Re: [PATCH v6 03/20] perf record: Introduce thread local variable

Re: [PATCH v6 03/20] perf record: Introduce thread local variable

[Riccardo Mancini]

Hi,

thank you very much for your work for adding threading capabilites to perf
record. 
I did some testing on your entire patchset, especially checking for memory
issues using ASan. This is just the first of a couple of emails to point out
some issues I found.
I will also do additional tests in the future.

On Wed, 2021-05-26 at 13:52 +0300, Alexey Bayduraev wrote:
SNIP
> @@ -2220,18 +2275,20 @@ static int __cmd_record(struct record *rec, int argc,
> const char **argv)
>                 goto out_child;
>         }
>  
> -       if (!quiet)
> -               fprintf(stderr, "[ perf record: Woken up %ld times to write data
> ]\n", waking);
> -
>         if (target__none(&rec->opts.target))
>                 record__synthesize_workload(rec, true);
>  
>  out_child:
> +       record__stop_threads(rec, &waking);
> +out_free_threads:
>         record__free_thread_data(rec);
>         evlist__finalize_ctlfd(rec->evlist);
>         record__mmap_read_all(rec, true);
>         record__aio_mmap_read_sync(rec);

record__mmap_read_all should be moved before record__free_thread_data since it
uses the thread_data that's just been freed.
Furthermore, record__mmap_read_all should also be moved before the
out_free_threads label, since it cannot be called unless record__start_threads
succeeded, otherwise thread would be NULL and will cause a segfault (it happens
if there is an error somewhere else in perf, for example).

In my tests the following order works, but it should be double checked for
possible side-effects of this order change.

out_child:
        record__stop_threads(rec, &waking);
        record__mmap_read_all(rec, true);
out_free_threads:
        record__free_thread_data(rec);
        evlist__finalize_ctlfd(rec->evlist);
        record__aio_mmap_read_sync(rec);

Thanks,
Riccardo

> +       if (!quiet)
> +               fprintf(stderr, "[ perf record: Woken up %ld times to write data
> ]\n", waking);
> +
>         if (rec->session->bytes_transferred && rec->session->bytes_compressed) {
>                 ratio = (float)rec->session->bytes_transferred/(float)rec-
> >session->bytes_compressed;
>                 session->header.env.comp_ratio = ratio + 0.5;
SNIP

Re: [PATCH v6 05/20] perf record: Start threads in the beginning of trace streaming

[Riccardo Mancini]

Hi,

On Wed, 2021-05-26 at 13:52 +0300, Alexey Bayduraev wrote:
> Start thread in detached state because its management is implemented
> via messaging to avoid any scaling issues. Block signals prior thread
> start so only main tool thread would be notified on external async
> signals during data collection. Thread affinity mask is used to assign
> eligible cpus for the thread to run. Wait and sync on thread start using
> thread ack pipe.
> 
> Signed-off-by: Alexey Bayduraev <alexey.v.bayduraev@linux.intel.com>
> ---
>  tools/perf/builtin-record.c | 106 +++++++++++++++++++++++++++++++++++-
>  1 file changed, 105 insertions(+), 1 deletion(-)
> 
> diff --git a/tools/perf/builtin-record.c b/tools/perf/builtin-record.c
> index 838c1f779849..88fad12cbe5b 100644
> --- a/tools/perf/builtin-record.c
> +++ b/tools/perf/builtin-record.c
> @@ -1423,6 +1423,66 @@ static void record__thread_munmap_filtered(struct
> fdarray *fda, int fd,
>                 perf_mmap__put(map);
>  }
>  
> +static void *record__thread(void *arg)
> +{
> +       enum thread_msg msg = THREAD_MSG__READY;
> +       bool terminate = false;
> +       struct fdarray *pollfd;
> +       int err, ctlfd_pos;
> +
> +       thread = arg;
> +       thread->tid = syscall(SYS_gettid);
> +
> +       err = write(thread->pipes.ack[1], &msg, sizeof(msg));
> +       if (err == -1)
> +               pr_err("threads[%d]: failed to notify on start: %s", thread-
> >tid, strerror(errno));
> +
> +       pr_debug("threads[%d]: started on cpu=%d\n", thread->tid,
> sched_getcpu());
> +
> +       pollfd = &thread->pollfd;
> +       ctlfd_pos = thread->ctlfd_pos;
> +
> +       for (;;) {
