opensbi: Upgrade from v0.4 to v0.5
[qemu/ar7.git] / util / qsp.c
blob62265417fd7dfd3b709ee3b52e7234499f039572
1 /*
2 * qsp.c - QEMU Synchronization Profiler
4 * Copyright (C) 2018, Emilio G. Cota <cota@braap.org>
6 * License: GNU GPL, version 2 or later.
7 * See the COPYING file in the top-level directory.
9 * QSP profiles the time spent in synchronization primitives, which can
10 * help diagnose performance problems, e.g. scalability issues when
11 * contention is high.
13 * The primitives currently supported are mutexes, recursive mutexes and
14 * condition variables. Note that not all related functions are intercepted;
15 * instead we profile only those functions that can have a performance impact,
16 * either due to blocking (e.g. cond_wait, mutex_lock) or cache line
17 * contention (e.g. mutex_lock, mutex_trylock).
19 * QSP's design focuses on speed and scalability. This is achieved
20 * by having threads do their profiling entirely on thread-local data.
21 * The appropriate thread-local data is found via a QHT, i.e. a concurrent hash
22 * table. To aggregate data in order to generate a report, we iterate over
23 * all entries in the hash table. Depending on the number of threads and
24 * synchronization objects this might be expensive, but note that it is
25 * very rarely called -- reports are generated only when requested by users.
27 * Reports are generated as a table where each row represents a call site. A
28 * call site is the triplet formed by the __file__ and __LINE__ of the caller
29 * as well as the address of the "object" (i.e. mutex, rec. mutex or condvar)
30 * being operated on. Optionally, call sites that operate on different objects
31 * of the same type can be coalesced, which can be particularly useful when
32 * profiling dynamically-allocated objects.
34 * Alternative designs considered:
36 * - Use an off-the-shelf profiler such as mutrace. This is not a viable option
37 * for us because QEMU has __malloc_hook set (by one of the libraries it
38 * uses); leaving this hook unset is required to avoid deadlock in mutrace.
40 * - Use a glib HT for each thread, protecting each HT with its own lock.
41 * This isn't simpler than the current design, and is 10% slower in the
42 * atomic_add-bench microbenchmark (-m option).
44 * - For reports, just use a binary tree as we aggregate data, instead of having
45 * an intermediate hash table. This would simplify the code only slightly, but
46 * would perform badly if there were many threads and objects to track.
48 * - Wrap operations on qsp entries with RCU read-side critical sections, so
49 * that qsp_reset() can delete entries. Unfortunately, the overhead of calling
50 * rcu_read_lock/unlock slows down atomic_add-bench -m by 24%. Having
51 * a snapshot that is updated on qsp_reset() avoids this overhead.
53 * Related Work:
54 * - Lennart Poettering's mutrace: http://0pointer.de/blog/projects/mutrace.html
55 * - Lozi, David, Thomas, Lawall and Muller. "Remote Core Locking: Migrating
56 * Critical-Section Execution to Improve the Performance of Multithreaded
57 * Applications", USENIX ATC'12.
60 #include "qemu/osdep.h"
61 #include "qemu/qemu-print.h"
62 #include "qemu/thread.h"
63 #include "qemu/timer.h"
64 #include "qemu/qht.h"
65 #include "qemu/rcu.h"
66 #include "qemu/xxhash.h"
68 enum QSPType {
69 QSP_MUTEX,
70 QSP_BQL_MUTEX,
71 QSP_REC_MUTEX,
72 QSP_CONDVAR,
75 struct QSPCallSite {
76 const void *obj;
77 const char *file; /* i.e. __FILE__; shortened later */
78 int line;
79 enum QSPType type;
81 typedef struct QSPCallSite QSPCallSite;
83 struct QSPEntry {
84 void *thread_ptr;
85 const QSPCallSite *callsite;
86 uint64_t n_acqs;
87 uint64_t ns;
88 unsigned int n_objs; /* count of coalesced objs; only used for reporting */
90 typedef struct QSPEntry QSPEntry;
92 struct QSPSnapshot {
93 struct rcu_head rcu;
94 struct qht ht;
96 typedef struct QSPSnapshot QSPSnapshot;
98 /* initial sizing for hash tables */
99 #define QSP_INITIAL_SIZE 64
101 /* If this file is moved, QSP_REL_PATH should be updated accordingly */
102 #define QSP_REL_PATH "util/qsp.c"
104 /* this file's full path. Used to present all call sites with relative paths */
105 static size_t qsp_qemu_path_len;
107 /* the address of qsp_thread gives us a unique 'thread ID' */
108 static __thread int qsp_thread;
111 * Call sites are the same for all threads, so we track them in a separate hash
112 * table to save memory.
