1fe9959a2SEmilio G. Cota /* 2fe9959a2SEmilio G. Cota * qsp.c - QEMU Synchronization Profiler 3fe9959a2SEmilio G. Cota * 4fe9959a2SEmilio G. Cota * Copyright (C) 2018, Emilio G. Cota <cota@braap.org> 5fe9959a2SEmilio G. Cota * 6fe9959a2SEmilio G. Cota * License: GNU GPL, version 2 or later. 7fe9959a2SEmilio G. Cota * See the COPYING file in the top-level directory. 8fe9959a2SEmilio G. Cota * 9fe9959a2SEmilio G. Cota * QSP profiles the time spent in synchronization primitives, which can 10fe9959a2SEmilio G. Cota * help diagnose performance problems, e.g. scalability issues when 11fe9959a2SEmilio G. Cota * contention is high. 12fe9959a2SEmilio G. Cota * 13fe9959a2SEmilio G. Cota * The primitives currently supported are mutexes, recursive mutexes and 14fe9959a2SEmilio G. Cota * condition variables. Note that not all related functions are intercepted; 15fe9959a2SEmilio G. Cota * instead we profile only those functions that can have a performance impact, 16fe9959a2SEmilio G. Cota * either due to blocking (e.g. cond_wait, mutex_lock) or cache line 17fe9959a2SEmilio G. Cota * contention (e.g. mutex_lock, mutex_trylock). 18fe9959a2SEmilio G. Cota * 19fe9959a2SEmilio G. Cota * QSP's design focuses on speed and scalability. This is achieved 20fe9959a2SEmilio G. Cota * by having threads do their profiling entirely on thread-local data. 21fe9959a2SEmilio G. Cota * The appropriate thread-local data is found via a QHT, i.e. a concurrent hash 22fe9959a2SEmilio G. Cota * table. To aggregate data in order to generate a report, we iterate over 23fe9959a2SEmilio G. Cota * all entries in the hash table. Depending on the number of threads and 24fe9959a2SEmilio G. Cota * synchronization objects this might be expensive, but note that it is 25fe9959a2SEmilio G. Cota * very rarely called -- reports are generated only when requested by users. 26fe9959a2SEmilio G. Cota * 27fe9959a2SEmilio G. Cota * Reports are generated as a table where each row represents a call site. A 28fe9959a2SEmilio G. Cota * call site is the triplet formed by the __file__ and __LINE__ of the caller 29fe9959a2SEmilio G. Cota * as well as the address of the "object" (i.e. mutex, rec. mutex or condvar) 30d557de4aSEmilio G. Cota * being operated on. Optionally, call sites that operate on different objects 31d557de4aSEmilio G. Cota * of the same type can be coalesced, which can be particularly useful when 32d557de4aSEmilio G. Cota * profiling dynamically-allocated objects. 33fe9959a2SEmilio G. Cota * 34fe9959a2SEmilio G. Cota * Alternative designs considered: 35fe9959a2SEmilio G. Cota * 36fe9959a2SEmilio G. Cota * - Use an off-the-shelf profiler such as mutrace. This is not a viable option 37fe9959a2SEmilio G. Cota * for us because QEMU has __malloc_hook set (by one of the libraries it 38fe9959a2SEmilio G. Cota * uses); leaving this hook unset is required to avoid deadlock in mutrace. 39fe9959a2SEmilio G. Cota * 40fe9959a2SEmilio G. Cota * - Use a glib HT for each thread, protecting each HT with its own lock. 41fe9959a2SEmilio G. Cota * This isn't simpler than the current design, and is 10% slower in the 42fe9959a2SEmilio G. Cota * atomic_add-bench microbenchmark (-m option). 43fe9959a2SEmilio G. Cota * 44fe9959a2SEmilio G. Cota * - For reports, just use a binary tree as we aggregate data, instead of having 45fe9959a2SEmilio G. Cota * an intermediate hash table. This would simplify the code only slightly, but 46fe9959a2SEmilio G. Cota * would perform badly if there were many threads and objects to track. 47fe9959a2SEmilio G. Cota * 48996e8d9aSEmilio G. Cota * - Wrap operations on qsp entries with RCU read-side critical sections, so 49996e8d9aSEmilio G. Cota * that qsp_reset() can delete entries. Unfortunately, the overhead of calling 50996e8d9aSEmilio G. Cota * rcu_read_lock/unlock slows down atomic_add-bench -m by 24%. Having 51996e8d9aSEmilio G. Cota * a snapshot that is updated on qsp_reset() avoids this overhead. 52996e8d9aSEmilio G. Cota * 53fe9959a2SEmilio G. Cota * Related Work: 54fe9959a2SEmilio G. Cota * - Lennart Poettering's mutrace: http://0pointer.de/blog/projects/mutrace.html 55fe9959a2SEmilio G. Cota * - Lozi, David, Thomas, Lawall and Muller. "Remote Core Locking: Migrating 56fe9959a2SEmilio G. Cota * Critical-Section Execution to Improve the Performance of Multithreaded 57fe9959a2SEmilio G. Cota * Applications", USENIX ATC'12. 58fe9959a2SEmilio G. Cota */ 59fe9959a2SEmilio G. Cota #include "qemu/osdep.h" 60fe9959a2SEmilio G. Cota #include "qemu/thread.h" 61fe9959a2SEmilio G. Cota #include "qemu/timer.h" 62fe9959a2SEmilio G. Cota #include "qemu/qht.h" 63996e8d9aSEmilio G. Cota #include "qemu/rcu.h" 64*fe656e31SEmilio G. Cota #include "qemu/xxhash.h" 65fe9959a2SEmilio G. Cota 66fe9959a2SEmilio G. Cota enum QSPType { 67fe9959a2SEmilio G. Cota QSP_MUTEX, 68cb764d06SEmilio G. Cota QSP_BQL_MUTEX, 69fe9959a2SEmilio G. Cota QSP_REC_MUTEX, 70fe9959a2SEmilio G. Cota QSP_CONDVAR, 71fe9959a2SEmilio G. Cota }; 72fe9959a2SEmilio G. Cota 73fe9959a2SEmilio G. Cota struct QSPCallSite { 74fe9959a2SEmilio G. Cota const void *obj; 75fe9959a2SEmilio G. Cota const char *file; /* i.e. __FILE__; shortened later */ 76fe9959a2SEmilio G. Cota int line; 77fe9959a2SEmilio