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 */ 59ac7ff4cfSMarkus Armbruster 60fe9959a2SEmilio G. Cota #include "qemu/osdep.h" 61ac7ff4cfSMarkus Armbruster #include "qemu/qemu-print.h" 62fe9959a2SEmilio G. Cota #include "qemu/thread.h" 63fe9959a2SEmilio G. Cota #include "qemu/timer.h" 64fe9959a2SEmilio G. Cota #include "qemu/qht.h" 65996e8d9aSEmilio G. Cota #include "qemu/rcu.h" 66fe656e31SEmilio G. Cota #include "qemu/xxhash.h" 67fe9959a2SEmilio G. Cota 68fe9959a2SEmilio G. Cota enum QSPType { 69fe9959a2SEmilio G. Cota QSP_MUTEX, 70cb764d06SEmilio G. Cota QSP_BQL_MUTEX, 71fe9959a2SEmilio G. Cota QSP_REC_MUTEX, 72fe9959a2SEmilio G. Cota QSP_CONDVAR, 73fe9959a2SEmilio G. Cota }; 74fe9959a2SEmilio G. Cota 75fe9959a2SEmilio G. Cota struct QSPCallSite { 76fe9959a2SEmilio G. Cota const void *obj; 77fe9959a2SEmilio G. Cota const char *file; /* i.e. __FILE__; shortened later */ 78fe9959a2SEmilio G. Cota int line; 79fe9959a2SEmilio G. Cota enum QSPType type; 80fe9959a2SEmilio G. Cota }; 81fe9959a2SEmilio G. Cota typedef struct QSPCallSite QSPCallSite; 82fe9959a2SEmilio G. Cota 83fe9959a2SEmilio G. Cota struct QSPEntry { 84fe9959a2SEmilio G. Cota void *thread_ptr; 85fe9959a2SEmilio G. Cota const QSPCallSite *callsite; 86fe9959a2SEmilio G. Cota uint64_t n_acqs; 87fe9959a2SEmilio G. Cota uint64_t ns; 88d557de4aSEmilio G. Cota unsigned int n_objs; /* count of coalesced objs; only used for reporting */ 89fe9959a2SEmilio G. Cota }; 90fe9959a2SEmilio G. Cota typedef struct QSPEntry QSPEntry; 91fe9959a2SEmilio G. Cota 92996e8d9aSEmilio G. Cota struct QSPSnapshot { 93996e8d9aSEmilio G. Cota struct rcu_head rcu; 94996e8d9aSEmilio G. Cota struct qht ht; 95996e8d9aSEmilio G. Cota }; 96996e8d9aSEmilio G. Cota typedef struct QSPSnapshot QSPSnapshot; 97996e8d9aSEmilio G. Cota 98fe9959a2SEmilio G. Cota /* initial sizing for hash tables */ 99fe9959a2SEmilio G. Cota #define QSP_INITIAL_SIZE 64 100fe9959a2SEmilio G. Cota 101fe9959a2SEmilio G. Cota /* If this file is moved, QSP_REL_PATH should be updated accordingly */ 102fe9959a2SEmilio G. Cota #define QSP_REL_PATH "util/qsp.c" 103fe9959a2SEmilio G. Cota 104fe9959a2SEmilio G. Cota /* this file's full path. Used to present all call sites with relative paths */ 105fe9959a2SEmilio G. Cota static size_t qsp_qemu_path_len; 106fe9959a2SEmilio G. Cota 107fe9959a2SEmilio G. Cota /* the address of qsp_thread gives us a unique 'thread ID' */ 108fe9959a2SEmilio G. Cota static __thread int qsp_thread; 109fe9959a2SEmilio G. Cota 110fe9959a2SEmilio G. Cota /* 111fe9959a2SEmilio G. Cota * Call sites are the same for all threads, so we track them in a separate hash 112fe9959a2SEmilio G. Cota * table to save memory. 113fe9959a2SEmilio G. Cota */ 114fe9959a2SEmilio G. Cota static struct qht qsp_callsite_ht; 115fe9959a2SEmilio G. Cota 116fe9959a2SEmilio G. Cota static struct qht qsp_ht; 117996e8d9aSEmilio G. Cota static QSPSnapshot *qsp_snapshot; 118fe9959a2SEmilio G. Cota static bool qsp_initialized, qsp_initializing; 119fe9959a2SEmilio G. Cota 120fe9959a2SEmilio G. Cota static const char * const qsp_typenames[] = { 121fe9959a2SEmilio G. Cota [QSP_MUTEX] = "mutex", 122cb764d06SEmilio G. Cota [QSP_BQL_MUTEX] = "BQL mutex", 123fe9959a2SEmilio G. Cota [QSP_REC_MUTEX] = "rec_mutex", 124fe9959a2SEmilio G. Cota [QSP_CONDVAR] = "condvar", 125fe9959a2SEmilio G. Cota }; 126fe9959a2SEmilio G. Cota 127cb764d06SEmilio G. Cota QemuMutexLockFunc qemu_bql_mutex_lock_func = qemu_mutex_lock_impl; 128fe9959a2SEmilio G. Cota QemuMutexLockFunc qemu_mutex_lock_func = qemu_mutex_lock_impl; 129fe9959a2SEmilio G. Cota QemuMutexTrylockFunc qemu_mutex_trylock_func = qemu_mutex_trylock_impl; 130fe9959a2SEmilio G. Cota QemuRecMutexLockFunc qemu_rec_mutex_lock_func = qemu_rec_mutex_lock_impl; 131fe9959a2SEmilio G. Cota QemuRecMutexTrylockFunc qemu_rec_mutex_trylock_func = 132fe9959a2SEmilio G. Cota qemu_rec_mutex_trylock_impl; 133fe9959a2SEmilio G. Cota QemuCondWaitFunc qemu_cond_wait_func = qemu_cond_wait_impl; 134*3dcc9c6eSYury Kotov QemuCondTimedWaitFunc qemu_cond_timedwait_func = qemu_cond_timedwait_impl; 135fe9959a2SEmilio G. Cota 136fe9959a2SEmilio G. Cota /* 137fe9959a2SEmilio G. Cota * It pays off to _not_ hash callsite->file; hashing a string is slow, and 138fe9959a2SEmilio G. Cota * without it we still get a pretty unique hash. 139fe9959a2SEmilio G. Cota */ 140fe9959a2SEmilio G. Cota static inline 141c971d8faSEmilio G. Cota uint32_t do_qsp_callsite_hash(const QSPCallSite *callsite, uint64_t ab) 142fe9959a2SEmilio G. Cota { 143c971d8faSEmilio G. Cota uint64_t cd = (uint64_t)(uintptr_t)callsite->obj; 144fe9959a2SEmilio G. Cota uint32_t e = callsite->line; 145fe9959a2SEmilio G. Cota uint32_t f = callsite->type; 146fe9959a2SEmilio G. Cota 147c971d8faSEmilio G. Cota return qemu_xxhash6(ab, cd, e, f); 