[v1,4/6] perf threads: Move threads to its own files

Message ID 20240214063708.972376-5-irogers@google.com
State New
Headers
Series Thread memory improvements and fixes |

Commit Message

Ian Rogers Feb. 14, 2024, 6:37 a.m. UTC
  Move threads out of machine and move thread_rb_node into the C
file. This hides the implementation of threads from the rest of the
code allowing for it to be refactored.

Locking discipline is tightened up in this change.

Signed-off-by: Ian Rogers <irogers@google.com>
---
 tools/perf/util/Build                 |   1 +
 tools/perf/util/bpf_lock_contention.c |   8 +-
 tools/perf/util/machine.c             | 287 ++++----------------------
 tools/perf/util/machine.h             |  20 +-
 tools/perf/util/thread.c              |   2 +-
 tools/perf/util/thread.h              |   6 -
 tools/perf/util/threads.c             | 244 ++++++++++++++++++++++
 tools/perf/util/threads.h             |  35 ++++
 8 files changed, 325 insertions(+), 278 deletions(-)
 create mode 100644 tools/perf/util/threads.c
 create mode 100644 tools/perf/util/threads.h
  

Comments

Namhyung Kim Feb. 27, 2024, 7:07 a.m. UTC | #1
On Tue, Feb 13, 2024 at 10:37 PM Ian Rogers <irogers@google.com> wrote:
>
> Move threads out of machine and move thread_rb_node into the C
> file. This hides the implementation of threads from the rest of the
> code allowing for it to be refactored.
>
> Locking discipline is tightened up in this change.

Doesn't look like a simple code move.  Can we split the locking
change from the move to make the reviewer's life a bit easier? :)

