pull __packed macro, right?
Thanks!
Ian Rogers
reviewerRe: [PATCH v2 4/6] perf cpumap: Fix alignment for masks in event encoding
Quoted 20 lines · Alexander Gordeev
On Tue, Jun 14, 2022 at 07:33:51AM -0700, Ian Rogers wrote:
Hi Ian,
Although it is committed, still have a question.
tools/lib/perf/include/perf/event.h line 9
Quoted 8 lines
index
e7758707cadd..d2d32589758a100644tools/lib/perf/include/perf/event.h +1
-0@@ -6,6 +6,7 @@#include <linux/types.h> #include <linux/limits.h> #include <linux/bpf.h> +#include <linux/compiler.h> Is it correct approach to include it into user-exposed headers? AFAICT
headers_install.shstrips #include <linux/compiler.h> and compiler*.h itself do not get installed with makeheaders_install.
Quoted 7 lines
[...]
Quoted 1 line
+struct
__packedperf_record_cpu_map_data{And it is only needed to pull
__packedmacro, right?Thanks!
Hi Alexander,
I can see your point about compiler.h, it is true that it is just being used for the __packed definition. Why don't you write your proposed change as a patch and send it to LKML? libperf is more of an experimental library than a stable API. We don't currently have any build tests for things outside of the kernel tree.
Thanks,
Ian
Ian Rogers
reviewerPATCH v2 5/6 perf events: Prefer union over variable length array
In reply to Ian RogersIt is possible for casts to introduce alignment issues, prefer a union for perf_record_event_update.
--- tools/lib/perf/include/perf/event.h | 11 ++++++++++- tools/perf/tests/event_update.c | 14 ++++---------- tools/perf/util/header.c | 24 ++++++++---------------- tools/perf/util/synthetic-events.c | 12 +++++------- 4 files changed, 27 insertions(+), 34 deletions(-)
tools/lib/perf/include/perf/event.h +10 -1
@@ -221,7 +221,16 @@ struct perf_record_event_update {struct perf_event_header header;__u64 type;__u64 id; -char data[]; +union { +/* Used when type == PERF_EVENT_UPDATE__SCALE. */ +struct perf_record_event_update_scale scale; +/* Used when type == PERF_EVENT_UPDATE__UNIT. */ +char unit[0 ]; +/* Used when type == PERF_EVENT_UPDATE__NAME. */ +char name[0 ]; +/* Used when type == PERF_EVENT_UPDATE__CPUS. */ +struct perf_record_event_update_cpus cpus; + }; };#define MAX_EVENT_NAME 64
tools/perf/tests/event_update.c +4 -10
@@ -21,7 +21,7 @@ static int process_event_unit(struct perf_tool *tool __maybe_unused,TEST_ASSERT_VAL ("wrong id" , ev->id ==123 );TEST_ASSERT_VAL ("wrong id" , ev->type ==PERF_EVENT_UPDATE__UNIT ); -TEST_ASSERT_VAL ("wrong unit" , !strcmp (ev->data,"KRAVA" )); +TEST_ASSERT_VAL ("wrong unit" , !strcmp (ev->unit,"KRAVA" ));return 0 ; }
@@ -31,13 +31,10 @@ static int process_event_scale(struct perf_tool *tool __maybe_unused, struct machine *machine __maybe_unused) {struct perf_record_event_update *ev = (struct perf_record_event_update *)event; -struct perf_record_event_update_scale *ev_data; - - ev_data = (struct perf_record_event_update_scale *)ev->data;TEST_ASSERT_VAL ("wrong id" , ev->id ==123 );TEST_ASSERT_VAL ("wrong id" , ev->type ==PERF_EVENT_UPDATE__SCALE ); -TEST_ASSERT_VAL ("wrong scale" , ev_data->scale ==0.123 ); +TEST_ASSERT_VAL ("wrong scale" , ev->scale.scale ==0.123 );return 0 ; }
@@ -56,7 +53,7 @@ static int process_event_name(struct perf_tool *tool,TEST_ASSERT_VAL ("wrong id" , ev->id ==123 );TEST_ASSERT_VAL ("wrong id" , ev->type ==PERF_EVENT_UPDATE__NAME ); -TEST_ASSERT_VAL ("wrong name" , !strcmp (ev->data, tmp->name)); +TEST_ASSERT_VAL ("wrong name" , !strcmp (ev->name, tmp->name));return 0 ; }
@@ -66,12 +63,9 @@ static int process_event_cpus(struct perf_tool *tool __maybe_unused, struct machine *machine __maybe_unused) {struct perf_record_event_update *ev = (struct perf_record_event_update *)event; -struct perf_record_event_update_cpus *ev_data;struct perf_cpu_map *map; - ev_data = (struct perf_record_event_update_cpus *) ev->data; - - map =cpu_map__new_data (&ev_data->cpus); + map =cpu_map__new_data (&ev->cpus.cpus);TEST_ASSERT_VAL ("wrong id" , ev->id ==123 );TEST_ASSERT_VAL ("wrong type" , ev->type ==PERF_EVENT_UPDATE__CPUS );
tools/perf/util/header.c +8 -16
@@ -4264,8 +4264,6 @@ int perf_event__process_feature(struct perf_session *session,size_t perf_event__fprintf_event_update (union perf_event *event,FILE *fp) {struct perf_record_event_update *ev = &event->event_update; -struct perf_record_event_update_scale *ev_scale; -struct perf_record_event_update_cpus *ev_cpus;struct perf_cpu_map *map;size_t ret;
@@ -4273,20 +4271,18 @@ size_t perf_event__fprintf_event_update(union perf_event *event, FILE *fp)switch (ev->type) {case PERF_EVENT_UPDATE__SCALE : - ev_scale = (struct perf_record_event_update_scale *)ev->data; - ret +=fprintf (fp,"... scale: %f\n" , ev_scale->scale); + ret +=fprintf (fp,"... scale: %f\n" , ev->scale.scale);break ;case PERF_EVENT_UPDATE__UNIT : - ret +=fprintf (fp,"... unit: %s\n" , ev->data); + ret +=fprintf (fp,"... unit: %s\n" , ev->unit);break ;case PERF_EVENT_UPDATE__NAME : - ret +=fprintf (fp,"... name: %s\n" , ev->data); + ret +=fprintf (fp,"... name: %s\n" , ev->name);break ;case PERF_EVENT_UPDATE__CPUS : - ev_cpus = (struct perf_record_event_update_cpus *)ev->data; ret +=fprintf (fp,"... " ); - map =cpu_map__new_data (&ev_cpus->cpus); + map =cpu_map__new_data (&ev->cpus.cpus);if (map) ret +=cpu_map__fprintf (map, fp);else
@@ -4343,8 +4339,6 @@ int perf_event__process_event_update(struct perf_tool *tool __maybe_unused, struct evlist **pevlist) {struct perf_record_event_update *ev = &event->event_update; -struct perf_record_event_update_scale *ev_scale; -struct perf_record_event_update_cpus *ev_cpus;struct evlist *evlist;struct evsel *evsel;struct perf_cpu_map *map;
@@ -4361,19 +4355,17 @@ int perf_event__process_event_update(struct perf_tool *tool __maybe_unused,switch (ev->type) {case PERF_EVENT_UPDATE__UNIT :free ((char *)evsel->unit); - evsel->unit =strdup (ev->data); + evsel->unit =strdup (ev->unit);break ;case PERF_EVENT_UPDATE__NAME :free (evsel->name); - evsel->name =strdup (ev->data); + evsel->name =strdup (ev->name);break ;case PERF_EVENT_UPDATE__SCALE : - ev_scale = (struct perf_record_event_update_scale *)ev->data; - evsel->scale = ev_scale->scale; + evsel->scale = ev->scale.scale;break ;case PERF_EVENT_UPDATE__CPUS : - ev_cpus = (struct perf_record_event_update_cpus *)ev->data; - map =cpu_map__new_data (&ev_cpus->cpus); + map =cpu_map__new_data (&ev->cpus.cpus);if (map) {perf_cpu_map__put (evsel->core.own_cpus); evsel->core.own_cpus = map;