148fe9959a2SEmilio G. Cota } 149fe9959a2SEmilio G. Cota 150fe9959a2SEmilio G. Cota static inline 151fe9959a2SEmilio G. Cota uint32_t qsp_callsite_hash(const QSPCallSite *callsite) 152fe9959a2SEmilio G. Cota { 153fe9959a2SEmilio G. Cota return do_qsp_callsite_hash(callsite, 0); 154fe9959a2SEmilio G. Cota } 155fe9959a2SEmilio G. Cota 156fe9959a2SEmilio G. Cota static inline uint32_t do_qsp_entry_hash(const QSPEntry *entry, uint64_t a) 157fe9959a2SEmilio G. Cota { 158fe9959a2SEmilio G. Cota return do_qsp_callsite_hash(entry->callsite, a); 159fe9959a2SEmilio G. Cota } 160fe9959a2SEmilio G. Cota 161fe9959a2SEmilio G. Cota static uint32_t qsp_entry_hash(const QSPEntry *entry) 162fe9959a2SEmilio G. Cota { 163fe9959a2SEmilio G. Cota return do_qsp_entry_hash(entry, (uint64_t)(uintptr_t)entry->thread_ptr); 164fe9959a2SEmilio G. Cota } 165fe9959a2SEmilio G. Cota 166fe9959a2SEmilio G. Cota static uint32_t qsp_entry_no_thread_hash(const QSPEntry *entry) 167fe9959a2SEmilio G. Cota { 168fe9959a2SEmilio G. Cota return do_qsp_entry_hash(entry, 0); 169fe9959a2SEmilio G. Cota } 170fe9959a2SEmilio G. Cota 171d557de4aSEmilio G. Cota /* without the objects we need to hash the file name to get a decent hash */ 172d557de4aSEmilio G. Cota static uint32_t qsp_entry_no_thread_obj_hash(const QSPEntry *entry) 173d557de4aSEmilio G. Cota { 174d557de4aSEmilio G. Cota const QSPCallSite *callsite = entry->callsite; 175c971d8faSEmilio G. Cota uint64_t ab = g_str_hash(callsite->file); 176c971d8faSEmilio G. Cota uint64_t cd = callsite->line; 177d557de4aSEmilio G. Cota uint32_t e = callsite->type; 178d557de4aSEmilio G. Cota 179c971d8faSEmilio G. Cota return qemu_xxhash5(ab, cd, e); 180d557de4aSEmilio G. Cota } 181d557de4aSEmilio G. Cota 182fe9959a2SEmilio G. Cota static bool qsp_callsite_cmp(const void *ap, const void *bp) 183fe9959a2SEmilio G. Cota { 184fe9959a2SEmilio G. Cota const QSPCallSite *a = ap; 185fe9959a2SEmilio G. Cota const QSPCallSite *b = bp; 186fe9959a2SEmilio G. Cota 187fe9959a2SEmilio G. Cota return a == b || 188fe9959a2SEmilio G. Cota (a->obj == b->obj && 189fe9959a2SEmilio G. Cota a->line == b->line && 190fe9959a2SEmilio G. Cota a->type == b->type && 191fe9959a2SEmilio G. Cota (a->file == b->file || !strcmp(a->file, b->file))); 192fe9959a2SEmilio G. Cota } 193fe9959a2SEmilio G. Cota 194d557de4aSEmilio G. Cota static bool qsp_callsite_no_obj_cmp(const void *ap, const void *bp) 195d557de4aSEmilio G. Cota { 196d557de4aSEmilio G. Cota const QSPCallSite *a = ap; 197d557de4aSEmilio G. Cota const QSPCallSite *b = bp; 198d557de4aSEmilio G. Cota 199d557de4aSEmilio G. Cota return a == b || 200d557de4aSEmilio G. Cota (a->line == b->line && 201d557de4aSEmilio G. Cota a->type == b->type && 202d557de4aSEmilio G. Cota (a->file == b->file || !strcmp(a->file, b->file))); 203d557de4aSEmilio G. Cota } 204d557de4aSEmilio G. Cota 205fe9959a2SEmilio G. Cota static bool qsp_entry_no_thread_cmp(const void *ap, const void *bp) 206fe9959a2SEmilio G. Cota { 207fe9959a2SEmilio G. Cota const QSPEntry *a = ap; 208fe9959a2SEmilio G. Cota const QSPEntry *b = bp; 209fe9959a2SEmilio G. Cota 210fe9959a2SEmilio G. Cota return qsp_callsite_cmp(a->callsite, b->callsite); 211fe9959a2SEmilio G. Cota } 212fe9959a2SEmilio G. Cota 213d557de4aSEmilio G. Cota static bool qsp_entry_no_thread_obj_cmp(const void *ap, const void *bp) 214d557de4aSEmilio G. Cota { 215d557de4aSEmilio G. Cota const QSPEntry *a = ap; 216d557de4aSEmilio G. Cota const QSPEntry *b = bp; 217d557de4aSEmilio G. Cota 218d557de4aSEmilio G. Cota return qsp_callsite_no_obj_cmp(a->callsite, b->callsite); 219d557de4aSEmilio G. Cota } 220d557de4aSEmilio G. Cota 221fe9959a2SEmilio G. Cota static bool qsp_entry_cmp(const void *ap, const void *bp) 222fe9959a2SEmilio G. Cota { 223fe9959a2SEmilio G. Cota const QSPEntry *a = ap; 224fe9959a2SEmilio G. Cota const QSPEntry *b = bp; 225fe9959a2SEmilio G. Cota 226fe9959a2SEmilio G. Cota return a->thread_ptr == b->thread_ptr && 227fe9959a2SEmilio G. Cota qsp_callsite_cmp(a->callsite, b->callsite); 228fe9959a2SEmilio G. Cota } 229fe9959a2SEmilio G. Cota 230fe9959a2SEmilio G. Cota /* 231fe9959a2SEmilio G. Cota * Normally we'd call this from a constructor function, but we want it to work 232fe9959a2SEmilio G. Cota * via libutil as well. 233fe9959a2SEmilio G. Cota */ 234fe9959a2SEmilio G. Cota static void qsp_do_init(void) 235fe9959a2SEmilio G. Cota { 236fe9959a2SEmilio G. Cota /* make sure this file's path in the tree is up to date with QSP_REL_PATH */ 237fe9959a2SEmilio G. Cota g_assert(strstr(__FILE__, QSP_REL_PATH)); 238fe9959a2SEmilio G. Cota qsp_qemu_path_len = strlen(__FILE__) - strlen(QSP_REL_PATH); 239fe9959a2SEmilio G. Cota 240fe9959a2SEmilio G. Cota qht_init(&qsp_ht, qsp_entry_cmp, QSP_INITIAL_SIZE, 241fe9959a2SEmilio G. Cota QHT_MODE_AUTO_RESIZE | QHT_MODE_RAW_MUTEXES); 242fe9959a2SEmilio G. Cota qht_init(&qsp_callsite_ht, qsp_callsite_cmp, QSP_INITIAL_SIZE, 243fe9959a2SEmilio G. Cota QHT_MODE_AUTO_RESIZE | QHT_MODE_RAW_MUTEXES); 244fe9959a2SEmilio G. Cota } 245fe9959a2SEmilio G. Cota 246fe9959a2SEmilio G. Cota static __attribute__((noinline)) void qsp_init__slowpath(void) 247fe9959a2SEmilio G. Cota { 248fe9959a2SEmilio G. Cota if (atomic_cmpxchg(&qsp_initializing, false, true) == false) { 249fe9959a2SEmilio G. Cota qsp_do_init(); 250fe9959a2SEmilio G. Cota atomic_set(&qsp_initialized, true); 251fe9959a2SEmilio G. Cota } else { 252fe9959a2SEmilio G. Cota while (!atomic_read(&qsp_initialized)) { 253fe9959a2SEmilio G. Cota cpu_relax(); 254fe9959a2SEmilio G. Cota } 255fe9959a2SEmilio G. Cota } 256fe9959a2SEmilio G. Cota } 257fe9959a2SEmilio G. Cota 258fe9959a2SEmilio G. Cota /* qsp_init() must be called from _all_ exported functions */ 259fe9959a2SEmilio G. Cota static inline void qsp_init(void) 260fe9959a2SEmilio G. Cota { 261fe9959a2SEmilio G. Cota if (likely(atomic_read(&qsp_initialized))) { 262fe9959a2SEmilio G. Cota return; 263fe9959a2SEmilio G. Cota } 264fe9959a2SEmilio G. Cota qsp_init__slowpath(); 265fe9959a2SEmilio G. Cota } 266fe9959a2SEmilio G. Cota 267fe9959a2SEmilio G. Cota static QSPCallSite *qsp_callsite_find(const QSPCallSite *orig) 268fe9959a2SEmilio G. Cota { 269fe9959a2SEmilio G. Cota QSPCallSite *callsite; 270fe9959a2SEmilio G. Cota uint32_t hash; 271fe9959a2SEmilio G. Cota 272fe9959a2SEmilio G. Cota hash = qsp_callsite_hash(orig); 273fe9959a2SEmilio G. Cota callsite = qht_lookup(&qsp_callsite_ht, orig, hash); 274fe9959a2SEmilio G. Cota if (callsite == NULL) { 275fe9959a2SEmilio G. Cota void *existing = NULL; 276fe9959a2SEmilio G. Cota 277fe9959a2SEmilio G. Cota callsite = g_new(QSPCallSite, 1); 278fe9959a2SEmilio G. Cota memcpy(callsite, orig, sizeof(*callsite)); 279fe9959a2SEmilio G. Cota qht_insert(&qsp_callsite_ht, callsite, hash, &existing); 280fe9959a2SEmilio G. Cota if (unlikely(existing)) { 281fe9959a2SEmilio G. Cota g_free(callsite); 282fe9959a2SEmilio G. Cota callsite = existing; 283fe9959a2SEmilio G. Cota } 284fe9959a2SEmilio G. Cota } 285fe9959a2SEmilio G. Cota return callsite; 286fe9959a2SEmilio G. Cota } 287fe9959a2SEmilio G. Cota 288fe9959a2SEmilio G. Cota static QSPEntry * 289fe9959a2SEmilio G. Cota qsp_entry_create(struct qht *ht, const QSPEntry *entry, uint32_t hash) 290fe9959a2SEmilio G. Cota { 291fe9959a2SEmilio G. Cota QSPEntry *e; 292fe9959a2SEmilio G. Cota void *existing = NULL; 293fe9959a2SEmilio G. Cota 294fe9959a2SEmilio G. Cota e = g_new0(QSPEntry, 1); 295fe9959a2SEmilio G. Cota e->thread_ptr = entry->thread_ptr; 296fe9959a2SEmilio G. Cota e->callsite = qsp_callsite_find(entry->callsite); 297fe9959a2SEmilio G. Cota 298fe9959a2SEmilio G. Cota qht_insert(ht, e, hash, &existing); 299fe9959a2SEmilio G. Cota if (unlikely(existing)) { 300fe9959a2SEmilio G. Cota g_free(e); 301fe9959a2SEmilio G. Cota e = existing; 302fe9959a2SEmilio G. Cota } 303fe9959a2SEmilio G. Cota return e; 304fe9959a2SEmilio G. Cota } 305fe9959a2SEmilio G. Cota 306fe9959a2SEmilio G. Cota static QSPEntry * 307fe9959a2SEmilio G. Cota qsp_entry_find(struct qht *ht, const QSPEntry *entry, uint32_t hash) 308fe9959a2SEmilio G. Cota { 309fe9959a2SEmilio G. Cota QSPEntry *e; 310fe9959a2SEmilio G. Cota 311fe9959a2SEmilio G. Cota e = qht_lookup(ht, entry, hash); 312fe9959a2SEmilio G. Cota if (e == NULL) { 313fe9959a2SEmilio G. Cota e = qsp_entry_create(ht, entry, hash); 314fe9959a2SEmilio G. Cota } 315fe9959a2SEmilio G. Cota return e; 316fe9959a2SEmilio G. Cota } 317fe9959a2SEmilio G. Cota 318fe9959a2SEmilio G. Cota /* 319fe9959a2SEmilio G. Cota * Note: Entries are never removed, so callers do not have to be in an RCU 320fe9959a2SEmilio G. Cota * read-side critical section. 321fe9959a2SEmilio G. Cota */ 322fe9959a2SEmilio G. Cota static QSPEntry *qsp_entry_get(const void *obj, const char *file, int line, 323fe9959a2SEmilio G. Cota enum QSPType type) 324fe9959a2SEmilio G. Cota { 325fe9959a2SEmilio G. Cota QSPCallSite callsite = { 326fe9959a2SEmilio G. Cota .obj = obj, 327fe9959a2SEmilio G. Cota .file = file, 328fe9959a2SEmilio G. Cota .line = line, 329fe9959a2SEmilio G. Cota .type = type, 330fe9959a2SEmilio G. Cota }; 331fe9959a2SEmilio G. Cota QSPEntry orig; 332fe9959a2SEmilio G. Cota uint32_t hash; 333fe9959a2SEmilio G. Cota 334fe9959a2SEmilio G. Cota qsp_init(); 335fe9959a2SEmilio G. Cota 336fe9959a2SEmilio G. Cota orig.thread_ptr = &qsp_thread; 337fe9959a2SEmilio G. Cota orig.callsite = &callsite; 338fe9959a2SEmilio G. Cota 339fe9959a2SEmilio G. Cota hash = qsp_entry_hash(&orig); 340fe9959a2SEmilio G. Cota return qsp_entry_find(&qsp_ht, &orig, hash); 341fe9959a2SEmilio G. Cota } 342fe9959a2SEmilio G. Cota 343fe9959a2SEmilio G. Cota /* 344fe9959a2SEmilio G. Cota * @e is in the global hash table; it is only written to by the current thread, 345fe9959a2SEmilio G. Cota * so we write to it atomically (as in "write once") to prevent torn reads. 