Thanks,
Namhyung

>
> Signed-off-by: Ian Rogers <irogers@google.com>
> ---
>  tools/perf/util/Build                 |   1 +
>  tools/perf/util/bpf_lock_contention.c |   8 +-
>  tools/perf/util/machine.c             | 287 ++++----------------------
>  tools/perf/util/machine.h             |  20 +-
>  tools/perf/util/thread.c              |   2 +-
>  tools/perf/util/thread.h              |   6 -
>  tools/perf/util/threads.c             | 244 ++++++++++++++++++++++
>  tools/perf/util/threads.h             |  35 ++++
>  8 files changed, 325 insertions(+), 278 deletions(-)
>  create mode 100644 tools/perf/util/threads.c
>  create mode 100644 tools/perf/util/threads.h
>
> diff --git a/tools/perf/util/Build b/tools/perf/util/Build
> index 8027f450fa3e..a0e8cd68d490 100644
> --- a/tools/perf/util/Build
> +++ b/tools/perf/util/Build
> @@ -71,6 +71,7 @@ perf-y += ordered-events.o
>  perf-y += namespaces.o
>  perf-y += comm.o
>  perf-y += thread.o
> +perf-y += threads.o
>  perf-y += thread_map.o
>  perf-y += parse-events-flex.o
>  perf-y += parse-events-bison.o
> diff --git a/tools/perf/util/bpf_lock_contention.c b/tools/perf/util/bpf_lock_contention.c
> index 31ff19afc20c..3992c8a9fd96 100644
> --- a/tools/perf/util/bpf_lock_contention.c
> +++ b/tools/perf/util/bpf_lock_contention.c
> @@ -210,7 +210,7 @@ static const char *lock_contention_get_name(struct lock_contention *con,
>
>                 /* do not update idle comm which contains CPU number */
>                 if (pid) {
> -                       struct thread *t = __machine__findnew_thread(machine, /*pid=*/-1, pid);
> +                       struct thread *t = machine__findnew_thread(machine, /*pid=*/-1, pid);
>
>                         if (t == NULL)
>                                 return name;
> @@ -302,9 +302,9 @@ int lock_contention_read(struct lock_contention *con)
>                 return -1;
>
>         if (con->aggr_mode == LOCK_AGGR_TASK) {
> -               struct thread *idle = __machine__findnew_thread(machine,
> -                                                               /*pid=*/0,
> -                                                               /*tid=*/0);
> +               struct thread *idle = machine__findnew_thread(machine,
> +                                                             /*pid=*/0,
> +                                                             /*tid=*/0);
>                 thread__set_comm(idle, "swapper", /*timestamp=*/0);
>         }
>
> diff --git a/tools/perf/util/machine.c b/tools/perf/util/machine.c
> index e072b2115b64..e668a97255f8 100644
> --- a/tools/perf/util/machine.c
> +++ b/tools/perf/util/machine.c
> @@ -43,9 +43,6 @@
>  #include <linux/string.h>
>  #include <linux/zalloc.h>
>
> -static void __machine__remove_thread(struct machine *machine, struct thread_rb_node *nd,
> -                                    struct thread *th, bool lock);
> -
>  static struct dso *machine__kernel_dso(struct machine *machine)
>  {
>         return map__dso(machine->vmlinux_map);
> @@ -58,35 +55,6 @@ static void dsos__init(struct dsos *dsos)
>         init_rwsem(&dsos->lock);
>  }
>
> -static void machine__threads_init(struct machine *machine)
> -{
> -       int i;
> -
> -       for (i = 0; i < THREADS__TABLE_SIZE; i++) {
> -               struct threads *threads = &machine->threads[i];
> -               threads->entries = RB_ROOT_CACHED;
> -               init_rwsem(&threads->lock);
> -               threads->nr = 0;
> -               threads->last_match = NULL;
> -       }
> -}
> -
> -static int thread_rb_node__cmp_tid(const void *key, const struct rb_node *nd)
> -{
> -       int to_find = (int) *((pid_t *)key);
> -
> -       return to_find - (int)thread__tid(rb_entry(nd, struct thread_rb_node, rb_node)->thread);
> -}
> -
> -static struct thread_rb_node *thread_rb_node__find(const struct thread *th,
> -                                                  struct rb_root *tree)
> -{
> -       pid_t to_find = thread__tid(th);
> -       struct rb_node *nd = rb_find(&to_find, tree, thread_rb_node__cmp_tid);
> -
> -       return rb_entry(nd, struct thread_rb_node, rb_node);
> -}
> -
>  static int machine__set_mmap_name(struct machine *machine)
>  {
>         if (machine__is_host(machine))
> @@ -120,7 +88,7 @@ int machine__init(struct machine *machine, const char *root_dir, pid_t pid)
>         RB_CLEAR_NODE(&machine->rb_node);
>         dsos__init(&machine->dsos);
>
> -       machine__threads_init(machine);
> +       threads__init(&machine->threads);
>
>         machine->vdso_info = NULL;
>         machine->env = NULL;
> @@ -221,27 +189,11 @@ static void dsos__exit(struct dsos *dsos)
>
>  void machine__delete_threads(struct machine *machine)
>  {
> -       struct rb_node *nd;
> -       int i;
> -
> -       for (i = 0; i < THREADS__TABLE_SIZE; i++) {
> -               struct threads *threads = &machine->threads[i];
> -               down_write(&threads->lock);
> -               nd = rb_first_cached(&threads->entries);
> -               while (nd) {
> -                       struct thread_rb_node *trb = rb_entry(nd, struct thread_rb_node, rb_node);
> -
> -                       nd = rb_next(nd);
> -                       __machine__remove_thread(machine, trb, trb->thread, false);
> -               }
> -               up_write(&threads->lock);
> -       }
> +       threads__remove_all_threads(&machine->threads);
>  }
>
>  void machine__exit(struct machine *machine)
>  {
> -       int i;
> -
>         if (machine == NULL)
>                 return;
>
> @@ -254,12 +206,7 @@ void machine__exit(struct machine *machine)
>         zfree(&machine->current_tid);
>         zfree(&machine->kallsyms_filename);
>
> -       machine__delete_threads(machine);
> -       for (i = 0; i < THREADS__TABLE_SIZE; i++) {
> -               struct threads *threads = &machine->threads[i];
> -
> -               exit_rwsem(&threads->lock);
> -       }
> +       threads__exit(&machine->threads);
>  }
>
>  void machine__delete(struct machine *machine)
> @@ -526,7 +473,7 @@ static void machine__update_thread_pid(struct machine *machine,
>         if (thread__pid(th) == thread__tid(th))
>                 return;
>
> -       leader = __machine__findnew_thread(machine, thread__pid(th), thread__pid(th));
> +       leader = machine__findnew_thread(machine, thread__pid(th), thread__pid(th));
>         if (!leader)
>                 goto out_err;
>
> @@ -560,160 +507,55 @@ static void machine__update_thread_pid(struct machine *machine,
>         goto out_put;
>  }
>
> -/*
> - * Front-end cache - TID lookups come in blocks,
> - * so most of the time we dont have to look up
> - * the full rbtree:
> - */
> -static struct thread*
> -__threads__get_last_match(struct threads *threads, struct machine *machine,
> -                         int pid, int tid)
> -{
> -       struct thread *th;
> -
> -       th = threads->last_match;
> -       if (th != NULL) {
> -               if (thread__tid(th) == tid) {
> -                       machine__update_thread_pid(machine, th, pid);
> -                       return thread__get(th);
> -               }
> -               thread__put(threads->last_match);
> -               threads->last_match = NULL;
> -       }
> -
> -       return NULL;
> -}
> -
> -static struct thread*
> -threads__get_last_match(struct threads *threads, struct machine *machine,
> -                       int pid, int tid)
> -{
> -       struct thread *th = NULL;
> -
> -       if (perf_singlethreaded)
> -               th = __threads__get_last_match(threads, machine, pid, tid);
> -
> -       return th;
> -}
> -
> -static void
> -__threads__set_last_match(struct threads *threads, struct thread *th)
> -{
> -       thread__put(threads->last_match);
> -       threads->last_match = thread__get(th);
> -}
> -
> -static void
> -threads__set_last_match(struct threads *threads, struct thread *th)
> -{
> -       if (perf_singlethreaded)
> -               __threads__set_last_match(threads, th);
> -}
> -
>  /*
>   * Caller must eventually drop thread->refcnt returned with a successful
>   * lookup/new thread inserted.
>   */
> -static struct thread *____machine__findnew_thread(struct machine *machine,
> -                                                 struct threads *threads,
> -                                                 pid_t pid, pid_t tid,
> -                                                 bool create)
> +static struct thread *__machine__findnew_thread(struct machine *machine,
> +                                               pid_t pid,
> +                                               pid_t tid,
> +                                               bool create)
>  {
> -       struct rb_node **p = &threads->entries.rb_root.rb_node;
> -       struct rb_node *parent = NULL;
> -       struct thread *th;
> -       struct thread_rb_node *nd;
> -       bool leftmost = true;
> +       struct thread *th = threads__find(&machine->threads, tid);
> +       bool created;
>
> -       th = threads__get_last_match(threads, machine, pid, tid);
> -       if (th)
> +       if (th) {
> +               machine__update_thread_pid(machine, th, pid);
>                 return th;
> -
> -       while (*p != NULL) {
> -               parent = *p;
> -               th = rb_entry(parent, struct thread_rb_node, rb_node)->thread;
> -
> -               if (thread__tid(th) == tid) {
> -                       threads__set_last_match(threads, th);
> -                       machine__update_thread_pid(machine, th, pid);
> -                       return thread__get(th);
> -               }
> -
> -               if (tid < thread__tid(th))
> -                       p = &(*p)->rb_left;
> -               else {
> -                       p = &(*p)->rb_right;
> -                       leftmost = false;
> -               }
>         }
> -
>         if (!create)
>                 return NULL;
>
> -       th = thread__new(pid, tid);
> -       if (th == NULL)
> -               return NULL;
> -
> -       nd = malloc(sizeof(*nd));
> -       if (nd == NULL) {
> -               thread__put(th);
> -               return NULL;
> -       }
> -       nd->thread = th;
> -
> -       rb_link_node(&nd->rb_node, parent, p);
> -       rb_insert_color_cached(&nd->rb_node, &threads->entries, leftmost);
> -       /*
> -        * We have to initialize maps separately after rb tree is updated.
> -        *
> -        * The reason is that we call machine__findnew_thread within
> -        * thread__init_maps to find the thread leader and that would screwed
> -        * the rb tree.
> -        */
> -       if (thread__init_maps(th, machine)) {
> -               pr_err("Thread init failed thread %d\n", pid);
> -               rb_erase_cached(&nd->rb_node, &threads->entries);
> -               RB_CLEAR_NODE(&nd->rb_node);
> -               free(nd);
> -               thread__put(th);
> -               return NULL;
> -       }
> -       /*
> -        * It is now in the rbtree, get a ref
> -        */
> -       threads__set_last_match(threads, th);
> -       ++threads->nr;
> -
> -       return thread__get(th);
> -}
> +       th = threads__findnew(&machine->threads, pid, tid, &created);
> +       if (created) {
> +               /*
> +                * We have to initialize maps separately after rb tree is
> +                * updated.
> +                *
> +                * The reason is that we call machine__findnew_thread within
> +                * thread__init_maps to find the thread leader and that would
> +                * screwed the rb tree.
> +                */
> +               if (thread__init_maps(th, machine)) {
> +                       pr_err("Thread init failed thread %d\n", pid);
> +                       threads__remove(&machine->threads, th);
> +                       thread__put(th);
> +                       return NULL;
> +               }
> +       } else
> +               machine__update_thread_pid(machine, th, pid);
>
> -struct thread *__machine__findnew_thread(struct machine *machine, pid_t pid, pid_t tid)
> -{
> -       return ____machine__findnew_thread(machine, machine__threads(machine, tid), pid, tid, true);
> +       return th;
>  }
>
> -struct thread *machine__findnew_thread(struct machine *machine, pid_t pid,
> -                                      pid_t tid)
> +struct thread *machine__findnew_thread(struct machine *machine, pid_t pid, pid_t tid)
>  {
> -       struct threads *threads = machine__threads(machine, tid);
> -       struct thread *th;
> -
> -       down_write(&threads->lock);
> -       th = __machine__findnew_thread(machine, pid, tid);
> -       up_write(&threads->lock);
> -       return th;
> +       return __machine__findnew_thread(machine, pid, tid, /*create=*/true);
>  }
>
> -struct thread *machine__find_thread(struct machine *machine, pid_t pid,
> -                                   pid_t tid)
> +struct thread *machine__find_thread(struct machine *machine, pid_t pid, pid_t tid)
>  {
> -       struct threads *threads = machine__threads(machine, tid);
> -       struct thread *th;
> -
> -       down_read(&threads->lock);
> -       th =  ____machine__findnew_thread(machine, threads, pid, tid, false);
> -       up_read(&threads->lock);
> -       return th;
> +       return __machine__findnew_thread(machine, pid, tid, /*create=*/false);
>  }
>
>  /*
> @@ -1127,23 +969,13 @@ static int machine_fprintf_cb(struct thread *thread, void *data)
>         return 0;
>  }
>
> -static size_t machine__threads_nr(const struct machine *machine)
> -{
> -       size_t nr = 0;
> -
> -       for (int i = 0; i < THREADS__TABLE_SIZE; i++)
> -               nr += machine->threads[i].nr;
> -
> -       return nr;
> -}
> -
>  size_t machine__fprintf(struct machine *machine, FILE *fp)
>  {
>         struct machine_fprintf_cb_args args = {
>                 .fp = fp,
>                 .printed = 0,
>         };
> -       size_t ret = fprintf(fp, "Threads: %zu\n", machine__threads_nr(machine));
> +       size_t ret = fprintf(fp, "Threads: %zu\n", threads__nr(&machine->threads));
>
>         machine__for_each_thread(machine, machine_fprintf_cb, &args);
>         return ret + args.printed;
> @@ -2069,36 +1901,9 @@ int machine__process_mmap_event(struct machine *machine, union perf_event *event
>         return 0;
>  }
>
> -static void __machine__remove_thread(struct machine *machine, struct thread_rb_node *nd,
> -                                    struct thread *th, bool lock)
> -{
> -       struct threads *threads = machine__threads(machine, thread__tid(th));
> -
> -       if (!nd)
> -               nd = thread_rb_node__find(th, &threads->entries.rb_root);
> -
> -       if (threads->last_match && RC_CHK_EQUAL(threads->last_match, th))
> -               threads__set_last_match(threads, NULL);
> -
> -       if (lock)
> -               down_write(&threads->lock);
> -
> -       BUG_ON(refcount_read(thread__refcnt(th)) == 0);
> -
> -       thread__put(nd->thread);
> -       rb_erase_cached(&nd->rb_node, &threads->entries);
> -       RB_CLEAR_NODE(&nd->rb_node);
> -       --threads->nr;
> -
> -       free(nd);
> -
> -       if (lock)
> -               up_write(&threads->lock);
> -}
> -
>  void machine__remove_thread(struct machine *machine, struct thread *th)
>  {
> -       return __machine__remove_thread(machine, NULL, th, true);
> +       return threads__remove(&machine->threads, th);
>  }
>
>  int machine__process_fork_event(struct machine *machine, union perf_event *event,
> @@ -3232,23 +3037,7 @@ int machine__for_each_thread(struct machine *machine,
>                              int (*fn)(struct thread *thread, void *p),
>                              void *priv)
>  {
> -       struct threads *threads;
> -       struct rb_node *nd;
> -       int rc = 0;
> -       int i;
> -
> -       for (i = 0; i < THREADS__TABLE_SIZE; i++) {
> -               threads = &machine->threads[i];
> -               for (nd = rb_first_cached(&threads->entries); nd;
> -                    nd = rb_next(nd)) {
> -                       struct thread_rb_node *trb = rb_entry(nd, struct thread_rb_node, rb_node);
> -
> -                       rc = fn(trb->thread, priv);
> -                       if (rc != 0)
> -                               return rc;
> -               }
> -       }
> -       return rc;
> +       return threads__for_each_thread(&machine->threads, fn, priv);
>  }
>
>  int machines__for_each_thread(struct machines *machines,
> diff --git a/tools/perf/util/machine.h b/tools/perf/util/machine.h
> index b738ce84817b..e28c787616fe 100644
> --- a/tools/perf/util/machine.h
> +++ b/tools/perf/util/machine.h
> @@ -7,6 +7,7 @@
>  #include "maps.h"
>  #include "dsos.h"
>  #include "rwsem.h"
> +#include "threads.h"
>
>  struct addr_location;
>  struct branch_stack;
> @@ -28,16 +29,6 @@ extern const char *ref_reloc_sym_names[];
>
>  struct vdso_info;
>
> -#define THREADS__TABLE_BITS    8
> -#define THREADS__TABLE_SIZE    (1 << THREADS__TABLE_BITS)
> -
> -struct threads {
> -       struct rb_root_cached  entries;
> -       struct rw_semaphore    lock;
> -       unsigned int           nr;
> -       struct thread          *last_match;
> -};
> -
>  struct machine {
>         struct rb_node    rb_node;
>         pid_t             pid;
> @@ -48,7 +39,7 @@ struct machine {
>         char              *root_dir;
>         char              *mmap_name;
>         char              *kallsyms_filename;
> -       struct threads    threads[THREADS__TABLE_SIZE];
> +       struct threads    threads;
>         struct vdso_info  *vdso_info;
>         struct perf_env   *env;
>         struct dsos       dsos;
> @@ -69,12 +60,6 @@ struct machine {
>         bool              trampolines_mapped;
>  };
>
> -static inline struct threads *machine__threads(struct machine *machine, pid_t tid)
> -{
> -       /* Cast it to handle tid == -1 */
> -       return &machine->threads[(unsigned int)tid % THREADS__TABLE_SIZE];
> -}
> -
>  /*
>   * The main kernel (vmlinux) map
>   */
> @@ -220,7 +205,6 @@ bool machine__is(struct machine *machine, const char *arch);
>  bool machine__normalized_is(struct machine *machine, const char *arch);
>  int machine__nr_cpus_avail(struct machine *machine);
>
> -struct thread *__machine__findnew_thread(struct machine *machine, pid_t pid, pid_t tid);
>  struct thread *machine__findnew_thread(struct machine *machine, pid_t pid, pid_t tid);
>
>  struct dso *machine__findnew_dso_id(struct machine *machine, const char *filename, struct dso_id *id);
> diff --git a/tools/perf/util/thread.c b/tools/perf/util/thread.c
> index c59ab4d79163..1aa8962dcf52 100644
> --- a/tools/perf/util/thread.c
> +++ b/tools/perf/util/thread.c
> @@ -26,7 +26,7 @@ int thread__init_maps(struct thread *thread, struct machine *machine)
>         if (pid == thread__tid(thread) || pid == -1) {
>                 thread__set_maps(thread, maps__new(machine));
>         } else {
> -               struct thread *leader = __machine__findnew_thread(machine, pid, pid);
> +               struct thread *leader = machine__findnew_thread(machine, pid, pid);
>
>                 if (leader) {
>                         thread__set_maps(thread, maps__get(thread__maps(leader)));
> diff --git a/tools/perf/util/thread.h b/tools/perf/util/thread.h
> index 0df775b5c110..4b8f3e9e513b 100644
> --- a/tools/perf/util/thread.h
> +++ b/tools/perf/util/thread.h
> @@ -3,7 +3,6 @@
>  #define __PERF_THREAD_H
>
>  #include <linux/refcount.h>
> -#include <linux/rbtree.h>
>  #include <linux/list.h>
>  #include <stdio.h>
>  #include <unistd.h>
> @@ -30,11 +29,6 @@ struct lbr_stitch {
>         struct callchain_cursor_node    *prev_lbr_cursor;
>  };
>
> -struct thread_rb_node {
> -       struct rb_node rb_node;
> -       struct thread *thread;
> -};
> -
>  DECLARE_RC_STRUCT(thread) {
>         /** @maps: mmaps associated with this thread. */
>         struct maps             *maps;
> diff --git a/tools/perf/util/threads.c b/tools/perf/util/threads.c
> new file mode 100644
> index 000000000000..d984ec939c7b
> --- /dev/null
> +++ b/tools/perf/util/threads.c
> @@ -0,0 +1,244 @@
> +// SPDX-License-Identifier: GPL-2.0
> +#include "threads.h"
> +#include "machine.h"
> +#include "thread.h"
> +
> +struct thread_rb_node {
> +       struct rb_node rb_node;
> +       struct thread *thread;
> +};
> +
> +static struct threads_table_entry *threads__table(struct threads *threads, pid_t tid)
> +{
> +       /* Cast it to handle tid == -1 */
> +       return &threads->table[(unsigned int)tid % THREADS__TABLE_SIZE];
> +}
> +
> +void threads__init(struct threads *threads)
> +{
> +       for (int i = 0; i < THREADS__TABLE_SIZE; i++) {
> +               struct threads_table_entry *table = &threads->table[i];
> +
> +               table->entries = RB_ROOT_CACHED;
> +               init_rwsem(&table->lock);
> +               table->nr = 0;
> +               table->last_match = NULL;
> +       }
> +}
> +
> +void threads__exit(struct threads *threads)
> +{
> +       threads__remove_all_threads(threads);
> +       for (int i = 0; i < THREADS__TABLE_SIZE; i++) {
> +               struct threads_table_entry *table = &threads->table[i];
> +
> +               exit_rwsem(&table->lock);
> +       }
> +}
> +
> +size_t threads__nr(struct threads *threads)
> +{
> +       size_t nr = 0;
> +
> +       for (int i = 0; i < THREADS__TABLE_SIZE; i++) {
> +               struct threads_table_entry *table = &threads->table[i];
> +
> +               down_read(&table->lock);
> +               nr += table->nr;
> +               up_read(&table->lock);
> +       }
> +       return nr;
> +}
> +
> +/*
> + * Front-end cache - TID lookups come in blocks,
> + * so most of the time we dont have to look up
> + * the full rbtree:
> + */
> +static struct thread *__threads_table_entry__get_last_match(struct threads_table_entry *table,
> +                                                           pid_t tid)
> +{
> +       struct thread *th, *res = NULL;
> +
> +       th = table->last_match;
> +       if (th != NULL) {
> +               if (thread__tid(th) == tid)
> +                       res = thread__get(th);
> +       }
> +       return res;
> +}
> +
> +static void __threads_table_entry__set_last_match(struct threads_table_entry *table,
> +                                                 struct thread *th)
> +{
> +       thread__put(table->last_match);
> +       table->last_match = thread__get(th);
> +}
> +
> +static void threads_table_entry__set_last_match(struct threads_table_entry *table,
> +                                               struct thread *th)
> +{
> +       down_write(&table->lock);
> +       __threads_table_entry__set_last_match(table, th);
> +       up_write(&table->lock);
> +}
> +
> +struct thread *threads__find(struct threads *threads, pid_t tid)
> +{
> +       struct threads_table_entry *table  = threads__table(threads, tid);
> +       struct rb_node **p;
> +       struct thread *res = NULL;
> +
> +       down_read(&table->lock);
> +       res = __threads_table_entry__get_last_match(table, tid);
> +       if (res)
> +               return res;
> +
> +       p = &table->entries.rb_root.rb_node;
> +       while (*p != NULL) {
> +               struct rb_node *parent = *p;
> +               struct thread *th = rb_entry(parent, struct thread_rb_node, rb_node)->thread;
> +
> +               if (thread__tid(th) == tid) {
> +                       res = thread__get(th);
> +                       break;
> +               }
> +
> +               if (tid < thread__tid(th))
> +                       p = &(*p)->rb_left;
> +               else
> +                       p = &(*p)->rb_right;
> +       }
> +       up_read(&table->lock);
> +       if (res)
> +               threads_table_entry__set_last_match(table, res);
> +       return res;
> +}
> +
> +struct thread *threads__findnew(struct threads *threads, pid_t pid, pid_t tid, bool *created)
> +{
> +       struct threads_table_entry *table  = threads__table(threads, tid);
> +       struct rb_node **p;
> +       struct rb_node *parent = NULL;
> +       struct thread *res = NULL;
> +       struct thread_rb_node *nd;
> +       bool leftmost = true;
> +
> +       *created = false;
> +       down_write(&table->lock);
> +       p = &table->entries.rb_root.rb_node;
> +       while (*p != NULL) {
> +               struct thread *th;
> +
> +               parent = *p;
> +               th = rb_entry(parent, struct thread_rb_node, rb_node)->thread;
> +
> +               if (thread__tid(th) == tid) {
> +                       __threads_table_entry__set_last_match(table, th);
> +                       res = thread__get(th);
> +                       goto out_unlock;
> +               }
> +
> +               if (tid < thread__tid(th))
> +                       p = &(*p)->rb_left;
> +               else {
> +                       leftmost = false;
> +                       p = &(*p)->rb_right;
> +               }
> +       }
> +       nd = malloc(sizeof(*nd));
> +       if (nd == NULL)
> +               goto out_unlock;
> +       res = thread__new(pid, tid);
> +       if (!res)
> +               free(nd);
> +       else {
> +               *created = true;
> +               nd->thread = thread__get(res);
> +               rb_link_node(&nd->rb_node, parent, p);
> +               rb_insert_color_cached(&nd->rb_node, &table->entries, leftmost);
> +               ++table->nr;
> +               __threads_table_entry__set_last_match(table, res);
> +       }
> +out_unlock:
> +       up_write(&table->lock);
> +       return res;
> +}
> +
> +void threads__remove_all_threads(struct threads *threads)
> +{
> +       for (int i = 0; i < THREADS__TABLE_SIZE; i++) {
> +               struct threads_table_entry *table = &threads->table[i];
> +               struct rb_node *nd;
> +
> +               down_write(&table->lock);
> +               __threads_table_entry__set_last_match(table, NULL);
> +               nd = rb_first_cached(&table->entries);
> +               while (nd) {
> +                       struct thread_rb_node *trb = rb_entry(nd, struct thread_rb_node, rb_node);
> +
> +                       nd = rb_next(nd);
> +                       thread__put(trb->thread);
> +                       rb_erase_cached(&trb->rb_node, &table->entries);
> +                       RB_CLEAR_NODE(&trb->rb_node);
> +                       --table->nr;
> +
> +                       free(trb);
> +               }
> +               assert(table->nr == 0);
> +               up_write(&table->lock);
> +       }
> +}
> +
> +void threads__remove(struct threads *threads, struct thread *thread)
> +{
> +       struct rb_node **p;
> +       struct threads_table_entry *table  = threads__table(threads, thread__tid(thread));
> +       pid_t tid = thread__tid(thread);
> +
> +       down_write(&table->lock);
> +       if (table->last_match && RC_CHK_EQUAL(table->last_match, thread))
> +               __threads_table_entry__set_last_match(table, NULL);
> +
> +       p = &table->entries.rb_root.rb_node;
> +       while (*p != NULL) {
> +               struct rb_node *parent = *p;
> +               struct thread_rb_node *nd = rb_entry(parent, struct thread_rb_node, rb_node);
> +               struct thread *th = nd->thread;
> +
> +               if (RC_CHK_EQUAL(th, thread)) {
> +                       thread__put(nd->thread);
> +                       rb_erase_cached(&nd->rb_node, &table->entries);
> +                       RB_CLEAR_NODE(&nd->rb_node);
> +                       --table->nr;
> +                       free(nd);
> +                       break;
> +               }
> +
> +               if (tid < thread__tid(th))
> +                       p = &(*p)->rb_left;
> +               else
> +                       p = &(*p)->rb_right;
> +       }
> +       up_write(&table->lock);
> +}
> +
> +int threads__for_each_thread(struct threads *threads,
> +                            int (*fn)(struct thread *thread, void *data),
> +                            void *data)
> +{
> +       for (int i = 0; i < THREADS__TABLE_SIZE; i++) {
> +               struct threads_table_entry *table = &threads->table[i];
> +               struct rb_node *nd;
> +
> +               for (nd = rb_first_cached(&table->entries); nd; nd = rb_next(nd)) {
> +                       struct thread_rb_node *trb = rb_entry(nd, struct thread_rb_node, rb_node);
> +                       int rc = fn(trb->thread, data);
> +
> +                       if (rc != 0)
> +                               return rc;
> +               }
> +       }
> +       return 0;
> +
> +}
> diff --git a/tools/perf/util/threads.h b/tools/perf/util/threads.h
> new file mode 100644
> index 000000000000..ed67de627578
> --- /dev/null
> +++ b/tools/perf/util/threads.h
> @@ -0,0 +1,35 @@
> +/* SPDX-License-Identifier: GPL-2.0 */
> +#ifndef __PERF_THREADS_H
> +#define __PERF_THREADS_H
> +
> +#include <linux/rbtree.h>
> +#include "rwsem.h"
> +
> +struct thread;
> +
> +#define THREADS__TABLE_BITS    8
> +#define THREADS__TABLE_SIZE    (1 << THREADS__TABLE_BITS)
> +
> +struct threads_table_entry {
> +       struct rb_root_cached  entries;
> +       struct rw_semaphore    lock;
> +       unsigned int           nr;
> +       struct thread          *last_match;
> +};
> +
> +struct threads {
> +       struct threads_table_entry table[THREADS__TABLE_SIZE];
> +};
> +
> +void threads__init(struct threads *threads);
> +void threads__exit(struct threads *threads);
> +size_t threads__nr(struct threads *threads);
> +struct thread *threads__find(struct threads *threads, pid_t tid);
> +struct thread *threads__findnew(struct threads *threads, pid_t pid, pid_t tid, bool *created);
> +void threads__remove_all_threads(struct threads *threads);
> +void threads__remove(struct threads *threads, struct thread *thread);
> +int threads__for_each_thread(struct threads *threads,
> +                            int (*fn)(struct thread *thread, void *data),
> +                            void *data);
> +
> +#endif /* __PERF_THREADS_H */
> --
> 2.43.0.687.g38aa6559b0-goog
>
  
