Current implementation of cpu_set_t type by glibc has internal cpu
mask size limitation of no more than 1024 CPUs. This limitation confines
NUMA awareness of Perf tool in record mode, thru --affinity option,
to the first 1024 CPUs on machines with larger amount of CPUs.
This patch set enables Perf tool to overcome 1024 CPUs limitation by
using a dedicated struct mmap_cpu_mask type and applying tool's bitmap
API operations to manipulate affinity masks of the tool's thread and
the mmaped data buffers.
tools bitmap API has been extended with bitmap_free() function and
bitmap_equal() operation whose implementation is derived from the
kernel one.
---
Alexey Budankov (3):
tools bitmap: implement bitmap_equal() operation at bitmap API
perf mmap: declare type for cpu mask of arbitrary length
perf record: adapt affinity to machines with #CPUs > 1K
tools/include/linux/bitmap.h | 30 +++++++++++++++++++++++++++
tools/lib/bitmap.c | 15 ++++++++++++++
tools/perf/builtin-record.c | 27 ++++++++++++++++++------
tools/perf/util/mmap.c | 40 ++++++++++++++++++++++++++++++------
tools/perf/util/mmap.h | 13 +++++++++++-
5 files changed, 112 insertions(+), 13 deletions(-)
---
Changes in v3:
- implemented perf_mmap__print_cpu_mask() function
- use perf_mmap__print_cpu_mask() to log thread and mmap cpus masks
when verbose level is equal to 2
Changes in v2:
- implemented bitmap_free() for symmetry with bitmap_alloc()
- capitalized MMAP_CPU_MASK_BYTES() macro
- returned -1 from perf_mmap__setup_affinity_mask()
- implemented releasing of masks using bitmap_free()
- moved debug printing under -vv option
---
Testing:
tools/perf/perf record -vv --affinity=cpu -- ls
thread mask[8]: empty
...
------------------------------------------------------------
perf_event_attr:
...
mmap size 528384B
0x7fddecf760b8: mmap mask[8]: 0
0x7fddecf86180: mmap mask[8]: 1
0x7fddecf96248: mmap mask[8]: 2
0x7fddecfa6310: mmap mask[8]: 3
0x7fddecfb63d8: mmap mask[8]: 4
0x7fddecfc64a0: mmap mask[8]: 5
0x7fddecfd6568: mmap mask[8]: 6
0x7fddecfe6630: mmap mask[8]: 7
------------------------------------------------------------
perf_event_attr:
...
Synthesizing TSC conversion information
0x7fddecf760b8: thread mask[8]: 0
0x7fddecf86180: thread mask[8]: 1
0x7fddecf96248: thread mask[8]: 2
arch copy Documentation init kernel MAINTAINERS modules.builtin.modinfo perf.data scripts System.map vmlinux
block COPYING drivers ipc lbuild Makefile modules.order perf.data.old security tools vmlinux.o
certs CREDITS fs Kbuild lib mm Module.symvers README sound usr
config-5.2.7-100.fc29.x86_64 crypto include Kconfig LICENSES modules.builtin net samples stdio virt
0x7fddecfa6310: thread mask[8]: 3
0x7fddecfb63d8: thread mask[8]: 4
0x7fddecfc64a0: thread mask[8]: 5
0x7fddecfd6568: thread mask[8]: 6
0x7fddecfe6630: thread mask[8]: 7
0x7fddecf760b8: thread mask[8]: 0
0x7fddecf86180: thread mask[8]: 1
0x7fddecf96248: thread mask[8]: 2
0x7fddecfa6310: thread mask[8]: 3
0x7fddecfb63d8: thread mask[8]: 4
0x7fddecfc64a0: thread mask[8]: 5
0x7fddecfd6568: thread mask[8]: 6
0x7fddecfe6630: thread mask[8]: 7
[ perf record: Woken up 0 times to write data ]
0x7fddecf760b8: thread mask[8]: 0
0x7fddecf86180: thread mask[8]: 1
0x7fddecf96248: thread mask[8]: 2
0x7fddecfa6310: thread mask[8]: 3
0x7fddecfb63d8: thread mask[8]: 4
0x7fddecfc64a0: thread mask[8]: 5
0x7fddecfd6568: thread mask[8]: 6
0x7fddecfe6630: thread mask[8]: 7
Looking at the vmlinux_path (8 entries long)
Using vmlinux for symbols
[ perf record: Captured and wrote 0.014 MB perf.data (8 samples) ]
--
2.20.1
Extend tools bitmap API with bitmap_equal() implementation.
The implementation has been derived from the kernel.
Extend tools bitmap API with bitmap_free() implementation for
symmetry with bitmap_alloc() function.