114 static struct qht qsp_callsite_ht;
116 static struct qht qsp_ht;
117 static QSPSnapshot *qsp_snapshot;
118 static bool qsp_initialized, qsp_initializing;
120 static const char * const qsp_typenames[] = {
121 [QSP_MUTEX] = "mutex",
122 [QSP_BQL_MUTEX] = "BQL mutex",
123 [QSP_REC_MUTEX] = "rec_mutex",
124 [QSP_CONDVAR] = "condvar",
127 QemuMutexLockFunc qemu_bql_mutex_lock_func = qemu_mutex_lock_impl;
128 QemuMutexLockFunc qemu_mutex_lock_func = qemu_mutex_lock_impl;
129 QemuMutexTrylockFunc qemu_mutex_trylock_func = qemu_mutex_trylock_impl;
130 QemuRecMutexLockFunc qemu_rec_mutex_lock_func = qemu_rec_mutex_lock_impl;
131 QemuRecMutexTrylockFunc qemu_rec_mutex_trylock_func =
132 qemu_rec_mutex_trylock_impl;
133 QemuCondWaitFunc qemu_cond_wait_func = qemu_cond_wait_impl;
134 QemuCondTimedWaitFunc qemu_cond_timedwait_func = qemu_cond_timedwait_impl;
137 * It pays off to _not_ hash callsite->file; hashing a string is slow, and
138 * without it we still get a pretty unique hash.
140 static inline
141 uint32_t do_qsp_callsite_hash(const QSPCallSite *callsite, uint64_t ab)
143 uint64_t cd = (uint64_t)(uintptr_t)callsite->obj;
144 uint32_t e = callsite->line;
145 uint32_t f = callsite->type;
147 return qemu_xxhash6(ab, cd, e, f);
150 static inline
151 uint32_t qsp_callsite_hash(const QSPCallSite *callsite)
153 return do_qsp_callsite_hash(callsite, 0);
156 static inline uint32_t do_qsp_entry_hash(const QSPEntry *entry, uint64_t a)
158 return do_qsp_callsite_hash(entry->callsite, a);
161 static uint32_t qsp_entry_hash(const QSPEntry *entry)
163 return do_qsp_entry_hash(entry, (uint64_t)(uintptr_t)entry->thread_ptr);
166 static uint32_t qsp_entry_no_thread_hash(const QSPEntry *entry)
168 return do_qsp_entry_hash(entry, 0);
171 /* without the objects we need to hash the file name to get a decent hash */
172 static uint32_t qsp_entry_no_thread_obj_hash(const QSPEntry *entry)
174 const QSPCallSite *callsite = entry->callsite;
175 uint64_t ab = g_str_hash(callsite->file);
176 uint64_t cd = callsite->line;
177 uint32_t e = callsite->type;
179 return qemu_xxhash5(ab, cd, e);
182 static bool qsp_callsite_cmp(const void *ap, const void *bp)
184 const QSPCallSite *a = ap;
185 const QSPCallSite *b = bp;
187 return a == b ||
188 (a->obj == b->obj &&
189 a->line == b->line &&
190 a->type == b->type &&
191 (a->file == b->file || !strcmp(a->file, b->file)));
194 static bool qsp_callsite_no_obj_cmp(const void *ap, const void *bp)
196 const QSPCallSite *a = ap;
197 const QSPCallSite *b = bp;
199 return a == b ||
200 (a->line == b->line &&
201 a->type == b->type &&
202 (a->file == b->file || !strcmp(a->file, b->file)));
205 static bool qsp_entry_no_thread_cmp(const void *ap, const void *bp)
207 const QSPEntry *a = ap;
208 const QSPEntry *b = bp;
210 return qsp_callsite_cmp(a->callsite, b->callsite);
213 static bool qsp_entry_no_thread_obj_cmp(const void *ap, const void *bp)
215 const QSPEntry *a = ap;
216 const QSPEntry *b = bp;
218 return qsp_callsite_no_obj_cmp(a->callsite, b->callsite);
221 static bool qsp_entry_cmp(const void *ap, const void *bp)
223 const QSPEntry *a = ap;
224 const QSPEntry *b = bp;
226 return a->thread_ptr == b->thread_ptr &&
227 qsp_callsite_cmp(a->callsite, b->callsite);
231 * Normally we'd call this from a constructor function, but we want it to work
232 * via libutil as well.