G. Cota enum QSPType type; 78fe9959a2SEmilio G. Cota }; 79fe9959a2SEmilio G. Cota typedef struct QSPCallSite QSPCallSite; 80fe9959a2SEmilio G. Cota 81fe9959a2SEmilio G. Cota struct QSPEntry { 82fe9959a2SEmilio G. Cota void *thread_ptr; 83fe9959a2SEmilio G. Cota const QSPCallSite *callsite; 84fe9959a2SEmilio G. Cota uint64_t n_acqs; 85fe9959a2SEmilio G. Cota uint64_t ns; 86d557de4aSEmilio G. Cota unsigned int n_objs; /* count of coalesced objs; only used for reporting */ 87fe9959a2SEmilio G. Cota }; 88fe9959a2SEmilio G. Cota typedef struct QSPEntry QSPEntry; 89fe9959a2SEmilio G. Cota 90996e8d9aSEmilio G. Cota struct QSPSnapshot { 91996e8d9aSEmilio G. Cota struct rcu_head rcu; 92996e8d9aSEmilio G. Cota struct qht ht; 93996e8d9aSEmilio G. Cota }; 94996e8d9aSEmilio G. Cota typedef struct QSPSnapshot QSPSnapshot; 95996e8d9aSEmilio G. Cota 96fe9959a2SEmilio G. Cota /* initial sizing for hash tables */ 97fe9959a2SEmilio G. Cota #define QSP_INITIAL_SIZE 64 98fe9959a2SEmilio G. Cota 99fe9959a2SEmilio G. Cota /* If this file is moved, QSP_REL_PATH should be updated accordingly */ 100fe9959a2SEmilio G. Cota #define QSP_REL_PATH "util/qsp.c" 101fe9959a2SEmilio G. Cota 102fe9959a2SEmilio G. Cota /* this file's full path. Used to present all call sites with relative paths */ 103fe9959a2SEmilio G. Cota static size_t qsp_qemu_path_len; 104fe9959a2SEmilio G. Cota 105fe9959a2SEmilio G. Cota /* the address of qsp_thread gives us a unique 'thread ID' */ 106fe9959a2SEmilio G. Cota static __thread int qsp_thread; 107fe9959a2SEmilio G. Cota 108fe9959a2SEmilio G. Cota /* 109fe9959a2SEmilio G. Cota * Call sites are the same for all threads, so we track them in a separate hash 110fe9959a2SEmilio G. Cota * table to save memory. 111fe9959a2SEmilio G. Cota */ 112fe9959a2SEmilio G. Cota static struct qht qsp_callsite_ht; 113fe9959a2SEmilio G. Cota 114fe9959a2SEmilio G. Cota static struct qht qsp_ht; 115996e8d9aSEmilio G. Cota static QSPSnapshot *qsp_snapshot; 116fe9959a2SEmilio G. Cota static bool qsp_initialized, qsp_initializing; 117fe9959a2SEmilio G. Cota 118fe9959a2SEmilio G. Cota static const char * const qsp_typenames[] = { 119fe9959a2SEmilio G. Cota [QSP_MUTEX] = "mutex", 120cb764d06SEmilio G. Cota [QSP_BQL_MUTEX] = "BQL mutex", 121fe9959a2SEmilio G. Cota [QSP_REC_MUTEX] = "rec_mutex", 122fe9959a2SEmilio G. Cota [QSP_CONDVAR] = "condvar", 123fe9959a2SEmilio G. Cota }; 124fe9959a2SEmilio G. Cota 125cb764d06SEmilio G. Cota QemuMutexLockFunc qemu_bql_mutex_lock_func = qemu_mutex_lock_impl; 126fe9959a2SEmilio G. Cota QemuMutexLockFunc qemu_mutex_lock_func = qemu_mutex_lock_impl; 127fe9959a2SEmilio G. Cota QemuMutexTrylockFunc qemu_mutex_trylock_func = qemu_mutex_trylock_impl; 128fe9959a2SEmilio G. Cota QemuRecMutexLockFunc qemu_rec_mutex_lock_func = qemu_rec_mutex_lock_impl; 129fe9959a2SEmilio G. Cota QemuRecMutexTrylockFunc qemu_rec_mutex_trylock_func = 130fe9959a2SEmilio G. Cota qemu_rec_mutex_trylock_impl; 131fe9959a2SEmilio G. Cota QemuCondWaitFunc qemu_cond_wait_func = qemu_cond_wait_impl; 132fe9959a2SEmilio G. Cota 133fe9959a2SEmilio G. Cota /* 134fe9959a2SEmilio G. Cota * It pays off to _not_ hash callsite->file; hashing a string is slow, and 135fe9959a2SEmilio G. Cota * without it we still get a pretty unique hash. 136fe9959a2SEmilio G. Cota */ 137fe9959a2SEmilio G. Cota static inline 138c971d8faSEmilio G. Cota uint32_t do_qsp_callsite_hash(const QSPCallSite *callsite, uint64_t ab) 139fe9959a2SEmilio G. Cota { 140c971d8faSEmilio G. Cota uint64_t cd = (uint64_t)(uintptr_t)callsite->obj; 141fe9959a2SEmilio G. Cota uint32_t e = callsite->line; 142fe9959a2SEmilio G. Cota uint32_t f = callsite->type; 143fe9959a2SEmilio G. Cota 144c971d8faSEmilio G. Cota return qemu_xxhash6(ab, cd, e, f); 145fe9959a2SEmilio G. Cota } 146fe9959a2SEmilio G. Cota 147fe9959a2SEmilio G. Cota static inline 148fe9959a2SEmilio G. Cota uint32_t qsp_callsite_hash(const QSPCallSite *callsite) 149fe9959a2SEmilio G. Cota { 150fe9959a2SEmilio G. Cota return do_qsp_callsite_hash(callsite, 0); 151fe9959a2SEmilio G. Cota } 152fe9959a2SEmilio G. Cota 153fe9959a2SEmilio G. Cota static inline uint32_t do_qsp_entry_hash(const QSPEntry *entry, uint64_t a) 154fe9959a2SEmilio G. Cota { 155fe9959a2SEmilio G. Cota return do_qsp_callsite_hash(entry->callsite, a); 156fe9959a2SEmilio G. Cota } 157fe9959a2SEmilio G. Cota 158fe9959a2SEmilio G. Cota static uint32_t qsp_entry_hash(const QSPEntry *entry) 159fe9959a2SEmilio G. Cota { 160fe9959a2SEmilio G. Cota return do_qsp_entry_hash(entry, (uint64_t)(uintptr_t)entry->thread_ptr); 161fe9959a2SEmilio G. Cota } 162fe9959a2SEmilio G. Cota 163fe9959a2SEmilio G. Cota static uint32_t qsp_entry_no_thread_hash(const QSPEntry *entry) 164fe9959a2SEmilio G. Cota { 165fe9959a2SEmilio G. Cota return do_qsp_entry_hash(entry, 0); 166fe9959a2SEmilio G. Cota } 167fe9959a2SEmilio G. Cota 168d557de4aSEmilio G. Cota /* without the objects we need to hash the file name to get a decent hash */ 169d557de4aSEmilio G. Cota static uint32_t qsp_entry_no_thread_obj_hash(const QSPEntry *entry) 170d557de4aSEmilio G. Cota { 171d557de4aSEmilio G. Cota const QSPCallSite *callsite = entry->callsite; 172c971d8faSEmilio G. Cota uint64_t ab = g_str_hash(callsite->file); 173c971d8faSEmilio G. Cota uint64_t cd = callsite->line; 174d557de4aSEmilio G. Cota uint32_t e = callsite->type; 175d557de4aSEmilio G. Cota 176c971d8faSEmilio G. Cota return qemu_xxhash5(ab, cd, e); 177d557de4aSEmilio G. Cota } 178d557de4aSEmilio G. Cota 179fe9959a2SEmilio G. Cota static bool qsp_callsite_cmp(const void *ap, const void *bp) 180fe9959a2SEmilio G. Cota { 181fe9959a2SEmilio G. Cota const QSPCallSite *a = ap; 182fe9959a2SEmilio G. Cota const QSPCallSite *b = bp; 183fe9959a2SEmilio G. Cota 184fe9959a2SEmilio G. Cota return a == b || 185fe9959a2SEmilio G. Cota (a->obj == b->obj && 186fe9959a2SEmilio G. Cota a->line == b->line && 187fe9959a2SEmilio G. Cota a->type == b->type && 188fe9959a2SEmilio G. Cota (a->file == b->file || !strcmp(a->file, b->file))); 189fe9959a2SEmilio G. Cota } 190fe9959a2SEmilio G. Cota 191d557de4aSEmilio G. Cota static bool qsp_callsite_no_obj_cmp(const void *ap, const void *bp) 192d557de4aSEmilio G. Cota { 193d557de4aSEmilio G. Cota const QSPCallSite *a = ap; 194d557de4aSEmilio G. Cota const QSPCallSite *b = bp; 195d557de4aSEmilio G. Cota 196d557de4aSEmilio G. Cota return a == b || 197d557de4aSEmilio G. Cota (a->line == b->line && 198d557de4aSEmilio G. Cota a->type == b->type && 199d557de4aSEmilio G. Cota (a->file == b->file || !strcmp(a->file, b->file))); 200d557de4aSEmilio G. Cota } 201d557de4aSEmilio G. Cota 202fe9959a2SEmilio G. Cota static bool qsp_entry_no_thread_cmp(const void *ap, const void *bp) 203fe9959a2SEmilio G. Cota { 204fe9959a2SEmilio G. Cota const QSPEntry *a = ap; 205fe9959a2SEmilio G. Cota const QSPEntry *b = bp; 206fe9959a2SEmilio G. Cota 207fe9959a2SEmilio G. Cota return qsp_callsite_cmp(a->callsite, b->callsite); 208fe9959a2SEmilio G. Cota } 209fe9959a2SEmilio G. Cota 210d557de4aSEmilio G. Cota static bool qsp_entry_no_thread_obj_cmp(const void *ap, const void *bp) 211d557de4aSEmilio G. Cota { 212d557de4aSEmilio G. Cota const QSPEntry *a = ap; 213d557de4aSEmilio G. Cota const QSPEntry *b = bp; 214d557de4aSEmilio G. Cota 215d557de4aSEmilio G. Cota return qsp_callsite_no_obj_cmp(a->callsite, b->callsite); 216d557de4aSEmilio G. Cota } 217d557de4aSEmilio G. Cota 218fe9959a2SEmilio G. Cota static bool qsp_entry_cmp(const void *ap, const void *bp) 219fe9959a2SEmilio G. Cota { 220fe9959a2SEmilio G. Cota const QSPEntry *a = ap; 221fe9959a2SEmilio G. Cota const QSPEntry *b = bp; 222fe9959a2SEmilio G. Cota 223fe9959a2SEmilio G. Cota return a->thread_ptr == b->thread_ptr && 224fe9959a2SEmilio G. Cota qsp_callsite_cmp(a->callsite, b->callsite); 225fe9959a2SEmilio G. Cota } 226fe9959a2SEmilio G. Cota 227fe9959a2SEmilio G. Cota /* 228fe9959a2SEmilio G. Cota * Normally we'd call this from a constructor function, but we want it to work 229fe9959a2SEmilio G. Cota * via libutil as well. 230fe9959a2SEmilio G. Cota */ 231fe9959a2SEmilio G. Cota static void qsp_do_init(void) 232fe9959a2SEmilio G. Cota { 233fe9959a2SEmilio G. Cota /* make sure this file's path in the tree is up to date with QSP_REL_PATH */ 234fe9959a2SEmilio G. Cota g_assert(strstr(__FILE__, QSP_REL_PATH)); 235fe9959a2SEmilio G. Cota qsp_qemu_path_len = strlen(__FILE__) - strlen(QSP_REL_PATH); 236fe9959a2SEmilio G. Cota 237fe9959a2SEmilio G. Cota qht_init(&qsp_ht, qsp_entry_cmp, QSP_INITIAL_SIZE, 238fe9959a2SEmilio G. Cota QHT_MODE_AUTO_RESIZE | QHT_MODE_RAW_MUTEXES); 239fe9959a2SEmilio G. Cota qht_init(&qsp_callsite_ht, qsp_callsite_cmp, QSP_INITIAL_SIZE, 240fe9959a2SEmilio G. Cota QHT_MODE_AUTO_RESIZE | QHT_MODE_RAW_MUTEXES); 241fe9959a2SEmilio G. Cota } 242fe9959a2SEmilio G. Cota 243fe9959a2SEmilio G. Cota static __attribute__((noinline)) void qsp_init__slowpath(void) 244fe9959a2SEmilio G. Cota { 245fe9959a2SEmilio G. Cota if (atomic_cmpxchg(&qsp_initializing, false, true) == false) { 246fe9959a2SEmilio G. Cota qsp_do_init(); 247fe9959a2SEmilio G. Cota atomic_set(&qsp_initialized, true); 248fe9959a2SEmilio G. Cota } else { 249fe9959a2SEmilio G. Cota while (!atomic_read(&qsp_initialized)) { 250fe9959a2SEmilio G. Cota cpu_relax(); 251fe9959a2SEmilio G. Cota } 252fe9959a2SEmilio G. Cota } 253fe9959a2SEmilio G. Cota } 254fe9959a2SEmilio G. Cota 255fe9959a2SEmilio G. Cota /* qsp_init() must be called from _all_ exported functions */ 256fe9959a2SEmilio G. Cota static inline void qsp_init(void) 257fe9959a2SEmilio G. Cota { 258fe9959a2SEmilio G. Cota if (likely(atomic_read(&qsp_initialized))) { 259fe9959a2SEmilio G. Cota return; 260fe9959a2SEmilio G. Cota } 261fe9959a2SEmilio G. Cota qsp_init__slowpath(); 262fe9959a2SEmilio G. Cota } 263fe9959a2SEmilio G. Cota 264fe9959a2SEmilio G. Cota static QSPCallSite *qsp_callsite_find(const QSPCallSite *orig) 265fe9959a2SEmilio G. Cota { 266fe9959a2SEmilio G. Cota QSPCallSite *callsite; 267fe9959a2SEmilio G. Cota uint32_t hash; 268fe9959a2SEmilio G. Cota 269fe9959a2SEmilio G. Cota hash = qsp_callsite_hash(orig); 270fe9959a2SEmilio G. Cota callsite = qht_lookup(&qsp_callsite_ht, orig, hash); 271fe9959a2SEmilio G. Cota if (callsite == NULL) { 272fe9959a2SEmilio G. Cota void *existing = NULL; 273fe9959a2SEmilio G. Cota 274fe9959a2SEmilio G. Cota callsite = g_new(QSPCallSite, 1); 275fe9959a2SEmilio G. Cota memcpy(callsite, orig, sizeof(*callsite)); 