346fe9959a2SEmilio G. Cota */ 347fe9959a2SEmilio G. Cota static inline void do_qsp_entry_record(QSPEntry *e, int64_t delta, bool acq) 348fe9959a2SEmilio G. Cota { 349ac8c7748SEmilio G. Cota atomic_set_u64(&e->ns, e->ns + delta); 350fe9959a2SEmilio G. Cota if (acq) { 351ac8c7748SEmilio G. Cota atomic_set_u64(&e->n_acqs, e->n_acqs + 1); 352fe9959a2SEmilio G. Cota } 353fe9959a2SEmilio G. Cota } 354fe9959a2SEmilio G. Cota 355fe9959a2SEmilio G. Cota static inline void qsp_entry_record(QSPEntry *e, int64_t delta) 356fe9959a2SEmilio G. Cota { 357fe9959a2SEmilio G. Cota do_qsp_entry_record(e, delta, true); 358fe9959a2SEmilio G. Cota } 359fe9959a2SEmilio G. Cota 360fe9959a2SEmilio G. Cota #define QSP_GEN_VOID(type_, qsp_t_, func_, impl_) \ 361fe9959a2SEmilio G. Cota static void func_(type_ *obj, const char *file, int line) \ 362fe9959a2SEmilio G. Cota { \ 363fe9959a2SEmilio G. Cota QSPEntry *e; \ 364fe9959a2SEmilio G. Cota int64_t t0, t1; \ 365fe9959a2SEmilio G. Cota \ 366fe9959a2SEmilio G. Cota t0 = get_clock(); \ 367fe9959a2SEmilio G. Cota impl_(obj, file, line); \ 368fe9959a2SEmilio G. Cota t1 = get_clock(); \ 369fe9959a2SEmilio G. Cota \ 370fe9959a2SEmilio G. Cota e = qsp_entry_get(obj, file, line, qsp_t_); \ 371fe9959a2SEmilio G. Cota qsp_entry_record(e, t1 - t0); \ 372fe9959a2SEmilio G. Cota } 373fe9959a2SEmilio G. Cota 374fe9959a2SEmilio G. Cota #define QSP_GEN_RET1(type_, qsp_t_, func_, impl_) \ 375fe9959a2SEmilio G. Cota static int func_(type_ *obj, const char *file, int line) \ 376fe9959a2SEmilio G. Cota { \ 377fe9959a2SEmilio G. Cota QSPEntry *e; \ 378fe9959a2SEmilio G. Cota int64_t t0, t1; \ 379fe9959a2SEmilio G. Cota int err; \ 380fe9959a2SEmilio G. Cota \ 381fe9959a2SEmilio G. Cota t0 = get_clock(); \ 382fe9959a2SEmilio G. Cota err = impl_(obj, file, line); \ 383fe9959a2SEmilio G. Cota t1 = get_clock(); \ 384fe9959a2SEmilio G. Cota \ 385fe9959a2SEmilio G. Cota e = qsp_entry_get(obj, file, line, qsp_t_); \ 386fe9959a2SEmilio G. Cota do_qsp_entry_record(e, t1 - t0, !err); \ 387fe9959a2SEmilio G. Cota return err; \ 388fe9959a2SEmilio G. Cota } 389fe9959a2SEmilio G. Cota 390cb764d06SEmilio G. Cota QSP_GEN_VOID(QemuMutex, QSP_BQL_MUTEX, qsp_bql_mutex_lock, qemu_mutex_lock_impl) 391fe9959a2SEmilio G. Cota QSP_GEN_VOID(QemuMutex, QSP_MUTEX, qsp_mutex_lock, qemu_mutex_lock_impl) 392fe9959a2SEmilio G. Cota QSP_GEN_RET1(QemuMutex, QSP_MUTEX, qsp_mutex_trylock, qemu_mutex_trylock_impl) 393fe9959a2SEmilio G. Cota 394fe9959a2SEmilio G. Cota QSP_GEN_VOID(QemuRecMutex, QSP_REC_MUTEX, qsp_rec_mutex_lock, 395fe9959a2SEmilio G. Cota qemu_rec_mutex_lock_impl) 396fe9959a2SEmilio G. Cota QSP_GEN_RET1(QemuRecMutex, QSP_REC_MUTEX, qsp_rec_mutex_trylock, 397fe9959a2SEmilio G. Cota qemu_rec_mutex_trylock_impl) 398fe9959a2SEmilio G. Cota 399fe9959a2SEmilio G. Cota #undef QSP_GEN_RET1 400fe9959a2SEmilio G. Cota #undef QSP_GEN_VOID 401fe9959a2SEmilio G. Cota 402fe9959a2SEmilio G. Cota static void 403fe9959a2SEmilio G. Cota qsp_cond_wait(QemuCond *cond, QemuMutex *mutex, const char *file, int line) 404fe9959a2SEmilio G. Cota { 405fe9959a2SEmilio G. Cota QSPEntry *e; 406fe9959a2SEmilio G. Cota int64_t t0, t1; 407fe9959a2SEmilio G. Cota 408fe9959a2SEmilio G. Cota t0 = get_clock(); 409fe9959a2SEmilio G. Cota qemu_cond_wait_impl(cond, mutex, file, line); 410fe9959a2SEmilio G. Cota t1 = get_clock(); 411fe9959a2SEmilio G. Cota 412fe9959a2SEmilio G. Cota e = qsp_entry_get(cond, file, line, QSP_CONDVAR); 413fe9959a2SEmilio G. Cota qsp_entry_record(e, t1 - t0); 414fe9959a2SEmilio G. Cota } 415fe9959a2SEmilio G. Cota 416*3dcc9c6eSYury Kotov static bool 417*3dcc9c6eSYury Kotov qsp_cond_timedwait(QemuCond *cond, QemuMutex *mutex, int ms, 418*3dcc9c6eSYury Kotov const char *file, int line) 419*3dcc9c6eSYury Kotov { 420*3dcc9c6eSYury Kotov QSPEntry *e; 421*3dcc9c6eSYury Kotov int64_t t0, t1; 422*3dcc9c6eSYury Kotov bool ret; 423*3dcc9c6eSYury Kotov 424*3dcc9c6eSYury Kotov t0 = get_clock(); 425*3dcc9c6eSYury Kotov ret = qemu_cond_timedwait_impl(cond, mutex, ms, file, line); 426*3dcc9c6eSYury Kotov t1 = get_clock(); 427*3dcc9c6eSYury Kotov 428*3dcc9c6eSYury Kotov e = qsp_entry_get(cond, file, line, QSP_CONDVAR); 429*3dcc9c6eSYury Kotov qsp_entry_record(e, t1 - t0); 430*3dcc9c6eSYury Kotov return ret; 431*3dcc9c6eSYury Kotov } 432*3dcc9c6eSYury Kotov 433fe9959a2SEmilio G. Cota bool qsp_is_enabled(void) 434fe9959a2SEmilio G. Cota { 435fe9959a2SEmilio G. Cota return atomic_read(&qemu_mutex_lock_func) == qsp_mutex_lock; 436fe9959a2SEmilio G. Cota } 437fe9959a2SEmilio G. Cota 438fe9959a2SEmilio G. Cota void qsp_enable(void) 