Ian Rogers Feb. 27, 2024, 7:24 a.m. UTC | #2
On Mon, Feb 26, 2024 at 11:07 PM Namhyung Kim <namhyung@kernel.org> wrote:
>
> On Tue, Feb 13, 2024 at 10:37 PM Ian Rogers <irogers@google.com> wrote:
> >
> > Move threads out of machine and move thread_rb_node into the C
> > file. This hides the implementation of threads from the rest of the
> > code allowing for it to be refactored.
> >
> > Locking discipline is tightened up in this change.
>
> Doesn't look like a simple code move.  Can we split the locking
> change from the move to make the reviewer's life a bit easier? :)

Not sure I follow. Take threads_nr as an example.

The old code is in machine.c, so:
-static size_t machine__threads_nr(const struct machine *machine)
-{
-       size_t nr = 0;
-
-       for (int i = 0; i < THREADS__TABLE_SIZE; i++)
-               nr += machine->threads[i].nr;
-
-       return nr;
-}

The new code is in threads.c:
+size_t threads__nr(struct threads *threads)
+{
+       size_t nr = 0;
+
+       for (int i = 0; i < THREADS__TABLE_SIZE; i++) {
+               struct threads_table_entry *table = &threads->table[i];
+
+               down_read(&table->lock);
+               nr += table->nr;
+               up_read(&table->lock);
+       }
+       return nr;
+}

So it is a copy paste from one file to the other. The only difference
is that the old code failed to take a lock when reading "nr" so the
locking is added. I wanted to make sure all the functions in threads.c
were properly correct wrt locking, semaphore creation and destruction,
etc.  We could have a broken threads.c and fix it in the next change,
but given that's a bug it could make bisection more difficult.
Ultimately I thought the locking changes were small enough to not
warrant being on their own compared to the advantages of having a sane
threads abstraction.