276fe9959a2SEmilio G. Cota qht_insert(&qsp_callsite_ht, callsite, hash, &existing); 277fe9959a2SEmilio G. Cota if (unlikely(existing)) { 278fe9959a2SEmilio G. Cota g_free(callsite); 279fe9959a2SEmilio G. Cota callsite = existing; 280fe9959a2SEmilio G. Cota } 281fe9959a2SEmilio G. Cota } 282fe9959a2SEmilio G. Cota return callsite; 283fe9959a2SEmilio G. Cota } 284fe9959a2SEmilio G. Cota 285fe9959a2SEmilio G. Cota static QSPEntry * 286fe9959a2SEmilio G. Cota qsp_entry_create(struct qht *ht, const QSPEntry *entry, uint32_t hash) 287fe9959a2SEmilio G. Cota { 288fe9959a2SEmilio G. Cota QSPEntry *e; 289fe9959a2SEmilio G. Cota void *existing = NULL; 290fe9959a2SEmilio G. Cota 291fe9959a2SEmilio G. Cota e = g_new0(QSPEntry, 1); 292fe9959a2SEmilio G. Cota e->thread_ptr = entry->thread_ptr; 293fe9959a2SEmilio G. Cota e->callsite = qsp_callsite_find(entry->callsite); 294fe9959a2SEmilio G. Cota 295fe9959a2SEmilio G. Cota qht_insert(ht, e, hash, &existing); 296fe9959a2SEmilio G. Cota if (unlikely(existing)) { 297fe9959a2SEmilio G. Cota g_free(e); 298fe9959a2SEmilio G. Cota e = existing; 299fe9959a2SEmilio G. Cota } 300fe9959a2SEmilio G. Cota return e; 301fe9959a2SEmilio G. Cota } 302fe9959a2SEmilio G. Cota 303fe9959a2SEmilio G. Cota static QSPEntry * 304fe9959a2SEmilio G. Cota qsp_entry_find(struct qht *ht, const QSPEntry *entry, uint32_t hash) 305fe9959a2SEmilio G. Cota { 306fe9959a2SEmilio G. Cota QSPEntry *e; 307fe9959a2SEmilio G. Cota 308fe9959a2SEmilio G. Cota e = qht_lookup(ht, entry, hash); 309fe9959a2SEmilio G. Cota if (e == NULL) { 310fe9959a2SEmilio G. Cota e = qsp_entry_create(ht, entry, hash); 311fe9959a2SEmilio G. Cota } 312fe9959a2SEmilio G. Cota return e; 313fe9959a2SEmilio G. Cota } 314fe9959a2SEmilio G. Cota 315fe9959a2SEmilio G. Cota /* 316fe9959a2SEmilio G. Cota * Note: Entries are never removed, so callers do not have to be in an RCU 317fe9959a2SEmilio G. Cota * read-side critical section. 318fe9959a2SEmilio G. Cota */ 319fe9959a2SEmilio G. Cota static QSPEntry *qsp_entry_get(const void *obj, const char *file, int line, 320fe9959a2SEmilio G. Cota enum QSPType type) 321fe9959a2SEmilio G. Cota { 322fe9959a2SEmilio G. Cota QSPCallSite callsite = { 323fe9959a2SEmilio G. Cota .obj = obj, 324fe9959a2SEmilio G. Cota .file = file, 325fe9959a2SEmilio G. Cota .line = line, 326fe9959a2SEmilio G. Cota .type = type, 327fe9959a2SEmilio G. Cota }; 328fe9959a2SEmilio G. Cota QSPEntry orig; 329fe9959a2SEmilio G. Cota uint32_t hash; 330fe9959a2SEmilio G. Cota 331fe9959a2SEmilio G. Cota qsp_init(); 332fe9959a2SEmilio G. Cota 333fe9959a2SEmilio G. Cota orig.thread_ptr = &qsp_thread; 334fe9959a2SEmilio G. Cota orig.callsite = &callsite; 335fe9959a2SEmilio G. Cota 336fe9959a2SEmilio G. Cota hash = qsp_entry_hash(&orig); 337fe9959a2SEmilio G. Cota return qsp_entry_find(&qsp_ht, &orig, hash); 338fe9959a2SEmilio G. Cota } 339fe9959a2SEmilio G. Cota 340fe9959a2SEmilio G. Cota /* 341fe9959a2SEmilio G. Cota * @e is in the global hash table; it is only written to by the current thread, 342fe9959a2SEmilio G. Cota * so we write to it atomically (as in "write once") to prevent torn reads. 343fe9959a2SEmilio G. Cota */ 344fe9959a2SEmilio G. Cota static inline void do_qsp_entry_record(QSPEntry *e, int64_t delta, bool acq) 345fe9959a2SEmilio G. Cota { 346ac8c7748SEmilio G. Cota atomic_set_u64(&e->ns, e->ns + delta); 347fe9959a2SEmilio G. Cota if (acq) { 348ac8c7748SEmilio G. Cota atomic_set_u64(&e->n_acqs, e->n_acqs + 1); 349fe9959a2SEmilio G. Cota } 350fe9959a2SEmilio G. Cota } 351fe9959a2SEmilio G. Cota 352fe9959a2SEmilio G. Cota static inline void qsp_entry_record(QSPEntry *e, int64_t delta) 353fe9959a2SEmilio G. Cota { 354fe9959a2SEmilio G. Cota do_qsp_entry_record(e, delta, true); 355fe9959a2SEmilio G. Cota } 356fe9959a2SEmilio G. Cota 357fe9959a2SEmilio G. Cota #define QSP_GEN_VOID(type_, qsp_t_, func_, impl_) \ 358fe9959a2SEmilio G. Cota static void func_(type_ *obj, const char *file, int line) \ 359fe9959a2SEmilio G. Cota { \ 360fe9959a2SEmilio G. Cota QSPEntry *e; \ 361fe9959a2SEmilio G. Cota int64_t t0, t1; \ 362fe9959a2SEmilio G. Cota \ 363fe9959a2SEmilio G. Cota t0 = get_clock(); \ 364fe9959a2SEmilio G. Cota impl_(obj, file, line); \ 365fe9959a2SEmilio G. Cota t1 = get_clock(); \ 366fe9959a2SEmilio G. Cota \ 367fe9959a2SEmilio G. Cota e = qsp_entry_get(obj, file, line, qsp_t_); \ 368fe9959a2SEmilio G. Cota qsp_entry_record(e, t1 - t0); \ 369fe9959a2SEmilio G. Cota } 370fe9959a2SEmilio G. Cota 371fe9959a2SEmilio G. Cota #define QSP_GEN_RET1(type_, qsp_t_, func_, impl_) \ 372fe9959a2SEmilio G. Cota static int func_(type_ *obj, const char *file, int line) \ 373fe9959a2SEmilio G. Cota { \ 374fe9959a2SEmilio G. Cota QSPEntry *e; \ 375fe9959a2SEmilio G. Cota int64_t t0, t1; \ 376fe9959a2SEmilio G. Cota int err; \ 377fe9959a2SEmilio G. Cota \ 378fe9959a2SEmilio G. Cota t0 = get_clock(); \ 379fe9959a2SEmilio G. Cota err = impl_(obj, file, line); \ 380fe9959a2SEmilio G. Cota t1 = get_clock(); \ 381fe9959a2SEmilio G. Cota \ 382fe9959a2SEmilio G. Cota e = qsp_entry_get(obj, file, line, qsp_t_); \ 383fe9959a2SEmilio G. Cota do_qsp_entry_record(e, t1 - t0, !err); \ 384fe9959a2SEmilio G. Cota return err; \ 