439fe9959a2SEmilio G. Cota { 440fe9959a2SEmilio G. Cota atomic_set(&qemu_mutex_lock_func, qsp_mutex_lock); 441fe9959a2SEmilio G. Cota atomic_set(&qemu_mutex_trylock_func, qsp_mutex_trylock); 442cb764d06SEmilio G. Cota atomic_set(&qemu_bql_mutex_lock_func, qsp_bql_mutex_lock); 443fe9959a2SEmilio G. Cota atomic_set(&qemu_rec_mutex_lock_func, qsp_rec_mutex_lock); 444fe9959a2SEmilio G. Cota atomic_set(&qemu_rec_mutex_trylock_func, qsp_rec_mutex_trylock); 445fe9959a2SEmilio G. Cota atomic_set(&qemu_cond_wait_func, qsp_cond_wait); 446*3dcc9c6eSYury Kotov atomic_set(&qemu_cond_timedwait_func, qsp_cond_timedwait); 447fe9959a2SEmilio G. Cota } 448fe9959a2SEmilio G. Cota 449fe9959a2SEmilio G. Cota void qsp_disable(void) 450fe9959a2SEmilio G. Cota { 451fe9959a2SEmilio G. Cota atomic_set(&qemu_mutex_lock_func, qemu_mutex_lock_impl); 452fe9959a2SEmilio G. Cota atomic_set(&qemu_mutex_trylock_func, qemu_mutex_trylock_impl); 453cb764d06SEmilio G. Cota atomic_set(&qemu_bql_mutex_lock_func, qemu_mutex_lock_impl); 454fe9959a2SEmilio G. Cota atomic_set(&qemu_rec_mutex_lock_func, qemu_rec_mutex_lock_impl); 455fe9959a2SEmilio G. Cota atomic_set(&qemu_rec_mutex_trylock_func, qemu_rec_mutex_trylock_impl); 456fe9959a2SEmilio G. Cota atomic_set(&qemu_cond_wait_func, qemu_cond_wait_impl); 457*3dcc9c6eSYury Kotov atomic_set(&qemu_cond_timedwait_func, qemu_cond_timedwait_impl); 458fe9959a2SEmilio G. Cota } 459fe9959a2SEmilio G. Cota 460fe9959a2SEmilio G. Cota static gint qsp_tree_cmp(gconstpointer ap, gconstpointer bp, gpointer up) 461fe9959a2SEmilio G. Cota { 462fe9959a2SEmilio G. Cota const QSPEntry *a = ap; 463fe9959a2SEmilio G. Cota const QSPEntry *b = bp; 4640a22777cSEmilio G. Cota enum QSPSortBy sort_by = *(enum QSPSortBy *)up; 465fe9959a2SEmilio G. Cota const QSPCallSite *ca; 466fe9959a2SEmilio G. Cota const QSPCallSite *cb; 467fe9959a2SEmilio G. Cota 4680a22777cSEmilio G. Cota switch (sort_by) { 4690a22777cSEmilio G. Cota case QSP_SORT_BY_TOTAL_WAIT_TIME: 470fe9959a2SEmilio G. Cota if (a->ns > b->ns) { 471fe9959a2SEmilio G. Cota return -1; 472fe9959a2SEmilio G. Cota } else if (a->ns < b->ns) { 473fe9959a2SEmilio G. Cota return 1; 474fe9959a2SEmilio G. Cota } 4750a22777cSEmilio G. Cota break; 4760a22777cSEmilio G. Cota case QSP_SORT_BY_AVG_WAIT_TIME: 4770a22777cSEmilio G. Cota { 4780a22777cSEmilio G. Cota double avg_a = a->n_acqs ? a->ns / a->n_acqs : 0; 4790a22777cSEmilio G. Cota double avg_b = b->n_acqs ? b->ns / b->n_acqs : 0; 4800a22777cSEmilio G. Cota 4810a22777cSEmilio G. Cota if (avg_a > avg_b) { 4820a22777cSEmilio G. Cota return -1; 4830a22777cSEmilio G. Cota } else if (avg_a < avg_b) { 4840a22777cSEmilio G. Cota return 1; 4850a22777cSEmilio G. Cota } 4860a22777cSEmilio G. Cota break; 4870a22777cSEmilio G. Cota } 4880a22777cSEmilio G. Cota default: 4890a22777cSEmilio G. Cota g_assert_not_reached(); 4900a22777cSEmilio G. Cota } 4910a22777cSEmilio G. Cota 492fe9959a2SEmilio G. Cota ca = a->callsite; 493fe9959a2SEmilio G. Cota cb = b->callsite; 494fe9959a2SEmilio G. Cota /* Break the tie with the object's address */ 495fe9959a2SEmilio G. Cota if (ca->obj < cb->obj) { 496fe9959a2SEmilio G. Cota return -1; 497fe9959a2SEmilio G. Cota } else if (ca->obj > cb->obj) { 498fe9959a2SEmilio G. Cota return 1; 499fe9959a2SEmilio G. Cota } else { 500fe9959a2SEmilio G. Cota int cmp; 501fe9959a2SEmilio G. Cota 502fe9959a2SEmilio G. Cota /* same obj. Break the tie with the callsite's file */ 503fe9959a2SEmilio G. Cota cmp = strcmp(ca->file, cb->file); 504fe9959a2SEmilio G. Cota if (cmp) { 505fe9959a2SEmilio G. Cota return cmp; 506fe9959a2SEmilio G. Cota } 507fe9959a2SEmilio G. Cota /* same callsite file. Break the tie with the callsite's line */ 508fe9959a2SEmilio G. Cota g_assert(ca->line != cb->line); 509fe9959a2SEmilio G. Cota if (ca->line < cb->line) { 510fe9959a2SEmilio G. Cota return -1; 511fe9959a2SEmilio G. Cota } else if (ca->line > cb->line) { 512fe9959a2SEmilio G. Cota return 1; 513fe9959a2SEmilio G. Cota } else { 514fe9959a2SEmilio G. Cota /* break the tie with the callsite's type */ 515fe9959a2SEmilio G. Cota return cb->type - ca->type; 516fe9959a2SEmilio G. Cota } 517fe9959a2SEmilio G. Cota } 518fe9959a2SEmilio G. Cota } 519fe9959a2SEmilio G. Cota 52078255ba2SEmilio G. Cota static void qsp_sort(void *p, uint32_t h, void *userp) 521fe9959a2SEmilio G. Cota { 522fe9959a2SEmilio G. Cota QSPEntry *e = p; 523fe9959a2SEmilio G. Cota GTree *tree = userp; 524fe9959a2SEmilio G. Cota 525fe9959a2SEmilio G. Cota g_tree_insert(tree, e, NULL); 526fe9959a2SEmilio G. Cota } 527fe9959a2SEmilio G. Cota 52878255ba2SEmilio G. Cota static void qsp_aggregate(void *p, uint32_t h, void *up) 529fe9959a2SEmilio G. Cota { 530fe9959a2SEmilio G. Cota struct qht *ht = up; 531fe9959a2SEmilio G. Cota const QSPEntry *e = p; 532fe9959a2SEmilio G. Cota QSPEntry *agg; 533fe9959a2SEmilio G. Cota uint32_t hash; 534fe9959a2SEmilio G. Cota 535fe9959a2SEmilio G. Cota hash = qsp_entry_no_thread_hash(e); 536fe9959a2SEmilio G. Cota agg = qsp_entry_find(ht, e, hash); 537ac8c7748SEmilio G. Cota /* 538ac8c7748SEmilio G. Cota * The entry is in the global hash table; read from it atomically (as in 539ac8c7748SEmilio G. Cota * "read once"). 