xref: /openbmc/qemu/util/qsp.c (revision 3d9569b8)
1 /*
2  * qsp.c - QEMU Synchronization Profiler
3  *
4  * Copyright (C) 2018, Emilio G. Cota <cota@braap.org>
5  *
6  * License: GNU GPL, version 2 or later.
7  *   See the COPYING file in the top-level directory.
8  *
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.
12  *
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).
18  *
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.
26  *
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.
33  *
34  * Alternative designs considered:
35  *
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.
39  *
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).
43  *
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.
47  *
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.
52  *
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.
58  */
59 
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"
67 
68 enum QSPType {
69     QSP_MUTEX,
70     QSP_BQL_MUTEX,
71     QSP_REC_MUTEX,
72     QSP_CONDVAR,
73 };
74 
75 struct QSPCallSite {
76     const void *obj;
77     const char *file; /* i.e. __FILE__; shortened later */
78     int line;
79     enum QSPType type;
80 };
81 typedef struct QSPCallSite QSPCallSite;
82 
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 */
89 };
90 typedef struct QSPEntry QSPEntry;
91 
92 struct QSPSnapshot {
93     struct rcu_head rcu;
94     struct qht ht;
95 };
96 typedef struct QSPSnapshot QSPSnapshot;
97 
98 /* initial sizing for hash tables */
99 #define QSP_INITIAL_SIZE 64
100 
101 /* If this file is moved, QSP_REL_PATH should be updated accordingly */
102 #define QSP_REL_PATH "util/qsp.c"
103 
104 /* this file's full path. Used to present all call sites with relative paths */
105 static size_t qsp_qemu_path_len;
106 
107 /* the address of qsp_thread gives us a unique 'thread ID' */
108 static __thread int qsp_thread;
109 
110 /*
111  * Call sites are the same for all threads, so we track them in a separate hash
112  * table to save memory.
113  */
114 static struct qht qsp_callsite_ht;
115 
116 static struct qht qsp_ht;
117 static QSPSnapshot *qsp_snapshot;
118 static bool qsp_initialized, qsp_initializing;
119 
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",
125 };
126 
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 
135 /*
136  * It pays off to _not_ hash callsite->file; hashing a string is slow, and
137  * without it we still get a pretty unique hash.
138  */
139 static inline
140 uint32_t do_qsp_callsite_hash(const QSPCallSite *callsite, uint64_t ab)
141 {
142     uint64_t cd = (uint64_t)(uintptr_t)callsite->obj;
143     uint32_t e = callsite->line;
144     uint32_t f = callsite->type;
145 
146     return qemu_xxhash6(ab, cd, e, f);
147 }
148 
149 static inline
150 uint32_t qsp_callsite_hash(const QSPCallSite *callsite)
151 {
152     return do_qsp_callsite_hash(callsite, 0);
153 }
154 
155 static inline uint32_t do_qsp_entry_hash(const QSPEntry *entry, uint64_t a)
156 {
157     return do_qsp_callsite_hash(entry->callsite, a);
158 }
159 
160 static uint32_t qsp_entry_hash(const QSPEntry *entry)
161 {
162     return do_qsp_entry_hash(entry, (uint64_t)(uintptr_t)entry->thread_ptr);
163 }
164 
165 static uint32_t qsp_entry_no_thread_hash(const QSPEntry *entry)
166 {
167     return do_qsp_entry_hash(entry, 0);
168 }
169 
170 /* without the objects we need to hash the file name to get a decent hash */
171 static uint32_t qsp_entry_no_thread_obj_hash(const QSPEntry *entry)
172 {
173     const QSPCallSite *callsite = entry->callsite;
174     uint64_t ab = g_str_hash(callsite->file);
175     uint64_t cd = callsite->line;
176     uint32_t e = callsite->type;
177 
178     return qemu_xxhash5(ab, cd, e);
179 }
180 
181 static bool qsp_callsite_cmp(const void *ap, const void *bp)
182 {
183     const QSPCallSite *a = ap;
184     const QSPCallSite *b = bp;
185 
186     return a == b ||
187         (a->obj == b->obj &&
188          a->line == b->line &&
189          a->type == b->type &&
190          (a->file == b->file || !strcmp(a->file, b->file)));
191 }
192 
193 static bool qsp_callsite_no_obj_cmp(const void *ap, const void *bp)
194 {
195     const QSPCallSite *a = ap;
196     const QSPCallSite *b = bp;
197 
198     return a == b ||
199         (a->line == b->line &&
200          a->type == b->type &&
201          (a->file == b->file || !strcmp(a->file, b->file)));
202 }
203 
204 static bool qsp_entry_no_thread_cmp(const void *ap, const void *bp)
205 {
206     const QSPEntry *a = ap;
207     const QSPEntry *b = bp;
208 
209     return qsp_callsite_cmp(a->callsite, b->callsite);
210 }
211 
212 static bool qsp_entry_no_thread_obj_cmp(const void *ap, const void *bp)
213 {
214     const QSPEntry *a = ap;
215     const QSPEntry *b = bp;
216 
217     return qsp_callsite_no_obj_cmp(a->callsite, b->callsite);
218 }
219 
220 static bool qsp_entry_cmp(const void *ap, const void *bp)
221 {
222     const QSPEntry *a = ap;
223     const QSPEntry *b = bp;
224 
225     return a->thread_ptr == b->thread_ptr &&
226         qsp_callsite_cmp(a->callsite, b->callsite);
227 }
228 
229 /*
230  * Normally we'd call this from a constructor function, but we want it to work
231  * via libutil as well.