> +               unsigned long long hits = thread->samples;
> +
> +               if (record__mmap_read_all(thread->rec, false) < 0 ||
> terminate)
> +                       break;
> +
> +               if (hits == thread->samples) {
> +
> +                       err = fdarray__poll(pollfd, -1);
> +                       /*
> +                        * Propagate error, only if there's any. Ignore
> positive
> +                        * number of returned events and interrupt error.
> +                        */
> +                       if (err > 0 || (err < 0 && errno == EINTR))
> +                               err = 0;
> +                       thread->waking++;
> +
> +                       if (fdarray__filter(pollfd, POLLERR | POLLHUP,
> +                                           record__thread_munmap_filtered,
> NULL) == 0)
> +                               break;
> +               }
> +
> +               if (pollfd->entries[ctlfd_pos].revents & POLLHUP) {
> +                       terminate = true;
> +                       close(thread->pipes.msg[0]);
> +                       pollfd->entries[ctlfd_pos].fd = -1;
> +                       pollfd->entries[ctlfd_pos].events = 0;
> +               }
> +
> +               pollfd->entries[ctlfd_pos].revents = 0;
> +       }
> +       record__mmap_read_all(thread->rec, true);
> +
> +       err = write(thread->pipes.ack[1], &msg, sizeof(msg));
> +       if (err == -1)
> +               pr_err("threads[%d]: failed to notify on termination: %s",
> +                      thread->tid, strerror(errno));
> +
> +       return NULL;
> +}
> +
>  static void record__init_features(struct record *rec)
>  {
>         struct perf_session *session = rec->session;
> @@ -1886,13 +1946,57 @@ static int record__terminate_thread(struct thread_data
> *thread_data)
>  
>  static int record__start_threads(struct record *rec)
>  {
> +       int t, tt, ret = 0, nr_threads = rec->nr_threads;
>         struct thread_data *thread_data = rec->thread_data;
> +       sigset_t full, mask;
> +       pthread_t handle;
> +       pthread_attr_t attrs;
> +
> +       sigfillset(&full);
> +       if (sigprocmask(SIG_SETMASK, &full, &mask)) {
> +               pr_err("Failed to block signals on threads start: %s\n",
> strerror(errno));
> +               return -1;
> +       }
> +
> +       pthread_attr_init(&attrs);
> +       pthread_attr_setdetachstate(&attrs, PTHREAD_CREATE_DETACHED);
> +
> +       for (t = 1; t < nr_threads; t++) {
> +               enum thread_msg msg = THREAD_MSG__UNDEFINED;
> +
> +               pthread_attr_setaffinity_np(&attrs,
> +                                          
> MMAP_CPU_MASK_BYTES(&(thread_data[t].mask->affinity)),
> +                                           (cpu_set_t *)(thread_data[t].mask-
> >affinity.bits));
> +
> +               if (pthread_create(&handle, &attrs, record__thread,
> &thread_data[t])) {
> +                       for (tt = 1; tt < t; tt++)
> +                               record__terminate_thread(&thread_data[t]);
> +                       pr_err("Failed to start threads: %s\n",
> strerror(errno));
> +                       ret = -1;
> +                       goto out_err;
> +               }
> +
> +               if (read(thread_data[t].pipes.ack[0], &msg, sizeof(msg)) > 0)
> +                       pr_debug2("threads[%d]: sent %s\n", rec-
> >thread_data[t].tid,
> +                                thread_msg_tags[msg]);
> +       }
> +
> +       if (nr_threads > 1) {
> +               sched_setaffinity(0, MMAP_CPU_MASK_BYTES(&thread_data[0].mask-
> >affinity),
> +                                 (cpu_set_t *)thread_data[0].mask-
> >affinity.bits);
> +       }
>  
>         thread = &thread_data[0];
>  
>         pr_debug("threads[%d]: started on cpu=%d\n", thread->tid,
> sched_getcpu());
>  
> -       return 0;
> +out_err:
> +       if (sigprocmask(SIG_SETMASK, &mask, NULL)) {
> +               pr_err("Failed to unblock signals on threads start: %s\n",
> strerror(errno));
> +               ret = -1;
> +       }
> +
> +       return ret;
>  }