Thanks,
Ian

> Thanks,
> Namhyung
>
> >
> > Signed-off-by: Ian Rogers <irogers@google.com>
> > ---
> >  tools/perf/util/Build                 |   1 +
> >  tools/perf/util/bpf_lock_contention.c |   8 +-
> >  tools/perf/util/machine.c             | 287 ++++----------------------
> >  tools/perf/util/machine.h             |  20 +-
> >  tools/perf/util/thread.c              |   2 +-
> >  tools/perf/util/thread.h              |   6 -
> >  tools/perf/util/threads.c             | 244 ++++++++++++++++++++++
> >  tools/perf/util/threads.h             |  35 ++++
> >  8 files changed, 325 insertions(+), 278 deletions(-)
> >  create mode 100644 tools/perf/util/threads.c
> >  create mode 100644 tools/perf/util/threads.h
> >
> > diff --git a/tools/perf/util/Build b/tools/perf/util/Build
> > index 8027f450fa3e..a0e8cd68d490 100644
> > --- a/tools/perf/util/Build
> > +++ b/tools/perf/util/Build
> > @@ -71,6 +71,7 @@ perf-y += ordered-events.o
> >  perf-y += namespaces.o
> >  perf-y += comm.o
> >  perf-y += thread.o
> > +perf-y += threads.o
> >  perf-y += thread_map.o
> >  perf-y += parse-events-flex.o
> >  perf-y += parse-events-bison.o
> > diff --git a/tools/perf/util/bpf_lock_contention.c b/tools/perf/util/bpf_lock_contention.c
> > index 31ff19afc20c..3992c8a9fd96 100644
> > --- a/tools/perf/util/bpf_lock_contention.c
> > +++ b/tools/perf/util/bpf_lock_contention.c
> > @@ -210,7 +210,7 @@ static const char *lock_contention_get_name(struct lock_contention *con,
> >
> >                 /* do not update idle comm which contains CPU number */
> >                 if (pid) {
> > -                       struct thread *t = __machine__findnew_thread(machine, /*pid=*/-1, pid);
> > +                       struct thread *t = machine__findnew_thread(machine, /*pid=*/-1, pid);
> >
> >                         if (t == NULL)
> >                                 return name;
> > @@ -302,9 +302,9 @@ int lock_contention_read(struct lock_contention *con)
> >                 return -1;
> >
> >         if (con->aggr_mode == LOCK_AGGR_TASK) {
> > -               struct thread *idle = __machine__findnew_thread(machine,
> > -                                                               /*pid=*/0,
> > -                                                               /*tid=*/0);
> > +               struct thread *idle = machine__findnew_thread(machine,
> > +                                                             /*pid=*/0,
> > +                                                             /*tid=*/0);
> >                 thread__set_comm(idle, "swapper", /*timestamp=*/0);
> >         }
> >
> > diff --git a/tools/perf/util/machine.c b/tools/perf/util/machine.c
> > index e072b2115b64..e668a97255f8 100644
> > --- a/tools/perf/util/machine.c
> > +++ b/tools/perf/util/machine.c
> > @@ -43,9 +43,6 @@
> >  #include <linux/string.h>
> >  #include <linux/zalloc.h>
> >
> > -static void __machine__remove_thread(struct machine *machine, struct thread_rb_node *nd,
> > -                                    struct thread *th, bool lock);
> > -
> >  static struct dso *machine__kernel_dso(struct machine *machine)
> >  {
> >         return map__dso(machine->vmlinux_map);
> > @@ -58,35 +55,6 @@ static void dsos__init(struct dsos *dsos)
> >         init_rwsem(&dsos->lock);
> >  }
> >
> > -static void machine__threads_init(struct machine *machine)
> > -{
> > -       int i;
> > -
> > -       for (i = 0; i < THREADS__TABLE_SIZE; i++) {
> > -               struct threads *threads = &machine->threads[i];
> > -               threads->entries = RB_ROOT_CACHED;
> > -               init_rwsem(&threads->lock);
> > -               threads->nr = 0;
> > -               threads->last_match = NULL;
> > -       }
> > -}
> > -
> > -static int thread_rb_node__cmp_tid(const void *key, const struct rb_node *nd)
> > -{
> > -       int to_find = (int) *((pid_t *)key);
> > -
> > -       return to_find - (int)thread__tid(rb_entry(nd, struct thread_rb_node, rb_node)->thread);
> > -}
> > -
> > -static struct thread_rb_node *thread_rb_node__find(const struct thread *th,
> > -                                                  struct rb_root *tree)
> > -{
> > -       pid_t to_find = thread__tid(th);
> > -       struct rb_node *nd = rb_find(&to_find, tree, thread_rb_node__cmp_tid);
> > -
> > -       return rb_entry(nd, struct thread_rb_node, rb_node);
> > -}
> > -
> >  static int machine__set_mmap_name(struct machine *machine)
> >  {
> >         if (machine__is_host(machine))
> > @@ -120,7 +88,7 @@ int machine__init(struct machine *machine, const char *root_dir, pid_t pid)
> >         RB_CLEAR_NODE(&machine->rb_node);
> >         dsos__init(&machine->dsos);
> >
> > -       machine__threads_init(machine);
> > +       threads__init(&machine->threads);
> >
> >         machine->vdso_info = NULL;
> >         machine->env = NULL;
> > @@ -221,27 +189,11 @@ static void dsos__exit(struct dsos *dsos)
> >
> >  void machine__delete_threads(struct machine *machine)
> >  {
> > -       struct rb_node *nd;
> > -       int i;
> > -
> > -       for (i = 0; i < THREADS__TABLE_SIZE; i++) {
> > -               struct threads *threads = &machine->threads[i];
> > -               down_write(&threads->lock);
> > -               nd = rb_first_cached(&threads->entries);
> > -               while (nd) {
> > -                       struct thread_rb_node *trb = rb_entry(nd, struct thread_rb_node, rb_node);
> > -
> > -                       nd = rb_next(nd);
> > -                       __machine__remove_thread(machine, trb, trb->thread, false);
> > -               }
> > -               up_write(&threads->lock);
> > -       }
> > +       threads__remove_all_threads(&machine->threads);
> >  }
> >
> >  void machine__exit(struct machine *machine)
> >  {
> > -       int i;
> > -
> >         if (machine == NULL)
> >                 return;
> >
> > @@ -254,12 +206,7 @@ void machine__exit(struct machine *machine)
> >         zfree(&machine->current_tid);
> >         zfree(&machine->kallsyms_filename);
> >
> > -       machine__delete_threads(machine);
> > -       for (i = 0; i < THREADS__TABLE_SIZE; i++) {
> > -               struct threads *threads = &machine->threads[i];
> > -
> > -               exit_rwsem(&threads->lock);
> > -       }
> > +       threads__exit(&machine->threads);
> >  }
> >
> >  void machine__delete(struct machine *machine)
> > @@ -526,7 +473,7 @@ static void machine__update_thread_pid(struct machine *machine,
> >         if (thread__pid(th) == thread__tid(th))
> >                 return;
> >
> > -       leader = __machine__findnew_thread(machine, thread__pid(th), thread__pid(th));
> > +       leader = machine__findnew_thread(machine, thread__pid(th), thread__pid(th));
> >         if (!leader)
> >                 goto out_err;
> >
> > @@ -560,160 +507,55 @@ static void machine__update_thread_pid(struct machine *machine,
> >         goto out_put;
> >  }
> >
> > -/*
> > - * Front-end cache - TID lookups come in blocks,
> > - * so most of the time we dont have to look up
> > - * the full rbtree:
> > - */
> > -static struct thread*
> > -__threads__get_last_match(struct threads *threads, struct machine *machine,
> > -                         int pid, int tid)
> > -{
> > -       struct thread *th;
> > -
> > -       th = threads->last_match;
> > -       if (th != NULL) {
> > -               if (thread__tid(th) == tid) {
> > -                       machine__update_thread_pid(machine, th, pid);
> > -                       return thread__get(th);
> > -               }
> > -               thread__put(threads->last_match);
> > -               threads->last_match = NULL;
> > -       }
> > -
> > -       return NULL;
> > -}
> > -
> > -static struct thread*
> > -threads__get_last_match(struct threads *threads, struct machine *machine,
> > -                       int pid, int tid)
> > -{
> > -       struct thread *th = NULL;
> > -
> > -       if (perf_singlethreaded)
> > -               th = __threads__get_last_match(threads, machine, pid, tid);
> > -
> > -       return th;
> > -}
> > -
> > -static void
> > -__threads__set_last_match(struct threads *threads, struct thread *th)
> > -{
> > -       thread__put(threads->last_match);
> > -       threads->last_match = thread__get(th);
> > -}
> > -
> > -static void
> > -threads__set_last_match(struct threads *threads, struct thread *th)
> > -{
> > -       if (perf_singlethreaded)
> > -               __threads__set_last_match(threads, th);
> > -}
> > -
> >  /*
> >   * Caller must eventually drop thread->refcnt returned with a successful
> >   * lookup/new thread inserted.
> >   */
> > -static struct thread *____machine__findnew_thread(struct machine *machine,
> > -                                                 struct threads *threads,
> > -                                                 pid_t pid, pid_t tid,
> > -                                                 bool create)
> > +static struct thread *__machine__findnew_thread(struct machine *machine,
> > +                                               pid_t pid,
> > +                                               pid_t tid,
> > +                                               bool create)
> >  {
> > -       struct rb_node **p = &threads->entries.rb_root.rb_node;
> > -       struct rb_node *parent = NULL;
> > -       struct thread *th;
> > -       struct thread_rb_node *nd;
> > -       bool leftmost = true;
> > +       struct thread *th = threads__find(&machine->threads, tid);
> > +       bool created;
> >
> > -       th = threads__get_last_match(threads, machine, pid, tid);
> > -       if (th)
> > +       if (th) {
> > +               machine__update_thread_pid(machine, th, pid);
> >                 return th;
> > -
> > -       while (*p != NULL) {
> > -               parent = *p;
> > -               th = rb_entry(parent, struct thread_rb_node, rb_node)->thread;
> > -
> > -               if (thread__tid(th) == tid) {
> > -                       threads__set_last_match(threads, th);
> > -                       machine__update_thread_pid(machine, th, pid);
> > -                       return thread__get(th);
> > -               }
> > -
> > -               if (tid < thread__tid(th))
> > -                       p = &(*p)->rb_left;
> > -               else {
> > -                       p = &(*p)->rb_right;
> > -                       leftmost = false;
> > -               }
> >         }
> > -
> >         if (!create)
> >                 return NULL;
> >
> > -       th = thread__new(pid, tid);
> > -       if (th == NULL)
> > -               return NULL;
> > -
> > -       nd = malloc(sizeof(*nd));
> > -       if (nd == NULL) {
> > -               thread__put(th);
> > -               return NULL;
> > -       }
> > -       nd->thread = th;
> > -
> > -       rb_link_node(&nd->rb_node, parent, p);
> > -       rb_insert_color_cached(&nd->rb_node, &threads->entries, leftmost);
> > -       /*
> > -        * We have to initialize maps separately after rb tree is updated.
> > -        *
> > -        * The reason is that we call machine__findnew_thread within
> > -        * thread__init_maps to find the thread leader and that would screwed
> > -        * the rb tree.
> > -        */
> > -       if (thread__init_maps(th, machine)) {
> > -               pr_err("Thread init failed thread %d\n", pid);
> > -               rb_erase_cached(&nd->rb_node, &threads->entries);
> > -               RB_CLEAR_NODE(&nd->rb_node);
> > -               free(nd);
> > -               thread__put(th);
> > -               return NULL;
> > -       }
> > -       /*
> > -        * It is now in the rbtree, get a ref
> > -        */
> > -       threads__set_last_match(threads, th);
> > -       ++threads->nr;
> > -
> > -       return thread__get(th);
> > -}
> > +       th = threads__findnew(&machine->threads, pid, tid, &created);
> > +       if (created) {
> > +               /*
> > +                * We have to initialize maps separately after rb tree is
> > +                * updated.
> > +                *
> > +                * The reason is that we call machine__findnew_thread within
> > +                * thread__init_maps to find the thread leader and that would
> > +                * screwed the rb tree.
> > +                */
> > +               if (thread__init_maps(th, machine)) {
> > +                       pr_err("Thread init failed thread %d\n", pid);
> > +                       threads__remove(&machine->threads, th);
> > +                       thread__put(th);
> > +                       return NULL;
> > +               }
> > +       } else
> > +               machine__update_thread_pid(machine, th, pid);
> >
> > -struct thread *__machine__findnew_thread(struct machine *machine, pid_t pid, pid_t tid)
> > -{
> > -       return ____machine__findnew_thread(machine, machine__threads(machine, tid), pid, tid, true);
> > +       return th;
> >  }
> >
> > -struct thread *machine__findnew_thread(struct machine *machine, pid_t pid,
> > -                                      pid_t tid)
> > +struct thread *machine__findnew_thread(struct machine *machine, pid_t pid, pid_t tid)
> >  {
> > -       struct threads *threads = machine__threads(machine, tid);
> > -       struct thread *th;
> > -
> > -       down_write(&threads->lock);
> > -       th = __machine__findnew_thread(machine, pid, tid);
> > -       up_write(&threads->lock);
> > -       return th;
> > +       return __machine__findnew_thread(machine, pid, tid, /*create=*/true);
> >  }
> >
> > -struct thread *machine__find_thread(struct machine *machine, pid_t pid,
> > -                                   pid_t tid)
> > +struct thread *machine__find_thread(struct machine *machine, pid_t pid, pid_t tid)
> >  {
> > -       struct threads *threads = machine__threads(machine, tid);
> > -       struct thread *th;
> > -
> > -       down_read(&threads->lock);
> > -       th =  ____machine__findnew_thread(machine, threads, pid, tid, false);
> > -       up_read(&threads->lock);
> > -       return th;
> > +       return __machine__findnew_thread(machine, pid, tid, /*create=*/false);
> >  }
> >
> >  /*
> > @@ -1127,23 +969,13 @@ static int machine_fprintf_cb(struct thread *thread, void *data)
> >         return 0;
> >  }
> >
> > -static size_t machine__threads_nr(const struct machine *machine)
> > -{
> > -       size_t nr = 0;
> > -
> > -       for (int i = 0; i < THREADS__TABLE_SIZE; i++)
> > -               nr += machine->threads[i].nr;
> > -
> > -       return nr;
> > -}
> > -
> >  size_t machine__fprintf(struct machine *machine, FILE *fp)
> >  {
> >         struct machine_fprintf_cb_args args = {
> >                 .fp = fp,
> >                 .printed = 0,
> >         };
> > -       size_t ret = fprintf(fp, "Threads: %zu\n", machine__threads_nr(machine));
> > +       size_t ret = fprintf(fp, "Threads: %zu\n", threads__nr(&machine->threads));
> >
> >         machine__for_each_thread(machine, machine_fprintf_cb, &args);
> >         return ret + args.printed;
> > @@ -2069,36 +1901,9 @@ int machine__process_mmap_event(struct machine *machine, union perf_event *event
> >         return 0;
> >  }
> >
> > -static void __machine__remove_thread(struct machine *machine, struct thread_rb_node *nd,
> > -                                    struct thread *th, bool lock)
> > -{
> > -       struct threads *threads = machine__threads(machine, thread__tid(th));
> > -
> > -       if (!nd)
> > -               nd = thread_rb_node__find(th, &threads->entries.rb_root);
> > -
> > -       if (threads->last_match && RC_CHK_EQUAL(threads->last_match, th))
> > -               threads__set_last_match(threads, NULL);
> > -
> > -       if (lock)
> > -               down_write(&threads->lock);
> > -
> > -       BUG_ON(refcount_read(thread__refcnt(th)) == 0);
> > -
> > -       thread__put(nd->thread);
> > -       rb_erase_cached(&nd->rb_node, &threads->entries);
> > -       RB_CLEAR_NODE(&nd->rb_node);
> > -       --threads->nr;
> > -
> > -       free(nd);
> > -
> > -       if (lock)
> > -               up_write(&threads->lock);
> > -}
> > -
> >  void machine__remove_thread(struct machine *machine, struct thread *th)
> >  {
> > -       return __machine__remove_thread(machine, NULL, th, true);
> > +       return threads__remove(&machine->threads, th);
> >  }
> >
> >  int machine__process_fork_event(struct machine *machine, union perf_event *event,
> > @@ -3232,23 +3037,7 @@ int machine__for_each_thread(struct machine *machine,
> >                              int (*fn)(struct thread *thread, void *p),
> >                              void *priv)
> >  {
> > -       struct threads *threads;
> > -       struct rb_node *nd;
> > -       int rc = 0;
> > -       int i;
> > -
> > -       for (i = 0; i < THREADS__TABLE_SIZE; i++) {
> > -               threads = &machine->threads[i];
> > -               for (nd = rb_first_cached(&threads->entries); nd;
> > -                    nd = rb_next(nd)) {
> > -                       struct thread_rb_node *trb = rb_entry(nd, struct thread_rb_node, rb_node);
> > -
> > -                       rc = fn(trb->thread, priv);
> > -                       if (rc != 0)
> > -                               return rc;
> > -               }
> > -       }
> > -       return rc;
> > +       return threads__for_each_thread(&machine->threads, fn, priv);
> >  }
> >
> >  int machines__for_each_thread(struct machines *machines,
> > diff --git a/tools/perf/util/machine.h b/tools/perf/util/machine.h
> > index b738ce84817b..e28c787616fe 100644
> > --- a/tools/perf/util/machine.h
> > +++ b/tools/perf/util/machine.h
> > @@ -7,6 +7,7 @@
> >  #include "maps.h"
> >  #include "dsos.h"
> >  #include "rwsem.h"
> > +#include "threads.h"
> >
> >  struct addr_location;
> >  struct branch_stack;
> > @@ -28,16 +29,6 @@ extern const char *ref_reloc_sym_names[];
> >
> >  struct vdso_info;
> >
> > -#define THREADS__TABLE_BITS    8
> > -#define THREADS__TABLE_SIZE    (1 << THREADS__TABLE_BITS)
> > -
> > -struct threads {
> > -       struct rb_root_cached  entries;
> > -       struct rw_semaphore    lock;
> > -       unsigned int           nr;
> > -       struct thread          *last_match;
> > -};
> > -
> >  struct machine {
> >         struct rb_node    rb_node;
> >         pid_t             pid;
> > @@ -48,7 +39,7 @@ struct machine {
> >         char              *root_dir;
> >         char              *mmap_name;
> >         char              *kallsyms_filename;
> > -       struct threads    threads[THREADS__TABLE_SIZE];
> > +       struct threads    threads;
> >         struct vdso_info  *vdso_info;
> >         struct perf_env   *env;
> >         struct dsos       dsos;
> > @@ -69,12 +60,6 @@ struct machine {
> >         bool              trampolines_mapped;
> >  };
> >
> > -static inline struct threads *machine__threads(struct machine *machine, pid_t tid)
> > -{
> > -       /* Cast it to handle tid == -1 */
> > -       return &machine->threads[(unsigned int)tid % THREADS__TABLE_SIZE];
> > -}
> > -
> >  /*
> >   * The main kernel (vmlinux) map
> >   */
> > @@ -220,7 +205,6 @@ bool machine__is(struct machine *machine, const char *arch);
> >  bool machine__normalized_is(struct machine *machine, const char *arch);
> >  int machine__nr_cpus_avail(struct machine *machine);
> >
> > -struct thread *__machine__findnew_thread(struct machine *machine, pid_t pid, pid_t tid);
> >  struct thread *machine__findnew_thread(struct machine *machine, pid_t pid, pid_t tid);
> >
> >  struct dso *machine__findnew_dso_id(struct machine *machine, const char *filename, struct dso_id *id);
> > diff --git a/tools/perf/util/thread.c b/tools/perf/util/thread.c
> > index c59ab4d79163..1aa8962dcf52 100644
> > --- a/tools/perf/util/thread.c
> > +++ b/tools/perf/util/thread.c
> > @@ -26,7 +26,7 @@ int thread__init_maps(struct thread *thread, struct machine *machine)
> >         if (pid == thread__tid(thread) || pid == -1) {
> >                 thread__set_maps(thread, maps__new(machine));
> >         } else {
> > -               struct thread *leader = __machine__findnew_thread(machine, pid, pid);
> > +               struct thread *leader = machine__findnew_thread(machine, pid, pid);
> >
> >                 if (leader) {
> >                         thread__set_maps(thread, maps__get(thread__maps(leader)));
> > diff --git a/tools/perf/util/thread.h b/tools/perf/util/thread.h
> > index 0df775b5c110..4b8f3e9e513b 100644
> > --- a/tools/perf/util/thread.h
> > +++ b/tools/perf/util/thread.h
> > @@ -3,7 +3,6 @@
> >  #define __PERF_THREAD_H
> >
> >  #include <linux/refcount.h>
> > -#include <linux/rbtree.h>
> >  #include <linux/list.h>
> >  #include <stdio.h>
> >  #include <unistd.h>
> > @@ -30,11 +29,6 @@ struct lbr_stitch {
> >         struct callchain_cursor_node    *prev_lbr_cursor;
> >  };
> >
> > -struct thread_rb_node {
> > -       struct rb_node rb_node;
> > -       struct thread *thread;
> > -};
> > -
> >  DECLARE_RC_STRUCT(thread) {
> >         /** @maps: mmaps associated with this thread. */
> >         struct maps             *maps;
> > diff --git a/tools/perf/util/threads.c b/tools/perf/util/threads.c
> > new file mode 100644
> > index 000000000000..d984ec939c7b
> > --- /dev/null
> > +++ b/tools/perf/util/threads.c
> > @@ -0,0 +1,244 @@
> > +// SPDX-License-Identifier: GPL-2.0
> > +#include "threads.h"
> > +#include "machine.h"
> > +#include "thread.h"
> > +
> > +struct thread_rb_node {
> > +       struct rb_node rb_node;
> > +       struct thread *thread;
> > +};
> > +
> > +static struct threads_table_entry *threads__table(struct threads *threads, pid_t tid)
> > +{
> > +       /* Cast it to handle tid == -1 */
> > +       return &threads->table[(unsigned int)tid % THREADS__TABLE_SIZE];
> > +}
> > +
> > +void threads__init(struct threads *threads)
> > +{
> > +       for (int i = 0; i < THREADS__TABLE_SIZE; i++) {
> > +               struct threads_table_entry *table = &threads->table[i];
> > +
> > +               table->entries = RB_ROOT_CACHED;
> > +               init_rwsem(&table->lock);
> > +               table->nr = 0;
> > +               table->last_match = NULL;
> > +       }
> > +}
> > +
> > +void threads__exit(struct threads *threads)
> > +{
> > +       threads__remove_all_threads(threads);
> > +       for (int i = 0; i < THREADS__TABLE_SIZE; i++) {
> > +               struct threads_table_entry *table = &threads->table[i];
> > +
> > +               exit_rwsem(&table->lock);
> > +       }
> > +}
> > +
> > +size_t threads__nr(struct threads *threads)
> > +{
> > +       size_t nr = 0;
> > +
> > +       for (int i = 0; i < THREADS__TABLE_SIZE; i++) {
> > +               struct threads_table_entry *table = &threads->table[i];
> > +
> > +               down_read(&table->lock);
> > +               nr += table->nr;
> > +               up_read(&table->lock);
> > +       }
> > +       return nr;
> > +}
> > +
> > +/*
> > + * Front-end cache - TID lookups come in blocks,
> > + * so most of the time we dont have to look up
> > + * the full rbtree:
> > + */
> > +static struct thread *__threads_table_entry__get_last_match(struct threads_table_entry *table,
> > +                                                           pid_t tid)
> > +{
> > +       struct thread *th, *res = NULL;
> > +
> > +       th = table->last_match;
> > +       if (th != NULL) {
> > +               if (thread__tid(th) == tid)
> > +                       res = thread__get(th);
> > +       }
> > +       return res;
> > +}
> > +
> > +static void __threads_table_entry__set_last_match(struct threads_table_entry *table,
> > +                                                 struct thread *th)
> > +{
> > +       thread__put(table->last_match);
> > +       table->last_match = thread__get(th);
> > +}
> > +
> > +static void threads_table_entry__set_last_match(struct threads_table_entry *table,
> > +                                               struct thread *th)
> > +{
> > +       down_write(&table->lock);
> > +       __threads_table_entry__set_last_match(table, th);
> > +       up_write(&table->lock);
> > +}
> > +
> > +struct thread *threads__find(struct threads *threads, pid_t tid)
> > +{
> > +       struct threads_table_entry *table  = threads__table(threads, tid);
> > +       struct rb_node **p;
> > +       struct thread *res = NULL;
> > +
> > +       down_read(&table->lock);
> > +       res = __threads_table_entry__get_last_match(table, tid);
> > +       if (res)
> > +               return res;
> > +
> > +       p = &table->entries.rb_root.rb_node;
> > +       while (*p != NULL) {
> > +               struct rb_node *parent = *p;
> > +               struct thread *th = rb_entry(parent, struct thread_rb_node, rb_node)->thread;
> > +
> > +               if (thread__tid(th) == tid) {
> > +                       res = thread__get(th);
> > +                       break;
> > +               }
> > +
> > +               if (tid < thread__tid(th))
> > +                       p = &(*p)->rb_left;
> > +               else
> > +                       p = &(*p)->rb_right;
> > +       }
> > +       up_read(&table->lock);
> > +       if (res)
> > +               threads_table_entry__set_last_match(table, res);
> > +       return res;
> > +}
> > +
> > +struct thread *threads__findnew(struct threads *threads, pid_t pid, pid_t tid, bool *created)
> > +{
> > +       struct threads_table_entry *table  = threads__table(threads, tid);
> > +       struct rb_node **p;
> > +       struct rb_node *parent = NULL;
> > +       struct thread *res = NULL;
> > +       struct thread_rb_node *nd;
> > +       bool leftmost = true;
> > +
> > +       *created = false;
> > +       down_write(&table->lock);
> > +       p = &table->entries.rb_root.rb_node;
> > +       while (*p != NULL) {
> > +               struct thread *th;
> > +
> > +               parent = *p;
> > +               th = rb_entry(parent, struct thread_rb_node, rb_node)->thread;
> > +
> > +               if (thread__tid(th) == tid) {
> > +                       __threads_table_entry__set_last_match(table, th);
> > +                       res = thread__get(th);
> > +                       goto out_unlock;
> > +               }
> > +
> > +               if (tid < thread__tid(th))
> > +                       p = &(*p)->rb_left;
> > +               else {
> > +                       leftmost = false;
> > +                       p = &(*p)->rb_right;
> > +               }
> > +       }
> > +       nd = malloc(sizeof(*nd));
> > +       if (nd == NULL)
> > +               goto out_unlock;
> > +       res = thread__new(pid, tid);
> > +       if (!res)
> > +               free(nd);
> > +       else {
> > +               *created = true;
> > +               nd->thread = thread__get(res);
> > +               rb_link_node(&nd->rb_node, parent, p);
> > +               rb_insert_color_cached(&nd->rb_node, &table->entries, leftmost);
> > +               ++table->nr;
> > +               __threads_table_entry__set_last_match(table, res);
> > +       }
> > +out_unlock:
> > +       up_write(&table->lock);
> > +       return res;
> > +}
> > +
> > +void threads__remove_all_threads(struct threads *threads)
> > +{
> > +       for (int i = 0; i < THREADS__TABLE_SIZE; i++) {
> > +               struct threads_table_entry *table = &threads->table[i];
> > +               struct rb_node *nd;
> > +
> > +               down_write(&table->lock);
> > +               __threads_table_entry__set_last_match(table, NULL);
> > +               nd = rb_first_cached(&table->entries);
> > +               while (nd) {
> > +                       struct thread_rb_node *trb = rb_entry(nd, struct thread_rb_node, rb_node);
> > +
> > +                       nd = rb_next(nd);
> > +                       thread__put(trb->thread);
> > +                       rb_erase_cached(&trb->rb_node, &table->entries);
> > +                       RB_CLEAR_NODE(&trb->rb_node);
> > +                       --table->nr;
> > +
> > +                       free(trb);
> > +               }
> > +               assert(table->nr == 0);
> > +               up_write(&table->lock);
> > +       }
> > +}
> > +
> > +void threads__remove(struct threads *threads, struct thread *thread)
> > +{
> > +       struct rb_node **p;
> > +       struct threads_table_entry *table  = threads__table(threads, thread__tid(thread));
> > +       pid_t tid = thread__tid(thread);
> > +
> > +       down_write(&table->lock);
> > +       if (table->last_match && RC_CHK_EQUAL(table->last_match, thread))
> > +               __threads_table_entry__set_last_match(table, NULL);
> > +
> > +       p = &table->entries.rb_root.rb_node;
> > +       while (*p != NULL) {
> > +               struct rb_node *parent = *p;
> > +               struct thread_rb_node *nd = rb_entry(parent, struct thread_rb_node, rb_node);
> > +               struct thread *th = nd->thread;
> > +
> > +               if (RC_CHK_EQUAL(th, thread)) {
> > +                       thread__put(nd->thread);
> > +                       rb_erase_cached(&nd->rb_node, &table->entries);
> > +                       RB_CLEAR_NODE(&nd->rb_node);
> > +                       --table->nr;
> > +                       free(nd);
> > +                       break;
> > +               }
> > +
> > +               if (tid < thread__tid(th))
> > +                       p = &(*p)->rb_left;
> > +               else
> > +                       p = &(*p)->rb_right;
> > +       }
> > +       up_write(&table->lock);
> > +}
> > +
> > +int threads__for_each_thread(struct threads *threads,
> > +                            int (*fn)(struct thread *thread, void *data),
> > +                            void *data)
> > +{
> > +       for (int i = 0; i < THREADS__TABLE_SIZE; i++) {
> > +               struct threads_table_entry *table = &threads->table[i];
> > +               struct rb_node *nd;
> > +
> > +               for (nd = rb_first_cached(&table->entries); nd; nd = rb_next(nd)) {
> > +                       struct thread_rb_node *trb = rb_entry(nd, struct thread_rb_node, rb_node);
> > +                       int rc = fn(trb->thread, data);
> > +
> > +                       if (rc != 0)
> > +                               return rc;
> > +               }
> > +       }
> > +       return 0;
> > +
> > +}
> > diff --git a/tools/perf/util/threads.h b/tools/perf/util/threads.h
> > new file mode 100644
> > index 000000000000..ed67de627578
> > --- /dev/null
> > +++ b/tools/perf/util/threads.h
> > @@ -0,0 +1,35 @@
> > +/* SPDX-License-Identifier: GPL-2.0 */
> > +#ifndef __PERF_THREADS_H
> > +#define __PERF_THREADS_H
> > +
> > +#include <linux/rbtree.h>
> > +#include "rwsem.h"
> > +
> > +struct thread;
> > +
> > +#define THREADS__TABLE_BITS    8
> > +#define THREADS__TABLE_SIZE    (1 << THREADS__TABLE_BITS)
> > +
> > +struct threads_table_entry {
> > +       struct rb_root_cached  entries;
> > +       struct rw_semaphore    lock;
> > +       unsigned int           nr;
> > +       struct thread          *last_match;
> > +};
> > +
> > +struct threads {
> > +       struct threads_table_entry table[THREADS__TABLE_SIZE];
> > +};
> > +
> > +void threads__init(struct threads *threads);
> > +void threads__exit(struct threads *threads);
> > +size_t threads__nr(struct threads *threads);
> > +struct thread *threads__find(struct threads *threads, pid_t tid);
> > +struct thread *threads__findnew(struct threads *threads, pid_t pid, pid_t tid, bool *created);
> > +void threads__remove_all_threads(struct threads *threads);
> > +void threads__remove(struct threads *threads, struct thread *thread);
> > +int threads__for_each_thread(struct threads *threads,
> > +                            int (*fn)(struct thread *thread, void *data),
> > +                            void *data);
> > +
> > +#endif /* __PERF_THREADS_H */
> > --
> > 2.43.0.687.g38aa6559b0-goog
> >
  