232  */
233 static void qsp_do_init(void)
234 {
235     /* make sure this file's path in the tree is up to date with QSP_REL_PATH */
236     g_assert(strstr(__FILE__, QSP_REL_PATH));
237     qsp_qemu_path_len = strlen(__FILE__) - strlen(QSP_REL_PATH);
238 
239     qht_init(&qsp_ht, qsp_entry_cmp, QSP_INITIAL_SIZE,
240              QHT_MODE_AUTO_RESIZE | QHT_MODE_RAW_MUTEXES);
241     qht_init(&qsp_callsite_ht, qsp_callsite_cmp, QSP_INITIAL_SIZE,
242              QHT_MODE_AUTO_RESIZE | QHT_MODE_RAW_MUTEXES);
243 }
244 
245 static __attribute__((noinline)) void qsp_init__slowpath(void)
246 {
247     if (atomic_cmpxchg(&qsp_initializing, false, true) == false) {
248         qsp_do_init();
249         atomic_set(&qsp_initialized, true);
250     } else {
251         while (!atomic_read(&qsp_initialized)) {
252             cpu_relax();
253         }
254     }
255 }
256 
257 /* qsp_init() must be called from _all_ exported functions */
258 static inline void qsp_init(void)
259 {
260     if (likely(atomic_read(&qsp_initialized))) {
261         return;
262     }
263     qsp_init__slowpath();
264 }
265 
266 static QSPCallSite *qsp_callsite_find(const QSPCallSite *orig)
267 {
268     QSPCallSite *callsite;
269     uint32_t hash;
270 
271     hash = qsp_callsite_hash(orig);
272     callsite = qht_lookup(&qsp_callsite_ht, orig, hash);
273     if (callsite == NULL) {
274         void *existing = NULL;
275 
276         callsite = g_new(QSPCallSite, 1);
277         memcpy(callsite, orig, sizeof(*callsite));
278         qht_insert(&qsp_callsite_ht, callsite, hash, &existing);
279         if (unlikely(existing)) {
280             g_free(callsite);
281             callsite = existing;
282         }
283     }
284     return callsite;
285 }
286 
287 static QSPEntry *
288 qsp_entry_create(struct qht *ht, const QSPEntry *entry, uint32_t hash)
289 {
290     QSPEntry *e;
291     void *existing = NULL;
292 
293     e = g_new0(QSPEntry, 1);
294     e->thread_ptr = entry->thread_ptr;
295     e->callsite = qsp_callsite_find(entry->callsite);
296 
297     qht_insert(ht, e, hash, &existing);
298     if (unlikely(existing)) {
299         g_free(e);
300         e = existing;
301     }
302     return e;
303 }
304 
305 static QSPEntry *
306 qsp_entry_find(struct qht *ht, const QSPEntry *entry, uint32_t hash)
307 {
308     QSPEntry *e;
309 
310     e = qht_lookup(ht, entry, hash);
311     if (e == NULL) {
312         e = qsp_entry_create(ht, entry, hash);
313     }
314     return e;
315 }
316 
317 /*
318  * Note: Entries are never removed, so callers do not have to be in an RCU
319  * read-side critical section.