ASan complains of a memory leak of the attrs, since pthread_attr_destroy is
missing. It could be added just after out_err label.

Thanks,
Riccardo

>  
>  static int record__stop_threads(struct record *rec, unsigned long *waking)

Re: [PATCH v6 10/20] perf record: Introduce --threads=<spec> command line option

[Riccardo Mancini]

Hi,

the parameter provided to --thread is not checked. If an invalid value is
inserted, it will crash.

On Wed, 2021-05-26 at 13:53 +0300, Alexey Bayduraev wrote:
> Provide --threads option in perf record command line interface.
> The option can have a value in the form of masks that specify
> cpus to be monitored with data streaming threads and its layout
> in system topology. The masks can be filtered using cpu mask
> provided via -C option.
> 
> The specification value can be user defined list of masks. Masks
> separated by colon define cpus to be monitored by one thread and
> affinity mask of that thread is separated by slash. For example:
> <cpus mask 1>/<affinity mask 1>:<cpu mask 2>/<affinity mask 2>
> specifies parallel threads layout that consists of two threads
> with corresponding assigned cpus to be monitored.
> 
> The specification value can be a string e.g. "cpu", "core" or
> "socket" meaning creation of data streaming thread for every
> cpu or core or socket to monitor distinct cpus or cpus grouped
> by core or socket.
> 
> The option provided with no or empty value defaults to per-cpu
> parallel threads layout creating data streaming thread for every
> cpu being monitored.
> 
> Feature design and implementation are based on prototypes [1], [2].
> 
> [1] git clone https://git.kernel.org/pub/scm/linux/kernel/git/jolsa/perf.git -
> b perf/record_threads
> [2] https://lore.kernel.org/lkml/20180913125450.21342-1-jolsa@kernel.org/
> 
> Suggested-by: Jiri Olsa <jolsa@kernel.org>
> Suggested-by: Namhyung Kim <namhyung@kernel.org>
> Acked-by: Andi Kleen <ak@linux.intel.com>
> Signed-off-by: Alexey Bayduraev <alexey.v.bayduraev@linux.intel.com>
> ---
>  tools/perf/builtin-record.c | 347 +++++++++++++++++++++++++++++++++++-
>  tools/perf/util/record.h    |   1 +
>  2 files changed, 346 insertions(+), 2 deletions(-)
> 
> diff --git a/tools/perf/builtin-record.c b/tools/perf/builtin-record.c
> index e118efe21ba7..a54d72475629 100644
> --- a/tools/perf/builtin-record.c
> +++ b/tools/perf/builtin-record.c
> @@ -51,6 +51,7 @@
>  #include "util/evlist-hybrid.h"
>  #include "asm/bug.h"
>  #include "perf.h"
> +#include "cputopo.h"
>  
>  #include <errno.h>
>  #include <inttypes.h>
> @@ -122,6 +123,20 @@ static const char *thread_msg_tags[THREAD_MSG__MAX] = {
>         "UNDEFINED", "READY"
>  };
>  
> +enum thread_spec {
> +       THREAD_SPEC__UNDEFINED = 0,
> +       THREAD_SPEC__CPU,
> +       THREAD_SPEC__CORE,
> +       THREAD_SPEC__SOCKET,
> +       THREAD_SPEC__NUMA,
> +       THREAD_SPEC__USER,
> +       THREAD_SPEC__MAX,
> +};
> +
> +static const char *thread_spec_tags[THREAD_SPEC__MAX] = {
> +       "undefined", "cpu", "core", "socket", "numa", "user"
> +};
> +
>  struct record {
>         struct perf_tool        tool;
>         struct record_opts      opts;
> @@ -2721,6 +2736,70 @@ static void record__thread_mask_free(struct thread_mask
> *mask)
>         record__mmap_cpu_mask_free(&mask->affinity);
>  }
>  
> +static int record__thread_mask_or(struct thread_mask *dest, struct
> thread_mask *src1,
> +                                  struct thread_mask *src2)
> +{
> +       if (src1->maps.nbits != src2->maps.nbits ||
> +           dest->maps.nbits != src1->maps.nbits ||
> +           src1->affinity.nbits != src2->affinity.nbits ||
> +           dest->affinity.nbits != src1->affinity.nbits)
> +               return -EINVAL;
> +
> +       bitmap_or(dest->maps.bits, src1->maps.bits,
> +                 src2->maps.bits, src1->maps.nbits);
> +       bitmap_or(dest->affinity.bits, src1->affinity.bits,
> +                 src2->affinity.bits, src1->affinity.nbits);
> +
> +       return 0;
> +}
> +
> +static int record__thread_mask_intersects(struct thread_mask *mask_1, struct
> thread_mask *mask_2)
> +{
> +       int res1, res2;
> +
> +       if (mask_1->maps.nbits != mask_2->maps.nbits ||
> +           mask_1->affinity.nbits != mask_2->affinity.nbits)
> +               return -EINVAL;
> +
> +       res1 = bitmap_intersects(mask_1->maps.bits, mask_2->maps.bits,
> +                                mask_1->maps.nbits);
> +       res2 = bitmap_intersects(mask_1->affinity.bits, mask_2->affinity.bits,
> +                                mask_1->affinity.nbits);
> +       if (res1 || res2)
> +               return 1;
> +
> +       return 0;
> +}
> +
> +static int record__parse_threads(const struct option *opt, const char *str,
> int unset)
> +{
> +       int s;
> +       struct record_opts *opts = opt->value;
> +
> +       if (unset || !str || !strlen(str)) {
> +               opts->threads_spec = THREAD_SPEC__CPU;
> +       } else {
> +               for (s = 1; s < THREAD_SPEC__MAX; s++) {
> +                       if (s == THREAD_SPEC__USER) {
> +                               opts->threads_user_spec = strdup(str);
> +                               opts->threads_spec = THREAD_SPEC__USER;
> +                               break;
> +                       }
> +                       if (!strncasecmp(str, thread_spec_tags[s],
> strlen(thread_spec_tags[s]))) {
> +                               opts->threads_spec = s;
> +                               break;
> +                       }
> +               }
> +       }
> +