Signed-off-by: Alexey Budankov <[email protected]>
---
Changes in v2:
- implemented bitmap_free() for symmetry with bitmap_alloc()
---
tools/include/linux/bitmap.h | 30 ++++++++++++++++++++++++++++++
tools/lib/bitmap.c | 15 +++++++++++++++
2 files changed, 45 insertions(+)
diff --git a/tools/include/linux/bitmap.h b/tools/include/linux/bitmap.h
index 05dca5c203f3..477a1cae513f 100644
--- a/tools/include/linux/bitmap.h
+++ b/tools/include/linux/bitmap.h
@@ -15,6 +15,8 @@ void __bitmap_or(unsigned long *dst, const unsigned long *bitmap1,
const unsigned long *bitmap2, int bits);
int __bitmap_and(unsigned long *dst, const unsigned long *bitmap1,
const unsigned long *bitmap2, unsigned int bits);
+int __bitmap_equal(const unsigned long *bitmap1,
+ const unsigned long *bitmap2, unsigned int bits);
void bitmap_clear(unsigned long *map, unsigned int start, int len);
#define BITMAP_FIRST_WORD_MASK(start) (~0UL << ((start) & (BITS_PER_LONG - 1)))
@@ -123,6 +125,15 @@ static inline unsigned long *bitmap_alloc(int nbits)
return calloc(1, BITS_TO_LONGS(nbits) * sizeof(unsigned long));
}
+/*
+ * bitmap_free - Free bitmap
+ * @bitmap: pointer to bitmap
+ */
+static inline void bitmap_free(unsigned long *bitmap)
+{
+ free(bitmap);
+}
+
/*
* bitmap_scnprintf - print bitmap list into buffer
* @bitmap: bitmap
@@ -148,4 +159,23 @@ static inline int bitmap_and(unsigned long *dst, const unsigned long *src1,
return __bitmap_and(dst, src1, src2, nbits);
}
+#ifdef __LITTLE_ENDIAN
+#define BITMAP_MEM_ALIGNMENT 8
+#else
+#define BITMAP_MEM_ALIGNMENT (8 * sizeof(unsigned long))
+#endif
+#define BITMAP_MEM_MASK (BITMAP_MEM_ALIGNMENT - 1)
+#define IS_ALIGNED(x, a) (((x) & ((typeof(x))(a) - 1)) == 0)
+
+static inline int bitmap_equal(const unsigned long *src1,
+ const unsigned long *src2, unsigned int nbits)
+{
+ if (small_const_nbits(nbits))
+ return !((*src1 ^ *src2) & BITMAP_LAST_WORD_MASK(nbits));
+ if (__builtin_constant_p(nbits & BITMAP_MEM_MASK) &&
+ IS_ALIGNED(nbits, BITMAP_MEM_ALIGNMENT))
+ return !memcmp(src1, src2, nbits / 8);
+ return __bitmap_equal(src1, src2, nbits);
+}
+
#endif /* _PERF_BITOPS_H */
diff --git a/tools/lib/bitmap.c b/tools/lib/bitmap.c
index 38494782be06..5043747ef6c5 100644
--- a/tools/lib/bitmap.c
+++ b/tools/lib/bitmap.c
@@ -71,3 +71,18 @@ int __bitmap_and(unsigned long *dst, const unsigned long *bitmap1,
BITMAP_LAST_WORD_MASK(bits));
return result != 0;
}
+
+int __bitmap_equal(const unsigned long *bitmap1,
+ const unsigned long *bitmap2, unsigned int bits)
+{
+ unsigned int k, lim = bits/BITS_PER_LONG;
+ for (k = 0; k < lim; ++k)
+ if (bitmap1[k] != bitmap2[k])
+ return 0;
+
+ if (bits % BITS_PER_LONG)
+ if ((bitmap1[k] ^ bitmap2[k]) & BITMAP_LAST_WORD_MASK(bits))
+ return 0;
+
+ return 1;
+}
--
2.20.1
Declare a dedicated struct map_cpu_mask type for cpu masks of
arbitrary length. Mask is available thru bits pointer and the
mask length is kept in nbits field. MMAP_CPU_MASK_BYTES() macro
returns mask storage size in bytes. perf_mmap__print_cpu_mask()
function can be used to log text representation of the mask.