320  */
321 static QSPEntry *qsp_entry_get(const void *obj, const char *file, int line,
322                                enum QSPType type)
323 {
324     QSPCallSite callsite = {
325         .obj = obj,
326         .file = file,
327         .line = line,
328         .type = type,
329     };
330     QSPEntry orig;
331     uint32_t hash;
332 
333     qsp_init();
334 
335     orig.thread_ptr = &qsp_thread;
336     orig.callsite = &callsite;
337 
338     hash = qsp_entry_hash(&orig);
339     return qsp_entry_find(&qsp_ht, &orig, hash);
340 }
341 
342 /*
343  * @e is in the global hash table; it is only written to by the current thread,
344  * so we write to it atomically (as in "write once") to prevent torn reads.
345  */
346 static inline void do_qsp_entry_record(QSPEntry *e, int64_t delta, bool acq)
347 {
348     atomic_set_u64(&e->ns, e->ns + delta);
349     if (acq) {
350         atomic_set_u64(&e->n_acqs, e->n_acqs + 1);
351     }
352 }
353 
354 static inline void qsp_entry_record(QSPEntry *e, int64_t delta)
355 {
356     do_qsp_entry_record(e, delta, true);
357 }
358 
359 #define QSP_GEN_VOID(type_, qsp_t_, func_, impl_)                       \
360     static void func_(type_ *obj, const char *file, int line)           \
361     {                                                                   \
362         QSPEntry *e;                                                    \
363         int64_t t0, t1;                                                 \
364                                                                         \
365         t0 = get_clock();                                               \
366         impl_(obj, file, line);                                         \
367         t1 = get_clock();                                               \
368                                                                         \
369         e = qsp_entry_get(obj, file, line, qsp_t_);                     \
370         qsp_entry_record(e, t1 - t0);                                   \
371     }
372 
373 #define QSP_GEN_RET1(type_, qsp_t_, func_, impl_)                       \
374     static int func_(type_ *obj, const char *file, int line)            \
375     {                                                                   \
376         QSPEntry *e;                                                    \
377         int64_t t0, t1;                                                 \
378         int err;                                                        \
379                                                                         \
380         t0 = get_clock();                                               \
381         err = impl_(obj, file, line);                                   \
382         t1 = get_clock();                                               \
383                                                                         \
384         e = qsp_entry_get(obj, file, line, qsp_t_);                     \
385         do_qsp_entry_record(e, t1 - t0, !err);                          \
386         return err;                                                     \
387     }
388 
389 QSP_GEN_VOID(QemuMutex, QSP_BQL_MUTEX, qsp_bql_mutex_lock, qemu_mutex_lock_impl)
390 QSP_GEN_VOID(QemuMutex, QSP_MUTEX, qsp_mutex_lock, qemu_mutex_lock_impl)
391 QSP_GEN_RET1(QemuMutex, QSP_MUTEX, qsp_mutex_trylock, qemu_mutex_trylock_impl)
392 
393 QSP_GEN_VOID(QemuRecMutex, QSP_REC_MUTEX, qsp_rec_mutex_lock,
394              qemu_rec_mutex_lock_impl)
395 QSP_GEN_RET1(QemuRecMutex, QSP_REC_MUTEX, qsp_rec_mutex_trylock,
396              qemu_rec_mutex_trylock_impl)
397 
398 #undef QSP_GEN_RET1
399 #undef QSP_GEN_VOID
400 
401 static void