tools/perf/util/synthetic-events.c +5 -7
@@ -1848,7 +1848,7 @@ int perf_event__synthesize_event_update_unit(struct perf_tool *tool, struct evseif (ev ==NULL )return -ENOMEM ; -strlcpy (ev->data, evsel->unit, size +1 ); +strlcpy (ev->unit, evsel->unit, size +1 ); err =process (tool, (union perf_event *)ev,NULL ,NULL );free (ev);return err;
@@ -1865,8 +1865,7 @@ int perf_event__synthesize_event_update_scale(struct perf_tool *tool, struct evsif (ev ==NULL )return -ENOMEM ; - ev_data = (struct perf_record_event_update_scale *)ev->data; - ev_data->scale = evsel->scale; + ev->scale.scale = evsel->scale; err =process (tool, (union perf_event *)ev,NULL ,NULL );free (ev);return err;
@@ -1883,7 +1882,7 @@ int perf_event__synthesize_event_update_name(struct perf_tool *tool, struct evseif (ev ==NULL )return -ENOMEM ; -strlcpy (ev->data, evsel->name, len +1 ); +strlcpy (ev->name, evsel->name, len +1 ); err =process (tool, (union perf_event *)ev,NULL ,NULL );free (ev);return err;
@@ -1892,7 +1891,7 @@ int perf_event__synthesize_event_update_name(struct perf_tool *tool, struct evseint perf_event__synthesize_event_update_cpus (struct perf_tool *tool,struct evsel *evsel,perf_event__handler_t process) { -size_t size =sizeof (struct perf_record_event_update ); +size_t size =sizeof (struct perf_event_header ) +sizeof (u64) +sizeof (u64);struct perf_record_event_update *ev;int max, err;u16 type;
@@ -1909,8 +1908,7 @@ int perf_event__synthesize_event_update_cpus(struct perf_tool *tool, struct evse ev->type =PERF_EVENT_UPDATE__CPUS ; ev->id = evsel->core.id[0 ]; -cpu_map_data__synthesize ((struct perf_record_cpu_map_data *)ev->data, - evsel->core.own_cpus, type, max); +cpu_map_data__synthesize (&ev->cpus.cpus, evsel->core.own_cpus, type, max); err =process (tool, (union perf_event *)ev,NULL ,NULL );free (ev);
Signature · 1 line
2.36.1.476.g0c4daa206d-goog
Ian Rogers
reviewerPATCH v2 6/6 perf cpumap: Add range data encoding
In reply to Ian RogersOften cpumaps encode a range of all CPUs, add a compact encoding that doesn't require a bit mask or list of all CPUs.
--- tools/lib/perf/include/perf/event.h | 14 +++ tools/perf/tests/cpumap.c | 52 ++++++++-- tools/perf/util/cpumap.c | 31 +++++- tools/perf/util/session.c | 5 + tools/perf/util/synthetic-events.c | 151 ++++++++++++++-------------- 5 files changed, 166 insertions(+), 87 deletions(-)
tools/lib/perf/include/perf/event.h +14 -0
@@ -152,6 +152,7 @@ struct perf_record_header_attr {enum {PERF_CPU_MAP__CPUS =0 ,PERF_CPU_MAP__MASK =1 , +PERF_CPU_MAP__RANGE_CPUS =2 , };/*
@@ -185,6 +186,17 @@ struct perf_record_mask_cpu_map64 {__u64 mask[]; }; +/* +* An encoding of a CPU map for a range starting at start_cpu through to +* end_cpu. If any_cpu is 1, an any CPU (-1) value (aka dummy value) is present. +*/ +struct perf_record_range_cpu_map { +__u8 any_cpu; +__u8 __pad; +__u16 start_cpu; +__u16 end_cpu; +}; +struct __packed perf_record_cpu_map_data {__u16 type;union {
@@ -194,6 +206,8 @@ struct __packed perf_record_cpu_map_data {struct perf_record_mask_cpu_map32 mask32_data;/* Used when type == PERF_CPU_MAP__MASK and long_size == 8. */ struct perf_record_mask_cpu_map64 mask64_data; +/* Used when type == PERF_CPU_MAP__RANGE_CPUS. */ +struct perf_record_range_cpu_map range_cpu_data; }; };
tools/perf/tests/cpumap.c +43 -9
91 lines not on this page · raw message
tools/perf/util/cpumap.c +29 -2
@@ -112,12 +112,39 @@ static struct perf_cpu_map *cpu_map__from_mask(const struct perf_record_cpu_map_ } +static struct perf_cpu_map *cpu_map__from_range (const struct perf_record_cpu_map_data *data) +{ +struct perf_cpu_map *map; +unsigned int i =0 ; + + map =perf_cpu_map__empty_new (data->range_cpu_data.end_cpu - + data->range_cpu_data.start_cpu +1 + data->range_cpu_data.any_cpu); +if (!map) +return NULL ; + +if (data->range_cpu_data.any_cpu) + map->map[i++].cpu =-1 ; + +for (int cpu = data->range_cpu_data.start_cpu; cpu <= data->range_cpu_data.end_cpu; + i++, cpu++) + map->map[i].cpu = cpu; + +return map; +} +struct perf_cpu_map *cpu_map__new_data (const struct perf_record_cpu_map_data *data) { -if (data->type ==PERF_CPU_MAP__CPUS ) +switch (data->type) { +case PERF_CPU_MAP__CPUS :return cpu_map__from_entries (data); -else +case PERF_CPU_MAP__MASK :return cpu_map__from_mask (data); +case PERF_CPU_MAP__RANGE_CPUS : +return cpu_map__from_range (data); +default : +pr_err ("cpu_map__new_data unknown type %d\n" , data->type); +return NULL ; + } }size_t cpu_map__fprintf (struct perf_cpu_map *map,FILE *fp)
tools/perf/util/session.c +5 -0
@@ -941,6 +941,11 @@ static void perf_event__cpu_map_swap(union perf_event *event,default :pr_err ("cpu_map swap: unsupported long size\n" ); } +break ; +case PERF_CPU_MAP__RANGE_CPUS : + data->range_cpu_data.start_cpu =bswap_16 (data->range_cpu_data.start_cpu); + data->range_cpu_data.end_cpu =bswap_16 (data->range_cpu_data.end_cpu); +break ;default :break ; }
tools/perf/util/synthetic-events.c +75 -76
215 lines not on this page · raw message
Signature · 1 line
2.36.1.476.g0c4daa206d-goog
Re: [PATCH v2 6/6] perf cpumap: Add range data encoding
On Tue, Jun 14, 2022 at 07:33:53AM -0700, Ian Rogers wrote:
SNIP
Quoted 31 lines
tools/perf/util/synthetic-events.c +7
-18-static size_t cpus_size (const struct perf_cpu_map *map) +static void synthesize_range_cpus (struct synthesize_cpu_map_data *data) { -return sizeof (struct cpu_map_entries ) +perf_cpu_map__nr (map) *sizeof (u16); + data->data->type =PERF_CPU_MAP__RANGE_CPUS ; + data->data->range_cpu_data.any_cpu = data->has_any_cpu; + data->data->range_cpu_data.start_cpu = data->min_cpu; + data->data->range_cpu_data.end_cpu = data->max_cpu; } -static size_t mask_size (const struct perf_cpu_map *map,int *max) -{ - *max =perf_cpu_map__max (map).cpu; -return sizeof (struct perf_record_mask_cpu_map32 ) +BITS_TO_U32 (*max) *sizeof (__u32); -} - -static void *cpu_map_data__alloc (const struct perf_cpu_map *map,size_t *size, -u16 *type,int *max) +static void *cpu_map_data__alloc (struct synthesize_cpu_map_data *syn_data, +size_t header_size) {size_t size_cpus, size_mask; -bool is_dummy =perf_cpu_map__empty (map); -/* -* Both array and mask data have variable size based -* on the number of cpus and their actual values. -* The size of the 'struct perf_record_cpu_map_data' is: -* -* array = size of 'struct cpu_map_entries' + -* number of cpus * sizeof(u64)