234 static void qsp_do_init(void)
236 /* make sure this file's path in the tree is up to date with QSP_REL_PATH */
237 g_assert(strstr(__FILE__, QSP_REL_PATH));
238 qsp_qemu_path_len = strlen(__FILE__) - strlen(QSP_REL_PATH);
240 qht_init(&qsp_ht, qsp_entry_cmp, QSP_INITIAL_SIZE,
241 QHT_MODE_AUTO_RESIZE | QHT_MODE_RAW_MUTEXES);
242 qht_init(&qsp_callsite_ht, qsp_callsite_cmp, QSP_INITIAL_SIZE,
243 QHT_MODE_AUTO_RESIZE | QHT_MODE_RAW_MUTEXES);
246 static __attribute__((noinline)) void qsp_init__slowpath(void)
248 if (atomic_cmpxchg(&qsp_initializing, false, true) == false) {
249 qsp_do_init();
250 atomic_set(&qsp_initialized, true);
251 } else {
252 while (!atomic_read(&qsp_initialized)) {
253 cpu_relax();
258 /* qsp_init() must be called from _all_ exported functions */
259 static inline void qsp_init(void)
261 if (likely(atomic_read(&qsp_initialized))) {
262 return;
264 qsp_init__slowpath();
267 static QSPCallSite *qsp_callsite_find(const QSPCallSite *orig)
269 QSPCallSite *callsite;
270 uint32_t hash;
272 hash = qsp_callsite_hash(orig);
273 callsite = qht_lookup(&qsp_callsite_ht, orig, hash);
274 if (callsite == NULL) {
275 void *existing = NULL;
277 callsite = g_new(QSPCallSite, 1);
278 memcpy(callsite, orig, sizeof(*callsite));
279 qht_insert(&qsp_callsite_ht, callsite, hash, &existing);
280 if (unlikely(existing)) {
281 g_free(callsite);
282 callsite = existing;
285 return callsite;
288 static QSPEntry *
289 qsp_entry_create(struct qht *ht, const QSPEntry *entry, uint32_t hash)
291 QSPEntry *e;
292 void *existing = NULL;
294 e = g_new0(QSPEntry, 1);
295 e->thread_ptr = entry->thread_ptr;
296 e->callsite = qsp_callsite_find(entry->callsite);
298 qht_insert(ht, e, hash, &existing);
299 if (unlikely(existing)) {
300 g_free(e);
301 e = existing;
303 return e;
306 static QSPEntry *
307 qsp_entry_find(struct qht *ht, const QSPEntry *entry, uint32_t hash)
309 QSPEntry *e;
311 e = qht_lookup(ht, entry, hash);
312 if (e == NULL) {
313 e = qsp_entry_create(ht, entry, hash);
315 return e;
319 * Note: Entries are never removed, so callers do not have to be in an RCU
320 * read-side critical section.
322 static QSPEntry *qsp_entry_get(const void *obj, const char *file, int line,
323 enum QSPType type)
325 QSPCallSite callsite = {
326 .obj = obj,
327 .file = file,
328 .line = line,
329 .type = type,
331 QSPEntry orig;
332 uint32_t hash;
334 qsp_init();
336 orig.thread_ptr = &qsp_thread;
337 orig.callsite = &callsite;
339 hash = qsp_entry_hash(&orig);
340 return qsp_entry_find(&qsp_ht, &orig, hash);
344 * @e is in the global hash table; it is only written to by the current thread,
345 * so we write to it atomically (as in "write once") to prevent torn reads.