402 qsp_cond_wait(QemuCond *cond, QemuMutex *mutex, const char *file, int line)
403 {
404     QSPEntry *e;
405     int64_t t0, t1;
406 
407     t0 = get_clock();
408     qemu_cond_wait_impl(cond, mutex, file, line);
409     t1 = get_clock();
410 
411     e = qsp_entry_get(cond, file, line, QSP_CONDVAR);
412     qsp_entry_record(e, t1 - t0);
413 }
414 
415 bool qsp_is_enabled(void)
416 {
417     return atomic_read(&qemu_mutex_lock_func) == qsp_mutex_lock;
418 }
419 
420 void qsp_enable(void)
421 {
422     atomic_set(&qemu_mutex_lock_func, qsp_mutex_lock);
423     atomic_set(&qemu_mutex_trylock_func, qsp_mutex_trylock);
424     atomic_set(&qemu_bql_mutex_lock_func, qsp_bql_mutex_lock);
425     atomic_set(&qemu_rec_mutex_lock_func, qsp_rec_mutex_lock);
426     atomic_set(&qemu_rec_mutex_trylock_func, qsp_rec_mutex_trylock);
427     atomic_set(&qemu_cond_wait_func, qsp_cond_wait);
428 }
429 
430 void qsp_disable(void)
431 {
432     atomic_set(&qemu_mutex_lock_func, qemu_mutex_lock_impl);
433     atomic_set(&qemu_mutex_trylock_func, qemu_mutex_trylock_impl);
434     atomic_set(&qemu_bql_mutex_lock_func, qemu_mutex_lock_impl);
435     atomic_set(&qemu_rec_mutex_lock_func, qemu_rec_mutex_lock_impl);
436     atomic_set(&qemu_rec_mutex_trylock_func, qemu_rec_mutex_trylock_impl);
437     atomic_set(&qemu_cond_wait_func, qemu_cond_wait_impl);
438 }
439 
440 static gint qsp_tree_cmp(gconstpointer ap, gconstpointer bp, gpointer up)
441 {
442     const QSPEntry *a = ap;
443     const QSPEntry *b = bp;
444     enum QSPSortBy sort_by = *(enum QSPSortBy *)up;
445     const QSPCallSite *ca;
446     const QSPCallSite *cb;
447 
448     switch (sort_by) {
449     case QSP_SORT_BY_TOTAL_WAIT_TIME:
450         if (a->ns > b->ns) {
451             return -1;
452         } else if (a->ns < b->ns) {
453             return 1;
454         }
455         break;
456     case QSP_SORT_BY_AVG_WAIT_TIME:
457     {
458         double avg_a = a->n_acqs ? a->ns / a->n_acqs : 0;
459         double avg_b = b->n_acqs ? b->ns / b->n_acqs : 0;
460 
461         if (avg_a > avg_b) {
462             return -1;
463         } else if (avg_a < avg_b) {
464             return 1;
465         }
466         break;
467     }
468     default:
469         g_assert_not_reached();
470     }
471 
472     ca = a->callsite;
473     cb = b->callsite;
474     /* Break the tie with the object's address */
475     if (ca->obj < cb->obj) {
476         return -1;
477     } else if (ca->obj > cb->obj) {
478         return 1;
479     } else {
480         int cmp;
481 
482         /* same obj. Break the tie with the callsite's file */
483         cmp = strcmp(ca->file, cb->file);
484         if (cmp) {
485             return cmp;
486         }
487         /* same callsite file. Break the tie with the callsite's line */
488         g_assert(ca->line != cb->line);
489         if (ca->line < cb->line) {
490             return -1;
491         } else if (ca->line > cb->line) {
492             return 1;
493         } else {
494             /* break the tie with the callsite's type */
495             return cb->type - ca->type;
496         }
497     }
498 }
499 
500 static void qsp_sort(void *p, uint32_t h, void *userp)
501 {
502     QSPEntry *e = p;
503     GTree *tree = userp;
504 
505     g_tree_insert(tree, e, NULL);
506 }
507 
508 static void qsp_aggregate(void *p, uint32_t h, void *up)
509 {
510     struct qht *ht = up;
511     const QSPEntry *e = p;
512     QSPEntry *agg;
513     uint32_t hash;
514 
515     hash = qsp_entry_no_thread_hash(e);
516     agg = qsp_entry_find(ht, e, hash);
517     /*
518      * The entry is in the global hash table; read from it atomically (as in
519      * "read once").