540ac8c7748SEmilio G. Cota */ 541ac8c7748SEmilio G. Cota agg->ns += atomic_read_u64(&e->ns); 542ac8c7748SEmilio G. Cota agg->n_acqs += atomic_read_u64(&e->n_acqs); 543fe9959a2SEmilio G. Cota } 544fe9959a2SEmilio G. Cota 54578255ba2SEmilio G. Cota static void qsp_iter_diff(void *p, uint32_t hash, void *htp) 546996e8d9aSEmilio G. Cota { 547996e8d9aSEmilio G. Cota struct qht *ht = htp; 548996e8d9aSEmilio G. Cota QSPEntry *old = p; 549996e8d9aSEmilio G. Cota QSPEntry *new; 550996e8d9aSEmilio G. Cota 551996e8d9aSEmilio G. Cota new = qht_lookup(ht, old, hash); 552996e8d9aSEmilio G. Cota /* entries are never deleted, so we must have this one */ 553996e8d9aSEmilio G. Cota g_assert(new != NULL); 554996e8d9aSEmilio G. Cota /* our reading of the stats happened after the snapshot was taken */ 555996e8d9aSEmilio G. Cota g_assert(new->n_acqs >= old->n_acqs); 556996e8d9aSEmilio G. Cota g_assert(new->ns >= old->ns); 557996e8d9aSEmilio G. Cota 558996e8d9aSEmilio G. Cota new->n_acqs -= old->n_acqs; 559996e8d9aSEmilio G. Cota new->ns -= old->ns; 560996e8d9aSEmilio G. Cota 561996e8d9aSEmilio G. Cota /* No point in reporting an empty entry */ 562996e8d9aSEmilio G. Cota if (new->n_acqs == 0 && new->ns == 0) { 563996e8d9aSEmilio G. Cota bool removed = qht_remove(ht, new, hash); 564996e8d9aSEmilio G. Cota 565996e8d9aSEmilio G. Cota g_assert(removed); 566996e8d9aSEmilio G. Cota g_free(new); 567996e8d9aSEmilio G. Cota } 568996e8d9aSEmilio G. Cota } 569996e8d9aSEmilio G. Cota 570996e8d9aSEmilio G. Cota static void qsp_diff(struct qht *orig, struct qht *new) 571996e8d9aSEmilio G. Cota { 572996e8d9aSEmilio G. Cota qht_iter(orig, qsp_iter_diff, new); 573996e8d9aSEmilio G. Cota } 574996e8d9aSEmilio G. Cota 57578255ba2SEmilio G. Cota static void qsp_iter_callsite_coalesce(void *p, uint32_t h, void *htp) 576d557de4aSEmilio G. Cota { 577d557de4aSEmilio G. Cota struct qht *ht = htp; 578d557de4aSEmilio G. Cota QSPEntry *old = p; 579d557de4aSEmilio G. Cota QSPEntry *e; 580d557de4aSEmilio G. Cota uint32_t hash; 581d557de4aSEmilio G. Cota 582d557de4aSEmilio G. Cota hash = qsp_entry_no_thread_obj_hash(old); 583d557de4aSEmilio G. Cota e = qht_lookup(ht, old, hash); 584d557de4aSEmilio G. Cota if (e == NULL) { 585d557de4aSEmilio G. Cota e = qsp_entry_create(ht, old, hash); 586d557de4aSEmilio G. Cota e->n_objs = 1; 587d557de4aSEmilio G. Cota } else if (e->callsite->obj != old->callsite->obj) { 588d557de4aSEmilio G. Cota e->n_objs++; 589d557de4aSEmilio G. Cota } 590d557de4aSEmilio G. Cota e->ns += old->ns; 591d557de4aSEmilio G. Cota e->n_acqs += old->n_acqs; 592d557de4aSEmilio G. Cota } 593d557de4aSEmilio G. Cota 59478255ba2SEmilio G. Cota static void qsp_ht_delete(void *p, uint32_t h, void *htp) 595996e8d9aSEmilio G. Cota { 596996e8d9aSEmilio G. Cota g_free(p); 597996e8d9aSEmilio G. Cota } 598996e8d9aSEmilio G. Cota 599d557de4aSEmilio G. Cota static void qsp_mktree(GTree *tree, bool callsite_coalesce) 600fe9959a2SEmilio G. Cota { 601996e8d9aSEmilio G. Cota QSPSnapshot *snap; 602d557de4aSEmilio G. Cota struct qht ht, coalesce_ht; 603d557de4aSEmilio G. Cota struct qht *htp; 604fe9959a2SEmilio G. Cota 605996e8d9aSEmilio G. Cota /* 606996e8d9aSEmilio G. Cota * First, see if there's a prior snapshot, so that we read the global hash 607996e8d9aSEmilio G. Cota * table _after_ the snapshot has been created, which guarantees that 608996e8d9aSEmilio G. Cota * the entries we'll read will be a superset of the snapshot's entries. 609996e8d9aSEmilio G. Cota * 610996e8d9aSEmilio G. Cota * We must remain in an RCU read-side critical section until we're done 611996e8d9aSEmilio G. Cota * with the snapshot. 612996e8d9aSEmilio G. Cota */ 613996e8d9aSEmilio G. Cota rcu_read_lock(); 614996e8d9aSEmilio G. Cota snap = atomic_rcu_read(&qsp_snapshot); 615996e8d9aSEmilio G. Cota 616fe9959a2SEmilio G. Cota /* Aggregate all results from the global hash table into a local one */ 617fe9959a2SEmilio G. Cota qht_init(&ht, qsp_entry_no_thread_cmp, QSP_INITIAL_SIZE, 618fe9959a2SEmilio G. Cota QHT_MODE_AUTO_RESIZE | QHT_MODE_RAW_MUTEXES); 619fe9959a2SEmilio G. Cota qht_iter(&qsp_ht, qsp_aggregate, &ht); 620fe9959a2SEmilio G. Cota 621996e8d9aSEmilio G. Cota /* compute the difference wrt the snapshot, if any */ 622996e8d9aSEmilio G. Cota if (snap) { 623996e8d9aSEmilio G. Cota qsp_diff(&snap->ht, &ht); 624996e8d9aSEmilio G. Cota } 625996e8d9aSEmilio G. Cota /* done with the