Namhyung Kim Feb. 27, 2024, 5:31 p.m. UTC | #3
On Mon, Feb 26, 2024 at 11:24 PM Ian Rogers <irogers@google.com> wrote:
>
> On Mon, Feb 26, 2024 at 11:07 PM Namhyung Kim <namhyung@kernel.org> wrote:
> >
> > On Tue, Feb 13, 2024 at 10:37 PM Ian Rogers <irogers@google.com> wrote:
> > >
> > > Move threads out of machine and move thread_rb_node into the C
> > > file. This hides the implementation of threads from the rest of the
> > > code allowing for it to be refactored.
> > >
> > > Locking discipline is tightened up in this change.
> >
> > Doesn't look like a simple code move.  Can we split the locking
> > change from the move to make the reviewer's life a bit easier? :)
>
> Not sure I follow. Take threads_nr as an example.
>
> The old code is in machine.c, so:
> -static size_t machine__threads_nr(const struct machine *machine)
> -{
> -       size_t nr = 0;
> -
> -       for (int i = 0; i < THREADS__TABLE_SIZE; i++)
> -               nr += machine->threads[i].nr;
> -
> -       return nr;
> -}
>
> The new code is in threads.c:
> +size_t threads__nr(struct threads *threads)
> +{
> +       size_t nr = 0;
> +
> +       for (int i = 0; i < THREADS__TABLE_SIZE; i++) {
> +               struct threads_table_entry *table = &threads->table[i];
> +
> +               down_read(&table->lock);
> +               nr += table->nr;
> +               up_read(&table->lock);
> +       }
> +       return nr;
> +}
>
> So it is a copy paste from one file to the other. The only difference
> is that the old code failed to take a lock when reading "nr" so the
> locking is added. I wanted to make sure all the functions in threads.c
> were properly correct wrt locking, semaphore creation and destruction,
> etc.  We could have a broken threads.c and fix it in the next change,
> but given that's a bug it could make bisection more difficult.
> Ultimately I thought the locking changes were small enough to not
> warrant being on their own compared to the advantages of having a sane
> threads abstraction.

I can see some other differences like machine__findnew_thread()
which I think is due to the locking change.  Maybe we can fix the
problem before moving the code and let the code move simple.

Thanks,
Namhyung
  
Ian Rogers Feb. 27, 2024, 7:02 p.m. UTC | #4
On Tue, Feb 27, 2024 at 9:31 AM Namhyung Kim <namhyung@kernel.org> wrote:
>
> On Mon, Feb 26, 2024 at 11:24 PM Ian Rogers <irogers@google.com> wrote:
> >
> > On Mon, Feb 26, 2024 at 11:07 PM Namhyung Kim <namhyung@kernel.org> wrote:
> > >
> > > On Tue, Feb 13, 2024 at 10:37 PM Ian Rogers <irogers@google.com> wrote:
> > > >
> > > > Move threads out of machine and move thread_rb_node into the C
> > > > file. This hides the implementation of threads from the rest of the
> > > > code allowing for it to be refactored.
> > > >
> > > > Locking discipline is tightened up in this change.
> > >
> > > Doesn't look like a simple code move.  Can we split the locking
> > > change from the move to make the reviewer's life a bit easier? :)
> >
> > Not sure I follow. Take threads_nr as an example.
> >
> > The old code is in machine.c, so:
> > -static size_t machine__threads_nr(const struct machine *machine)
> > -{
> > -       size_t nr = 0;
> > -
> > -       for (int i = 0; i < THREADS__TABLE_SIZE; i++)
> > -               nr += machine->threads[i].nr;
> > -
> > -       return nr;
> > -}
> >
> > The new code is in threads.c:
> > +size_t threads__nr(struct threads *threads)
> > +{
> > +       size_t nr = 0;
> > +
> > +       for (int i = 0; i < THREADS__TABLE_SIZE; i++) {
> > +               struct threads_table_entry *table = &threads->table[i];
> > +
> > +               down_read(&table->lock);
> > +               nr += table->nr;
> > +               up_read(&table->lock);
> > +       }
> > +       return nr;
> > +}
> >
> > So it is a copy paste from one file to the other. The only difference
> > is that the old code failed to take a lock when reading "nr" so the
> > locking is added. I wanted to make sure all the functions in threads.c
> > were properly correct wrt locking, semaphore creation and destruction,
> > etc.  We could have a broken threads.c and fix it in the next change,
> > but given that's a bug it could make bisection more difficult.
> > Ultimately I thought the locking changes were small enough to not
> > warrant being on their own compared to the advantages of having a sane
> > threads abstraction.
>
> I can see some other differences like machine__findnew_thread()
> which I think is due to the locking change.  Maybe we can fix the
> problem before moving the code and let the code move simple.

I'll see what I can split out in v2. I don't think findnew will change
and the nr change is trivial. In the previous code the lock is taken
before calling __machine__findnew_thread, which doesn't make sense
when we try to abstract inside of threads, where it should
take/release the lock in the threads and not the machine code. Moving
the lock to __machine__findnew_thread doesn't really make sense as the
__ implies the lock is already held.

Thanks,
Ian

> Thanks,
> Namhyung
  
Arnaldo Carvalho de Melo Feb. 27, 2024, 7:17 p.m. UTC | #5
On Tue, Feb 27, 2024 at 09:31:33AM -0800, Namhyung Kim wrote:
> I can see some other differences like machine__findnew_thread()
> which I think is due to the locking change.  Maybe we can fix the
> problem before moving the code and let the code move simple.

I was going to suggest that, agreed.

We may start doing a refactoring, then find a bug, at that point we
first fix the problem them go back to refactoring.

- Arnaldo
  
Ian Rogers Feb. 27, 2024, 9:42 p.m. UTC | #6
On Tue, Feb 27, 2024 at 11:17 AM Arnaldo Carvalho de Melo
<acme@kernel.org> wrote:
>
> On Tue, Feb 27, 2024 at 09:31:33AM -0800, Namhyung Kim wrote:
> > I can see some other differences like machine__findnew_thread()
> > which I think is due to the locking change.  Maybe we can fix the
> > problem before moving the code and let the code move simple.
>
> I was going to suggest that, agreed.
>
> We may start doing a refactoring, then find a bug, at that point we
> first fix the problem them go back to refactoring.

Sure I do this all the time. Your typical complaint on the v+1 patch
set is to move the bug fixes to the front of the changes. On the v+2
patch set the bug fixes get applied but not the rest of the patch
series, etc.

Here we are refactoring code for an rb-tree implementation of threads
and worrying about its correctness. There's no indication it's not
correct, it is largely copy and paste, there is also good evidence in
the locking disciple it is more correct. The next patch deletes that
implementation, replacing it with a hash table. Were I not trying to
break things apart I could squash those 2 patches together, but I've
tried to do the right thing. Now we're trying to micro correct, break
apart, etc. a state that gets deleted. A reviewer could equally
criticise this being 2 changes rather than 1, and the cognitive load
of having to look at code that gets deleted. At some point it is a
judgement call, and I think this patch is actually the right size. I
think what is missing here is some motivation in the commit message to
the findnew refactoring and so I'll add that.

Thanks,
Ian
  
Namhyung Kim Feb. 28, 2024, 6:39 a.m. UTC | #7
On Tue, Feb 27, 2024 at 1:42 PM Ian Rogers <irogers@google.com> wrote:
>
> On Tue, Feb 27, 2024 at 11:17 AM Arnaldo Carvalho de Melo
> <acme@kernel.org> wrote:
> >
> > On Tue, Feb 27, 2024 at 09:31:33AM -0800, Namhyung Kim wrote:
> > > I can see some other differences like machine__findnew_thread()
> > > which I think is due to the locking change.  Maybe we can fix the
> > > problem before moving the code and let the code move simple.
> >
> > I was going to suggest that, agreed.
> >
> > We may start doing a refactoring, then find a bug, at that point we
> > first fix the problem them go back to refactoring.
>
> Sure I do this all the time. Your typical complaint on the v+1 patch
> set is to move the bug fixes to the front of the changes. On the v+2
> patch set the bug fixes get applied but not the rest of the patch
> series, etc.
>
> Here we are refactoring code for an rb-tree implementation of threads
> and worrying about its correctness. There's no indication it's not
> correct, it is largely copy and paste, there is also good evidence in
> the locking disciple it is more correct. The next patch deletes that
> implementation, replacing it with a hash table. Were I not trying to
> break things apart I could squash those 2 patches together, but I've
> tried to do the right thing. Now we're trying to micro correct, break
> apart, etc. a state that gets deleted. A reviewer could equally
> criticise this being 2 changes rather than 1, and the cognitive load
> of having to look at code that gets deleted. At some point it is a
> judgement call, and I think this patch is actually the right size. I
> think what is missing here is some motivation in the commit message to
> the findnew refactoring and so I'll add that.