520      */
521     agg->ns += atomic_read_u64(&e->ns);
522     agg->n_acqs += atomic_read_u64(&e->n_acqs);
523 }
524 
525 static void qsp_iter_diff(void *p, uint32_t hash, void *htp)
526 {
527     struct qht *ht = htp;
528     QSPEntry *old = p;
529     QSPEntry *new;
530 
531     new = qht_lookup(ht, old, hash);
532     /* entries are never deleted, so we must have this one */
533     g_assert(new != NULL);
534     /* our reading of the stats happened after the snapshot was taken */
535     g_assert(new->n_acqs >= old->n_acqs);
536     g_assert(new->ns >= old->ns);
537 
538     new->n_acqs -= old->n_acqs;
539     new->ns -= old->ns;
540 
541     /* No point in reporting an empty entry */
542     if (new->n_acqs == 0 && new->ns == 0) {
543         bool removed = qht_remove(ht, new, hash);
544 
545         g_assert(removed);
546         g_free(new);
547     }
548 }
549 
550 static void qsp_diff(struct qht *orig, struct qht *new)
551 {
552     qht_iter(orig, qsp_iter_diff, new);
553 }
554 
555 static void qsp_iter_callsite_coalesce(void *p, uint32_t h, void *htp)
556 {
557     struct qht *ht = htp;
558     QSPEntry *old = p;
559     QSPEntry *e;
560     uint32_t hash;
561 
562     hash = qsp_entry_no_thread_obj_hash(old);
563     e = qht_lookup(ht, old, hash);
564     if (e == NULL) {
565         e = qsp_entry_create(ht, old, hash);
566         e->n_objs = 1;
567     } else if (e->callsite->obj != old->callsite->obj) {
568         e->n_objs++;
569     }
570     e->ns += old->ns;
571     e->n_acqs += old->n_acqs;
572 }
573 
574 static void qsp_ht_delete(void *p, uint32_t h, void *htp)
575 {
576     g_free(p);
577 }
578 
579 static void qsp_mktree(GTree *tree, bool callsite_coalesce)
580 {
581     QSPSnapshot *snap;
582     struct qht ht, coalesce_ht;
583     struct qht *htp;
584 
585     /*
586      * First, see if there's a prior snapshot, so that we read the global hash
587      * table _after_ the snapshot has been created, which guarantees that
588      * the entries we'll read will be a superset of the snapshot's entries.
589      *
590      * We must remain in an RCU read-side critical section until we're done
591      * with the snapshot.
592      */
593     rcu_read_lock();
594     snap = atomic_rcu_read(&qsp_snapshot);
595 
596     /* Aggregate all results from the global hash table into a local one */
597     qht_init(&ht, qsp_entry_no_thread_cmp, QSP_INITIAL_SIZE,
598              QHT_MODE_AUTO_RESIZE | QHT_MODE_RAW_MUTEXES);
599     qht_iter(&qsp_ht, qsp_aggregate, &ht);
600 
601     /* compute the difference wrt the snapshot, if any */
602     if (snap) {
603         qsp_diff(&snap->ht, &ht);
604     }
605     /* done with the snapshot; RCU can reclaim it */
606     rcu_read_unlock();
607 
608     htp = &ht;
609     if (callsite_coalesce) {
610         qht_init(&coalesce_ht, qsp_entry_no_thread_obj_cmp, QSP_INITIAL_SIZE,
611                  QHT_MODE_AUTO_RESIZE | QHT_MODE_RAW_MUTEXES);
612         qht_iter(&ht, qsp_iter_callsite_coalesce, &coalesce_ht);
613 
614         /* free the previous hash table, and point htp to coalesce_ht */
615         qht_iter(&ht, qsp_ht_delete, NULL);
616         qht_destroy(&ht);
617         htp = &coalesce_ht;
618     }
619 
620     /* sort the hash table elements by using a tree */
621     qht_iter(htp, qsp_sort, tree);
622 
623     /* free the hash table, but keep the elements (those are in the tree now) */
624     qht_destroy(htp);
625 }
626 
627 /* free string with g_free */
628 static char *qsp_at(const QSPCallSite *callsite)
629 {
630     GString *s = g_string_new(NULL);
631     const char *shortened;
632 
633     /* remove the absolute path to qemu */
634     if (unlikely(strlen(callsite->file) < qsp_qemu_path_len)) {
635         shortened = callsite->file;
636     } else {
637         shortened = callsite->file + qsp_qemu_path_len;
638     }
639     g_string_append_printf(s, "%s:%u", shortened, callsite->line);
640     return g_string_free(s, FALSE);
641 }
642 
643 struct QSPReportEntry {
644     const void *obj;
645     char *callsite_at;
646     const char *typename;
647     double time_s;
648     double ns_avg;
649     uint64_t n_acqs;
650     unsigned int n_objs;
651 };
652 typedef struct QSPReportEntry QSPReportEntry;
653 
654 struct QSPReport {
655     QSPReportEntry *entries;
656     size_t n_entries;