347 static inline void do_qsp_entry_record(QSPEntry *e, int64_t delta, bool acq)
349 atomic_set_u64(&e->ns, e->ns + delta);
350 if (acq) {
351 atomic_set_u64(&e->n_acqs, e->n_acqs + 1);
355 static inline void qsp_entry_record(QSPEntry *e, int64_t delta)
357 do_qsp_entry_record(e, delta, true);
360 #define QSP_GEN_VOID(type_, qsp_t_, func_, impl_) \
361 static void func_(type_ *obj, const char *file, int line) \
363 QSPEntry *e; \
364 int64_t t0, t1; \
366 t0 = get_clock(); \
367 impl_(obj, file, line); \
368 t1 = get_clock(); \
370 e = qsp_entry_get(obj, file, line, qsp_t_); \
371 qsp_entry_record(e, t1 - t0); \
374 #define QSP_GEN_RET1(type_, qsp_t_, func_, impl_) \
375 static int func_(type_ *obj, const char *file, int line) \
377 QSPEntry *e; \
378 int64_t t0, t1; \
379 int err; \
381 t0 = get_clock(); \
382 err = impl_(obj, file, line); \
383 t1 = get_clock(); \
385 e = qsp_entry_get(obj, file, line, qsp_t_); \
386 do_qsp_entry_record(e, t1 - t0, !err); \
387 return err; \
390 QSP_GEN_VOID(QemuMutex, QSP_BQL_MUTEX, qsp_bql_mutex_lock, qemu_mutex_lock_impl)
391 QSP_GEN_VOID(QemuMutex, QSP_MUTEX, qsp_mutex_lock, qemu_mutex_lock_impl)
392 QSP_GEN_RET1(QemuMutex, QSP_MUTEX, qsp_mutex_trylock, qemu_mutex_trylock_impl)
394 QSP_GEN_VOID(QemuRecMutex, QSP_REC_MUTEX, qsp_rec_mutex_lock,
395 qemu_rec_mutex_lock_impl)
396 QSP_GEN_RET1(QemuRecMutex, QSP_REC_MUTEX, qsp_rec_mutex_trylock,
397 qemu_rec_mutex_trylock_impl)
399 #undef QSP_GEN_RET1
400 #undef QSP_GEN_VOID
402 static void
403 qsp_cond_wait(QemuCond *cond, QemuMutex *mutex, const char *file, int line)
405 QSPEntry *e;
406 int64_t t0, t1;
408 t0 = get_clock();
409 qemu_cond_wait_impl(cond, mutex, file, line);
410 t1 = get_clock();
412 e = qsp_entry_get(cond, file, line, QSP_CONDVAR);
413 qsp_entry_record(e, t1 - t0);
416 static bool
417 qsp_cond_timedwait(QemuCond *cond, QemuMutex *mutex, int ms,
418 const char *file, int line)
420 QSPEntry *e;
421 int64_t t0, t1;
422 bool ret;
424 t0 = get_clock();
425 ret = qemu_cond_timedwait_impl(cond, mutex, ms, file, line);
426 t1 = get_clock();
428 e = qsp_entry_get(cond, file, line, QSP_CONDVAR);
429 qsp_entry_record(e, t1 - t0);
430 return ret;
433 bool qsp_is_enabled(void)
435 return atomic_read(&qemu_mutex_lock_func) == qsp_mutex_lock;
438 void qsp_enable(void)
440 atomic_set(&qemu_mutex_lock_func, qsp_mutex_lock);
441 atomic_set(&qemu_mutex_trylock_func, qsp_mutex_trylock);
442 atomic_set(&qemu_bql_mutex_lock_func, qsp_bql_mutex_lock);
