| 1 | /* |
| 2 | * Copyright (C) 2020, Alex Bennée <alex.bennee@linaro.org> |
| 3 | * |
| 4 | * HW Profile - breakdown access patterns for IO to devices |
| 5 | * |
| 6 | * License: GNU GPL, version 2 or later. |
| 7 | * See the COPYING file in the top-level directory. |
| 8 | */ |
| 9 | |
| 10 | #include <inttypes.h> |
| 11 | #include <assert.h> |
| 12 | #include <stdlib.h> |
| 13 | #include <inttypes.h> |
| 14 | #include <string.h> |
| 15 | #include <unistd.h> |
| 16 | #include <stdio.h> |
| 17 | #include <glib.h> |
| 18 | |
| 19 | #include <qemu-plugin.h> |
| 20 | |
| 21 | QEMU_PLUGIN_EXPORT int qemu_plugin_version = QEMU_PLUGIN_VERSION; |
| 22 | |
| 23 | #define ARRAY_SIZE(x) (sizeof(x) / sizeof((x)[0])) |
| 24 | |
| 25 | typedef struct { |
| 26 | uint64_t cpu_read; |
| 27 | uint64_t cpu_write; |
| 28 | uint64_t reads; |
| 29 | uint64_t writes; |
| 30 | } IOCounts; |
| 31 | |
| 32 | typedef struct { |
| 33 | uint64_t off_or_pc; |
| 34 | IOCounts counts; |
| 35 | } IOLocationCounts; |
| 36 | |
| 37 | typedef struct { |
| 38 | const char *name; |
| 39 | uint64_t base; |
| 40 | IOCounts totals; |
| 41 | GHashTable *detail; |
| 42 | } DeviceCounts; |
| 43 | |
| 44 | static GMutex lock; |
| 45 | static GHashTable *devices; |
| 46 | static struct qemu_plugin_scoreboard *source_pc_scoreboard; |
| 47 | static qemu_plugin_u64 source_pc; |
| 48 | |
| 49 | /* track the access pattern to a piece of HW */ |
| 50 | static bool pattern; |
| 51 | /* track the source address of access to HW */ |
| 52 | static bool source; |
| 53 | /* track only matched regions of HW */ |
| 54 | static bool check_match; |
| 55 | static gchar **matches; |
| 56 | |
| 57 | static enum qemu_plugin_mem_rw rw = QEMU_PLUGIN_MEM_RW; |
| 58 | |
| 59 | static inline bool track_reads(void) |
| 60 | { |
| 61 | return rw == QEMU_PLUGIN_MEM_RW || rw == QEMU_PLUGIN_MEM_R; |
| 62 | } |
| 63 | |
| 64 | static inline bool track_writes(void) |
| 65 | { |
| 66 | return rw == QEMU_PLUGIN_MEM_RW || rw == QEMU_PLUGIN_MEM_W; |
| 67 | } |
| 68 | |
| 69 | static void plugin_init(void) |
| 70 | { |
| 71 | devices = g_hash_table_new(NULL, NULL); |
| 72 | } |
| 73 | |
| 74 | static gint sort_cmp(gconstpointer a, gconstpointer b, gpointer d) |
| 75 | { |
| 76 | DeviceCounts *ea = (DeviceCounts *) a; |
| 77 | DeviceCounts *eb = (DeviceCounts *) b; |
| 78 | return ea->totals.reads + ea->totals.writes > |
| 79 | eb->totals.reads + eb->totals.writes ? -1 : 1; |
| 80 | } |
| 81 | |
| 82 | static gint sort_loc(gconstpointer a, gconstpointer b, gpointer d) |
| 83 | { |
| 84 | IOLocationCounts *ea = (IOLocationCounts *) a; |
| 85 | IOLocationCounts *eb = (IOLocationCounts *) b; |
| 86 | return ea->off_or_pc > eb->off_or_pc; |
| 87 | } |
| 88 | |
| 89 | static void fmt_iocount_record(GString *s, IOCounts *rec) |
| 90 | { |
| 91 | if (track_reads()) { |