> +       pr_debug("threads_spec: %s", thread_spec_tags[opts->threads_spec]);
> +       if (opts->threads_spec == THREAD_SPEC__USER)
> +               pr_debug("=[%s]", opts->threads_user_spec);
> +       pr_debug("\n");
> +
> +       return 0;
> +}
> +
>  static int parse_output_max_size(const struct option *opt,
>                                  const char *str, int unset)
>  {
> @@ -3164,6 +3243,9 @@ static struct option __record_options[] = {
>                      "\t\t\t  Optionally send control command completion
> ('ack\\n') to ack-fd descriptor.\n"
>                      "\t\t\t  Alternatively, ctl-fifo / ack-fifo will be
> opened and used as ctl-fd / ack-fd.",
>                       parse_control_option),
> +       OPT_CALLBACK_OPTARG(0, "threads", &record.opts, NULL, "spec",
> +                           "write collected trace data into several data
> files using parallel threads",
> +                           record__parse_threads),
>         OPT_END()
>  };
>  
> @@ -3177,6 +3259,17 @@ static void record__mmap_cpu_mask_init(struct
> mmap_cpu_mask *mask, struct perf_c
>                 set_bit(cpus->map[c], mask->bits);
>  }
>  
> +static void record__mmap_cpu_mask_init_spec(struct mmap_cpu_mask *mask, char
> *mask_spec)
> +{
> +       struct perf_cpu_map *cpus;
> +
> +       cpus = perf_cpu_map__new(mask_spec);
> +       if (cpus) {
> +               record__mmap_cpu_mask_init(mask, cpus);
> +               perf_cpu_map__put(cpus);
> +       }
> +}
> +
>  static int record__alloc_thread_masks(struct record *rec, int nr_threads, int
> nr_bits)
>  {
>         int t, ret;
> @@ -3196,6 +3289,229 @@ static int record__alloc_thread_masks(struct record
> *rec, int nr_threads, int nr
>  
>         return 0;
>  }
> +
> +static int record__init_thread_cpu_masks(struct record *rec, struct
> perf_cpu_map *cpus)
> +{
> +       int t, ret, nr_cpus = perf_cpu_map__nr(cpus);
> +
> +       ret = record__alloc_thread_masks(rec, nr_cpus, cpu__max_cpu());
> +       if (ret)
> +               return ret;
> +
> +       rec->nr_threads = nr_cpus;
> +       pr_debug("threads: nr_threads=%d\n", rec->nr_threads);
> +
> +       for (t = 0; t < rec->nr_threads; t++) {
> +               set_bit(cpus->map[t], rec->thread_masks[t].maps.bits);
> +               pr_debug("thread_masks[%d]: maps mask [%d]\n", t, cpus-
> >map[t]);
> +               set_bit(cpus->map[t], rec->thread_masks[t].affinity.bits);
> +               pr_debug("thread_masks[%d]: affinity mask [%d]\n", t, cpus-
> >map[t]);
> +       }
> +
> +       return 0;
> +}
> +
> +static int record__init_thread_masks_spec(struct record *rec, struct
> perf_cpu_map *cpus,
> +                                         char **maps_spec, char
> **affinity_spec, u32 nr_spec)
> +{
> +       u32 s;
> +       int ret = 0, nr_threads = 0;
> +       struct mmap_cpu_mask cpus_mask;
> +       struct thread_mask thread_mask, full_mask, *prev_masks;
> +
> +       ret = record__mmap_cpu_mask_alloc(&cpus_mask, cpu__max_cpu());
> +       if (ret)
> +               goto out;
> +       record__mmap_cpu_mask_init(&cpus_mask, cpus);
> +       ret = record__thread_mask_alloc(&thread_mask, cpu__max_cpu());
> +       if (ret)
> +               goto out_free_cpu_mask;
> +       ret = record__thread_mask_alloc(&full_mask, cpu__max_cpu());
> +       if (ret)
> +               goto out_free_thread_mask;
> +       record__thread_mask_clear(&full_mask);
> +
> +       for (s = 0; s < nr_spec; s++) {
> +               record__thread_mask_clear(&thread_mask);
> +
> +               record__mmap_cpu_mask_init_spec(&thread_mask.maps,
> maps_spec[s]);
> +               record__mmap_cpu_mask_init_spec(&thread_mask.affinity,
> affinity_spec[s]);
> +
> +               if (!bitmap_and(thread_mask.maps.bits, thread_mask.maps.bits,
> +                               cpus_mask.bits, thread_mask.maps.nbits) ||
> +                   !bitmap_and(thread_mask.affinity.bits,
> thread_mask.affinity.bits,
> +                               cpus_mask.bits, thread_mask.affinity.nbits))
> +                       continue;
> +
> +               ret = record__thread_mask_intersects(&thread_mask,
> &full_mask);
> +               if (ret)
> +                       goto out_free_full_mask;
> +               record__thread_mask_or(&full_mask, &full_mask, &thread_mask);
> +
> +               prev_masks = rec->thread_masks;
> +               rec->thread_masks = realloc(rec->thread_masks,
> +                                           (nr_threads + 1) * sizeof(struct
> thread_mask));
> +               if (!rec->thread_masks) {
> +                       pr_err("Failed to allocate thread masks\n");
> +                       rec->thread_masks = prev_masks;
> +                       ret = -ENOMEM;
> +                       goto out_free_full_mask;
> +               }
> +               rec->thread_masks[nr_threads] = thread_mask;
> +               if (verbose) {
> +                       pr_debug("thread_masks[%d]: addr=", nr_threads);
> +                       mmap_cpu_mask__scnprintf(&rec-
> >thread_masks[nr_threads].maps, "maps");
> +                       pr_debug("thread_masks[%d]: addr=", nr_threads);
> +                       mmap_cpu_mask__scnprintf(&rec-
> >thread_masks[nr_threads].affinity,
> +                                                "affinity");
> +               }
> +               nr_threads++;
> +               ret = record__thread_mask_alloc(&thread_mask, cpu__max_cpu());
> +               if (ret)
> +                       goto out_free_full_mask;
> +       }
> +