443 atomic_set(&qemu_rec_mutex_lock_func, qsp_rec_mutex_lock);
444 atomic_set(&qemu_rec_mutex_trylock_func, qsp_rec_mutex_trylock);
445 atomic_set(&qemu_cond_wait_func, qsp_cond_wait);
446 atomic_set(&qemu_cond_timedwait_func, qsp_cond_timedwait);
449 void qsp_disable(void)
451 atomic_set(&qemu_mutex_lock_func, qemu_mutex_lock_impl);
452 atomic_set(&qemu_mutex_trylock_func, qemu_mutex_trylock_impl);
453 atomic_set(&qemu_bql_mutex_lock_func, qemu_mutex_lock_impl);
454 atomic_set(&qemu_rec_mutex_lock_func, qemu_rec_mutex_lock_impl);
455 atomic_set(&qemu_rec_mutex_trylock_func, qemu_rec_mutex_trylock_impl);
456 atomic_set(&qemu_cond_wait_func, qemu_cond_wait_impl);
457 atomic_set(&qemu_cond_timedwait_func, qemu_cond_timedwait_impl);
460 static gint qsp_tree_cmp(gconstpointer ap, gconstpointer bp, gpointer up)
462 const QSPEntry *a = ap;
463 const QSPEntry *b = bp;
464 enum QSPSortBy sort_by = *(enum QSPSortBy *)up;
465 const QSPCallSite *ca;
466 const QSPCallSite *cb;
468 switch (sort_by) {
469 case QSP_SORT_BY_TOTAL_WAIT_TIME:
470 if (a->ns > b->ns) {
471 return -1;
472 } else if (a->ns < b->ns) {
473 return 1;
475 break;
476 case QSP_SORT_BY_AVG_WAIT_TIME:
478 double avg_a = a->n_acqs ? a->ns / a->n_acqs : 0;
479 double avg_b = b->n_acqs ? b->ns / b->n_acqs : 0;
481 if (avg_a > avg_b) {
482 return -1;
483 } else if (avg_a < avg_b) {
484 return 1;
486 break;
488 default:
489 g_assert_not_reached();
492 ca = a->callsite;
493 cb = b->callsite;
494 /* Break the tie with the object's address */
495 if (ca->obj < cb->obj) {
496 return -1;
497 } else if (ca->obj > cb->obj) {
498 return 1;
499 } else {
500 int cmp;
502 /* same obj. Break the tie with the callsite's file */
503 cmp = strcmp(ca->file, cb->file);
504 if (cmp) {
505 return cmp;
507 /* same callsite file. Break the tie with the callsite's line */
508 g_assert(ca->line != cb->line);
509 if (ca->line < cb->line) {
510 return -1;
511 } else if (ca->line > cb->line) {
512 return 1;
513 } else {
514 /* break the tie with the callsite's type */
515 return cb->type - ca->type;
520 static void qsp_sort(void *p, uint32_t h, void *userp)
522 QSPEntry *e = p;
523 GTree *tree = userp;
525 g_tree_insert(tree, e, NULL);
528 static void qsp_aggregate(void *p, uint32_t h, void *up)
530 struct qht *ht = up;
531 const QSPEntry *e = p;
532 QSPEntry *agg;
533 uint32_t hash;
535 hash = qsp_entry_no_thread_hash(e);
536 agg = qsp_entry_find(ht, e, hash);
538 * The entry is in the global hash table; read from it atomically (as in
539 * "read once").