tools/perf/util/synthetic-events.c int perf_event__synthesize_thread_map2(struct perf_tool *tool, · line 1237
Quoted 10 lines
tools/perf/util/synthetic-events.c +2
-8-* -* mask = size of 'struct perf_record_record_cpu_map' + -* maximum cpu bit converted to size of longs -* -* and finally + the size of 'struct perf_record_cpu_map_data'. -*/ - size_cpus =cpus_size (map); - size_mask =mask_size (map, max); + syn_data->nr =perf_cpu_map__nr (syn_data->map); + syn_data->has_any_cpu = (perf_cpu_map__cpu (syn_data->map,0 ).cpu ==-1 ) ?1 :0 ;
I'm bit lost in the logic in here.. should it be '.cpu != -1' ? has_any_cpu is named as bool but used as index below ;-)
could you please keep/update the comment above and explain the conditions when each cpu_map type is taken
thanks,
jirka
Quoted 34 lines
tools/perf/util/synthetic-events.c +22
-9-if (is_dummy || (size_cpus < size_mask)) { - *size += size_cpus; - *type =PERF_CPU_MAP__CPUS ; - }else { - *size += size_mask; - *type =PERF_CPU_MAP__MASK ; + syn_data->min_cpu =perf_cpu_map__cpu (syn_data->map, syn_data->has_any_cpu).cpu; + syn_data->max_cpu =perf_cpu_map__max (syn_data->map).cpu; +if (syn_data->max_cpu - syn_data->min_cpu +1 == syn_data->nr - syn_data->has_any_cpu) { +/* A consecutive range of CPUs can be encoded using a range. */ +assert (sizeof (u16) +sizeof (struct perf_record_range_cpu_map ) ==sizeof (u64)); + syn_data->type =PERF_CPU_MAP__RANGE_CPUS ; + syn_data->size = header_size +sizeof (u64); +return zalloc (syn_data->size); } - *size +=sizeof (__u16);/* For perf_record_cpu_map_data.type. */ - *size =PERF_ALIGN (*size,sizeof (u64)); -return zalloc (*size); + size_cpus =sizeof (u16) +sizeof (struct cpu_map_entries ) + syn_data->nr *sizeof (u16); +/* Due to padding, the 4bytes per entry mask variant is always smaller. */ + size_mask =sizeof (u16) +sizeof (struct perf_record_mask_cpu_map32 ) + +BITS_TO_U32 (syn_data->max_cpu) *sizeof (__u32); +if (syn_data->has_any_cpu || size_cpus < size_mask) { +/* Follow the CPU map encoding. */ + syn_data->type =PERF_CPU_MAP__CPUS ; + syn_data->size = header_size +PERF_ALIGN (size_cpus,sizeof (u64)); +return zalloc (syn_data->size); + } +/* Encode using a bitmask. */ + syn_data->type =PERF_CPU_MAP__MASK ; + syn_data->size = header_size +PERF_ALIGN (size_mask,sizeof (u64)); +return zalloc (syn_data->size);
SNIP
Ian Rogers
reviewerRe: [PATCH v2 6/6] perf cpumap: Add range data encoding
Quoted 50 lines · Jiri Olsa
On Tue, Jun 14, 2022 at 07:33:53AM -0700, Ian Rogers wrote:
SNIP
Quoted 31 lines
tools/perf/util/synthetic-events.c +7
-18-static size_t cpus_size (const struct perf_cpu_map *map) +static void synthesize_range_cpus (struct synthesize_cpu_map_data *data) { -return sizeof (struct cpu_map_entries ) +perf_cpu_map__nr (map) *sizeof (u16); + data->data->type =PERF_CPU_MAP__RANGE_CPUS ; + data->data->range_cpu_data.any_cpu = data->has_any_cpu; + data->data->range_cpu_data.start_cpu = data->min_cpu; + data->data->range_cpu_data.end_cpu = data->max_cpu; } -static size_t mask_size (const struct perf_cpu_map *map,int *max) -{ - *max =perf_cpu_map__max (map).cpu; -return sizeof (struct perf_record_mask_cpu_map32 ) +BITS_TO_U32 (*max) *sizeof (__u32); -} - -static void *cpu_map_data__alloc (const struct perf_cpu_map *map,size_t *size, -u16 *type,int *max) +static void *cpu_map_data__alloc (struct synthesize_cpu_map_data *syn_data, +size_t header_size) {size_t size_cpus, size_mask; -bool is_dummy =perf_cpu_map__empty (map); -/* -* Both array and mask data have variable size based -* on the number of cpus and their actual values. -* The size of the 'struct perf_record_cpu_map_data' is: -* -* array = size of 'struct cpu_map_entries' + -* number of cpus * sizeof(u64) tools/perf/util/synthetic-events.c int perf_event__synthesize_thread_map2(struct perf_tool *tool, · line 1237
Quoted 10 lines
tools/perf/util/synthetic-events.c +2
-8-* -* mask = size of 'struct perf_record_record_cpu_map' + -* maximum cpu bit converted to size of longs -* -* and finally + the size of 'struct perf_record_cpu_map_data'. -*/ - size_cpus =cpus_size (map); - size_mask =mask_size (map, max); + syn_data->nr =perf_cpu_map__nr (syn_data->map); + syn_data->has_any_cpu = (perf_cpu_map__cpu (syn_data->map,0 ).cpu ==-1 ) ?1 :0 ;I'm bit lost in the logic in here.. should it be '.cpu != -1' ?
has_any_cpuis named as bool but used as index below ;-)
Quoted 2 lines
could you please keep/update the comment above and explain the conditions when each
cpu_maptype is taken
So this relates to the CPU map "empty" problem that I've been moaning about for a time with Adrian, Arnaldo, etc. The CPU map can contain -1 for the "any CPU" case in perf_event_open, it can also contain CPU values after this as the result of a merge. Having only -1 in the CPU map is referred to as a dummy (presumably because the dummy event uses it), what does it mean if you have the -1 and CPU values? Things get really messy when you look at what the definition of empty is. Here I've used the term from perf_event_open that -1 means this can be on any CPU. As the CPU map data is sorted we can take advantage that the "any CPU" value always comes first. If we have it we set the bit here. We also use the bit below so that we ignore the first array element if it has -1 in it. As the -1 adjusts the CPU map size, there are some other adjustments as well. Hopefully this makes things clearer and I want to rewrite parts of the CPU map API to make them clearer too - empty should mean == NULL or nr == 0, dummy is more confusing to me than "any CPU", etc.