Signed-off-by: Alexey Budankov <[email protected]>
---
Changes in v3:
- implemented perf_mmap__print_cpu_mask() function
Changes in v2:
- capitalized MMAP_CPU_MASK_BYTES() macro
---
tools/perf/util/mmap.c | 12 ++++++++++++
tools/perf/util/mmap.h | 11 +++++++++++
2 files changed, 23 insertions(+)
diff --git a/tools/perf/util/mmap.c b/tools/perf/util/mmap.c
index 063d1b93c53d..30ff7aef06f2 100644
--- a/tools/perf/util/mmap.c
+++ b/tools/perf/util/mmap.c
@@ -23,6 +23,18 @@
#include "mmap.h"
#include "../perf.h"
#include <internal/lib.h> /* page_size */
+#include <linux/bitmap.h>
+
+#define MASK_SIZE 1023
+void perf_mmap__print_cpu_mask(struct mmap_cpu_mask *mask, const char *tag)
+{
+ char buf[MASK_SIZE + 1];
+ size_t len;
+
+ len = bitmap_scnprintf(mask->bits, mask->nbits, buf, MASK_SIZE);
+ buf[len] = '\0';
+ pr_debug("%p: %s mask[%ld]: %s\n", mask, tag, mask->nbits, buf);
+}
size_t mmap__mmap_len(struct mmap *map)
{
diff --git a/tools/perf/util/mmap.h b/tools/perf/util/mmap.h
index bee4e83f7109..598e2def8a48 100644
--- a/tools/perf/util/mmap.h
+++ b/tools/perf/util/mmap.h
@@ -15,6 +15,15 @@
#include "event.h"
struct aiocb;
+
+struct mmap_cpu_mask {
+ unsigned long *bits;
+ size_t nbits;
+};
+
+#define MMAP_CPU_MASK_BYTES(m) \
+ (BITS_TO_LONGS(((struct mmap_cpu_mask *)m)->nbits) * sizeof(unsigned long))
+
/**
* struct mmap - perf's ring buffer mmap details
*
@@ -52,4 +61,6 @@ int perf_mmap__push(struct mmap *md, void *to,
size_t mmap__mmap_len(struct mmap *map);
+void perf_mmap__print_cpu_mask(struct mmap_cpu_mask *mask, const char *tag);
+
#endif /*__PERF_MMAP_H */
--
2.20.1
Use struct mmap_cpu_mask type for tool's thread and mmap data
buffers to overcome current 1024 CPUs mask size limitation of
cpu_set_t type.
Currently glibc cpu_set_t type has internal mask size limit
of 1024 CPUs. Moving to struct mmap_cpu_mask type allows
overcoming that limit. tools bitmap API is used to manipulate
objects of struct mmap_cpu_mask type.
Reported-by: Andi Kleen <[email protected]>
Signed-off-by: Alexey Budankov <[email protected]>
---
Changes in v3:
- use perf_mmap__print_cpu_mask() to log thread and mmap cpus masks
when verbose level is equal to 2
Changes in v2:
- returned -1 from perf_mmap__setup_affinity_mask()
- moved debug printing under -vv option
- implemented releasing of masks using bitmap_free()
---
tools/perf/builtin-record.c | 27 +++++++++++++++++++++------
tools/perf/util/mmap.c | 28 ++++++++++++++++++++++------
tools/perf/util/mmap.h | 2 +-
3 files changed, 44 insertions(+), 13 deletions(-)
diff --git a/tools/perf/builtin-record.c b/tools/perf/builtin-record.c
index b95c000c1ed9..b36699bf296b 100644
--- a/tools/perf/builtin-record.c
+++ b/tools/perf/builtin-record.c
@@ -62,6 +62,7 @@
#include <linux/string.h>
#include <linux/time64.h>
#include <linux/zalloc.h>
+#include <linux/bitmap.h>
struct switch_output {
bool enabled;
@@ -93,7 +94,7 @@ struct record {
bool timestamp_boundary;
struct switch_output switch_output;
unsigned long long samples;
- cpu_set_t affinity_mask;
+ struct mmap_cpu_mask affinity_mask;
unsigned long output_max_size; /* = 0: unlimited */
};
@@ -954,10 +955,15 @@ static struct perf_event_header finished_round_event = {
static void record__adjust_affinity(struct record *rec, struct mmap *map)
{
if (rec->opts.affinity != PERF_AFFINITY_SYS &&
- !CPU_EQUAL(&rec->affinity_mask, &map->affinity_mask)) {
- CPU_ZERO(&rec->affinity_mask);
- CPU_OR(&rec->affinity_mask, &rec->affinity_mask, &map->affinity_mask);
- sched_setaffinity(0, sizeof(rec->affinity_mask), &rec->affinity_mask);
+ !bitmap_equal(rec->affinity_mask.bits, map->affinity_mask.bits,
+ rec->affinity_mask.nbits)) {
+ bitmap_zero(rec->affinity_mask.bits, rec->affinity_mask.nbits);