Here it can be checked if nr_threads is > 0. Otherwise, print an error message
epxlaining the reason and return an error.
nr_threads can be 0 if the user inserts out-of-range CPUs in the spec, e.g. 
"--threads=15/17" on a 4 CPU machine.

> +       rec->nr_threads = nr_threads;
> +       pr_debug("threads: nr_threads=%d\n", rec->nr_threads);
> +
> +out_free_full_mask:
> +       record__thread_mask_free(&full_mask);
> +out_free_thread_mask:
> +       record__thread_mask_free(&thread_mask);
> +out_free_cpu_mask:
> +       record__mmap_cpu_mask_free(&cpus_mask);
> +out:
> +       return ret;
> +}
> +
> +static int record__init_thread_core_masks(struct record *rec, struct
> perf_cpu_map *cpus)
> +{
> +       int ret;
> +       struct cpu_topology *topo;
> +
> +       topo = cpu_topology__new();
> +       if (!topo)
> +               return -EINVAL;
> +
> +       ret = record__init_thread_masks_spec(rec, cpus, topo->thread_siblings,
> +                                            topo->thread_siblings, topo-
> >thread_sib);
> +       cpu_topology__delete(topo);
> +
> +       return ret;
> +}
> +
> +static int record__init_thread_socket_masks(struct record *rec, struct
> perf_cpu_map *cpus)
> +{
> +       int ret;
> +       struct cpu_topology *topo;
> +
> +       topo = cpu_topology__new();
> +       if (!topo)
> +               return -EINVAL;
> +
> +       ret = record__init_thread_masks_spec(rec, cpus, topo->core_siblings,
> +                                            topo->core_siblings, topo-
> >core_sib);
> +       cpu_topology__delete(topo);
> +
> +       return ret;
> +}
> +
> +static int record__init_thread_numa_masks(struct record *rec, struct
> perf_cpu_map *cpus)
> +{
> +       u32 s;
> +       int ret;
> +       char **spec;
> +       struct numa_topology *topo;
> +
> +       topo = numa_topology__new();
> +       if (!topo)
> +               return -EINVAL;
> +       spec = zalloc(topo->nr * sizeof(char *));
> +       if (!spec) {
> +               ret = -ENOMEM;
> +               goto out_delete_topo;
> +       }
> +       for (s = 0; s < topo->nr; s++)
> +               spec[s] = topo->nodes[s].cpus;
> +
> +       ret = record__init_thread_masks_spec(rec, cpus, spec, spec, topo->nr);
> +
> +       zfree(&spec);
> +
> +out_delete_topo:
> +       numa_topology__delete(topo);
> +
> +       return ret;
> +}
> +
> +static int record__init_thread_user_masks(struct record *rec, struct
> perf_cpu_map *cpus)
> +{
> +       int t, ret;
> +       u32 s, nr_spec = 0;
> +       char **maps_spec = NULL, **affinity_spec = NULL, **prev_spec;
> +       char *spec, *spec_ptr, *user_spec, *mask, *mask_ptr;
> +
> +       for (t = 0, user_spec = (char *)rec->opts.threads_user_spec; ; t++,
> user_spec = NULL) {
> +               spec = strtok_r(user_spec, ":", &spec_ptr);
> +               if (spec == NULL)
> +                       break;
> +               pr_debug(" spec[%d]: %s\n", t, spec);
> +               mask = strtok_r(spec, "/", &mask_ptr);
> +               if (mask == NULL)
> +                       break;
> +               pr_debug("  maps mask: %s\n", mask);
> +               prev_spec = maps_spec;
> +               maps_spec = realloc(maps_spec, (nr_spec + 1) * sizeof(char
> *));
> +               if (!maps_spec) {
> +                       pr_err("Failed to realloc maps_spec\n");
> +                       maps_spec = prev_spec;
> +                       ret = -ENOMEM;
> +                       goto out_free_all_specs;
> +               }
> +               maps_spec[nr_spec] = strdup(mask);
> +               if (!maps_spec[nr_spec]) {
> +                       pr_err("Failed to alloc maps_spec[%d]\n", nr_spec);
> +                       ret = -ENOMEM;
> +                       goto out_free_all_specs;
> +               }
> +               mask = strtok_r(NULL, "/", &mask_ptr);
> +               if (mask == NULL)

I think this should be a parse failure and return an error, not just skip to the
next one. 
Furthermore, maps_spec[nr_spec] should be freed before breaking/exiting.

Thanks,
Riccardo

> +                       break;
> +               pr_debug("  affinity mask: %s\n", mask);
> +               prev_spec = affinity_spec;
> +               affinity_spec = realloc(affinity_spec, (nr_spec + 1) *
> sizeof(char *));
> +               if (!affinity_spec) {
> +                       pr_err("Failed to realloc affinity_spec\n");
> +                       affinity_spec = prev_spec;
> +                       free(maps_spec[nr_spec]);
> +                       ret = -ENOMEM;
> +                       goto out_free_all_specs;
> +               }
> +               affinity_spec[nr_spec] = strdup(mask);
> +               if (!affinity_spec[nr_spec]) {
> +                       pr_err("Failed to alloc affinity_spec[%d]\n",
> nr_spec);
> +                       free(maps_spec[nr_spec]);
> +                       ret = -ENOMEM;
> +                       goto out_free_all_specs;
> +               }
> +               nr_spec++;
> +       }
> +
> +       ret = record__init_thread_masks_spec(rec, cpus, maps_spec,
> affinity_spec, nr_spec);
> +
> +out_free_all_specs:
> +       for (s = 0; s < nr_spec; s++) {
> +               if (maps_spec)
> +                       free(maps_spec[s]);
> +               if (affinity_spec)
> +                       free(affinity_spec[s]);
> +       }
> +       free(affinity_spec);
> +       free(maps_spec);
> +
> +       return ret;
> +}
> +
>  static int record__init_thread_default_masks(struct record *rec, struct
> perf_cpu_map *cpus)
>  {
>         int ret;
> @@ -3213,9 +3529,33 @@ static int record__init_thread_default_masks(struct
> record *rec, struct perf_cpu
>  
>  static int record__init_thread_masks(struct record *rec)
>  {
> +       int ret = 0;
>         struct perf_cpu_map *cpus = rec->evlist->core.cpus;
>  
> -       return record__init_thread_default_masks(rec, cpus);
> +       if (!record__threads_enabled(rec))
> +               return record__init_thread_default_masks(rec, cpus);
> +
> +       switch (rec->opts.threads_spec) {
> +       case THREAD_SPEC__CPU:
> +               ret = record__init_thread_cpu_masks(rec, cpus);
> +               break;
> +       case THREAD_SPEC__CORE:
> +               ret = record__init_thread_core_masks(rec, cpus);
> +               break;
> +       case THREAD_SPEC__SOCKET:
> +               ret = record__init_thread_socket_masks(rec, cpus);
> +               break;
> +       case THREAD_SPEC__NUMA:
> +               ret = record__init_thread_numa_masks(rec, cpus);
> +               break;
> +       case THREAD_SPEC__USER:
> +               ret = record__init_thread_user_masks(rec, cpus);
> +               break;
> +       default:
> +               break;
> +       }
> +
> +       return ret;
>  }
>  
>  static int record__fini_thread_masks(struct record *rec)
> @@ -3466,7 +3806,10 @@ int cmd_record(int argc, const char **argv)
>  
>         err = record__init_thread_masks(rec);
>         if (err) {
> -               pr_err("record__init_thread_masks failed, error %d\n", err);
> +               if (err > 0)
> +                       pr_err("ERROR: parallel data streaming masks (--
> threads) intersect.\n");
> +               else
> +                       pr_err("record__init_thread_masks failed, error %d\n",
> err);
>                 goto out;
>         }
>  
> diff --git a/tools/perf/util/record.h b/tools/perf/util/record.h
> index 4d68b7e27272..3da156498f47 100644
> --- a/tools/perf/util/record.h
> +++ b/tools/perf/util/record.h
> @@ -78,6 +78,7 @@ struct record_opts {
>         int           ctl_fd_ack;
>         bool          ctl_fd_close;
>         int           threads_spec;
> +       const char    *threads_user_spec;
>  };
>  
>  extern const char * const *record_usage;