Thanks,
Ian
Quoted 39 lines
thanks,
jirkaQuoted 34 lines
tools/perf/util/synthetic-events.c +22
-9-if (is_dummy || (size_cpus < size_mask)) { - *size += size_cpus; - *type =PERF_CPU_MAP__CPUS ; - }else { - *size += size_mask; - *type =PERF_CPU_MAP__MASK ; + syn_data->min_cpu =perf_cpu_map__cpu (syn_data->map, syn_data->has_any_cpu).cpu; + syn_data->max_cpu =perf_cpu_map__max (syn_data->map).cpu; +if (syn_data->max_cpu - syn_data->min_cpu +1 == syn_data->nr - syn_data->has_any_cpu) { +/* A consecutive range of CPUs can be encoded using a range. */ +assert (sizeof (u16) +sizeof (struct perf_record_range_cpu_map ) ==sizeof (u64)); + syn_data->type =PERF_CPU_MAP__RANGE_CPUS ; + syn_data->size = header_size +sizeof (u64); +return zalloc (syn_data->size); } - *size +=sizeof (__u16);/* For perf_record_cpu_map_data.type. */ - *size =PERF_ALIGN (*size,sizeof (u64)); -return zalloc (*size); + size_cpus =sizeof (u16) +sizeof (struct cpu_map_entries ) + syn_data->nr *sizeof (u16); +/* Due to padding, the 4bytes per entry mask variant is always smaller. */ + size_mask =sizeof (u16) +sizeof (struct perf_record_mask_cpu_map32 ) + +BITS_TO_U32 (syn_data->max_cpu) *sizeof (__u32); +if (syn_data->has_any_cpu || size_cpus < size_mask) { +/* Follow the CPU map encoding. */ + syn_data->type =PERF_CPU_MAP__CPUS ; + syn_data->size = header_size +PERF_ALIGN (size_cpus,sizeof (u64)); +return zalloc (syn_data->size); + } +/* Encode using a bitmask. */ + syn_data->type =PERF_CPU_MAP__MASK ; + syn_data->size = header_size +PERF_ALIGN (size_mask,sizeof (u64)); +return zalloc (syn_data->size);SNIP
Re: [PATCH v2 6/6] perf cpumap: Add range data encoding
In reply to Ian RogersOn Tue, Jun 14, 2022 at 07:33:53AM -0700, Ian Rogers wrote:
SNIP
Quoted 21 lines
tools/perf/util/synthetic-events.c +5
-4+ event =cpu_map_data__alloc (&syn_data,sizeof (struct perf_event_header ));if (!event)return NULL ; + syn_data.data = &event->data; event->header.type =PERF_RECORD_CPU_MAP ; - event->header.size = size; - event->data.type = type; - -cpu_map_data__synthesize (&event->data, map, type, max); + event->header.size = syn_data.size; +cpu_map_data__synthesize (&syn_data);return event; } + int perf_event__synthesize_cpu_map(struct perf_tool *tool, const struct perf_cpu_map *map, perf_event__handler_t process,@@ -1891,24 +1894,20 @@ int perf_event__synthesize_event_update_name(struct perf_tool *tool, struct evse int perf_event__synthesize_event_update_cpus(struct perf_tool *tool, struct evsel *evsel,
tools/perf/util/synthetic-events.c int perf_event__synthesize_event_update_name(struct perf_tool *tool, struct evse · line 1900
Quoted 10 lines
tools/perf/util/synthetic-events.c +1
-6perf_event__handler_t process) { -size_t size =sizeof (struct perf_event_header ) +sizeof (u64) +sizeof (u64); +struct synthesize_cpu_map_data syn_data = { .map = evsel->core.own_cpus };struct perf_record_event_update *ev; -int max, err; -u16 type; - -if (!evsel->core.own_cpus) -return 0 ;
all seems fine, just looks like we no longer do this check, might not be needed anymore, as that changed in past
thanks,
jirka
Quoted 22 lines
tools/perf/util/synthetic-events.c +5
-4+int err; - ev =cpu_map_data__alloc (evsel->core.own_cpus, &size, &type, &max); + ev =cpu_map_data__alloc (&syn_data,sizeof (struct perf_event_header ) +2 *sizeof (u64));if (!ev)return -ENOMEM ; + syn_data.data = &ev->cpus.cpus; ev->header.type =PERF_RECORD_EVENT_UPDATE ; - ev->header.size = (u16 )size; + ev->header.size = (u16 )syn_data.size; ev->type =PERF_EVENT_UPDATE__CPUS ; ev->id = evsel->core.id[0 ]; - -cpu_map_data__synthesize (&ev->cpus.cpus, evsel->core.own_cpus, type, max); +cpu_map_data__synthesize (&syn_data); err =process (tool, (union perf_event *)ev,NULL ,NULL );free (ev);Signature · 1 line
2.36.1.476.g0c4daa206d-goog
Ian Rogers
reviewerRe: [PATCH v2 6/6] perf cpumap: Add range data encoding
Quoted 31 lines · Jiri Olsa
On Tue, Jun 14, 2022 at 07:33:53AM -0700, Ian Rogers wrote:
SNIP
Quoted 26 lines
tools/perf/util/synthetic-events.c +6
-5+ event =cpu_map_data__alloc (&syn_data,sizeof (struct perf_event_header ));if (!event)return NULL ; + syn_data.data = &event->data; event->header.type =PERF_RECORD_CPU_MAP ; - event->header.size = size; - event->data.type = type; - -cpu_map_data__synthesize (&event->data, map, type, max); + event->header.size = syn_data.size; +cpu_map_data__synthesize (&syn_data);return event; } + int perf_event__synthesize_cpu_map(struct perf_tool *tool, const struct perf_cpu_map *map, perf_event__handler_t process,@@ -1891,24 +1894,20 @@ int perf_event__synthesize_event_update_name(struct perf_tool *tool, struct evseint perf_event__synthesize_event_update_cpus (struct perf_tool *tool,struct evsel *evsel,perf_event__handler_t process) { -size_t size =sizeof (struct perf_event_header ) +sizeof (u64) +sizeof (u64); +struct synthesize_cpu_map_data syn_data = { .map = evsel->core.own_cpus };struct perf_record_event_update *ev;
Quoted 8 lines
tools/perf/util/synthetic-events.c int perf_event__synthesize_event_update_name(struct perf_tool *tool, struct evse · line 1900
Quoted 5 lines
tools/perf/util/synthetic-events.c +0
-5-int max, err; -u16 type; - -if (!evsel->core.own_cpus) -return 0 ;all seems fine, just looks like we no longer do this check, might not be needed anymore, as that changed in past
This function is called in a test and in this file. The caller already does this test and so the check is redundant plus a little confusing:
https://git.kernel.org/pub/scm/linux/kernel/git/acme/linux.git/tree/tools/perf/util/synthetic-events.c?h=perf/core#n2062
As you say, it wasn't needed any more and so I removed it.
Thanks,
Ian
Quoted 25 lines
thanks,
jirkaQuoted 22 lines
tools/perf/util/synthetic-events.c +5
-5+int err; - ev =cpu_map_data__alloc (evsel->core.own_cpus, &size, &type, &max); + ev =cpu_map_data__alloc (&syn_data,sizeof (struct perf_event_header ) +2 *sizeof (u64));if (!ev)return -ENOMEM ; + syn_data.data = &ev->cpus.cpus; ev->header.type =PERF_RECORD_EVENT_UPDATE ; - ev->header.size = (u16 )size; + ev->header.size = (u16 )syn_data.size; ev->type =PERF_EVENT_UPDATE__CPUS ; ev->id = evsel->core.id[0 ]; - -cpu_map_data__synthesize (&ev->cpus.cpus, evsel->core.own_cpus, type, max); +cpu_map_data__synthesize (&syn_data); err =process (tool, (union perf_event *)ev,NULL ,NULL );free (ev); --2.36.1.476.g0c4daa206d-goog
Ian Rogers
reviewerRe: [PATCH v2 6/6] perf cpumap: Add range data encoding
Quoted 42 lines · Ian Rogers
Quoted 31 lines · Jiri Olsa
On Tue, Jun 14, 2022 at 07:33:53AM -0700, Ian Rogers wrote:
SNIP
Quoted 26 lines
tools/perf/util/synthetic-events.c +6
-5+ event =cpu_map_data__alloc (&syn_data,sizeof (struct perf_event_header ));if (!event)return NULL ; + syn_data.data = &event->data; event->header.type =PERF_RECORD_CPU_MAP ; - event->header.size = size; - event->data.type = type; - -cpu_map_data__synthesize (&event->data, map, type, max); + event->header.size = syn_data.size; +cpu_map_data__synthesize (&syn_data);return event; } + int perf_event__synthesize_cpu_map(struct perf_tool *tool, const struct perf_cpu_map *map, perf_event__handler_t process,@@ -1891,24 +1894,20 @@ int perf_event__synthesize_event_update_name(struct perf_tool *tool, struct evseint perf_event__synthesize_event_update_cpus (struct perf_tool *tool,struct evsel *evsel,perf_event__handler_t process) { -size_t size =sizeof (struct perf_event_header ) +sizeof (u64) +sizeof (u64); +struct synthesize_cpu_map_data syn_data = { .map = evsel->core.own_cpus };struct perf_record_event_update *ev;Quoted 8 lines
tools/perf/util/synthetic-events.c int perf_event__synthesize_event_update_name(struct perf_tool *tool, struct evse · line 1900
Quoted 5 lines
tools/perf/util/synthetic-events.c +0
-5-int max, err; -u16 type; - -if (!evsel->core.own_cpus) -return 0 ;all seems fine, just looks like we no longer do this check, might not be needed anymore, as that changed in past
Quoted 7 lines
This function is called in a test and in this file. The caller already does this test and so the check is redundant plus a little confusing:
https://git.kernel.org/pub/scm/linux/kernel/git/acme/linux.git/tree/tools/perf/util/synthetic-events.c?h=perf/core#n2062
As you say, it wasn't needed any more and so I removed it.Thanks,
Ian
Would be nice to land this imo. Anything outstanding?