385fe9959a2SEmilio G. Cota } 386fe9959a2SEmilio G. Cota 387cb764d06SEmilio G. Cota QSP_GEN_VOID(QemuMutex, QSP_BQL_MUTEX, qsp_bql_mutex_lock, qemu_mutex_lock_impl) 388fe9959a2SEmilio G. Cota QSP_GEN_VOID(QemuMutex, QSP_MUTEX, qsp_mutex_lock, qemu_mutex_lock_impl) 389fe9959a2SEmilio G. Cota QSP_GEN_RET1(QemuMutex, QSP_MUTEX, qsp_mutex_trylock, qemu_mutex_trylock_impl) 390fe9959a2SEmilio G. Cota 391fe9959a2SEmilio G. Cota QSP_GEN_VOID(QemuRecMutex, QSP_REC_MUTEX, qsp_rec_mutex_lock, 392fe9959a2SEmilio G. Cota qemu_rec_mutex_lock_impl) 393fe9959a2SEmilio G. Cota QSP_GEN_RET1(QemuRecMutex, QSP_REC_MUTEX, qsp_rec_mutex_trylock, 394fe9959a2SEmilio G. Cota qemu_rec_mutex_trylock_impl) 395fe9959a2SEmilio G. Cota 396fe9959a2SEmilio G. Cota #undef QSP_GEN_RET1 397fe9959a2SEmilio G. Cota #undef QSP_GEN_VOID 398fe9959a2SEmilio G. Cota 399fe9959a2SEmilio G. Cota static void 400fe9959a2SEmilio G. Cota qsp_cond_wait(QemuCond *cond, QemuMutex *mutex, const char *file, int line) 401fe9959a2SEmilio G. Cota { 402fe9959a2SEmilio G. Cota QSPEntry *e; 403fe9959a2SEmilio G. Cota int64_t t0, t1; 404fe9959a2SEmilio G. Cota 405fe9959a2SEmilio G. Cota t0 = get_clock(); 406fe9959a2SEmilio G. Cota qemu_cond_wait_impl(cond, mutex, file, line); 407fe9959a2SEmilio G. Cota t1 = get_clock(); 408fe9959a2SEmilio G. Cota 409fe9959a2SEmilio G. Cota e = qsp_entry_get(cond, file, line, QSP_CONDVAR); 410fe9959a2SEmilio G. Cota qsp_entry_record(e, t1 - t0); 411fe9959a2SEmilio G. Cota } 412fe9959a2SEmilio G. Cota 413fe9959a2SEmilio G. Cota bool qsp_is_enabled(void) 414fe9959a2SEmilio G. Cota { 415fe9959a2SEmilio G. Cota return atomic_read(&qemu_mutex_lock_func) == qsp_mutex_lock; 416fe9959a2SEmilio G. Cota } 417fe9959a2SEmilio G. Cota 418fe9959a2SEmilio G. Cota void qsp_enable(void) 419fe9959a2SEmilio G. Cota { 420fe9959a2SEmilio G. Cota atomic_set(&qemu_mutex_lock_func, qsp_mutex_lock); 421fe9959a2SEmilio G. Cota atomic_set(&qemu_mutex_trylock_func, qsp_mutex_trylock); 422cb764d06SEmilio G. Cota atomic_set(&qemu_bql_mutex_lock_func, qsp_bql_mutex_lock); 423fe9959a2SEmilio G. Cota atomic_set(&qemu_rec_mutex_lock_func, qsp_rec_mutex_lock); 424fe9959a2SEmilio G. Cota atomic_set(&qemu_rec_mutex_trylock_func, qsp_rec_mutex_trylock); 425fe9959a2SEmilio G. Cota atomic_set(&qemu_cond_wait_func, qsp_cond_wait); 426fe9959a2SEmilio G. Cota } 427fe9959a2SEmilio G. Cota 428fe9959a2SEmilio G. Cota void qsp_disable(void) 429fe9959a2SEmilio G. Cota { 430fe9959a2SEmilio G. Cota atomic_set(&qemu_mutex_lock_func, qemu_mutex_lock_impl); 431fe9959a2SEmilio G. Cota atomic_set(&qemu_mutex_trylock_func, qemu_mutex_trylock_impl); 432cb764d06SEmilio G. Cota atomic_set(&qemu_bql_mutex_lock_func, qemu_mutex_lock_impl); 433fe9959a2SEmilio G. Cota atomic_set(&qemu_rec_mutex_lock_func, qemu_rec_mutex_lock_impl); 434fe9959a2SEmilio G. Cota atomic_set(&qemu_rec_mutex_trylock_func, qemu_rec_mutex_trylock_impl); 435fe9959a2SEmilio G. Cota atomic_set(&qemu_cond_wait_func, qemu_cond_wait_impl); 436fe9959a2SEmilio G. Cota } 437fe9959a2SEmilio G. Cota 438fe9959a2SEmilio G. Cota static gint qsp_tree_cmp(gconstpointer ap, gconstpointer bp, gpointer up) 439fe9959a2SEmilio G. Cota { 440fe9959a2SEmilio G. Cota const QSPEntry *a = ap; 441fe9959a2SEmilio G. Cota const QSPEntry *b = bp; 4420a22777cSEmilio G. Cota enum QSPSortBy sort_by = *(enum QSPSortBy *)up; 443fe9959a2SEmilio G. Cota const QSPCallSite *ca; 444fe9959a2SEmilio G. Cota const QSPCallSite *cb; 445fe9959a2SEmilio G. Cota 4460a22777cSEmilio G. Cota switch (sort_by) { 4470a22777cSEmilio G. Cota case QSP_SORT_BY_TOTAL_WAIT_TIME: 448fe9959a2SEmilio G. Cota if (a->ns > b->ns) { 449fe9959a2SEmilio G. Cota return -1; 450fe9959a2SEmilio G. Cota } else if (a->ns < b->ns) { 451fe9959a2SEmilio G. Cota return 1; 452fe9959a2SEmilio G. Cota } 4530a22777cSEmilio G. Cota break; 4540a22777cSEmilio G. Cota case QSP_SORT_BY_AVG_WAIT_TIME: 4550a22777cSEmilio G. Cota { 4560a22777cSEmilio G. Cota double avg_a = a->n_acqs ? a->ns / a->n_acqs : 0; 4570a22777cSEmilio G. Cota double avg_b = b->n_acqs ? b->ns / b->n_acqs : 0; 4580a22777cSEmilio G. Cota 4590a22777cSEmilio G. Cota if (avg_a > avg_b) { 4600a22777cSEmilio G. Cota return -1; 4610a22777cSEmilio G. Cota } else if (avg_a < avg_b) { 4620a22777cSEmilio G. Cota return 1; 4630a22777cSEmilio G. Cota } 4640a22777cSEmilio G. Cota break; 4650a22777cSEmilio G. Cota } 4660a22777cSEmilio G. Cota default: 4670a22777cSEmilio G. Cota g_assert_not_reached(); 4680a22777cSEmilio G. Cota } 4690a22777cSEmilio G. Cota 470fe9959a2SEmilio G. Cota ca = a->callsite; 471fe9959a2SEmilio G. Cota cb = b->callsite; 472fe9959a2SEmilio G. Cota /* Break the tie with the object's address */ 473fe9959a2SEmilio G. Cota if (ca->obj < cb->obj) { 474fe9959a2SEmilio G. Cota return -1; 475fe9959a2SEmilio G. Cota } else if (ca->obj > cb->obj) { 476fe9959a2SEmilio G. Cota return 1; 477fe9959a2SEmilio G. Cota } else { 478fe9959a2SEmilio G. Cota int cmp; 479fe9959a2SEmilio G. Cota 480fe9959a2SEmilio G. Cota /* same obj. Break the tie with the callsite's file */ 481fe9959a2SEmilio G. Cota cmp = strcmp(ca->file, cb->file); 482fe9959a2SEmilio G. Cota if (cmp) { 483fe9959a2SEmilio G. Cota return cmp; 484fe9959a2SEmilio G. Cota } 