541 agg->ns += atomic_read_u64(&e->ns);
542 agg->n_acqs += atomic_read_u64(&e->n_acqs);
545 static void qsp_iter_diff(void *p, uint32_t hash, void *htp)
547 struct qht *ht = htp;
548 QSPEntry *old = p;
549 QSPEntry *new;
551 new = qht_lookup(ht, old, hash);
552 /* entries are never deleted, so we must have this one */
553 g_assert(new != NULL);
554 /* our reading of the stats happened after the snapshot was taken */
555 g_assert(new->n_acqs >= old->n_acqs);
556 g_assert(new->ns >= old->ns);
558 new->n_acqs -= old->n_acqs;
559 new->ns -= old->ns;
561 /* No point in reporting an empty entry */
562 if (new->n_acqs == 0 && new->ns == 0) {
563 bool removed = qht_remove(ht, new, hash);
565 g_assert(removed);
566 g_free(new);
570 static void qsp_diff(struct qht *orig, struct qht *new)
572 qht_iter(orig, qsp_iter_diff, new);
575 static void qsp_iter_callsite_coalesce(void *p, uint32_t h, void *htp)
577 struct qht *ht = htp;
578 QSPEntry *old = p;
579 QSPEntry *e;
580 uint32_t hash;
582 hash = qsp_entry_no_thread_obj_hash(old);
583 e = qht_lookup(ht, old, hash);
584 if (e == NULL) {
585 e = qsp_entry_create(ht, old, hash);
586 e->n_objs = 1;
587 } else if (e->callsite->obj != old->callsite->obj) {
588 e->n_objs++;
590 e->ns += old->ns;
591 e->n_acqs += old->n_acqs;
594 static void qsp_ht_delete(void *p, uint32_t h, void *htp)
596 g_free(p);
599 static void qsp_mktree(GTree *tree, bool callsite_coalesce)
601 QSPSnapshot *snap;
602 struct qht ht, coalesce_ht;
603 struct qht *htp;
606 * First, see if there's a prior snapshot, so that we read the global hash
607 * table _after_ the snapshot has been created, which guarantees that
608 * the entries we'll read will be a superset of the snapshot's entries.
610 * We must remain in an RCU read-side critical section until we're done
611 * with the snapshot.
613 rcu_read_lock();
614 snap = atomic_rcu_read(&qsp_snapshot);
616 /* Aggregate all results from the global hash table into a local one */
617 qht_init(&ht, qsp_entry_no_thread_cmp, QSP_INITIAL_SIZE,
618 QHT_MODE_AUTO_RESIZE | QHT_MODE_RAW_MUTEXES);
619 qht_iter(&qsp_ht, qsp_aggregate, &ht);
621 /* compute the difference wrt the snapshot, if any */
622 if (snap) {
623 qsp_diff(&snap->ht, &ht);
625 /* done with the snapshot; RCU can reclaim it */
626 rcu_read_unlock();
628 htp = &ht;
629 if (callsite_coalesce) {
630 qht_init(&coalesce_ht, qsp_entry_no_thread_obj_cmp, QSP_INITIAL_SIZE,
631 QHT_MODE_AUTO_RESIZE | QHT_MODE_RAW_MUTEXES);
632 qht_iter(&ht, qsp_iter_callsite_coalesce, &coalesce_ht);
634 /* free the previous hash table, and point htp to coalesce_ht */
635 qht_iter(&ht, qsp_ht_delete, NULL);
636 qht_destroy(&ht);
637 htp = &coalesce_ht;
640 /* sort the hash table elements by using a tree */
641 qht_iter(htp, qsp_sort, tree);
643 /* free the hash table, but keep the elements (those are in the tree now) */
644 qht_destroy(htp);
647 /* free string with g_free */
648 static char *qsp_at(const QSPCallSite *callsite)
650 GString *s = g_string_new(NULL);
651 const char *shortened;
653 /* remove the absolute path to qemu */
654 if (unlikely(strlen(callsite->file) < qsp_qemu_path_len)) {
655 shortened = callsite->file;
656 } else {
657 shortened = callsite->file + qsp_qemu_path_len;
659 g_string_append_printf(s, "%s:%u", shortened, callsite->line);
660 return g_string_free(s, FALSE);
663 struct QSPReportEntry {
664 const void *obj;
665 char *callsite_at;
666 const char *typename;
667 double time_s;
668 double ns_avg;
669 uint64_t n_acqs;
670 unsigned int n_objs;
672 typedef struct QSPReportEntry QSPReportEntry;