Thanks,
Ian
Quoted 25 lines
Quoted 25 lines
thanks,
jirkaQuoted 22 lines
tools/perf/util/synthetic-events.c +5
-5+int err; - ev =cpu_map_data__alloc (evsel->core.own_cpus, &size, &type, &max); + ev =cpu_map_data__alloc (&syn_data,sizeof (struct perf_event_header ) +2 *sizeof (u64));if (!ev)return -ENOMEM ; + syn_data.data = &ev->cpus.cpus; ev->header.type =PERF_RECORD_EVENT_UPDATE ; - ev->header.size = (u16 )size; + ev->header.size = (u16 )syn_data.size; ev->type =PERF_EVENT_UPDATE__CPUS ; ev->id = evsel->core.id[0 ]; - -cpu_map_data__synthesize (&ev->cpus.cpus, evsel->core.own_cpus, type, max); +cpu_map_data__synthesize (&syn_data); err =process (tool, (union perf_event *)ev,NULL ,NULL );free (ev); --2.36.1.476.g0c4daa206d-goog
Re: [PATCH v2 6/6] perf cpumap: Add range data encoding
Quoted 48 lines · Ian Rogers
Quoted 47 lines · Ian Rogers
Quoted 31 lines · Jiri Olsa
On Tue, Jun 14, 2022 at 07:33:53AM -0700, Ian Rogers wrote:
SNIP
Quoted 26 lines
tools/perf/util/synthetic-events.c +6
-5+ event =cpu_map_data__alloc (&syn_data,sizeof (struct perf_event_header ));if (!event)return NULL ; + syn_data.data = &event->data; event->header.type =PERF_RECORD_CPU_MAP ; - event->header.size = size; - event->data.type = type; - -cpu_map_data__synthesize (&event->data, map, type, max); + event->header.size = syn_data.size; +cpu_map_data__synthesize (&syn_data);return event; } + int perf_event__synthesize_cpu_map(struct perf_tool *tool, const struct perf_cpu_map *map, perf_event__handler_t process,@@ -1891,24 +1894,20 @@ int perf_event__synthesize_event_update_name(struct perf_tool *tool, struct evseint perf_event__synthesize_event_update_cpus (struct perf_tool *tool,struct evsel *evsel,perf_event__handler_t process) { -size_t size =sizeof (struct perf_event_header ) +sizeof (u64) +sizeof (u64); +struct synthesize_cpu_map_data syn_data = { .map = evsel->core.own_cpus };struct perf_record_event_update *ev;Quoted 8 lines
tools/perf/util/synthetic-events.c int perf_event__synthesize_event_update_name(struct perf_tool *tool, struct evse · line 1900
Quoted 5 lines
tools/perf/util/synthetic-events.c +0
-5-int max, err; -u16 type; - -if (!evsel->core.own_cpus) -return 0 ;all seems fine, just looks like we no longer do this check, might not be needed anymore, as that changed in past
This function is called in a test and in this file. The caller already does this test and so the check is redundant plus a little confusing:
https://git.kernel.org/pub/scm/linux/kernel/git/acme/linux.git/tree/tools/perf/util/synthetic-events.c?h=perf/core#n2062
As you say, it wasn't needed any more and so I removed it.
Quoted 4 lines
Quoted 2 lines
Thanks,
IanWould be nice to land this imo. Anything outstanding?
I'll check first hour tomorrow, today was a holiday down here.
- Arnaldo
Quoted 29 lines
Thanks,
IanQuoted 25 lines
Quoted 25 lines
thanks,
jirkaQuoted 22 lines
tools/perf/util/synthetic-events.c +5
-5+int err; - ev =cpu_map_data__alloc (evsel->core.own_cpus, &size, &type, &max); + ev =cpu_map_data__alloc (&syn_data,sizeof (struct perf_event_header ) +2 *sizeof (u64));if (!ev)return -ENOMEM ; + syn_data.data = &ev->cpus.cpus; ev->header.type =PERF_RECORD_EVENT_UPDATE ; - ev->header.size = (u16 )size; + ev->header.size = (u16 )syn_data.size; ev->type =PERF_EVENT_UPDATE__CPUS ; ev->id = evsel->core.id[0 ]; - -cpu_map_data__synthesize (&ev->cpus.cpus, evsel->core.own_cpus, type, max); +cpu_map_data__synthesize (&syn_data); err =process (tool, (union perf_event *)ev,NULL ,NULL );free (ev); --2.36.1.476.g0c4daa206d-goog
Arnaldo Carvalho de Melo
maintainerRe: [PATCH v2 6/6] perf cpumap: Add range data encoding
In reply to Ian RogersEm Wed, Sep 07, 2022 at 03:41:19PM -0700, Ian Rogers escreveu:
Quoted 48 lines · Ian Rogers
Quoted 47 lines · Jiri Olsa
Quoted 31 lines · Ian Rogers
SNIP
Quoted 26 lines
tools/perf/util/synthetic-events.c +6
-5+ event =cpu_map_data__alloc (&syn_data,sizeof (struct perf_event_header ));if (!event)return NULL ; + syn_data.data = &event->data; event->header.type =PERF_RECORD_CPU_MAP ; - event->header.size = size; - event->data.type = type; - -cpu_map_data__synthesize (&event->data, map, type, max); + event->header.size = syn_data.size; +cpu_map_data__synthesize (&syn_data);return event; } + int perf_event__synthesize_cpu_map(struct perf_tool *tool, const struct perf_cpu_map *map, perf_event__handler_t process,@@ -1891,24 +1894,20 @@ int perf_event__synthesize_event_update_name(struct perf_tool *tool, struct evseint perf_event__synthesize_event_update_cpus (struct perf_tool *tool,struct evsel *evsel,perf_event__handler_t process) { -size_t size =sizeof (struct perf_event_header ) +sizeof (u64) +sizeof (u64); +struct synthesize_cpu_map_data syn_data = { .map = evsel->core.own_cpus };struct perf_record_event_update *ev;Quoted 8 lines
tools/perf/util/synthetic-events.c int perf_event__synthesize_event_update_name(struct perf_tool *tool, struct evse · line 1900
Quoted 5 lines
tools/perf/util/synthetic-events.c +0
-5-int max, err; -u16 type; - -if (!evsel->core.own_cpus) -return 0 ;all seems fine, just looks like we no longer do this check, might not be needed anymore, as that changed in past
This function is called in a test and in this file. The caller already does this test and so the check is redundant plus a little confusing:
https://git.kernel.org/pub/scm/linux/kernel/git/acme/linux.git/tree/tools/perf/util/synthetic-events.c?h=perf/core#n2062
As you say, it wasn't needed any more and so I removed it.
Quoted 4 lines
Quoted 2 lines
Thanks,
IanWould be nice to land this imo. Anything outstanding?
So the last two patches were outstanding, 5/6 applied cleanly, 6/6 had some fuzzes, minimal stuff, applied, I'll push it out soon, please check.
- Arnaldo
Ian Rogers
reviewerQuoted 55 lines · Ian Rogers
A mask encoding of a cpu map is laid out as:
u16 nr
u16long_size
unsigned long mask[];
However, the mask may be 8-byte aligned meaning there is a 4-byte pad afterlong_size. This means 32-bit and 64-bit builds see the mask as being at different offsets. On top of this the structure is in the byte data[] encoded as:
u16 type
char data[]
This means the mask's struct isn't the required 4 or 8 byte aligned, but is offset by 2. Consequently the long reads and writes are causing undefined behavior as the alignment is broken.These changes do minor clean up with const, visibility of functions and using the constant time max function. It then adds 32 and 64-bit mask encoding variants, packed to match current alignment. Taking the address of a packed struct leads to unaligned data, so function arguments are altered to be passed the packed struct. To compact the mask encoding further and drop the padding, the 4-byte variant is preferred. Finally a new range encoding is added, that reduces the size of the common case of a range of CPUs to a single u64.