| 92 | g_string_append_printf(s, ", %"PRIx64", %"PRId64, |
| 93 | rec->cpu_read, rec->reads); |
| 94 | } |
| 95 | if (track_writes()) { |
| 96 | g_string_append_printf(s, ", %"PRIx64", %"PRId64, |
| 97 | rec->cpu_write, rec->writes); |
| 98 | } |
| 99 | } |
| 100 | |
| 101 | static void fmt_dev_record(GString *s, DeviceCounts *rec) |
| 102 | { |
| 103 | g_string_append_printf(s, "%s, 0x%"PRIx64, |
| 104 | rec->name, rec->base); |
| 105 | fmt_iocount_record(s, &rec->totals); |
| 106 | g_string_append_c(s, '\n'); |
| 107 | } |
| 108 | |
| 109 | static void plugin_exit(void *p) |
| 110 | { |
| 111 | g_autoptr(GString) report = g_string_new(""); |
| 112 | GList *counts; |
| 113 | |
| 114 | if (!(pattern || source)) { |
| 115 | g_string_printf(report, "Device, Address"); |
| 116 | if (track_reads()) { |
| 117 | g_string_append_printf(report, ", RCPUs, Reads"); |
| 118 | } |
| 119 | if (track_writes()) { |
| 120 | g_string_append_printf(report, ", WCPUs, Writes"); |
| 121 | } |
| 122 | g_string_append_c(report, '\n'); |
| 123 | } |
| 124 | |
| 125 | counts = g_hash_table_get_values(devices); |
| 126 | if (counts && g_list_next(counts)) { |
| 127 | GList *it; |
| 128 | |
| 129 | it = g_list_sort_with_data(counts, sort_cmp, NULL); |
| 130 | |
| 131 | while (it) { |
| 132 | DeviceCounts *rec = (DeviceCounts *) it->data; |
| 133 | if (rec->detail) { |
| 134 | GList *accesses = g_hash_table_get_values(rec->detail); |
| 135 | GList *io_it = g_list_sort_with_data(accesses, sort_loc, NULL); |
| 136 | const char *prefix = pattern ? "off" : "pc"; |
| 137 | g_string_append_printf(report, "%s @ 0x%"PRIx64"\n", |
| 138 | rec->name, rec->base); |
| 139 | while (io_it) { |
| 140 | IOLocationCounts *loc = (IOLocationCounts *) io_it->data; |
| 141 | g_string_append_printf(report, " %s:%08"PRIx64, |
| 142 | prefix, loc->off_or_pc); |
| 143 | fmt_iocount_record(report, &loc->counts); |
| 144 | g_string_append_c(report, '\n'); |
| 145 | io_it = io_it->next; |
| 146 | } |
| 147 | } else { |
| 148 | fmt_dev_record(report, rec); |
| 149 | } |
| 150 | it = it->next; |
| 151 | }; |
| 152 | g_list_free(it); |
| 153 | } |
| 154 | |
| 155 | qemu_plugin_outs(report->str); |
| 156 | } |
| 157 | |
| 158 | static DeviceCounts *new_count(const char *name, uint64_t base) |
| 159 | { |
| 160 | DeviceCounts *count = g_new0(DeviceCounts, 1); |
| 161 | count->name = name; |
| 162 | count->base = base; |
| 163 | if (pattern || source) { |
| 164 | count->detail = g_hash_table_new(g_int64_hash, g_int64_equal); |
| 165 | } |
| 166 | g_hash_table_insert(devices, (gpointer) name, count); |
| 167 | return count; |
| 168 | } |
| 169 | |
| 170 | static IOLocationCounts *new_location(GHashTable *table, uint64_t off_or_pc) |
| 171 | { |
| 172 | IOLocationCounts *loc = g_new0(IOLocationCounts, 1); |
| 173 | loc->off_or_pc = off_or_pc; |