+ bitmap_or(rec->affinity_mask.bits, rec->affinity_mask.bits,
+ map->affinity_mask.bits, rec->affinity_mask.nbits);
+ sched_setaffinity(0, MMAP_CPU_MASK_BYTES(&rec->affinity_mask),
+ (cpu_set_t *)rec->affinity_mask.bits);
+ if (verbose == 2)
+ perf_mmap__print_cpu_mask(&map->affinity_mask, "thread");
}
}
@@ -2389,7 +2395,6 @@ int cmd_record(int argc, const char **argv)
# undef REASON
#endif
- CPU_ZERO(&rec->affinity_mask);
rec->opts.affinity = PERF_AFFINITY_SYS;
rec->evlist = evlist__new();
@@ -2455,6 +2460,14 @@ int cmd_record(int argc, const char **argv)
symbol__init(NULL);
+ rec->affinity_mask.nbits = cpu__max_cpu();
+ rec->affinity_mask.bits = bitmap_alloc(rec->affinity_mask.nbits);
+ if (!rec->affinity_mask.bits) {
+ pr_err("Failed to allocate thread mask for %ld cpus\n", rec->affinity_mask.nbits);
+ return -ENOMEM;
+ }
+ pr_debug2("thread mask[%ld]: empty\n", rec->affinity_mask.nbits);
+
err = record__auxtrace_init(rec);
if (err)
goto out;
@@ -2569,6 +2582,8 @@ int cmd_record(int argc, const char **argv)
err = __cmd_record(&record, argc, argv);
out:
+ if (rec->affinity_mask.bits)
+ bitmap_free(rec->affinity_mask.bits);
evlist__delete(rec->evlist);
symbol__exit();
auxtrace_record__free(rec->itr);
diff --git a/tools/perf/util/mmap.c b/tools/perf/util/mmap.c
index 30ff7aef06f2..615d05870849 100644
--- a/tools/perf/util/mmap.c
+++ b/tools/perf/util/mmap.c
@@ -219,6 +219,9 @@ static void perf_mmap__aio_munmap(struct mmap *map __maybe_unused)
void mmap__munmap(struct mmap *map)
{
+ if (map->affinity_mask.bits)
+ bitmap_free(map->affinity_mask.bits);
+
perf_mmap__aio_munmap(map);
if (map->data != NULL) {
munmap(map->data, mmap__mmap_len(map));
@@ -227,7 +230,7 @@ void mmap__munmap(struct mmap *map)
auxtrace_mmap__munmap(&map->auxtrace_mmap);
}
-static void build_node_mask(int node, cpu_set_t *mask)
+static void build_node_mask(int node, struct mmap_cpu_mask *mask)
{
int c, cpu, nr_cpus;
const struct perf_cpu_map *cpu_map = NULL;
@@ -240,17 +243,23 @@ static void build_node_mask(int node, cpu_set_t *mask)
for (c = 0; c < nr_cpus; c++) {
cpu = cpu_map->map[c]; /* map c index to online cpu index */
if (cpu__get_node(cpu) == node)
- CPU_SET(cpu, mask);
+ set_bit(cpu, mask->bits);
}
}
-static void perf_mmap__setup_affinity_mask(struct mmap *map, struct mmap_params *mp)
+static int perf_mmap__setup_affinity_mask(struct mmap *map, struct mmap_params *mp)
{
- CPU_ZERO(&map->affinity_mask);
+ map->affinity_mask.nbits = cpu__max_cpu();
+ map->affinity_mask.bits = bitmap_alloc(map->affinity_mask.nbits);
+ if (!map->affinity_mask.bits)
+ return -1;
+
if (mp->affinity == PERF_AFFINITY_NODE && cpu__max_node() > 1)
build_node_mask(cpu__get_node(map->core.cpu), &map->affinity_mask);
else if (mp->affinity == PERF_AFFINITY_CPU)
- CPU_SET(map->core.cpu, &map->affinity_mask);
+ set_bit(map->core.cpu, map->affinity_mask.bits);
+
+ return 0;
}
int mmap__mmap(struct mmap *map, struct mmap_params *mp, int fd, int cpu)
@@ -261,7 +270,14 @@ int mmap__mmap(struct mmap *map, struct mmap_params *mp, int fd, int cpu)
return -1;
}
- perf_mmap__setup_affinity_mask(map, mp);
+ if (perf_mmap__setup_affinity_mask(map, mp)) {
+ pr_debug2("failed to alloc mmap affinity mask, error %d\n",
+ errno);
+ return -1;
+ }
+
+ if (verbose == 2)
+ perf_mmap__print_cpu_mask(&map->affinity_mask, "mmap");
map->core.flush = mp->flush;
diff --git a/tools/perf/util/mmap.h b/tools/perf/util/mmap.h
index 598e2def8a48..ab6449675d5b 100644
--- a/tools/perf/util/mmap.h
+++ b/tools/perf/util/mmap.h
@@ -40,7 +40,7 @@ struct mmap {
int nr_cblocks;
} aio;
#endif
- cpu_set_t affinity_mask;
+ struct mmap_cpu_mask affinity_mask;
void *data;
int comp_level;
};
--
2.20.1