Re: [PATCH v6 16/20] perf session: Introduce decompressor into trace reader object

[Riccardo Mancini]

Hi,

On Wed, 2021-05-26 at 13:53 +0300, Alexey Bayduraev wrote:
> Introduce decompressor into trace reader object so that decompression
> could be executed on per data file basis separately for every data
> file located in data directory.
> 
> Signed-off-by: Alexey Bayduraev <alexey.v.bayduraev@linux.intel.com>
> ---
>  tools/perf/util/session.c | 51 +++++++++++++++++++++++++++++----------
>  tools/perf/util/session.h |  1 +
>  2 files changed, 39 insertions(+), 13 deletions(-)
> 
> diff --git a/tools/perf/util/session.c b/tools/perf/util/session.c
> index 335c073bae87..6c9a682eb291 100644
> --- a/tools/perf/util/session.c
> +++ b/tools/perf/util/session.c
> @@ -73,6 +73,9 @@ struct reader {
>         u64              data_offset;
>         reader_cb_t      process;
>         bool             in_place_update;
> +       struct zstd_data zstd_data;
> +       struct decomp    *decomp;
> +       struct decomp    *decomp_last;
>         struct reader_state state;
>  };
>  
> @@ -85,7 +88,10 @@ static int perf_session__process_compressed_event(struct
> perf_session *session,
>         size_t decomp_size, src_size;
>         u64 decomp_last_rem = 0;
>         size_t mmap_len, decomp_len = session->header.env.comp_mmap_len;
> -       struct decomp *decomp, *decomp_last = session->decomp_last;
> +       struct decomp *decomp, *decomp_last = session->active_reader ?
> +               session->active_reader->decomp_last : session->decomp_last;
> +       struct zstd_data *zstd_data = session->active_reader ?
> +               &session->active_reader->zstd_data : &session->zstd_data;
>  
>         if (decomp_last) {
>                 decomp_last_rem = decomp_last->size - decomp_last->head;
> @@ -113,7 +119,7 @@ static int perf_session__process_compressed_event(struct
> perf_session *session,
>         src = (void *)event + sizeof(struct perf_record_compressed);
>         src_size = event->pack.header.size - sizeof(struct
> perf_record_compressed);
>  
> -       decomp_size = zstd_decompress_stream(&(session->zstd_data), src,
> src_size,
> +       decomp_size = zstd_decompress_stream(zstd_data, src, src_size,
>                                 &(decomp->data[decomp_last_rem]), decomp_len -
> decomp_last_rem);
>         if (!decomp_size) {
>                 munmap(decomp, mmap_len);
> @@ -123,12 +129,22 @@ static int perf_session__process_compressed_event(struct
> perf_session *session,
>  
>         decomp->size += decomp_size;
>  
> -       if (session->decomp == NULL) {
> -               session->decomp = decomp;
> -               session->decomp_last = decomp;
> +       if (session->active_reader) {
> +               if (session->active_reader->decomp == NULL) {
> +                       session->active_reader->decomp = decomp;
> +                       session->active_reader->decomp_last = decomp;
> +               } else {
> +                       session->active_reader->decomp_last->next = decomp;
> +                       session->active_reader->decomp_last = decomp;
> +               }
>         } else {
> -               session->decomp_last->next = decomp;
> -               session->decomp_last = decomp;
> +               if (session->decomp == NULL) {
> +                       session->decomp = decomp;
> +                       session->decomp_last = decomp;
> +               } else {
> +                       session->decomp_last->next = decomp;
> +                       session->decomp_last = decomp;
> +               }
>         }
>  
>         pr_debug("decomp (B): %zd to %zd\n", src_size, decomp_size);
> @@ -319,11 +335,10 @@ static void perf_session__delete_threads(struct
> perf_session *session)
>         machine__delete_threads(&session->machines.host);
>  }
>  
> -static void perf_session__release_decomp_events(struct perf_session *session)
> +static void perf_decomp__release_events(struct decomp *next)
>  {
> -       struct decomp *next, *decomp;
> +       struct decomp *decomp;
>         size_t mmap_len;
> -       next = session->decomp;
>         do {
>                 decomp = next;
>                 if (decomp == NULL)
> @@ -336,6 +351,8 @@ static void perf_session__release_decomp_events(struct
> perf_session *session)
>  
>  void perf_session__delete(struct perf_session *session)
>  {
> +       int r;
> +
>         if (session == NULL)
>                 return;
>         auxtrace__free(session);
> @@ -343,10 +360,12 @@ void perf_session__delete(struct perf_session *session)
>         perf_session__destroy_kernel_maps(session);
>         perf_session__delete_threads(session);
>         if (session->readers) {
> +               for (r = 0; r < session->nr_readers; r++)
> +                       perf_decomp__release_events(session-
> >readers[r].decomp);
>                 zfree(&session->readers);
>                 session->nr_readers = 0;
>         }
> -       perf_session__release_decomp_events(session);
> +       perf_decomp__release_events(session->decomp);
>         perf_env__exit(&session->header.env);
>         machines__exit(&session->machines);
>         if (session->data)
> @@ -2157,7 +2176,8 @@ static int __perf_session__process_decomp_events(struct
> perf_session *session)
>  {
>         s64 skip;
>         u64 size;
> -       struct decomp *decomp = session->decomp_last;
> +       struct decomp *decomp = session->active_reader ?
> +               session->active_reader->decomp_last : session->decomp_last;
>  
>         if (!decomp)
>                 return 0;
> @@ -2214,6 +2234,9 @@ reader__process_events(struct reader *rd, struct
> perf_session *session,
>  
>         memset(mmaps, 0, sizeof(st->mmaps));
>  
> +       if (zstd_init(&rd->zstd_data, 0))
> +               return -1;
> +
>         mmap_prot  = PROT_READ;
>         mmap_flags = MAP_SHARED;