On a 72 CPU (hyperthread) machine the original encoding of all CPUs is:
0x9a98[0x28]: event: 74
.
. ... raw event: size 40 bytes. 0000: 4a 00 00 00 00 00 28 00 01 00 02 00 08 00 00 00 J.....(......... . 0010: 00 00 ff ff ff ff ff ff ff ff ff 00 00 00 00 00 ................ . 0020: 00 00 00 00 00 00 00 00 ........0 0
0x9a98[0x28]:PERF_RECORD_CPU_MAPUsing the 4-byte encoding it is:
0x9a98@pipe [0x20]: event: 74
.
. ... raw event: size 32 bytes. 0000: 4a 00 00 00 00 00 20 00 01 00 03 00 04 00 ff ff J..... ......... . 0010: ff ff ff ff ff ff ff 00 00 00 00 00 00 00 00 00 ................0 0
0x9a98[0x20]:PERF_RECORD_CPU_MAPFinally, with the range encoding it is:
0x9ab8@pipe [0x10]: event: 74
.
. ... raw event: size 16 bytes
. 0000: 4a 00 00 00 00 00 10 00 02 00 00 00 00 00 47 00 J.............G.0 0
0x9ab8[0x10]:PERF_RECORD_CPU_MAPv2. Fixes a bug in the size computation of the update header
introduced by the last patch (Add range data encoding) and caught by address sanitizer.
Quoted 7 lines
Ian Rogers (6):
perf cpumap: Const map for max
perf cpumap: Synthetic events and const/static perf cpumap: Compute mask size in constant time perf cpumap: Fix alignment for masks in event encoding perf events: Prefer union over variable length array perf cpumap: Add range data encoding
Ping. There was some feedback on this change but nothing to create a v3. Feedback/acked-by/reviewed-by appreciated.
Thanks,
Ian
Quoted 17 lines
tools/lib/perf/cpumap.c | 2 +- tools/lib/perf/include/perf/cpumap.h | 2 +- tools/lib/perf/include/perf/event.h | 61 ++++++++- tools/perf/tests/cpumap.c | 71 ++++++++--- tools/perf/tests/event_update.c | 14 +-- tools/perf/util/cpumap.c | 111 +++++++++++++--- tools/perf/util/cpumap.h | 4 +- tools/perf/util/event.h | 4 - tools/perf/util/header.c | 24 ++-- tools/perf/util/session.c | 35 +++--- tools/perf/util/synthetic-events.c | 182 +++++++++++++-------------- tools/perf/util/synthetic-events.h | 2 +- 12 files changed, 327 insertions(+), 185 deletions(-)--
2.36.1.476.g0c4daa206d-goog
Ian Rogers
reviewerQuoted 65 lines · Ian Rogers
Quoted 55 lines · Ian Rogers
A mask encoding of a cpu map is laid out as:
u16 nr
u16long_size
unsigned long mask[];
However, the mask may be 8-byte aligned meaning there is a 4-byte pad afterlong_size. This means 32-bit and 64-bit builds see the mask as being at different offsets. On top of this the structure is in the byte data[] encoded as:
u16 type
char data[]
This means the mask's struct isn't the required 4 or 8 byte aligned, but is offset by 2. Consequently the long reads and writes are causing undefined behavior as the alignment is broken.These changes do minor clean up with const, visibility of functions and using the constant time max function. It then adds 32 and 64-bit mask encoding variants, packed to match current alignment. Taking the address of a packed struct leads to unaligned data, so function arguments are altered to be passed the packed struct. To compact the mask encoding further and drop the padding, the 4-byte variant is preferred. Finally a new range encoding is added, that reduces the size of the common case of a range of CPUs to a single u64.
On a 72 CPU (hyperthread) machine the original encoding of all CPUs is:
0x9a98[0x28]: event: 74
.
. ... raw event: size 40 bytes. 0000: 4a 00 00 00 00 00 28 00 01 00 02 00 08 00 00 00 J.....(......... . 0010: 00 00 ff ff ff ff ff ff ff ff ff 00 00 00 00 00 ................ . 0020: 00 00 00 00 00 00 00 00 ........0 0
0x9a98[0x28]:PERF_RECORD_CPU_MAPUsing the 4-byte encoding it is:
0x9a98@pipe [0x20]: event: 74
.
. ... raw event: size 32 bytes. 0000: 4a 00 00 00 00 00 20 00 01 00 03 00 04 00 ff ff J..... ......... . 0010: ff ff ff ff ff ff ff 00 00 00 00 00 00 00 00 00 ................0 0
0x9a98[0x20]:PERF_RECORD_CPU_MAPFinally, with the range encoding it is:
0x9ab8@pipe [0x10]: event: 74
.
. ... raw event: size 16 bytes
. 0000: 4a 00 00 00 00 00 10 00 02 00 00 00 00 00 47 00 J.............G.0 0
0x9ab8[0x10]:PERF_RECORD_CPU_MAPv2. Fixes a bug in the size computation of the update header
introduced by the last patch (Add range data encoding) and caught by address sanitizer.Quoted 7 lines
Ian Rogers (6):
perf cpumap: Const map for max
perf cpumap: Synthetic events and const/static perf cpumap: Compute mask size in constant time perf cpumap: Fix alignment for masks in event encoding perf events: Prefer union over variable length array perf cpumap: Add range data encoding
Quoted 2 lines
Ping. There was some feedback on this change but nothing to create a v3. Feedback/acked-by/reviewed-by appreciated.
Ping. Feedback appreciated.
Thanks,
Ian
Quoted 20 lines
Thanks,
IanQuoted 17 lines
tools/lib/perf/cpumap.c | 2 +- tools/lib/perf/include/perf/cpumap.h | 2 +- tools/lib/perf/include/perf/event.h | 61 ++++++++- tools/perf/tests/cpumap.c | 71 ++++++++--- tools/perf/tests/event_update.c | 14 +-- tools/perf/util/cpumap.c | 111 +++++++++++++--- tools/perf/util/cpumap.h | 4 +- tools/perf/util/event.h | 4 - tools/perf/util/header.c | 24 ++-- tools/perf/util/session.c | 35 +++--- tools/perf/util/synthetic-events.c | 182 +++++++++++++-------------- tools/perf/util/synthetic-events.h | 2 +- 12 files changed, 327 insertions(+), 185 deletions(-)--
2.36.1.476.g0c4daa206d-goog
Quoted 66 lines · Ian Rogers
Quoted 65 lines · Ian Rogers
Quoted 62 lines · Ian Rogers
A mask encoding of a cpu map is laid out as:
u16 nr
u16long_size
unsigned long mask[];
However, the mask may be 8-byte aligned meaning there is a 4-byte pad afterlong_size. This means 32-bit and 64-bit builds see the mask as being at different offsets. On top of this the structure is in the byte data[] encoded as:
u16 type
char data[]
This means the mask's struct isn't the required 4 or 8 byte aligned, but is offset by 2. Consequently the long reads and writes are causing undefined behavior as the alignment is broken.These changes do minor clean up with const, visibility of functions and using the constant time max function. It then adds 32 and 64-bit mask encoding variants, packed to match current alignment. Taking the address of a packed struct leads to unaligned data, so function arguments are altered to be passed the packed struct. To compact the mask encoding further and drop the padding, the 4-byte variant is preferred. Finally a new range encoding is added, that reduces the size of the common case of a range of CPUs to a single u64.
On a 72 CPU (hyperthread) machine the original encoding of all CPUs is:
0x9a98[0x28]: event: 74
.
. ... raw event: size 40 bytes. 0000: 4a 00 00 00 00 00 28 00 01 00 02 00 08 00 00 00 J.....(......... . 0010: 00 00 ff ff ff ff ff ff ff ff ff 00 00 00 00 00 ................ . 0020: 00 00 00 00 00 00 00 00 ........0 0
0x9a98[0x28]:PERF_RECORD_CPU_MAPUsing the 4-byte encoding it is:
0x9a98@pipe [0x20]: event: 74
.