657     size_t max_n_entries;
658 };
659 typedef struct QSPReport QSPReport;
660 
661 static gboolean qsp_tree_report(gpointer key, gpointer value, gpointer udata)
662 {
663     const QSPEntry *e = key;
664     QSPReport *report = udata;
665     QSPReportEntry *entry;
666 
667     if (report->n_entries == report->max_n_entries) {
668         return TRUE;
669     }
670     entry = &report->entries[report->n_entries];
671     report->n_entries++;
672 
673     entry->obj = e->callsite->obj;
674     entry->n_objs = e->n_objs;
675     entry->callsite_at = qsp_at(e->callsite);
676     entry->typename = qsp_typenames[e->callsite->type];
677     entry->time_s = e->ns * 1e-9;
678     entry->n_acqs = e->n_acqs;
679     entry->ns_avg = e->n_acqs ? e->ns / e->n_acqs : 0;
680     return FALSE;
681 }
682 
683 static void pr_report(const QSPReport *rep)
684 {
685     char *dashes;
686     size_t max_len = 0;
687     int callsite_len = 0;
688     int callsite_rspace;
689     int n_dashes;
690     size_t i;
691 
692     /* find out the maximum length of all 'callsite' fields */
693     for (i = 0; i < rep->n_entries; i++) {
694         const QSPReportEntry *e = &rep->entries[i];
695         size_t len = strlen(e->callsite_at);
696 
697         if (len > max_len) {
698             max_len = len;
699         }
700     }
701 
702     callsite_len = MAX(max_len, strlen("Call site"));
703     /* white space to leave to the right of "Call site" */
704     callsite_rspace = callsite_len - strlen("Call site");
705 
706     qemu_printf("Type               Object  Call site%*s  Wait Time (s)  "
707                 "       Count  Average (us)\n", callsite_rspace, "");
708 
709     /* build a horizontal rule with dashes */
710     n_dashes = 79 + callsite_rspace;
711     dashes = g_malloc(n_dashes + 1);
712     memset(dashes, '-', n_dashes);
713     dashes[n_dashes] = '\0';
714     qemu_printf("%s\n", dashes);
715 
716     for (i = 0; i < rep->n_entries; i++) {
717         const QSPReportEntry *e = &rep->entries[i];
718         GString *s = g_string_new(NULL);
719 
720         g_string_append_printf(s, "%-9s  ", e->typename);
721         if (e->n_objs > 1) {
722             g_string_append_printf(s, "[%12u]", e->n_objs);
723         } else {
724             g_string_append_printf(s, "%14p", e->obj);
725         }
726         g_string_append_printf(s, "  %s%*s  %13.5f  %12" PRIu64 "  %12.2f\n",
727                                e->callsite_at,
728                                callsite_len - (int)strlen(e->callsite_at), "",
729                                e->time_s, e->n_acqs, e->ns_avg * 1e-3);
730         qemu_printf("%s", s->str);
731         g_string_free(s, TRUE);
732     }
733 
734     qemu_printf("%s\n", dashes);
735     g_free(dashes);
736 }
737 
738 static void report_destroy(QSPReport *rep)
739 {
740     size_t i;
741 
742     for (i = 0; i < rep->n_entries; i++) {
743         QSPReportEntry *e = &rep->entries[i];
744 
745         g_free(e->callsite_at);
746     }
747     g_free(rep->entries);
748 }
749 
750 void qsp_report(size_t max, enum QSPSortBy sort_by,
751                 bool callsite_coalesce)
752 {
753     GTree *tree = g_tree_new_full(qsp_tree_cmp, &sort_by, g_free, NULL);
754     QSPReport rep;
755 
756     qsp_init();
757 
758     rep.entries = g_new0(QSPReportEntry, max);
759     rep.n_entries = 0;
760     rep.max_n_entries = max;
761 
762     qsp_mktree(tree, callsite_coalesce);
763     g_tree_foreach(tree, qsp_tree_report, &rep);
764     g_tree_destroy(tree);
765 
766     pr_report(&rep);
767     report_destroy(&rep);
768 }
769 
770 static void qsp_snapshot_destroy(QSPSnapshot *snap)
771 {
772     qht_iter(&snap->ht, qsp_ht_delete, NULL);
773     qht_destroy(&snap->ht);
774     g_free(snap);
775 }
776 
777 void qsp_reset(void)
778 {
779     QSPSnapshot *new = g_new(QSPSnapshot, 1);
780     QSPSnapshot *old;
781 
782     qsp_init();
783 
784     qht_init(&new->ht, qsp_entry_cmp, QSP_INITIAL_SIZE,
785              QHT_MODE_AUTO_RESIZE | QHT_MODE_RAW_MUTEXES);
786 
787     /* take a snapshot of the current state */
788     qht_iter(&qsp_ht, qsp_aggregate, &new->ht);
789 
790     /* replace the previous snapshot, if any */
791     old = atomic_xchg(&qsp_snapshot, new);
792     if (old) {
793         call_rcu(old, qsp_snapshot_destroy, rcu);
794     }
795 }
796