snapshot; RCU can reclaim it */ 626996e8d9aSEmilio G. Cota rcu_read_unlock(); 627996e8d9aSEmilio G. Cota 628d557de4aSEmilio G. Cota htp = &ht; 629d557de4aSEmilio G. Cota if (callsite_coalesce) { 630d557de4aSEmilio G. Cota qht_init(&coalesce_ht, qsp_entry_no_thread_obj_cmp, QSP_INITIAL_SIZE, 631d557de4aSEmilio G. Cota QHT_MODE_AUTO_RESIZE | QHT_MODE_RAW_MUTEXES); 632d557de4aSEmilio G. Cota qht_iter(&ht, qsp_iter_callsite_coalesce, &coalesce_ht); 633d557de4aSEmilio G. Cota 634d557de4aSEmilio G. Cota /* free the previous hash table, and point htp to coalesce_ht */ 635d557de4aSEmilio G. Cota qht_iter(&ht, qsp_ht_delete, NULL); 636d557de4aSEmilio G. Cota qht_destroy(&ht); 637d557de4aSEmilio G. Cota htp = &coalesce_ht; 638d557de4aSEmilio G. Cota } 639d557de4aSEmilio G. Cota 640fe9959a2SEmilio G. Cota /* sort the hash table elements by using a tree */ 641d557de4aSEmilio G. Cota qht_iter(htp, qsp_sort, tree); 642fe9959a2SEmilio G. Cota 643fe9959a2SEmilio G. Cota /* free the hash table, but keep the elements (those are in the tree now) */ 644d557de4aSEmilio G. Cota qht_destroy(htp); 645fe9959a2SEmilio G. Cota } 646fe9959a2SEmilio G. Cota 647fe9959a2SEmilio G. Cota /* free string with g_free */ 648fe9959a2SEmilio G. Cota static char *qsp_at(const QSPCallSite *callsite) 649fe9959a2SEmilio G. Cota { 650fe9959a2SEmilio G. Cota GString *s = g_string_new(NULL); 651fe9959a2SEmilio G. Cota const char *shortened; 652fe9959a2SEmilio G. Cota 653fe9959a2SEmilio G. Cota /* remove the absolute path to qemu */ 654fe9959a2SEmilio G. Cota if (unlikely(strlen(callsite->file) < qsp_qemu_path_len)) { 655fe9959a2SEmilio G. Cota shortened = callsite->file; 656fe9959a2SEmilio G. Cota } else { 657fe9959a2SEmilio G. Cota shortened = callsite->file + qsp_qemu_path_len; 658fe9959a2SEmilio G. Cota } 659fe9959a2SEmilio G. Cota g_string_append_printf(s, "%s:%u", shortened, callsite->line); 660fe9959a2SEmilio G. Cota return g_string_free(s, FALSE); 661fe9959a2SEmilio G. Cota } 662fe9959a2SEmilio G. Cota 663fe9959a2SEmilio G. Cota struct QSPReportEntry { 664fe9959a2SEmilio G. Cota const void *obj; 665fe9959a2SEmilio G. Cota char *callsite_at; 666fe9959a2SEmilio G. Cota const char *typename; 667fe9959a2SEmilio G. Cota double time_s; 668fe9959a2SEmilio G. Cota double ns_avg; 669fe9959a2SEmilio G. Cota uint64_t n_acqs; 670d557de4aSEmilio G. Cota unsigned int n_objs; 671fe9959a2SEmilio G. Cota }; 672fe9959a2SEmilio G. Cota typedef struct QSPReportEntry QSPReportEntry; 673fe9959a2SEmilio G. Cota 674fe9959a2SEmilio G. Cota struct QSPReport { 675fe9959a2SEmilio G. Cota QSPReportEntry *entries; 676fe9959a2SEmilio G. Cota size_t n_entries; 677fe9959a2SEmilio G. Cota size_t max_n_entries; 678fe9959a2SEmilio G. Cota }; 679fe9959a2SEmilio G. Cota typedef struct QSPReport QSPReport; 680fe9959a2SEmilio G. Cota 681fe9959a2SEmilio G. Cota static gboolean qsp_tree_report(gpointer key, gpointer value, gpointer udata) 682fe9959a2SEmilio G. Cota { 683fe9959a2SEmilio G. Cota const QSPEntry *e = key; 684fe9959a2SEmilio G. Cota QSPReport *report = udata; 685fe9959a2SEmilio G. Cota QSPReportEntry *entry; 686fe9959a2SEmilio G. Cota 687fe9959a2SEmilio G. Cota if (report->n_entries == report->max_n_entries) { 688fe9959a2SEmilio G. Cota return TRUE; 689fe9959a2SEmilio G. Cota } 690fe9959a2SEmilio G. Cota entry = &report->entries[report->n_entries]; 691fe9959a2SEmilio G. Cota report->n_entries++; 692fe9959a2SEmilio G. Cota 693fe9959a2SEmilio G. Cota entry->obj = e->callsite->obj; 694d557de4aSEmilio G. Cota entry->n_objs = e->n_objs; 695fe9959a2SEmilio G. Cota entry->callsite_at = qsp_at(e->callsite); 696fe9959a2SEmilio G. Cota entry->typename = qsp_typenames[e->callsite->type]; 697fe9959a2SEmilio G. Cota entry->time_s = e->ns * 1e-9; 698fe9959a2SEmilio G. Cota entry->n_acqs = e->n_acqs; 699fe9959a2SEmilio G. Cota entry->ns_avg = e->n_acqs ? e->ns / e->n_acqs : 0; 700fe9959a2SEmilio G. Cota return FALSE; 701fe9959a2SEmilio G. Cota } 702fe9959a2SEmilio G. Cota 703ac7ff4cfSMarkus Armbruster static void pr_report(const QSPReport *rep) 704fe9959a2SEmilio G. Cota { 705fe9959a2SEmilio G. Cota char *dashes; 706fe9959a2SEmilio G. Cota size_t max_len = 0; 707fe9959a2SEmilio G. Cota int callsite_len = 0; 708fe9959a2SEmilio G. Cota int callsite_rspace; 709fe9959a2SEmilio G. Cota int n_dashes; 710fe9959a2SEmilio G. Cota size_t i; 711fe9959a2SEmilio G. Cota 712fe9959a2SEmilio G. Cota /* find out the maximum length of all 'callsite' fields */ 713fe9959a2SEmilio G. Cota for (i = 0; i < rep->n_entries; i++) { 714fe9959a2SEmilio G. Cota const QSPReportEntry *e = &rep->entries[i]; 715fe9959a2SEmilio G. Cota size_t len = strlen(e->callsite_at); 716fe9959a2SEmilio G. Cota 717fe9959a2SEmilio G. Cota if (len > max_len) { 718fe9959a2SEmilio G. Cota max_len = len; 719fe9959a2SEmilio G. Cota } 720fe9959a2SEmilio G. Cota } 721fe9959a2SEmilio G. Cota 722fe9959a2SEmilio G. Cota callsite_len = MAX(max_len, strlen("Call site")); 723fe9959a2SEmilio G. Cota /* white space to leave to the right of "Call site" */ 724fe9959a2SEmilio G. Cota callsite_rspace = callsite_len - strlen("Call site"); 725fe9959a2SEmilio G. Cota 726ac7ff4cfSMarkus Armbruster qemu_printf("Type Object Call site%*s Wait Time (s) " 727fe9959a2SEmilio G. Cota " Count Average (us)\n", callsite_rspace, ""); 728fe9959a2SEmilio G. Cota 729fe9959a2SEmilio G. Cota /* build a horizontal rule with dashes */ 730fe9959a2SEmilio G. Cota n_dashes = 79 + callsite_rspace; 731fe9959a2SEmilio G. Cota dashes = g_malloc(n_dashes + 1); 732fe9959a2SEmilio G. Cota memset(dashes, '-', n_dashes); 733fe9959a2SEmilio G. Cota dashes[n_dashes] = '\0'; 734ac7ff4cfSMarkus Armbruster qemu_printf("%s\n", dashes); 735fe9959a2SEmilio G. Cota 736fe9959a2SEmilio G. Cota for (i = 0; i < rep->n_entries; i++) { 737fe9959a2SEmilio G. Cota const QSPReportEntry *e = &rep->entries[i]; 738d557de4aSEmilio G. Cota GString *s = g_string_new(NULL); 739fe9959a2SEmilio G. Cota 740d557de4aSEmilio G. Cota g_string_append_printf(s, "%-9s ", e->typename); 741d557de4aSEmilio G. Cota if (e->n_objs > 1) { 742d557de4aSEmilio G. Cota g_string_append_printf(s, "[%12u]", e->n_objs); 743d557de4aSEmilio G. Cota } else { 744d557de4aSEmilio G. Cota g_string_append_printf(s, "%14p", e->obj); 745d557de4aSEmilio G. Cota } 746d557de4aSEmilio G. Cota g_string_append_printf(s, " %s%*s %13.5f %12" PRIu64 " %12.2f\n", 747d557de4aSEmilio G. Cota e->callsite_at, 748d557de4aSEmilio G. Cota callsite_len - (int)strlen(e->callsite_at), "", 749d557de4aSEmilio G. Cota e->time_s, e->n_acqs, e->ns_avg * 1e-3); 750ac7ff4cfSMarkus Armbruster qemu_printf("%s", s->str); 751d557de4aSEmilio G. Cota g_string_free(s, TRUE); 752fe9959a2SEmilio G. Cota } 753fe9959a2SEmilio G. Cota 754ac7ff4cfSMarkus Armbruster qemu_printf("%s\n", dashes); 755fe9959a2SEmilio G. Cota g_free(dashes); 756fe9959a2SEmilio G. Cota } 757fe9959a2SEmilio G. Cota 758fe9959a2SEmilio G. Cota static void report_destroy(QSPReport *rep) 759fe9959a2SEmilio G. Cota { 760fe9959a2SEmilio G. Cota size_t i; 761fe9959a2SEmilio G. Cota 762fe9959a2SEmilio G. Cota for (i = 0; i < rep->n_entries; i++) { 763fe9959a2SEmilio G. Cota QSPReportEntry *e = &rep->entries[i]; 764fe9959a2SEmilio G. Cota 765fe9959a2SEmilio G. Cota g_free(e->callsite_at); 766fe9959a2SEmilio G. Cota } 767fe9959a2SEmilio G. Cota g_free(rep->entries); 768fe9959a2SEmilio G. Cota } 769fe9959a2SEmilio G. Cota 770ac7ff4cfSMarkus Armbruster void qsp_report(size_t max, enum QSPSortBy sort_by, 771ac7ff4cfSMarkus Armbruster bool callsite_coalesce) 772fe9959a2SEmilio G. Cota { 7730a22777cSEmilio G. Cota GTree *tree = g_tree_new_full(qsp_tree_cmp, &sort_by, g_free, NULL); 774fe9959a2SEmilio G. Cota QSPReport rep; 775fe9959a2SEmilio G. Cota 776fe9959a2SEmilio G. Cota qsp_init(); 777fe9959a2SEmilio G. Cota 778fe9959a2SEmilio G. Cota rep.entries = g_new0(QSPReportEntry, max); 779fe9959a2SEmilio G. Cota rep.n_entries = 0; 780fe9959a2SEmilio G. Cota rep.max_n_entries = max; 781fe9959a2SEmilio G. Cota 782d557de4aSEmilio G. Cota qsp_mktree(tree, callsite_coalesce); 783fe9959a2SEmilio G. Cota g_tree_foreach(tree, qsp_tree_report, &rep); 784fe9959a2SEmilio G. Cota g_tree_destroy(tree); 785fe9959a2SEmilio G. Cota 786ac7ff4cfSMarkus Armbruster pr_report(&rep); 787fe9959a2SEmilio G. Cota report_destroy(&rep); 788fe9959a2SEmilio G. Cota } 789996e8d9aSEmilio G. Cota 790996e8d9aSEmilio G. Cota static void qsp_snapshot_destroy(QSPSnapshot *snap) 791996e8d9aSEmilio G. Cota { 792996e8d9aSEmilio G. Cota qht_iter(&snap->ht, qsp_ht_delete, NULL); 793996e8d9aSEmilio G. Cota qht_destroy(&snap->ht); 794996e8d9aSEmilio G. Cota g_free(snap); 795996e8d9aSEmilio G. Cota } 796996e8d9aSEmilio G. Cota 797996e8d9aSEmilio G. Cota void qsp_reset(void) 798996e8d9aSEmilio G. Cota { 799996e8d9aSEmilio G. Cota QSPSnapshot *new = g_new(QSPSnapshot, 1); 800996e8d9aSEmilio G. Cota QSPSnapshot *old; 801996e8d9aSEmilio G. Cota 802996e8d9aSEmilio G. Cota qsp_init(); 803996e8d9aSEmilio G. Cota 804996e8d9aSEmilio G. Cota qht_init(&new->ht, qsp_entry_cmp, QSP_INITIAL_SIZE, 805996e8d9aSEmilio G. Cota QHT_MODE_AUTO_RESIZE | QHT_MODE_RAW_MUTEXES); 806996e8d9aSEmilio G. Cota 807996e8d9aSEmilio G. Cota /* take a snapshot of the current state */ 808996e8d9aSEmilio G. Cota qht_iter(&qsp_ht, qsp_aggregate, &new->ht); 809996e8d9aSEmilio G. Cota 810996e8d9aSEmilio G. Cota /* replace the previous snapshot, if any */ 811996e8d9aSEmilio G. Cota old = atomic_xchg(&qsp_snapshot, new); 812996e8d9aSEmilio G. Cota if (old) { 813996e8d9aSEmilio G. Cota call_rcu(old, qsp_snapshot_destroy, rcu); 814996e8d9aSEmilio G. Cota } 815996e8d9aSEmilio G. Cota } 816