. ... raw event: size 32 bytes. 0000: 4a 00 00 00 00 00 20 00 01 00 03 00 04 00 ff ff J..... ......... . 0010: ff ff ff ff ff ff ff 00 00 00 00 00 00 00 00 00 ................0 0
0x9a98[0x20]:PERF_RECORD_CPU_MAPFinally, with the range encoding it is:
0x9ab8@pipe [0x10]: event: 74
.
. ... raw event: size 16 bytes
. 0000: 4a 00 00 00 00 00 10 00 02 00 00 00 00 00 47 00 J.............G.0 0
0x9ab8[0x10]:PERF_RECORD_CPU_MAPv2. Fixes a bug in the size computation of the update header
introduced by the last patch (Add range data encoding) and caught by address sanitizer.Ian Rogers (6):
perf cpumap: Const map for max
perf cpumap: Synthetic events and const/static perf cpumap: Compute mask size in constant time perf cpumap: Fix alignment for masks in event encoding perf events: Prefer union over variable length array perf cpumap: Add range data encoding
Quoted 4 lines
Quoted 2 lines
Ping. There was some feedback on this change but nothing to create a v3. Feedback/acked-by/reviewed-by appreciated.
Ping. Feedback appreciated.
hi,
there's some unanswered feedback:
https://lore.kernel.org/linux-perf-users/YrwY3SP+jsTwrRBw@krava/
jirka
Ian Rogers
reviewerQuoted 71 lines · Jiri Olsa
Quoted 69 lines · Ian Rogers
Quoted 67 lines · Ian Rogers
Quoted 55 lines · Ian Rogers
A mask encoding of a cpu map is laid out as:
u16 nr
u16long_size
unsigned long mask[];
However, the mask may be 8-byte aligned meaning there is a 4-byte pad afterlong_size. This means 32-bit and 64-bit builds see the mask as being at different offsets. On top of this the structure is in the byte data[] encoded as:
u16 type
char data[]
This means the mask's struct isn't the required 4 or 8 byte aligned, but is offset by 2. Consequently the long reads and writes are causing undefined behavior as the alignment is broken.These changes do minor clean up with const, visibility of functions and using the constant time max function. It then adds 32 and 64-bit mask encoding variants, packed to match current alignment. Taking the address of a packed struct leads to unaligned data, so function arguments are altered to be passed the packed struct. To compact the mask encoding further and drop the padding, the 4-byte variant is preferred. Finally a new range encoding is added, that reduces the size of the common case of a range of CPUs to a single u64.
On a 72 CPU (hyperthread) machine the original encoding of all CPUs is:
0x9a98[0x28]: event: 74
.
. ... raw event: size 40 bytes. 0000: 4a 00 00 00 00 00 28 00 01 00 02 00 08 00 00 00 J.....(......... . 0010: 00 00 ff ff ff ff ff ff ff ff ff 00 00 00 00 00 ................ . 0020: 00 00 00 00 00 00 00 00 ........0 0
0x9a98[0x28]:PERF_RECORD_CPU_MAPUsing the 4-byte encoding it is:
0x9a98@pipe [0x20]: event: 74
.
. ... raw event: size 32 bytes. 0000: 4a 00 00 00 00 00 20 00 01 00 03 00 04 00 ff ff J..... ......... . 0010: ff ff ff ff ff ff ff 00 00 00 00 00 00 00 00 00 ................0 0
0x9a98[0x20]:PERF_RECORD_CPU_MAPFinally, with the range encoding it is:
0x9ab8@pipe [0x10]: event: 74
.
. ... raw event: size 16 bytes
. 0000: 4a 00 00 00 00 00 10 00 02 00 00 00 00 00 47 00 J.............G.0 0
0x9ab8[0x10]:PERF_RECORD_CPU_MAPv2. Fixes a bug in the size computation of the update header
introduced by the last patch (Add range data encoding) and caught by address sanitizer.Quoted 7 lines
Ian Rogers (6):
perf cpumap: Const map for max
perf cpumap: Synthetic events and const/static perf cpumap: Compute mask size in constant time perf cpumap: Fix alignment for masks in event encoding perf events: Prefer union over variable length array perf cpumap: Add range data encodingPing. There was some feedback on this change but nothing to create a v3. Feedback/acked-by/reviewed-by appreciated.
Quoted 7 lines
Quoted 1 line
Ping. Feedback appreciated.
hi,
there's some unanswered feedback:
https://lore.kernel.org/linux-perf-users/YrwY3SP+jsTwrRBw@krava/jirka
Thanks Jirka,
Was there a comment in particular? My reply was here:
https://lore.kernel.org/linux-perf-users/CAP-5=fU=BpP4OT2axZLSfRnKxQxmv-HXj8khBgmx3XQMS+abgA@mail.gmail.com/
I double checked, everyone of your comments was answered.
Thanks,
Ian
Quoted 78 lines · Ian Rogers
Quoted 77 lines · Jiri Olsa
Quoted 66 lines · Ian Rogers
Quoted 65 lines · Ian Rogers
Quoted 62 lines · Ian Rogers
A mask encoding of a cpu map is laid out as:
u16 nr
u16long_size
unsigned long mask[];
However, the mask may be 8-byte aligned meaning there is a 4-byte pad afterlong_size. This means 32-bit and 64-bit builds see the mask as being at different offsets. On top of this the structure is in the byte data[] encoded as:
u16 type
char data[]
This means the mask's struct isn't the required 4 or 8 byte aligned, but is offset by 2. Consequently the long reads and writes are causing undefined behavior as the alignment is broken.These changes do minor clean up with const, visibility of functions and using the constant time max function. It then adds 32 and 64-bit mask encoding variants, packed to match current alignment. Taking the address of a packed struct leads to unaligned data, so function arguments are altered to be passed the packed struct. To compact the mask encoding further and drop the padding, the 4-byte variant is preferred. Finally a new range encoding is added, that reduces the size of the common case of a range of CPUs to a single u64.
On a 72 CPU (hyperthread) machine the original encoding of all CPUs is:
0x9a98[0x28]: event: 74
.
. ... raw event: size 40 bytes. 0000: 4a 00 00 00 00 00 28 00 01 00 02 00 08 00 00 00 J.....(......... . 0010: 00 00 ff ff ff ff ff ff ff ff ff 00 00 00 00 00 ................ . 0020: 00 00 00 00 00 00 00 00 ........0 0
0x9a98[0x28]:PERF_RECORD_CPU_MAPUsing the 4-byte encoding it is:
0x9a98@pipe [0x20]: event: 74
.
. ... raw event: size 32 bytes. 0000: 4a 00 00 00 00 00 20 00 01 00 03 00 04 00 ff ff J..... ......... . 0010: ff ff ff ff ff ff ff 00 00 00 00 00 00 00 00 00 ................0 0
0x9a98[0x20]:PERF_RECORD_CPU_MAPFinally, with the range encoding it is:
0x9ab8@pipe [0x10]: event: 74
.
. ... raw event: size 16 bytes
. 0000: 4a 00 00 00 00 00 10 00 02 00 00 00 00 00 47 00 J.............G.0 0
0x9ab8[0x10]:PERF_RECORD_CPU_MAPv2. Fixes a bug in the size computation of the update header
introduced by the last patch (Add range data encoding) and caught by address sanitizer.Ian Rogers (6):
perf cpumap: Const map for max
perf cpumap: Synthetic events and const/static perf cpumap: Compute mask size in constant time perf cpumap: Fix alignment for masks in event encoding perf events: Prefer union over variable length array perf cpumap: Add range data encodingQuoted 4 lines
Quoted 2 lines
Ping. There was some feedback on this change but nothing to create a v3. Feedback/acked-by/reviewed-by appreciated.