485fe9959a2SEmilio G. Cota /* same callsite file. Break the tie with the callsite's line */ 486fe9959a2SEmilio G. Cota g_assert(ca->line != cb->line); 487fe9959a2SEmilio G. Cota if (ca->line < cb->line) { 488fe9959a2SEmilio G. Cota return -1; 489fe9959a2SEmilio G. Cota } else if (ca->line > cb->line) { 490fe9959a2SEmilio G. Cota return 1; 491fe9959a2SEmilio G. Cota } else { 492fe9959a2SEmilio G. Cota /* break the tie with the callsite's type */ 493fe9959a2SEmilio G. Cota return cb->type - ca->type; 494fe9959a2SEmilio G. Cota } 495fe9959a2SEmilio G. Cota } 496fe9959a2SEmilio G. Cota } 497fe9959a2SEmilio G. Cota 49878255ba2SEmilio G. Cota static void qsp_sort(void *p, uint32_t h, void *userp) 499fe9959a2SEmilio G. Cota { 500fe9959a2SEmilio G. Cota QSPEntry *e = p; 501fe9959a2SEmilio G. Cota GTree *tree = userp; 502fe9959a2SEmilio G. Cota 503fe9959a2SEmilio G. Cota g_tree_insert(tree, e, NULL); 504fe9959a2SEmilio G. Cota } 505fe9959a2SEmilio G. Cota 50678255ba2SEmilio G. Cota static void qsp_aggregate(void *p, uint32_t h, void *up) 507fe9959a2SEmilio G. Cota { 508fe9959a2SEmilio G. Cota struct qht *ht = up; 509fe9959a2SEmilio G. Cota const QSPEntry *e = p; 510fe9959a2SEmilio G. Cota QSPEntry *agg; 511fe9959a2SEmilio G. Cota uint32_t hash; 512fe9959a2SEmilio G. Cota 513fe9959a2SEmilio G. Cota hash = qsp_entry_no_thread_hash(e); 514fe9959a2SEmilio G. Cota agg = qsp_entry_find(ht, e, hash); 515ac8c7748SEmilio G. Cota /* 516ac8c7748SEmilio G. Cota * The entry is in the global hash table; read from it atomically (as in 517ac8c7748SEmilio G. Cota * "read once"). 518ac8c7748SEmilio G. Cota */ 519ac8c7748SEmilio G. Cota agg->ns += atomic_read_u64(&e->ns); 520ac8c7748SEmilio G. Cota agg->n_acqs += atomic_read_u64(&e->n_acqs); 521fe9959a2SEmilio G. Cota } 522fe9959a2SEmilio G. Cota 52378255ba2SEmilio G. Cota static void qsp_iter_diff(void *p, uint32_t hash, void *htp) 524996e8d9aSEmilio G. Cota { 525996e8d9aSEmilio G. Cota struct qht *ht = htp; 526996e8d9aSEmilio G. Cota QSPEntry *old = p; 527996e8d9aSEmilio G. Cota QSPEntry *new; 528996e8d9aSEmilio G. Cota 529996e8d9aSEmilio G. Cota new = qht_lookup(ht, old, hash); 530996e8d9aSEmilio G. Cota /* entries are never deleted, so we must have this one */ 531996e8d9aSEmilio G. Cota g_assert(new != NULL); 532996e8d9aSEmilio G. Cota /* our reading of the stats happened after the snapshot was taken */ 533996e8d9aSEmilio G. Cota g_assert(new->n_acqs >= old->n_acqs); 534996e8d9aSEmilio G. Cota g_assert(new->ns >= old->ns); 535996e8d9aSEmilio G. Cota 536996e8d9aSEmilio G. Cota new->n_acqs -= old->n_acqs; 537996e8d9aSEmilio G. Cota new->ns -= old->ns; 538996e8d9aSEmilio G. Cota 539996e8d9aSEmilio G. Cota /* No point in reporting an empty entry */ 540996e8d9aSEmilio G. Cota if (new->n_acqs == 0 && new->ns == 0) { 541996e8d9aSEmilio G. Cota bool removed = qht_remove(ht, new, hash); 542996e8d9aSEmilio G. Cota 543996e8d9aSEmilio G. Cota g_assert(removed); 544996e8d9aSEmilio G. Cota g_free(new); 545996e8d9aSEmilio G. Cota } 546996e8d9aSEmilio G. Cota } 547996e8d9aSEmilio G. Cota 548996e8d9aSEmilio G. Cota static void qsp_diff(struct qht *orig, struct qht *new) 549996e8d9aSEmilio G. Cota { 550996e8d9aSEmilio G. Cota qht_iter(orig, qsp_iter_diff, new); 551996e8d9aSEmilio G. Cota } 552996e8d9aSEmilio G. Cota 55378255ba2SEmilio G. Cota static void qsp_iter_callsite_coalesce(void *p, uint32_t h, void *htp) 554d557de4aSEmilio G. Cota { 555d557de4aSEmilio G. Cota struct qht *ht = htp; 556d557de4aSEmilio G. Cota QSPEntry *old = p; 557d557de4aSEmilio G. Cota QSPEntry *e; 558d557de4aSEmilio G. Cota uint32_t hash; 559d557de4aSEmilio G. Cota 560d557de4aSEmilio G. Cota hash = qsp_entry_no_thread_obj_hash(old); 561d557de4aSEmilio G. Cota e = qht_lookup(ht, old, hash); 562d557de4aSEmilio G. Cota if (e == NULL) { 563d557de4aSEmilio G. Cota e = qsp_entry_create(ht, old, hash); 564d557de4aSEmilio G. Cota e->n_objs = 1; 565d557de4aSEmilio G. Cota } else if (e->callsite->obj != old->callsite->obj) { 566d557de4aSEmilio G. Cota e->n_objs++; 567d557de4aSEmilio G. Cota } 568d557de4aSEmilio G. Cota e->ns += old->ns; 569d557de4aSEmilio G. Cota e->n_acqs += old->n_acqs; 570d557de4aSEmilio G. Cota } 571d557de4aSEmilio G. Cota 57278255ba2SEmilio G. Cota static void qsp_ht_delete(void *p, uint32_t h, void *htp) 573996e8d9aSEmilio G. Cota { 574996e8d9aSEmilio G. Cota g_free(p); 575996e8d9aSEmilio G. Cota } 576996e8d9aSEmilio G. Cota 577d557de4aSEmilio G. Cota static void qsp_mktree(GTree *tree, bool callsite_coalesce) 578fe9959a2SEmilio G. Cota { 579996e8d9aSEmilio G. Cota QSPSnapshot *snap; 580d557de4aSEmilio G. Cota struct qht ht, coalesce_ht; 581d557de4aSEmilio G. Cota struct qht *htp; 582fe9959a2SEmilio G. Cota 583996e8d9aSEmilio G. Cota /* 584996e8d9aSEmilio G. Cota * First, see if there's a prior snapshot, so that we read the global hash 585996e8d9aSEmilio G. Cota * table _after_ the snapshot has been created, which guarantees that 586996e8d9aSEmilio G. Cota * the entries we'll read will be a superset of the snapshot's entries. 587996e8d9aSEmilio G. Cota * 588996e8d9aSEmilio G. Cota * We must remain in an RCU read-side critical section until we're done 589996e8d9aSEmilio G. Cota * with the snapshot. 