I'm not against your approach and actually appreciate your effort
to split rb-tree refactoring and hash table introduction.  What I'm
asking is just to separate out the code moving.  I think you can do
whatever you want in the current file.  Once you have the final code
you can move it to its own file exactly the same.  When I look at this
commit, say a few years later, I won't expect a commit that says
moving something to a new file has other changes.

Thanks,
Namhyung
  
Ian Rogers Feb. 28, 2024, 7:24 a.m. UTC | #8
On Tue, Feb 27, 2024 at 10:40 PM Namhyung Kim <namhyung@kernel.org> wrote:
>
> On Tue, Feb 27, 2024 at 1:42 PM Ian Rogers <irogers@google.com> wrote:
> >
> > On Tue, Feb 27, 2024 at 11:17 AM Arnaldo Carvalho de Melo
> > <acme@kernel.org> wrote:
> > >
> > > On Tue, Feb 27, 2024 at 09:31:33AM -0800, Namhyung Kim wrote:
> > > > I can see some other differences like machine__findnew_thread()
> > > > which I think is due to the locking change.  Maybe we can fix the
> > > > problem before moving the code and let the code move simple.
> > >
> > > I was going to suggest that, agreed.
> > >
> > > We may start doing a refactoring, then find a bug, at that point we
> > > first fix the problem them go back to refactoring.
> >
> > Sure I do this all the time. Your typical complaint on the v+1 patch
> > set is to move the bug fixes to the front of the changes. On the v+2
> > patch set the bug fixes get applied but not the rest of the patch
> > series, etc.
> >
> > Here we are refactoring code for an rb-tree implementation of threads
> > and worrying about its correctness. There's no indication it's not
> > correct, it is largely copy and paste, there is also good evidence in
> > the locking disciple it is more correct. The next patch deletes that
> > implementation, replacing it with a hash table. Were I not trying to
> > break things apart I could squash those 2 patches together, but I've
> > tried to do the right thing. Now we're trying to micro correct, break
> > apart, etc. a state that gets deleted. A reviewer could equally
> > criticise this being 2 changes rather than 1, and the cognitive load
> > of having to look at code that gets deleted. At some point it is a
> > judgement call, and I think this patch is actually the right size. I
> > think what is missing here is some motivation in the commit message to
> > the findnew refactoring and so I'll add that.
>
> I'm not against your approach and actually appreciate your effort
> to split rb-tree refactoring and hash table introduction.  What I'm
> asking is just to separate out the code moving.  I think you can do
> whatever you want in the current file.  Once you have the final code
> you can move it to its own file exactly the same.  When I look at this
> commit, say a few years later, I won't expect a commit that says
> moving something to a new file has other changes.

The problem is that the code in machine treats the threads lock as if
it is a lock in machine. So there is __machine__findnew_thread which
implies the thread lock is held. This change is making threads its own
separate concept/collection and the lock belongs with that collection.
Most of the implementation of threads__findnew matches
__machine__findnew_thread, so we may be able to engineer a smaller
line diff by moving "__machine__findnew_thread" code into threads.c,
then renaming it to build the collection, etc. We could also build the
threads collection inside of machine and then in a separate change
move it to threads.[ch].  In the commit history this seems muddier
than just splitting out threads as a collection. Also, some of the API
design choices are motivated more by the hash table implementation of
the next patch than trying to have a good rbtree abstracted collection
of threads. Essentially it'd be engineering a collection of threads
but only with a view to delete it in the next patch. I don't think it
would be for the best and the commit history for deleted code is
unlikely to be looked upon.

Thanks,
Ian

> Thanks,
> Namhyung
  
Namhyung Kim Feb. 28, 2024, 11:43 p.m. UTC | #9
On Tue, Feb 27, 2024 at 11:24 PM Ian Rogers <irogers@google.com> wrote:
>
> On Tue, Feb 27, 2024 at 10:40 PM Namhyung Kim <namhyung@kernel.org> wrote:
> >
> > On Tue, Feb 27, 2024 at 1:42 PM Ian Rogers <irogers@google.com> wrote:
> > >
> > > On Tue, Feb 27, 2024 at 11:17 AM Arnaldo Carvalho de Melo
> > > <acme@kernel.org> wrote:
> > > >
> > > > On Tue, Feb 27, 2024 at 09:31:33AM -0800, Namhyung Kim wrote:
> > > > > I can see some other differences like machine__findnew_thread()
> > > > > which I think is due to the locking change.  Maybe we can fix the
> > > > > problem before moving the code and let the code move simple.
> > > >
> > > > I was going to suggest that, agreed.
> > > >
> > > > We may start doing a refactoring, then find a bug, at that point we
> > > > first fix the problem them go back to refactoring.
> > >
> > > Sure I do this all the time. Your typical complaint on the v+1 patch
> > > set is to move the bug fixes to the front of the changes. On the v+2
> > > patch set the bug fixes get applied but not the rest of the patch
> > > series, etc.
> > >
> > > Here we are refactoring code for an rb-tree implementation of threads
> > > and worrying about its correctness. There's no indication it's not
> > > correct, it is largely copy and paste, there is also good evidence in
> > > the locking disciple it is more correct. The next patch deletes that
> > > implementation, replacing it with a hash table. Were I not trying to
> > > break things apart I could squash those 2 patches together, but I've
> > > tried to do the right thing. Now we're trying to micro correct, break
> > > apart, etc. a state that gets deleted. A reviewer could equally
> > > criticise this being 2 changes rather than 1, and the cognitive load
> > > of having to look at code that gets deleted. At some point it is a
> > > judgement call, and I think this patch is actually the right size. I
> > > think what is missing here is some motivation in the commit message to
> > > the findnew refactoring and so I'll add that.
> >
> > I'm not against your approach and actually appreciate your effort
> > to split rb-tree refactoring and hash table introduction.  What I'm
> > asking is just to separate out the code moving.  I think you can do
> > whatever you want in the current file.  Once you have the final code
> > you can move it to its own file exactly the same.  When I look at this
> > commit, say a few years later, I won't expect a commit that says
> > moving something to a new file has other changes.
>
> The problem is that the code in machine treats the threads lock as if
> it is a lock in machine. So there is __machine__findnew_thread which
> implies the thread lock is held. This change is making threads its own
> separate concept/collection and the lock belongs with that collection.
> Most of the implementation of threads__findnew matches
> __machine__findnew_thread, so we may be able to engineer a smaller
> line diff by moving "__machine__findnew_thread" code into threads.c,
> then renaming it to build the collection, etc. We could also build the
> threads collection inside of machine and then in a separate change
> move it to threads.[ch].  In the commit history this seems muddier
> than just splitting out threads as a collection. Also, some of the API
> design choices are motivated more by the hash table implementation of
> the next patch than trying to have a good rbtree abstracted collection
> of threads. Essentially it'd be engineering a collection of threads
> but only with a view to delete it in the next patch. I don't think it
> would be for the best and the commit history for deleted code is
> unlikely to be looked upon.

I think the conversation is repeating. :)  Why not do this?

1. refactor threads code in machine.c and fix the locking
2. move threads code to its own file
3. use hash table in threads

Thanks,
Namhyung
  
Ian Rogers Feb. 29, 2024, 12:31 a.m. UTC | #10
On Wed, Feb 28, 2024 at 3:44 PM Namhyung Kim <namhyung@kernel.org> wrote:
>
> On Tue, Feb 27, 2024 at 11:24 PM Ian Rogers <irogers@google.com> wrote:
> >
> > On Tue, Feb 27, 2024 at 10:40 PM Namhyung Kim <namhyung@kernel.org> wrote:
> > >
> > > On Tue, Feb 27, 2024 at 1:42 PM Ian Rogers <irogers@google.com> wrote:
> > > >
> > > > On Tue, Feb 27, 2024 at 11:17 AM Arnaldo Carvalho de Melo
> > > > <acme@kernel.org> wrote:
> > > > >
> > > > > On Tue, Feb 27, 2024 at 09:31:33AM -0800, Namhyung Kim wrote:
> > > > > > I can see some other differences like machine__findnew_thread()
> > > > > > which I think is due to the locking change.  Maybe we can fix the
> > > > > > problem before moving the code and let the code move simple.
> > > > >
> > > > > I was going to suggest that, agreed.
> > > > >
> > > > > We may start doing a refactoring, then find a bug, at that point we
> > > > > first fix the problem them go back to refactoring.
> > > >
> > > > Sure I do this all the time. Your typical complaint on the v+1 patch
> > > > set is to move the bug fixes to the front of the changes. On the v+2
> > > > patch set the bug fixes get applied but not the rest of the patch
> > > > series, etc.
> > > >
> > > > Here we are refactoring code for an rb-tree implementation of threads
> > > > and worrying about its correctness. There's no indication it's not
> > > > correct, it is largely copy and paste, there is also good evidence in
> > > > the locking disciple it is more correct. The next patch deletes that
> > > > implementation, replacing it with a hash table. Were I not trying to
> > > > break things apart I could squash those 2 patches together, but I've
> > > > tried to do the right thing. Now we're trying to micro correct, break
> > > > apart, etc. a state that gets deleted. A reviewer could equally
> > > > criticise this being 2 changes rather than 1, and the cognitive load
> > > > of having to look at code that gets deleted. At some point it is a
> > > > judgement call, and I think this patch is actually the right size. I
> > > > think what is missing here is some motivation in the commit message to
> > > > the findnew refactoring and so I'll add that.
> > >
> > > I'm not against your approach and actually appreciate your effort
> > > to split rb-tree refactoring and hash table introduction.  What I'm
> > > asking is just to separate out the code moving.  I think you can do
> > > whatever you want in the current file.  Once you have the final code
> > > you can move it to its own file exactly the same.  When I look at this
> > > commit, say a few years later, I won't expect a commit that says
> > > moving something to a new file has other changes.
> >
> > The problem is that the code in machine treats the threads lock as if
> > it is a lock in machine. So there is __machine__findnew_thread which
> > implies the thread lock is held. This change is making threads its own
> > separate concept/collection and the lock belongs with that collection.
> > Most of the implementation of threads__findnew matches
> > __machine__findnew_thread, so we may be able to engineer a smaller
> > line diff by moving "__machine__findnew_thread" code into threads.c,
> > then renaming it to build the collection, etc. We could also build the
> > threads collection inside of machine and then in a separate change
> > move it to threads.[ch].  In the commit history this seems muddier
> > than just splitting out threads as a collection. Also, some of the API
> > design choices are motivated more by the hash table implementation of
> > the next patch than trying to have a good rbtree abstracted collection
> > of threads. Essentially it'd be engineering a collection of threads
> > but only with a view to delete it in the next patch. I don't think it
> > would be for the best and the commit history for deleted code is
> > unlikely to be looked upon.
>
> I think the conversation is repeating. :)  Why not do this?
>
> 1. refactor threads code in machine.c and fix the locking
> 2. move threads code to its own file
> 3. use hash table in threads

Step 3 is patch 5 in this series. I can split this patch into steps 1
and 2 as you say in v3. I'm not sure why step 2 is going to be of any
use as the code will match that of step 1, code that the next patch
then deletes. Anyway, it's a plan for v3.

Thanks,
Ian

> Thanks,
> Namhyung
  
David Laight Feb. 29, 2024, 9:59 p.m. UTC | #11
From: Ian Rogers
> Sent: 27 February 2024 07:24
> 
> On Mon, Feb 26, 2024 at 11:07 PM Namhyung Kim <namhyung@kernel.org> wrote:
> >
> > On Tue, Feb 13, 2024 at 10:37 PM Ian Rogers <irogers@google.com> wrote:
> > >
> > > Move threads out of machine and move thread_rb_node into the C
> > > file. This hides the implementation of threads from the rest of the
> > > code allowing for it to be refactored.
> > >
> > > Locking discipline is tightened up in this change.
> >
> > Doesn't look like a simple code move.  Can we split the locking
> > change from the move to make the reviewer's life a bit easier? :)
> 
> Not sure I follow. Take threads_nr as an example.
> 
> The old code is in machine.c, so:
> -static size_t machine__threads_nr(const struct machine *machine)
> -{
> -       size_t nr = 0;
> -
> -       for (int i = 0; i < THREADS__TABLE_SIZE; i++)
> -               nr += machine->threads[i].nr;
> -
> -       return nr;
> -}
> 
> The new code is in threads.c:
> +size_t threads__nr(struct threads *threads)
> +{
> +       size_t nr = 0;
> +
> +       for (int i = 0; i < THREADS__TABLE_SIZE; i++) {
> +               struct threads_table_entry *table = &threads->table[i];
> +
> +               down_read(&table->lock);
> +               nr += table->nr;
> +               up_read(&table->lock);
> +       }
> +       return nr;
> +}
> 
> So it is a copy paste from one file to the other. The only difference
> is that the old code failed to take a lock when reading "nr" so the
> locking is added. I wanted to make sure all the functions in threads.c
> were properly correct wrt locking, semaphore creation and destruction,
> etc.  We could have a broken threads.c and fix it in the next change,
> but given that's a bug it could make bisection more difficult.
> Ultimately I thought the locking changes were small enough to not
> warrant being on their own compared to the advantages of having a sane
> threads abstraction.

The lock is pretty much entirely pointless.
All it really does is slow the code down.
The most you could want is:
	nr += READ_ONCE(table->nr);
to avoid any hypothetical data tearing.