zstd_fini is never called on zstd_data.

Thanks,
Riccardo

> @@ -2257,12 +2280,13 @@ reader__process_events(struct reader *rd, struct
> perf_session *session,
>                 goto remap;
>         }
>  
> +       session->active_reader = rd;
>         size = event->header.size;
>  
>         skip = -EINVAL;
>  
>         if (size < sizeof(struct perf_event_header) ||
> -           (skip = rd->process(session, event, st->file_pos, rd->path)) < 0)
> {
> +           (skip = perf_session__process_event(session, event, st->file_pos,
> rd->path)) < 0) {
>                 pr_err("%#" PRIx64 " [%s] [%#x]: failed to process type: %d
> [%s]\n",
>                        st->file_offset + st->head, rd->path, event-
> >header.size,
>                        event->header.type, strerror(-skip));
> @@ -2289,6 +2313,7 @@ reader__process_events(struct reader *rd, struct
> perf_session *session,
>                 goto more;
>  
>  out:
> +       session->active_reader = NULL;
>         return err;
>  }
>  
> diff --git a/tools/perf/util/session.h b/tools/perf/util/session.h
> index 2815d00b5467..e0a8712f8770 100644
> --- a/tools/perf/util/session.h
> +++ b/tools/perf/util/session.h
> @@ -44,6 +44,7 @@ struct perf_session {
>         struct decomp           *decomp_last;
>         struct reader           *readers;
>         int                     nr_readers;
> +       struct reader           *active_reader;
>  };
>  
>  struct decomp {

Re: [PATCH v6 20/20] perf session: Load data directory files for analysis

[Riccardo Mancini]

Hi,

On Wed, 2021-05-26 at 13:53 +0300, Alexey Bayduraev wrote:
> Load data directory files and provide basic raw dump and aggregated
> analysis support of data directories in report mode, still with no
> memory consumption optimizations.
> 
> Design and implementation are based on the prototype [1], [2].
> 
> [1] git clone https://git.kernel.org/pub/scm/linux/kernel/git/jolsa/perf.git -
> b perf/record_threads
> [2] https://lore.kernel.org/lkml/20180913125450.21342-1-jolsa@kernel.org/
> 
> Suggested-by: Jiri Olsa <jolsa@kernel.org>
> Signed-off-by: Alexey Bayduraev <alexey.v.bayduraev@linux.intel.com>
> ---
>  tools/perf/util/session.c | 129 ++++++++++++++++++++++++++++++++++++++
>  1 file changed, 129 insertions(+)
> 
> diff --git a/tools/perf/util/session.c b/tools/perf/util/session.c
> index 041601810b85..dd4ef9749cd0 100644
> --- a/tools/perf/util/session.c
> +++ b/tools/perf/util/session.c
> @@ -65,6 +65,7 @@ struct reader_state {
>         u64      data_size;
>         u64      head;
>         bool     eof;
> +       u64      size;
>  };
>  
>  enum {
> @@ -2319,6 +2320,7 @@ reader__read_event(struct reader *rd, struct
> perf_session *session,
>         if (skip)
>                 size += skip;
>  
> +       st->size += size;
>         st->head += size;
>         st->file_pos += size;
>  
> @@ -2418,6 +2420,130 @@ static int __perf_session__process_events(struct
> perf_session *session)
>         return err;
>  }
>  
> +/*
> + * This function reads, merge and process directory data.
> + * It assumens the version 1 of directory data, where each
> + * data file holds per-cpu data, already sorted by kernel.
> + */
> +static int __perf_session__process_dir_events(struct perf_session *session)
> +{
> +       struct perf_data *data = session->data;
> +       struct perf_tool *tool = session->tool;
> +       int i, ret = 0, readers = 1;
> +       struct ui_progress prog;
> +       u64 total_size = perf_data__size(session->data);
> +       struct reader *rd;
> +
> +       perf_tool__fill_defaults(tool);
> +
> +       ui_progress__init_size(&prog, total_size, "Sorting events...");
> +
> +       for (i = 0; i < data->dir.nr; i++) {
> +               if (data->dir.files[i].size)
> +                       readers++;
> +       }
> +
> +       rd = session->readers = zalloc(readers * sizeof(struct reader));
> +       if (!rd)
> +               return -ENOMEM;
> +       session->nr_readers = readers;
> +       readers = 0;
> +
> +       rd[readers] = (struct reader) {
> +               .fd              = perf_data__fd(session->data),
> +               .path            = session->data->file.path,
> +               .data_size       = session->header.data_size,
> +               .data_offset     = session->header.data_offset,
> +               .in_place_update = session->data->in_place_update,
> +       };
> +       ret = reader__init(&rd[readers], NULL);
> +       if (ret)
> +               goto out_err;
> +       ret = reader__mmap(&rd[readers], session);
> +       if (ret != READER_OK) {
> +               if (ret == READER_EOF)
> +                       ret = -EINVAL;
> +               goto out_err;
> +       }
> +       readers++;
> +
> +       for (i = 0; i < data->dir.nr; i++) {
> +               if (data->dir.files[i].size) {
> +                       rd[readers] = (struct reader) {
> +                               .fd              = data->dir.files[i].fd,
> +                               .path            = data->dir.files[i].path,
> +                               .data_size       = data->dir.files[i].size,
> +                               .data_offset     = 0,
> +                               .in_place_update = session->data-
> >in_place_update,
> +                       };
> +                       ret = reader__init(&rd[readers], NULL);