Ping. Feedback appreciated.
hi,
there's some unanswered feedback:
https://lore.kernel.org/linux-perf-users/YrwY3SP+jsTwrRBw@krava/
Quoted 7 lines
Quoted 1 line
jirka
Thanks Jirka,
Was there a comment in particular? My reply was here:
https://lore.kernel.org/linux-perf-users/CAP-5=fU=BpP4OT2axZLSfRnKxQxmv-HXj8khBgmx3XQMS+abgA@mail.gmail.com/
I double checked, everyone of your comments was answered.
ugh sorry, seems it did not get into my inbox for some reason
jirka
Quoted 54 lines · Ian Rogers
A mask encoding of a cpu map is laid out as:
u16 nr
u16long_size
unsigned long mask[];
However, the mask may be 8-byte aligned meaning there is a 4-byte pad afterlong_size. This means 32-bit and 64-bit builds see the mask as being at different offsets. On top of this the structure is in the byte data[] encoded as:
u16 type
char data[]
This means the mask's struct isn't the required 4 or 8 byte aligned, but is offset by 2. Consequently the long reads and writes are causing undefined behavior as the alignment is broken.These changes do minor clean up with const, visibility of functions and using the constant time max function. It then adds 32 and 64-bit mask encoding variants, packed to match current alignment. Taking the address of a packed struct leads to unaligned data, so function arguments are altered to be passed the packed struct. To compact the mask encoding further and drop the padding, the 4-byte variant is preferred. Finally a new range encoding is added, that reduces the size of the common case of a range of CPUs to a single u64.
On a 72 CPU (hyperthread) machine the original encoding of all CPUs is:
0x9a98[0x28]: event: 74
.
. ... raw event: size 40 bytes. 0000: 4a 00 00 00 00 00 28 00 01 00 02 00 08 00 00 00 J.....(......... . 0010: 00 00 ff ff ff ff ff ff ff ff ff 00 00 00 00 00 ................ . 0020: 00 00 00 00 00 00 00 00 ........0 0
0x9a98[0x28]:PERF_RECORD_CPU_MAPUsing the 4-byte encoding it is:
0x9a98@pipe [0x20]: event: 74
.
. ... raw event: size 32 bytes. 0000: 4a 00 00 00 00 00 20 00 01 00 03 00 04 00 ff ff J..... ......... . 0010: ff ff ff ff ff ff ff 00 00 00 00 00 00 00 00 00 ................0 0
0x9a98[0x20]:PERF_RECORD_CPU_MAPFinally, with the range encoding it is:
0x9ab8@pipe [0x10]: event: 74
.
. ... raw event: size 16 bytes
. 0000: 4a 00 00 00 00 00 10 00 02 00 00 00 00 00 47 00 J.............G.0 0
0x9ab8[0x10]:PERF_RECORD_CPU_MAPv2. Fixes a bug in the size computation of the update header
introduced by the last patch (Add range data encoding) and caught by address sanitizer.
Quoted 7 lines
Ian Rogers (6):
perf cpumap: Const map for max
perf cpumap: Synthetic events and const/static perf cpumap: Compute mask size in constant time perf cpumap: Fix alignment for masks in event encoding perf events: Prefer union over variable length array perf cpumap: Add range data encoding
thanks,
jirka
Quoted 18 lines
tools/lib/perf/cpumap.c | 2 +- tools/lib/perf/include/perf/cpumap.h | 2 +- tools/lib/perf/include/perf/event.h | 61 ++++++++- tools/perf/tests/cpumap.c | 71 ++++++++--- tools/perf/tests/event_update.c | 14 +-- tools/perf/util/cpumap.c | 111 +++++++++++++--- tools/perf/util/cpumap.h | 4 +- tools/perf/util/event.h | 4 - tools/perf/util/header.c | 24 ++-- tools/perf/util/session.c | 35 +++--- tools/perf/util/synthetic-events.c | 182 +++++++++++++-------------- tools/perf/util/synthetic-events.h | 2 +- 12 files changed, 327 insertions(+), 185 deletions(-)Signature · 1 line
2.36.1.476.g0c4daa206d-goog
A message that is not in the archive (sent off-list or not yet received): CAP-5=fX-Ex1uv0hxCwDkkAyFV6VQNPRB5uSPpCDNgqu5ZV=bCA@mail.gmail.com
Em Sun, Aug 14, 2022 at 04:05:19PM -0700, Ian Rogers escreveu:
Quoted 3 lines
Hi,
It looks like we missed this one. Can it get into the next PR as a fix?
I've queued up to 4/6 for perf/urgent, 5 and 6 will go to perf/core once I merge perf/urgent there, i.e. after I send this to Linus.
- Arnaldo
Quoted 99 lines
Thanks,
IanQuoted 93 lines · Jiri Olsa
Quoted 23 lines · Ian Rogers
A mask encoding of a cpu map is laid out as:
u16 nr
u16long_size
unsigned long mask[];
However, the mask may be 8-byte aligned meaning there is a 4-byte pad afterlong_size. This means 32-bit and 64-bit builds see the mask as being at different offsets. On top of this the structure is in the byte data[] encoded as:
u16 type
char data[]
This means the mask's struct isn't the required 4 or 8 byte aligned, but is offset by 2. Consequently the long reads and writes are causing undefined behavior as the alignment is broken.These changes do minor clean up with const, visibility of functions and using the constant time max function. It then adds 32 and 64-bit mask encoding variants, packed to match current alignment. Taking the address of a packed struct leads to unaligned data, so function arguments are altered to be passed the packed struct. To compact the mask encoding further and drop the padding, the 4-byte variant is preferred. Finally a new range encoding is added, that reduces the size of the common case of a range of CPUs to a single u64.
Quoted 5 lines
On a 72 CPU (hyperthread) machine the original encoding of all CPUs is:
0x9a98[0x28]: event: 74
.
. ... raw event: size 40 bytes
. 0000: 4a 00 00 00 00 00 28 00 01 00 02 00 08 00 00 00J.....(.........
Quoted 1 line
. 0010: 00 00 ff ff ff ff ff ff ff ff ff 00 00 00 00 00
................
Quoted 10 lines
. 0020: 00 00 00 00 00 00 00 00 ........
0 0
0x9a98[0x28]:PERF_RECORD_CPU_MAPUsing the 4-byte encoding it is:
0x9a98@pipe [0x20]: event: 74
.
. ... raw event: size 32 bytes
. 0000: 4a 00 00 00 00 00 20 00 01 00 03 00 04 00 ff ff J..............
Quoted 1 line
. 0010: ff ff ff ff ff ff ff 00 00 00 00 00 00 00 00 00
................
Quoted 8 lines
0 0
0x9a98[0x20]:PERF_RECORD_CPU_MAPFinally, with the range encoding it is:
0x9ab8@pipe [0x10]: event: 74
.
. ... raw event: size 16 bytes
. 0000: 4a 00 00 00 00 00 10 00 02 00 00 00 00 00 47 00J.............G.
Quoted 7 lines
0 0
0x9ab8[0x10]:PERF_RECORD_CPU_MAPv2. Fixes a bug in the size computation of the update header
introduced by the last patch (Add range data encoding) and caught by address sanitizer.Quoted 7 lines
Ian Rogers (6):
perf cpumap: Const map for max
perf cpumap: Synthetic events and const/static perf cpumap: Compute mask size in constant time perf cpumap: Fix alignment for masks in event encoding perf events: Prefer union over variable length array perf cpumap: Add range data encodingAcked-by Jiri Olsa <[email protected]>thanks,
jirkaQuoted 18 lines
tools/lib/perf/cpumap.c | 2 +- tools/lib/perf/include/perf/cpumap.h | 2 +- tools/lib/perf/include/perf/event.h | 61 ++++++++- tools/perf/tests/cpumap.c | 71 ++++++++--- tools/perf/tests/event_update.c | 14 +-- tools/perf/util/cpumap.c | 111 +++++++++++++--- tools/perf/util/cpumap.h | 4 +- tools/perf/util/event.h | 4 - tools/perf/util/header.c | 24 ++-- tools/perf/util/session.c | 35 +++--- tools/perf/util/synthetic-events.c | 182 +++++++++++++-------------- tools/perf/util/synthetic-events.h | 2 +- 12 files changed, 327 insertions(+), 185 deletions(-)--
2.36.1.476.g0c4daa206d-goog
Signature · 1 line
- Arnaldo