	David

-
Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK
Registration No: 1397386 (Wales)
  
Ian Rogers March 1, 2024, 12:19 a.m. UTC | #12
On Thu, Feb 29, 2024 at 1:59 PM David Laight <David.Laight@aculab.com> wrote:
>
> From: Ian Rogers
> > Sent: 27 February 2024 07:24
> >
> > On Mon, Feb 26, 2024 at 11:07 PM Namhyung Kim <namhyung@kernel.org> wrote:
> > >
> > > On Tue, Feb 13, 2024 at 10:37 PM Ian Rogers <irogers@google.com> wrote:
> > > >
> > > > Move threads out of machine and move thread_rb_node into the C
> > > > file. This hides the implementation of threads from the rest of the
> > > > code allowing for it to be refactored.
> > > >
> > > > Locking discipline is tightened up in this change.
> > >
> > > Doesn't look like a simple code move.  Can we split the locking
> > > change from the move to make the reviewer's life a bit easier? :)
> >
> > Not sure I follow. Take threads_nr as an example.
> >
> > The old code is in machine.c, so:
> > -static size_t machine__threads_nr(const struct machine *machine)
> > -{
> > -       size_t nr = 0;
> > -
> > -       for (int i = 0; i < THREADS__TABLE_SIZE; i++)
> > -               nr += machine->threads[i].nr;
> > -
> > -       return nr;
> > -}
> >
> > The new code is in threads.c:
> > +size_t threads__nr(struct threads *threads)
> > +{
> > +       size_t nr = 0;
> > +
> > +       for (int i = 0; i < THREADS__TABLE_SIZE; i++) {
> > +               struct threads_table_entry *table = &threads->table[i];
> > +
> > +               down_read(&table->lock);
> > +               nr += table->nr;
> > +               up_read(&table->lock);
> > +       }
> > +       return nr;
> > +}
> >
> > So it is a copy paste from one file to the other. The only difference
> > is that the old code failed to take a lock when reading "nr" so the
> > locking is added. I wanted to make sure all the functions in threads.c
> > were properly correct wrt locking, semaphore creation and destruction,
> > etc.  We could have a broken threads.c and fix it in the next change,
> > but given that's a bug it could make bisection more difficult.
> > Ultimately I thought the locking changes were small enough to not
> > warrant being on their own compared to the advantages of having a sane
> > threads abstraction.
>
> The lock is pretty much entirely pointless.
> All it really does is slow the code down.
> The most you could want is:
>         nr += READ_ONCE(table->nr);
> to avoid any hypothetical data tearing.

Completely agreed, but as this is user space code I'm unclear on
thread sanitizer support for READ_ONCE. The code is also only called
in debug statements. The migration to a hashmap in the later patches
means the complexity of taking the lock is more justified, although
we're using libbpf's hashmap that has a size variable.

Thanks,
Ian

>         David
>
> -
> Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK
> Registration No: 1397386 (Wales)
  

Patch

diff --git a/tools/perf/util/Build b/tools/perf/util/Build
index 8027f450fa3e..a0e8cd68d490 100644
--- a/tools/perf/util/Build
+++ b/tools/perf/util/Build
@@ -71,6 +71,7 @@  perf-y += ordered-events.o
 perf-y += namespaces.o
 perf-y += comm.o
 perf-y += thread.o
+perf-y += threads.o
 perf-y += thread_map.o
 perf-y += parse-events-flex.o
 perf-y += parse-events-bison.o
diff --git a/tools/perf/util/bpf_lock_contention.c b/tools/perf/util/bpf_lock_contention.c
index 31ff19afc20c..3992c8a9fd96 100644
--- a/tools/perf/util/bpf_lock_contention.c
+++ b/tools/perf/util/bpf_lock_contention.c
@@ -210,7 +210,7 @@  static const char *lock_contention_get_name(struct lock_contention *con,
 
 		/* do not update idle comm which contains CPU number */
 		if (pid) {
-			struct thread *t = __machine__findnew_thread(machine, /*pid=*/-1, pid);
+			struct thread *t = machine__findnew_thread(machine, /*pid=*/-1, pid);
 
 			if (t == NULL)
 				return name;
@@ -302,9 +302,9 @@  int lock_contention_read(struct lock_contention *con)
 		return -1;
 
 	if (con->aggr_mode == LOCK_AGGR_TASK) {
-		struct thread *idle = __machine__findnew_thread(machine,
-								/*pid=*/0,
-								/*tid=*/0);
+		struct thread *idle = machine__findnew_thread(machine,
+							      /*pid=*/0,
+							      /*tid=*/0);
 		thread__set_comm(idle, "swapper", /*timestamp=*/0);
 	}
 
diff --git a/tools/perf/util/machine.c b/tools/perf/util/machine.c
index e072b2115b64..e668a97255f8 100644
--- a/tools/perf/util/machine.c
+++ b/tools/perf/util/machine.c
@@ -43,9 +43,6 @@ 
 #include <linux/string.h>
 #include <linux/zalloc.h>
 
-static void __machine__remove_thread(struct machine *machine, struct thread_rb_node *nd,
-				     struct thread *th, bool lock);
-
 static struct dso *machine__kernel_dso(struct machine *machine)
 {
 	return map__dso(machine->vmlinux_map);
@@ -58,35 +55,6 @@  static void dsos__init(struct dsos *dsos)
 	init_rwsem(&dsos->lock);
 }
 
-static void machine__threads_init(struct machine *machine)
-{
-	int i;
-
-	for (i = 0; i < THREADS__TABLE_SIZE; i++) {
-		struct threads *threads = &machine->threads[i];
-		threads->entries = RB_ROOT_CACHED;
-		init_rwsem(&threads->lock);
-		threads->nr = 0;
-		threads->last_match = NULL;
-	}
-}
-
-static int thread_rb_node__cmp_tid(const void *key, const struct rb_node *nd)
-{
-	int to_find = (int) *((pid_t *)key);
-
-	return to_find - (int)thread__tid(rb_entry(nd, struct thread_rb_node, rb_node)->thread);
-}
-
-static struct thread_rb_node *thread_rb_node__find(const struct thread *th,
-						   struct rb_root *tree)
-{
-	pid_t to_find = thread__tid(th);
-	struct rb_node *nd = rb_find(&to_find, tree, thread_rb_node__cmp_tid);
-
-	return rb_entry(nd, struct thread_rb_node, rb_node);
-}
-
 static int machine__set_mmap_name(struct machine *machine)
 {
 	if (machine__is_host(machine))
@@ -120,7 +88,7 @@  int machine__init(struct machine *machine, const char *root_dir, pid_t pid)
 	RB_CLEAR_NODE(&machine->rb_node);
 	dsos__init(&machine->dsos);
 
-	machine__threads_init(machine);
+	threads__init(&machine->threads);
 
 	machine->vdso_info = NULL;
 	machine->env = NULL;
@@ -221,27 +189,11 @@  static void dsos__exit(struct dsos *dsos)
 
 void machine__delete_threads(struct machine *machine)
 {
-	struct rb_node *nd;
-	int i;
-
-	for (i = 0; i < THREADS__TABLE_SIZE; i++) {
-		struct threads *threads = &machine->threads[i];
-		down_write(&threads->lock);
-		nd = rb_first_cached(&threads->entries);
-		while (nd) {
-			struct thread_rb_node *trb = rb_entry(nd, struct thread_rb_node, rb_node);
-
-			nd = rb_next(nd);
-			__machine__remove_thread(machine, trb, trb->thread, false);
-		}
-		up_write(&threads->lock);
-	}
+	threads__remove_all_threads(&machine->threads);
 }
 
 void machine__exit(struct machine *machine)
 {
-	int i;
-
 	if (machine == NULL)
 		return;
 
@@ -254,12 +206,7 @@  void machine__exit(struct machine *machine)
 	zfree(&machine->current_tid);
 	zfree(&machine->kallsyms_filename);
 
-	machine__delete_threads(machine);
-	for (i = 0; i < THREADS__TABLE_SIZE; i++) {
-		struct threads *threads = &machine->threads[i];
-
-		exit_rwsem(&threads->lock);
-	}
+	threads__exit(&machine->threads);
 }
 
 void machine__delete(struct machine *machine)
@@ -526,7 +473,7 @@  static void machine__update_thread_pid(struct machine *machine,
 	if (thread__pid(th) == thread__tid(th))
 		return;
 
-	leader = __machine__findnew_thread(machine, thread__pid(th), thread__pid(th));
+	leader = machine__findnew_thread(machine, thread__pid(th), thread__pid(th));
 	if (!leader)
 		goto out_err;
 
@@ -560,160 +507,55 @@  static void machine__update_thread_pid(struct machine *machine,
 	goto out_put;
 }
 
-/*
- * Front-end cache - TID lookups come in blocks,
- * so most of the time we dont have to look up
- * the full rbtree:
- */
-static struct thread*
-__threads__get_last_match(struct threads *threads, struct machine *machine,
-			  int pid, int tid)
-{
-	struct thread *th;
-
-	th = threads->last_match;
-	if (th != NULL) {
-		if (thread__tid(th) == tid) {
-			machine__update_thread_pid(machine, th, pid);
-			return thread__get(th);
-		}
-		thread__put(threads->last_match);
-		threads->last_match = NULL;
-	}
-
-	return NULL;
-}
-
-static struct thread*
-threads__get_last_match(struct threads *threads, struct machine *machine,
-			int pid, int tid)
-{
-	struct thread *th = NULL;
-
-	if (perf_singlethreaded)
-		th = __threads__get_last_match(threads, machine, pid, tid);
-
-	return th;
-}
-
-static void
-__threads__set_last_match(struct threads *threads, struct thread *th)
-{
-	thread__put(threads->last_match);
-	threads->last_match = thread__get(th);
-}
-
-static void
-threads__set_last_match(struct threads *threads, struct thread *th)
-{
-	if (perf_singlethreaded)
-		__threads__set_last_match(threads, th);
-}
-
 /*
  * Caller must eventually drop thread->refcnt returned with a successful
  * lookup/new thread inserted.
  */
-static struct thread *____machine__findnew_thread(struct machine *machine,
-						  struct threads *threads,
-						  pid_t pid, pid_t tid,
-						  bool create)
+static struct thread *__machine__findnew_thread(struct machine *machine,
+						pid_t pid,
+						pid_t tid,
+						bool create)
 {
-	struct rb_node **p = &threads->entries.rb_root.rb_node;
-	struct rb_node *parent = NULL;
-	struct thread *th;
-	struct thread_rb_node *nd;
-	bool leftmost = true;
+	struct thread *th = threads__find(&machine->threads, tid);
+	bool created;
 
-	th = threads__get_last_match(threads, machine, pid, tid);
-	if (th)
+	if (th) {
+		machine__update_thread_pid(machine, th, pid);
 		return th;
-
-	while (*p != NULL) {
-		parent = *p;
-		th = rb_entry(parent, struct thread_rb_node, rb_node)->thread;
-
-		if (thread__tid(th) == tid) {
-			threads__set_last_match(threads, th);
-			machine__update_thread_pid(machine, th, pid);
-			return thread__get(th);
-		}
-
-		if (tid < thread__tid(th))
-			p = &(*p)->rb_left;
-		else {
-			p = &(*p)->rb_right;
-			leftmost = false;
-		}
 	}
-
 	if (!create)
 		return NULL;
 
-	th = thread__new(pid, tid);
-	if (th == NULL)
-		return NULL;
-
-	nd = malloc(sizeof(*nd));
-	if (nd == NULL) {
-		thread__put(th);
-		return NULL;
-	}
-	nd->thread = th;
-
-	rb_link_node(&nd->rb_node, parent, p);
-	rb_insert_color_cached(&nd->rb_node, &threads->entries, leftmost);
-	/*
-	 * We have to initialize maps separately after rb tree is updated.
-	 *
-	 * The reason is that we call machine__findnew_thread within
-	 * thread__init_maps to find the thread leader and that would screwed
-	 * the rb tree.
-	 */
-	if (thread__init_maps(th, machine)) {
-		pr_err("Thread init failed thread %d\n", pid);
-		rb_erase_cached(&nd->rb_node, &threads->entries);
-		RB_CLEAR_NODE(&nd->rb_node);
-		free(nd);
-		thread__put(th);
-		return NULL;
-	}
-	/*
-	 * It is now in the rbtree, get a ref
-	 */
-	threads__set_last_match(threads, th);
-	++threads->nr;
-
-	return thread__get(th);
-}
+	th = threads__findnew(&machine->threads, pid, tid, &created);
+	if (created) {
+		/*
+		 * We have to initialize maps separately after rb tree is
+		 * updated.
+		 *
+		 * The reason is that we call machine__findnew_thread within
+		 * thread__init_maps to find the thread leader and that would
+		 * screwed the rb tree.
+		 */
+		if (thread__init_maps(th, machine)) {
+			pr_err("Thread init failed thread %d\n", pid);
+			threads__remove(&machine->threads, th);
+			thread__put(th);
+			return NULL;
+		}
+	} else
+		machine__update_thread_pid(machine, th, pid);
 
-struct thread *__machine__findnew_thread(struct machine *machine, pid_t pid, pid_t tid)
-{
-	return ____machine__findnew_thread(machine, machine__threads(machine, tid), pid, tid, true);
+	return th;
 }
 
-struct thread *machine__findnew_thread(struct machine *machine, pid_t pid,
-				       pid_t tid)
+struct thread *machine__findnew_thread(struct machine *machine, pid_t pid, pid_t tid)
 {
-	struct threads *threads = machine__threads(machine, tid);
-	struct thread *th;
-
-	down_write(&threads->lock);
-	th = __machine__findnew_thread(machine, pid, tid);
-	up_write(&threads->lock);
-	return th;
+	return __machine__findnew_thread(machine, pid, tid, /*create=*/true);
 }
 
-struct thread *machine__find_thread(struct machine *machine, pid_t pid,
-				    pid_t tid)
+struct thread *machine__find_thread(struct machine *machine, pid_t pid, pid_t tid)
 {
-	struct threads *threads = machine__threads(machine, tid);
-	struct thread *th;
-
-	down_read(&threads->lock);
-	th =  ____machine__findnew_thread(machine, threads, pid, tid, false);
-	up_read(&threads->lock);
-	return th;
+	return __machine__findnew_thread(machine, pid, tid, /*create=*/false);
 }
 
 /*
@@ -1127,23 +969,13 @@  static int machine_fprintf_cb(struct thread *thread, void *data)
 	return 0;
 }
 
-static size_t machine__threads_nr(const struct machine *machine)
-{
-	size_t nr = 0;
-
-	for (int i = 0; i < THREADS__TABLE_SIZE; i++)
-		nr += machine->threads[i].nr;
-
-	return nr;
-}
-
 size_t machine__fprintf(struct machine *machine, FILE *fp)
 {
 	struct machine_fprintf_cb_args args = {
 		.fp = fp,
 		.printed = 0,
 	};
-	size_t ret = fprintf(fp, "Threads: %zu\n", machine__threads_nr(machine));
+	size_t ret = fprintf(fp, "Threads: %zu\n", threads__nr(&machine->threads));
 
 	machine__for_each_thread(machine, machine_fprintf_cb, &args);
 	return ret + args.printed;
@@ -2069,36 +1901,9 @@  int machine__process_mmap_event(struct machine *machine, union perf_event *event
 	return 0;
 }
 