zstd_fini is never called on rd[readers].zstd_data
Maybe it can be done in perf_session__delete. For example, we could add a new
reader__fini function to do the cleanup of zstd data and
perf_decomp__release_events.

Thanks,
Riccardo

> +                       if (ret)
> +                               goto out_err;
> +                       ret = reader__mmap(&rd[readers], session);
> +                       if (ret != READER_OK) {
> +                               if (ret == READER_EOF)
> +                                       ret = -EINVAL;
> +                               goto out_err;
> +                       }
> +                       readers++;
> +               }
> +       }
> +
> +       i = 0;
> +
> +       while ((ret >= 0) && readers) {
> +               if (session_done())
> +                       return 0;
> +
> +               if (rd[i].state.eof) {
> +                       i = (i + 1) % session->nr_readers;
> +                       continue;
> +               }
> +
> +               ret = reader__read_event(&rd[i], session, &prog);
> +               if (ret < 0)
> +                       break;
> +               if (ret == READER_EOF) {
> +                       ret = reader__mmap(&rd[i], session);
> +                       if (ret < 0)
> +                               goto out_err;
> +                       if (ret == READER_EOF)
> +                               readers--;
> +               }
> +
> +               /*
> +                * Processing 10MBs of data from each reader in sequence,
> +                * because that's the way the ordered events sorting works
> +                * most efficiently.
> +                */
> +               if (rd[i].state.size >= 10*1024*1024) {
> +                       rd[i].state.size = 0;
> +                       i = (i + 1) % session->nr_readers;
> +               }
> +       }
> +
> +       ret = ordered_events__flush(&session->ordered_events,
> OE_FLUSH__FINAL);
> +       if (ret)
> +               goto out_err;
> +
> +       ret = perf_session__flush_thread_stacks(session);
> +out_err:
> +       ui_progress__finish();
> +
> +       if (!tool->no_warn)
> +               perf_session__warn_about_errors(session);
> +
> +       /*
> +        * We may switching perf.data output, make ordered_events
> +        * reusable.
> +        */
> +       ordered_events__reinit(&session->ordered_events);
> +
> +       session->one_mmap = false;
> +
> +       return ret;
> +}
> +
>  int perf_session__process_events(struct perf_session *session)
>  {
>         if (perf_session__register_idle_thread(session) < 0)
> @@ -2426,6 +2552,9 @@ int perf_session__process_events(struct perf_session
> *session)
>         if (perf_data__is_pipe(session->data))
>                 return __perf_session__process_pipe_events(session);
>  
> +       if (perf_data__is_dir(session->data))
> +               return __perf_session__process_dir_events(session);
> +
>         return __perf_session__process_events(session);
>  }
>  

Re: [PATCH v6 03/20] perf record: Introduce thread local variable

[Namhyung Kim]

Hi Riccardo,

On Thu, Jun 3, 2021 at 3:56 PM Riccardo Mancini <rickyman7@gmail.com> wrote:
>
> Hi,
>
> thank you very much for your work for adding threading capabilites to perf
> record.
> I did some testing on your entire patchset, especially checking for memory
> issues using ASan. This is just the first of a couple of emails to point out
> some issues I found.
> I will also do additional tests in the future.
>
> On Wed, 2021-05-26 at 13:52 +0300, Alexey Bayduraev wrote:
> SNIP
> > @@ -2220,18 +2275,20 @@ static int __cmd_record(struct record *rec, int argc,
> > const char **argv)
> >                 goto out_child;
> >         }
> >
> > -       if (!quiet)
> > -               fprintf(stderr, "[ perf record: Woken up %ld times to write data
> > ]\n", waking);
> > -
> >         if (target__none(&rec->opts.target))
> >                 record__synthesize_workload(rec, true);
> >
> >  out_child:
> > +       record__stop_threads(rec, &waking);
> > +out_free_threads:
> >         record__free_thread_data(rec);
> >         evlist__finalize_ctlfd(rec->evlist);
> >         record__mmap_read_all(rec, true);
> >         record__aio_mmap_read_sync(rec);
>
> record__mmap_read_all should be moved before record__free_thread_data since it
> uses the thread_data that's just been freed.
> Furthermore, record__mmap_read_all should also be moved before the
> out_free_threads label, since it cannot be called unless record__start_threads
> succeeded, otherwise thread would be NULL and will cause a segfault (it happens
> if there is an error somewhere else in perf, for example).
>
> In my tests the following order works, but it should be double checked for
> possible side-effects of this order change.
>
> out_child:
>         record__stop_threads(rec, &waking);
>         record__mmap_read_all(rec, true);
> out_free_threads:
>         record__free_thread_data(rec);
>         evlist__finalize_ctlfd(rec->evlist);
>         record__aio_mmap_read_sync(rec);

I wonder how it worked before.. maybe we should place
record__free_thread_data() far below.

Thanks,
Namhyung