674 struct QSPReport {
675 QSPReportEntry *entries;
676 size_t n_entries;
677 size_t max_n_entries;
679 typedef struct QSPReport QSPReport;
681 static gboolean qsp_tree_report(gpointer key, gpointer value, gpointer udata)
683 const QSPEntry *e = key;
684 QSPReport *report = udata;
685 QSPReportEntry *entry;
687 if (report->n_entries == report->max_n_entries) {
688 return TRUE;
690 entry = &report->entries[report->n_entries];
691 report->n_entries++;
693 entry->obj = e->callsite->obj;
694 entry->n_objs = e->n_objs;
695 entry->callsite_at = qsp_at(e->callsite);
696 entry->typename = qsp_typenames[e->callsite->type];
697 entry->time_s = e->ns * 1e-9;
698 entry->n_acqs = e->n_acqs;
699 entry->ns_avg = e->n_acqs ? e->ns / e->n_acqs : 0;
700 return FALSE;
703 static void pr_report(const QSPReport *rep)
705 char *dashes;
706 size_t max_len = 0;
707 int callsite_len = 0;
708 int callsite_rspace;
709 int n_dashes;
710 size_t i;
712 /* find out the maximum length of all 'callsite' fields */
713 for (i = 0; i < rep->n_entries; i++) {
714 const QSPReportEntry *e = &rep->entries[i];
715 size_t len = strlen(e->callsite_at);
717 if (len > max_len) {
718 max_len = len;
722 callsite_len = MAX(max_len, strlen("Call site"));
723 /* white space to leave to the right of "Call site" */
724 callsite_rspace = callsite_len - strlen("Call site");
726 qemu_printf("Type Object Call site%*s Wait Time (s) "
727 " Count Average (us)\n", callsite_rspace, "");
729 /* build a horizontal rule with dashes */
730 n_dashes = 79 + callsite_rspace;
731 dashes = g_malloc(n_dashes + 1);
732 memset(dashes, '-', n_dashes);
733 dashes[n_dashes] = '\0';
734 qemu_printf("%s\n", dashes);
736 for (i = 0; i < rep->n_entries; i++) {
737 const QSPReportEntry *e = &rep->entries[i];
738 GString *s = g_string_new(NULL);
740 g_string_append_printf(s, "%-9s ", e->typename);
741 if (e->n_objs > 1) {
742 g_string_append_printf(s, "[%12u]", e->n_objs);
743 } else {
744 g_string_append_printf(s, "%14p", e->obj);
746 g_string_append_printf(s, " %s%*s %13.5f %12" PRIu64 " %12.2f\n",
747 e->callsite_at,
748 callsite_len - (int)strlen(e->callsite_at), "",
749 e->time_s, e->n_acqs, e->ns_avg * 1e-3);
750 qemu_printf("%s", s->str);
751 g_string_free(s, TRUE);
754 qemu_printf("%s\n", dashes);
755 g_free(dashes);
758 static void report_destroy(QSPReport *rep)
760 size_t i;
762 for (i = 0; i < rep->n_entries; i++) {
763 QSPReportEntry *e = &rep->entries[i];
765 g_free(e->callsite_at);
767 g_free(rep->entries);
770 void qsp_report(size_t max, enum QSPSortBy sort_by,
771 bool callsite_coalesce)
773 GTree *tree = g_tree_new_full(qsp_tree_cmp, &sort_by, g_free, NULL);
774 QSPReport rep;
776 qsp_init();
778 rep.entries = g_new0(QSPReportEntry, max);
779 rep.n_entries = 0;
780 rep.max_n_entries = max;
782 qsp_mktree(tree, callsite_coalesce);
783 g_tree_foreach(tree, qsp_tree_report, &rep);
784 g_tree_destroy(tree);
786 pr_report(&rep);
787 report_destroy(&rep);
790 static void qsp_snapshot_destroy(QSPSnapshot *snap)
792 qht_iter(&snap->ht, qsp_ht_delete, NULL);
793 qht_destroy(&snap->ht);
794 g_free(snap);
797 void qsp_reset(void)
799 QSPSnapshot *new = g_new(QSPSnapshot, 1);
800 QSPSnapshot *old;
802 qsp_init();
804 qht_init(&new->ht, qsp_entry_cmp, QSP_INITIAL_SIZE,
805 QHT_MODE_AUTO_RESIZE | QHT_MODE_RAW_MUTEXES);
807 /* take a snapshot of the current state */
808 qht_iter(&qsp_ht, qsp_aggregate, &new->ht);
810 /* replace the previous snapshot, if any */
811 old = atomic_xchg(&qsp_snapshot, new);
812 if (old) {
813 call_rcu(old, qsp_snapshot_destroy, rcu);