-static void __machine__remove_thread(struct machine *machine, struct thread_rb_node *nd,
-				     struct thread *th, bool lock)
-{
-	struct threads *threads = machine__threads(machine, thread__tid(th));
-
-	if (!nd)
-		nd = thread_rb_node__find(th, &threads->entries.rb_root);
-
-	if (threads->last_match && RC_CHK_EQUAL(threads->last_match, th))
-		threads__set_last_match(threads, NULL);
-
-	if (lock)
-		down_write(&threads->lock);
-
-	BUG_ON(refcount_read(thread__refcnt(th)) == 0);
-
-	thread__put(nd->thread);
-	rb_erase_cached(&nd->rb_node, &threads->entries);
-	RB_CLEAR_NODE(&nd->rb_node);
-	--threads->nr;
-
-	free(nd);
-
-	if (lock)
-		up_write(&threads->lock);
-}
-
 void machine__remove_thread(struct machine *machine, struct thread *th)
 {
-	return __machine__remove_thread(machine, NULL, th, true);
+	return threads__remove(&machine->threads, th);
 }
 
 int machine__process_fork_event(struct machine *machine, union perf_event *event,
@@ -3232,23 +3037,7 @@  int machine__for_each_thread(struct machine *machine,
 			     int (*fn)(struct thread *thread, void *p),
 			     void *priv)
 {
-	struct threads *threads;
-	struct rb_node *nd;
-	int rc = 0;
-	int i;
-
-	for (i = 0; i < THREADS__TABLE_SIZE; i++) {
-		threads = &machine->threads[i];
-		for (nd = rb_first_cached(&threads->entries); nd;
-		     nd = rb_next(nd)) {
-			struct thread_rb_node *trb = rb_entry(nd, struct thread_rb_node, rb_node);
-
-			rc = fn(trb->thread, priv);
-			if (rc != 0)
-				return rc;
-		}
-	}
-	return rc;
+	return threads__for_each_thread(&machine->threads, fn, priv);
 }
 
 int machines__for_each_thread(struct machines *machines,
diff --git a/tools/perf/util/machine.h b/tools/perf/util/machine.h
index b738ce84817b..e28c787616fe 100644
--- a/tools/perf/util/machine.h
+++ b/tools/perf/util/machine.h
@@ -7,6 +7,7 @@ 
 #include "maps.h"
 #include "dsos.h"
 #include "rwsem.h"
+#include "threads.h"
 
 struct addr_location;
 struct branch_stack;
@@ -28,16 +29,6 @@  extern const char *ref_reloc_sym_names[];
 
 struct vdso_info;
 
-#define THREADS__TABLE_BITS	8
-#define THREADS__TABLE_SIZE	(1 << THREADS__TABLE_BITS)
-
-struct threads {
-	struct rb_root_cached  entries;
-	struct rw_semaphore    lock;
-	unsigned int	       nr;
-	struct thread	       *last_match;
-};
-
 struct machine {
 	struct rb_node	  rb_node;
 	pid_t		  pid;
@@ -48,7 +39,7 @@  struct machine {
 	char		  *root_dir;
 	char		  *mmap_name;
 	char		  *kallsyms_filename;
-	struct threads    threads[THREADS__TABLE_SIZE];
+	struct threads    threads;
 	struct vdso_info  *vdso_info;
 	struct perf_env   *env;
 	struct dsos	  dsos;
@@ -69,12 +60,6 @@  struct machine {
 	bool		  trampolines_mapped;
 };
 
-static inline struct threads *machine__threads(struct machine *machine, pid_t tid)
-{
-	/* Cast it to handle tid == -1 */
-	return &machine->threads[(unsigned int)tid % THREADS__TABLE_SIZE];
-}
-
 /*
  * The main kernel (vmlinux) map
  */
@@ -220,7 +205,6 @@  bool machine__is(struct machine *machine, const char *arch);
 bool machine__normalized_is(struct machine *machine, const char *arch);
 int machine__nr_cpus_avail(struct machine *machine);
 
-struct thread *__machine__findnew_thread(struct machine *machine, pid_t pid, pid_t tid);
 struct thread *machine__findnew_thread(struct machine *machine, pid_t pid, pid_t tid);
 
 struct dso *machine__findnew_dso_id(struct machine *machine, const char *filename, struct dso_id *id);
diff --git a/tools/perf/util/thread.c b/tools/perf/util/thread.c
index c59ab4d79163..1aa8962dcf52 100644
--- a/tools/perf/util/thread.c
+++ b/tools/perf/util/thread.c
@@ -26,7 +26,7 @@  int thread__init_maps(struct thread *thread, struct machine *machine)
 	if (pid == thread__tid(thread) || pid == -1) {
 		thread__set_maps(thread, maps__new(machine));
 	} else {
-		struct thread *leader = __machine__findnew_thread(machine, pid, pid);
+		struct thread *leader = machine__findnew_thread(machine, pid, pid);
 
 		if (leader) {
 			thread__set_maps(thread, maps__get(thread__maps(leader)));
diff --git a/tools/perf/util/thread.h b/tools/perf/util/thread.h
index 0df775b5c110..4b8f3e9e513b 100644
--- a/tools/perf/util/thread.h
+++ b/tools/perf/util/thread.h
@@ -3,7 +3,6 @@ 
 #define __PERF_THREAD_H
 
 #include <linux/refcount.h>
-#include <linux/rbtree.h>
 #include <linux/list.h>
 #include <stdio.h>
 #include <unistd.h>
@@ -30,11 +29,6 @@  struct lbr_stitch {
 	struct callchain_cursor_node	*prev_lbr_cursor;
 };
 
-struct thread_rb_node {
-	struct rb_node rb_node;
-	struct thread *thread;
-};
-
 DECLARE_RC_STRUCT(thread) {
 	/** @maps: mmaps associated with this thread. */
 	struct maps		*maps;
diff --git a/tools/perf/util/threads.c b/tools/perf/util/threads.c
new file mode 100644
index 000000000000..d984ec939c7b
--- /dev/null
+++ b/tools/perf/util/threads.c
@@ -0,0 +1,244 @@ 
+// SPDX-License-Identifier: GPL-2.0
+#include "threads.h"
+#include "machine.h"
+#include "thread.h"
+
+struct thread_rb_node {
+	struct rb_node rb_node;
+	struct thread *thread;
+};
+
+static struct threads_table_entry *threads__table(struct threads *threads, pid_t tid)
+{
+	/* Cast it to handle tid == -1 */
+	return &threads->table[(unsigned int)tid % THREADS__TABLE_SIZE];
+}
+
+void threads__init(struct threads *threads)
+{
+	for (int i = 0; i < THREADS__TABLE_SIZE; i++) {
+		struct threads_table_entry *table = &threads->table[i];
+
+		table->entries = RB_ROOT_CACHED;
+		init_rwsem(&table->lock);
+		table->nr = 0;
+		table->last_match = NULL;
+	}
+}
+
+void threads__exit(struct threads *threads)
+{
+	threads__remove_all_threads(threads);
+	for (int i = 0; i < THREADS__TABLE_SIZE; i++) {
+		struct threads_table_entry *table = &threads->table[i];
+
+		exit_rwsem(&table->lock);
+	}
+}
+
+size_t threads__nr(struct threads *threads)
+{
+	size_t nr = 0;
+
+	for (int i = 0; i < THREADS__TABLE_SIZE; i++) {
+		struct threads_table_entry *table = &threads->table[i];
+
+		down_read(&table->lock);
+		nr += table->nr;
+		up_read(&table->lock);
+	}
+	return nr;
+}
+
+/*
+ * Front-end cache - TID lookups come in blocks,
+ * so most of the time we dont have to look up
+ * the full rbtree:
+ */
+static struct thread *__threads_table_entry__get_last_match(struct threads_table_entry *table,
+							    pid_t tid)
+{
+	struct thread *th, *res = NULL;
+
+	th = table->last_match;
+	if (th != NULL) {
+		if (thread__tid(th) == tid)
+			res = thread__get(th);
+	}
+	return res;
+}
+
+static void __threads_table_entry__set_last_match(struct threads_table_entry *table,
+						  struct thread *th)
+{
+	thread__put(table->last_match);
+	table->last_match = thread__get(th);
+}
+
+static void threads_table_entry__set_last_match(struct threads_table_entry *table,
+						struct thread *th)
+{
+	down_write(&table->lock);
+	__threads_table_entry__set_last_match(table, th);
+	up_write(&table->lock);
+}
+
+struct thread *threads__find(struct threads *threads, pid_t tid)
+{
+	struct threads_table_entry *table  = threads__table(threads, tid);
+	struct rb_node **p;
+	struct thread *res = NULL;
+
+	down_read(&table->lock);
+	res = __threads_table_entry__get_last_match(table, tid);
+	if (res)
+		return res;
+
+	p = &table->entries.rb_root.rb_node;
+	while (*p != NULL) {
+		struct rb_node *parent = *p;
+		struct thread *th = rb_entry(parent, struct thread_rb_node, rb_node)->thread;
+
+		if (thread__tid(th) == tid) {
+			res = thread__get(th);
+			break;
+		}
+
+		if (tid < thread__tid(th))
+			p = &(*p)->rb_left;
+		else
+			p = &(*p)->rb_right;
+	}
+	up_read(&table->lock);
+	if (res)
+		threads_table_entry__set_last_match(table, res);
+	return res;
+}
+
+struct thread *threads__findnew(struct threads *threads, pid_t pid, pid_t tid, bool *created)
+{
+	struct threads_table_entry *table  = threads__table(threads, tid);
+	struct rb_node **p;
+	struct rb_node *parent = NULL;
+	struct thread *res = NULL;
+	struct thread_rb_node *nd;
+	bool leftmost = true;
+
+	*created = false;
+	down_write(&table->lock);
+	p = &table->entries.rb_root.rb_node;
+	while (*p != NULL) {
+		struct thread *th;
+
+		parent = *p;
+		th = rb_entry(parent, struct thread_rb_node, rb_node)->thread;
+
+		if (thread__tid(th) == tid) {
+			__threads_table_entry__set_last_match(table, th);
+			res = thread__get(th);
+			goto out_unlock;
+		}
+
+		if (tid < thread__tid(th))
+			p = &(*p)->rb_left;
+		else {
+			leftmost = false;
+			p = &(*p)->rb_right;
+		}
+	}
+	nd = malloc(sizeof(*nd));
+	if (nd == NULL)
+		goto out_unlock;
+	res = thread__new(pid, tid);
+	if (!res)
+		free(nd);
+	else {
+		*created = true;
+		nd->thread = thread__get(res);
+		rb_link_node(&nd->rb_node, parent, p);
+		rb_insert_color_cached(&nd->rb_node, &table->entries, leftmost);
+		++table->nr;
+		__threads_table_entry__set_last_match(table, res);
+	}
+out_unlock:
+	up_write(&table->lock);
+	return res;
+}
+
+void threads__remove_all_threads(struct threads *threads)
+{
+	for (int i = 0; i < THREADS__TABLE_SIZE; i++) {
+		struct threads_table_entry *table = &threads->table[i];
+		struct rb_node *nd;
+
+		down_write(&table->lock);
+		__threads_table_entry__set_last_match(table, NULL);
+		nd = rb_first_cached(&table->entries);
+		while (nd) {
+			struct thread_rb_node *trb = rb_entry(nd, struct thread_rb_node, rb_node);
+
+			nd = rb_next(nd);
+			thread__put(trb->thread);
+			rb_erase_cached(&trb->rb_node, &table->entries);
+			RB_CLEAR_NODE(&trb->rb_node);
+			--table->nr;
+
+			free(trb);
+		}
+		assert(table->nr == 0);
+		up_write(&table->lock);
+	}
+}
+
+void threads__remove(struct threads *threads, struct thread *thread)
+{
+	struct rb_node **p;
+	struct threads_table_entry *table  = threads__table(threads, thread__tid(thread));
+	pid_t tid = thread__tid(thread);
+
+	down_write(&table->lock);
+	if (table->last_match && RC_CHK_EQUAL(table->last_match, thread))
+		__threads_table_entry__set_last_match(table, NULL);
+
+	p = &table->entries.rb_root.rb_node;
+	while (*p != NULL) {
+		struct rb_node *parent = *p;
+		struct thread_rb_node *nd = rb_entry(parent, struct thread_rb_node, rb_node);
+		struct thread *th = nd->thread;
+
+		if (RC_CHK_EQUAL(th, thread)) {
+			thread__put(nd->thread);
+			rb_erase_cached(&nd->rb_node, &table->entries);
+			RB_CLEAR_NODE(&nd->rb_node);
+			--table->nr;
+			free(nd);
+			break;
+		}
+
+		if (tid < thread__tid(th))
+			p = &(*p)->rb_left;
+		else
+			p = &(*p)->rb_right;
+	}
+	up_write(&table->lock);
+}
+
+int threads__for_each_thread(struct threads *threads,
+			     int (*fn)(struct thread *thread, void *data),
+			     void *data)
+{
+	for (int i = 0; i < THREADS__TABLE_SIZE; i++) {
+		struct threads_table_entry *table = &threads->table[i];
+		struct rb_node *nd;
+
+		for (nd = rb_first_cached(&table->entries); nd; nd = rb_next(nd)) {
+			struct thread_rb_node *trb = rb_entry(nd, struct thread_rb_node, rb_node);
+			int rc = fn(trb->thread, data);
+
+			if (rc != 0)
+				return rc;
+		}
+	}
+	return 0;
+
+}
diff --git a/tools/perf/util/threads.h b/tools/perf/util/threads.h
new file mode 100644
index 000000000000..ed67de627578
--- /dev/null
+++ b/tools/perf/util/threads.h
@@ -0,0 +1,35 @@ 
+/* SPDX-License-Identifier: GPL-2.0 */
+#ifndef __PERF_THREADS_H
+#define __PERF_THREADS_H
+
+#include <linux/rbtree.h>
+#include "rwsem.h"
+
+struct thread;
+
+#define THREADS__TABLE_BITS	8
+#define THREADS__TABLE_SIZE	(1 << THREADS__TABLE_BITS)
+
+struct threads_table_entry {
+	struct rb_root_cached  entries;
+	struct rw_semaphore    lock;
+	unsigned int	       nr;
+	struct thread	       *last_match;
+};
+
+struct threads {
+	struct threads_table_entry table[THREADS__TABLE_SIZE];
+};
+
+void threads__init(struct threads *threads);
+void threads__exit(struct threads *threads);
+size_t threads__nr(struct threads *threads);
+struct thread *threads__find(struct threads *threads, pid_t tid);
+struct thread *threads__findnew(struct threads *threads, pid_t pid, pid_t tid, bool *created);
+void threads__remove_all_threads(struct threads *threads);
+void threads__remove(struct threads *threads, struct thread *thread);
+int threads__for_each_thread(struct threads *threads,
+			     int (*fn)(struct thread *thread, void *data),
+			     void *data);
+
+#endif	/* __PERF_THREADS_H */