590996e8d9aSEmilio G. Cota */ 591996e8d9aSEmilio G. Cota rcu_read_lock(); 592996e8d9aSEmilio G. Cota snap = atomic_rcu_read(&qsp_snapshot); 593996e8d9aSEmilio G. Cota 594fe9959a2SEmilio G. Cota /* Aggregate all results from the global hash table into a local one */ 595fe9959a2SEmilio G. Cota qht_init(&ht, qsp_entry_no_thread_cmp, QSP_INITIAL_SIZE, 596fe9959a2SEmilio G. Cota QHT_MODE_AUTO_RESIZE | QHT_MODE_RAW_MUTEXES); 597fe9959a2SEmilio G. Cota qht_iter(&qsp_ht, qsp_aggregate, &ht); 598fe9959a2SEmilio G. Cota 599996e8d9aSEmilio G. Cota /* compute the difference wrt the snapshot, if any */ 600996e8d9aSEmilio G. Cota if (snap) { 601996e8d9aSEmilio G. Cota qsp_diff(&snap->ht, &ht); 602996e8d9aSEmilio G. Cota } 603996e8d9aSEmilio G. Cota /* done with the snapshot; RCU can reclaim it */ 604996e8d9aSEmilio G. Cota rcu_read_unlock(); 605996e8d9aSEmilio G. Cota 606d557de4aSEmilio G. Cota htp = &ht; 607d557de4aSEmilio G. Cota if (callsite_coalesce) { 608d557de4aSEmilio G. Cota qht_init(&coalesce_ht, qsp_entry_no_thread_obj_cmp, QSP_INITIAL_SIZE, 609d557de4aSEmilio G. Cota QHT_MODE_AUTO_RESIZE | QHT_MODE_RAW_MUTEXES); 610d557de4aSEmilio G. Cota qht_iter(&ht, qsp_iter_callsite_coalesce, &coalesce_ht); 611d557de4aSEmilio G. Cota 612d557de4aSEmilio G. Cota /* free the previous hash table, and point htp to coalesce_ht */ 613d557de4aSEmilio G. Cota qht_iter(&ht, qsp_ht_delete, NULL); 614d557de4aSEmilio G. Cota qht_destroy(&ht); 615d557de4aSEmilio G. Cota htp = &coalesce_ht; 616d557de4aSEmilio G. Cota } 617d557de4aSEmilio G. Cota 618fe9959a2SEmilio G. Cota /* sort the hash table elements by using a tree */ 619d557de4aSEmilio G. Cota qht_iter(htp, qsp_sort, tree); 620fe9959a2SEmilio G. Cota 621fe9959a2SEmilio G. Cota /* free the hash table, but keep the elements (those are in the tree now) */ 622d557de4aSEmilio G. Cota qht_destroy(htp); 623fe9959a2SEmilio G. Cota } 624fe9959a2SEmilio G. Cota 625fe9959a2SEmilio G. Cota /* free string with g_free */ 626fe9959a2SEmilio G. Cota static char *qsp_at(const QSPCallSite *callsite) 627fe9959a2SEmilio G. Cota { 628fe9959a2SEmilio G. Cota GString *s = g_string_new(NULL); 629fe9959a2SEmilio G. Cota const char *shortened; 630fe9959a2SEmilio G. Cota 631fe9959a2SEmilio G. Cota /* remove the absolute path to qemu */ 632fe9959a2SEmilio G. Cota if (unlikely(strlen(callsite->file) < qsp_qemu_path_len)) { 633fe9959a2SEmilio G. Cota shortened = callsite->file; 634fe9959a2SEmilio G. Cota } else { 635fe9959a2SEmilio G. Cota shortened = callsite->file + qsp_qemu_path_len; 636fe9959a2SEmilio G. Cota } 637fe9959a2SEmilio G. Cota g_string_append_printf(s, "%s:%u", shortened, callsite->line); 638fe9959a2SEmilio G. Cota return g_string_free(s, FALSE); 639fe9959a2SEmilio G. Cota } 640fe9959a2SEmilio G. Cota 641fe9959a2SEmilio G. Cota struct QSPReportEntry { 642fe9959a2SEmilio G. Cota const void *obj; 643fe9959a2SEmilio G. Cota char *callsite_at; 644fe9959a2SEmilio G. Cota const char *typename; 645fe9959a2SEmilio G. Cota double time_s; 646fe9959a2SEmilio G. Cota double ns_avg; 647fe9959a2SEmilio G. Cota uint64_t n_acqs; 648d557de4aSEmilio G. Cota unsigned int n_objs; 649fe9959a2SEmilio G. Cota }; 650fe9959a2SEmilio G. Cota typedef struct QSPReportEntry QSPReportEntry; 651fe9959a2SEmilio G. Cota 652fe9959a2SEmilio G. Cota struct QSPReport { 653fe9959a2SEmilio G. Cota QSPReportEntry *entries; 654fe9959a2SEmilio G. Cota size_t n_entries; 655fe9959a2SEmilio G. Cota size_t max_n_entries; 656fe9959a2SEmilio G. Cota }; 657fe9959a2SEmilio G. Cota typedef struct QSPReport QSPReport; 658fe9959a2SEmilio G. Cota 659fe9959a2SEmilio G. Cota static gboolean qsp_tree_report(gpointer key, gpointer value, gpointer udata) 660fe9959a2SEmilio G. Cota { 661fe9959a2SEmilio G. Cota const QSPEntry *e = key; 662fe9959a2SEmilio G. Cota QSPReport *report = udata; 663fe9959a2SEmilio G. Cota QSPReportEntry *entry; 664fe9959a2SEmilio G. Cota 665fe9959a2SEmilio G. Cota if (report->n_entries == report->max_n_entries) { 666fe9959a2SEmilio G. Cota return TRUE; 667fe9959a2SEmilio G. Cota } 668fe9959a2SEmilio G. Cota entry = &report->entries[report->n_entries]; 669fe9959a2SEmilio G. Cota report->n_entries++; 670fe9959a2SEmilio G. Cota 671fe9959a2SEmilio G. Cota entry->obj = e->callsite->obj; 672d557de4aSEmilio G. Cota entry->n_objs = e->n_objs; 673fe9959a2SEmilio G. Cota entry->callsite_at = qsp_at(e->callsite); 674fe9959a2SEmilio G. Cota entry->typename = qsp_typenames[e->callsite->type]; 675fe9959a2SEmilio G. Cota entry->time_s = e->ns * 1e-9; 676fe9959a2SEmilio G. Cota entry->n_acqs = e->n_acqs; 677fe9959a2SEmilio G. Cota entry->ns_avg = e->n_acqs ? e->ns / e->n_acqs : 0; 678fe9959a2SEmilio G. Cota return FALSE; 679fe9959a2SEmilio G. Cota } 680fe9959a2SEmilio G. Cota 681fe9959a2SEmilio G. Cota static void 682fe9959a2SEmilio G. Cota pr_report(const QSPReport *rep, FILE *f, fprintf_function pr) 683fe9959a2SEmilio G. Cota { 684fe9959a2SEmilio G. Cota char *dashes; 685fe9959a2SEmilio G. Cota size_t max_len = 0; 686fe9959a2SEmilio G. Cota int callsite_len = 0; 687fe9959a2SEmilio G. Cota int callsite_rspace; 688fe9959a2SEmilio G. Cota int n_dashes; 689fe9959a2SEmilio G. Cota size_t i; 690fe9959a2SEmilio G. Cota 691fe9959a2SEmilio G. Cota /* find out the maximum length of all 'callsite' fields */ 692fe9959a2SEmilio G. Cota for (i = 0; i < rep->n_entries; i++) { 693fe9959a2SEmilio G. Cota const QSPReportEntry *e = &rep->entries[i]; 694fe9959a2SEmilio G. Cota size_t len = strlen(e->callsite_at); 695fe9959a2SEmilio G. Cota 696fe9959a2SEmilio G. Cota if (len > max_len) { 697fe9959a2SEmilio G. Cota max_len = len; 698fe9959a2SEmilio G. Cota } 699fe9959a2SEmilio G. Cota } 700fe9959a2SEmilio G. Cota 701fe9959a2SEmilio G. Cota callsite_len = MAX(max_len, strlen("Call site")); 702fe9959a2SEmilio G. Cota /* white space to leave to the right of "Call site" */ 703fe9959a2SEmilio G. Cota callsite_rspace = callsite_len - strlen("Call site"); 704fe9959a2SEmilio G. Cota 705fe9959a2SEmilio G. Cota pr(f, "Type Object Call site%*s Wait Time (s) " 706fe9959a2SEmilio G. Cota " Count Average (us)\n", callsite_rspace, ""); 707fe9959a2SEmilio G. Cota 708fe9959a2SEmilio G. Cota /* build a horizontal rule with dashes */ 709fe9959a2SEmilio G. Cota n_dashes = 79 + callsite_rspace; 710fe9959a2SEmilio G. Cota dashes = g_malloc(n_dashes + 1); 711fe9959a2SEmilio G. Cota memset(dashes, '-', n_dashes); 712fe9959a2SEmilio G. Cota dashes[n_dashes] = '\0'; 713fe9959a2SEmilio G. Cota pr(f, "%s\n", dashes); 714fe9959a2SEmilio G. Cota 715fe9959a2SEmilio G. Cota for (i = 0; i < rep->n_entries; i++) { 716fe9959a2SEmilio G. Cota const QSPReportEntry *e = &rep->entries[i]; 717d557de4aSEmilio G. Cota GString *s = g_string_new(NULL); 718fe9959a2SEmilio G. Cota 719d557de4aSEmilio G. Cota g_string_append_printf(s, "%-9s ", e->typename); 720d557de4aSEmilio G. Cota if (e->n_objs > 1) { 721d557de4aSEmilio G. Cota g_string_append_printf(s, "[%12u]", e->n_objs); 722d557de4aSEmilio G. Cota } else { 723d557de4aSEmilio G. Cota g_string_append_printf(s, "%14p", e->obj); 724d557de4aSEmilio G. Cota } 725d557de4aSEmilio G. Cota g_string_append_printf(s, " %s%*s %13.5f %12" PRIu64 " %12.2f\n", 726d557de4aSEmilio G. Cota e->callsite_at, 727d557de4aSEmilio G. Cota callsite_len - (int)strlen(e->callsite_at), "", 728d557de4aSEmilio G. Cota e->time_s, e->n_acqs, e->ns_avg * 1e-3); 729d557de4aSEmilio G. Cota pr(f, "%s", s->str); 730d557de4aSEmilio G. Cota g_string_free(s, TRUE); 731fe9959a2SEmilio G. Cota } 732fe9959a2SEmilio G. Cota 733fe9959a2SEmilio G. Cota pr(f, "%s\n", dashes); 734fe9959a2SEmilio G. Cota g_free(dashes); 735fe9959a2SEmilio G. Cota } 736fe9959a2SEmilio G. Cota 737fe9959a2SEmilio G. Cota static void report_destroy(QSPReport *rep) 738fe9959a2SEmilio G. Cota { 739fe9959a2SEmilio G. Cota size_t i; 740fe9959a2SEmilio G. Cota 741fe9959a2SEmilio G. Cota for (i = 0; i < rep->n_entries; i++) { 742fe9959a2SEmilio G. Cota QSPReportEntry *e = &rep->entries[i]; 743fe9959a2SEmilio G. Cota 744fe9959a2SEmilio G. Cota g_free(e->callsite_at); 745fe9959a2SEmilio G. Cota } 746fe9959a2SEmilio G. Cota g_free(rep->entries); 747fe9959a2SEmilio G. Cota } 748fe9959a2SEmilio G. Cota 7490a22777cSEmilio G. Cota void qsp_report(FILE *f, fprintf_function cpu_fprintf, size_t max, 750d557de4aSEmilio G. Cota enum QSPSortBy sort_by, bool callsite_coalesce) 751fe9959a2SEmilio G. Cota { 7520a22777cSEmilio G. Cota GTree *tree = g_tree_new_full(qsp_tree_cmp, &sort_by, g_free, NULL); 753fe9959a2SEmilio G. Cota QSPReport rep; 754fe9959a2SEmilio G. Cota 755fe9959a2SEmilio G. Cota qsp_init(); 756fe9959a2SEmilio G. Cota 757fe9959a2SEmilio G. Cota rep.entries = g_new0(QSPReportEntry, max); 758fe9959a2SEmilio G. Cota rep.n_entries = 0; 759fe9959a2SEmilio G. Cota rep.max_n_entries = max; 760fe9959a2SEmilio G. Cota 761d557de4aSEmilio G. Cota qsp_mktree(tree, callsite_coalesce); 762fe9959a2SEmilio G. Cota g_tree_foreach(tree, qsp_tree_report, &rep); 763fe9959a2SEmilio G. Cota g_tree_destroy(tree); 764fe9959a2SEmilio G. Cota 765fe9959a2SEmilio G. Cota pr_report(&rep, f, cpu_fprintf); 766fe9959a2SEmilio G. Cota report_destroy(&rep); 767fe9959a2SEmilio G. Cota } 768996e8d9aSEmilio G. Cota 769996e8d9aSEmilio G. Cota static void qsp_snapshot_destroy(QSPSnapshot *snap) 770996e8d9aSEmilio G. Cota { 771996e8d9aSEmilio G. Cota qht_iter(&snap->ht, qsp_ht_delete, NULL); 772996e8d9aSEmilio G. Cota qht_destroy(&snap->ht); 773996e8d9aSEmilio G. Cota g_free(snap); 774996e8d9aSEmilio G. Cota } 775996e8d9aSEmilio G. Cota 776996e8d9aSEmilio G. Cota void qsp_reset(void) 777996e8d9aSEmilio G. Cota { 778996e8d9aSEmilio G. Cota QSPSnapshot *new = g_new(QSPSnapshot, 1); 779996e8d9aSEmilio G. Cota QSPSnapshot *old; 780996e8d9aSEmilio G. Cota 781996e8d9aSEmilio G. Cota qsp_init(); 782996e8d9aSEmilio G. Cota 783996e8d9aSEmilio G. Cota qht_init(&new->ht, qsp_entry_cmp, QSP_INITIAL_SIZE, 784996e8d9aSEmilio G. Cota QHT_MODE_AUTO_RESIZE | QHT_MODE_RAW_MUTEXES); 785996e8d9aSEmilio G. Cota 786996e8d9aSEmilio G. Cota /* take a snapshot of the current state */ 787996e8d9aSEmilio G. Cota qht_iter(&qsp_ht, qsp_aggregate, &new->ht); 788996e8d9aSEmilio G. Cota 789996e8d9aSEmilio G. Cota /* replace the previous snapshot, if any */ 790996e8d9aSEmilio G. Cota old = atomic_xchg(&qsp_snapshot, new); 791996e8d9aSEmilio G. Cota if (old) { 792996e8d9aSEmilio G. Cota call_rcu(old, qsp_snapshot_destroy, rcu); 793996e8d9aSEmilio G. Cota } 794996e8d9aSEmilio G. Cota } 795