| 174 | g_hash_table_insert(table, &loc->off_or_pc, loc); |
| 175 | return loc; |
| 176 | } |
| 177 | |
| 178 | static void hwprofile_match_hit(DeviceCounts *rec, uint64_t off) |
| 179 | { |
| 180 | g_autoptr(GString) report = g_string_new("hwprofile: match @ offset"); |
| 181 | g_string_append_printf(report, "%"PRIx64", previous hits\n", off); |
| 182 | fmt_dev_record(report, rec); |
| 183 | qemu_plugin_outs(report->str); |
| 184 | } |
| 185 | |
| 186 | static void inc_count(IOCounts *count, bool is_write, unsigned int cpu_index) |
| 187 | { |
| 188 | if (is_write) { |
| 189 | count->writes++; |
| 190 | count->cpu_write |= (1 << cpu_index); |
| 191 | } else { |
| 192 | count->reads++; |
| 193 | count->cpu_read |= (1 << cpu_index); |
| 194 | } |
| 195 | } |
| 196 | |
| 197 | static void vcpu_haddr(unsigned int cpu_index, qemu_plugin_meminfo_t meminfo, |
| 198 | uint64_t vaddr, void *udata) |
| 199 | { |
| 200 | struct qemu_plugin_hwaddr *hwaddr = qemu_plugin_get_hwaddr(meminfo, vaddr); |
| 201 | |
| 202 | if (!hwaddr || !qemu_plugin_hwaddr_is_io(hwaddr)) { |
| 203 | return; |
| 204 | } else { |
| 205 | const char *name = qemu_plugin_hwaddr_device_name(hwaddr); |
| 206 | uint64_t off = qemu_plugin_hwaddr_phys_addr(hwaddr); |
| 207 | bool is_write = qemu_plugin_mem_is_store(meminfo); |
| 208 | DeviceCounts *counts; |
| 209 | |
| 210 | g_mutex_lock(&lock); |
| 211 | counts = (DeviceCounts *) g_hash_table_lookup(devices, name); |
| 212 | |
| 213 | if (!counts) { |
| 214 | uint64_t base = vaddr - off; |
| 215 | counts = new_count(name, base); |
| 216 | } |
| 217 | |
| 218 | if (check_match) { |
| 219 | if (g_strv_contains((const char * const *)matches, counts->name)) { |
| 220 | hwprofile_match_hit(counts, off); |
| 221 | inc_count(&counts->totals, is_write, cpu_index); |
| 222 | } |
| 223 | } else { |
| 224 | inc_count(&counts->totals, is_write, cpu_index); |
| 225 | } |
| 226 | |
| 227 | /* either track offsets or source of access */ |
| 228 | if (source) { |
| 229 | off = qemu_plugin_u64_get(source_pc, cpu_index); |
| 230 | } |
| 231 | |
| 232 | if (pattern || source) { |
| 233 | IOLocationCounts *io_count = g_hash_table_lookup(counts->detail, |
| 234 | &off); |
| 235 | if (!io_count) { |
| 236 | io_count = new_location(counts->detail, off); |
| 237 | } |
| 238 | inc_count(&io_count->counts, is_write, cpu_index); |
| 239 | } |
| 240 | |
| 241 | g_mutex_unlock(&lock); |
| 242 | } |
| 243 | } |
| 244 | |
| 245 | static void vcpu_tb_trans(struct qemu_plugin_tb *tb, void *userdata) |
| 246 | { |
| 247 | size_t n = qemu_plugin_tb_n_insns(tb); |
| 248 | size_t i; |
| 249 | |
| 250 | for (i = 0; i < n; i++) { |
| 251 | struct qemu_plugin_insn *insn = qemu_plugin_tb_get_insn(tb, i); |
| 252 | if (source) { |
| 253 | uint64_t pc = qemu_plugin_insn_vaddr(insn); |
| 254 | qemu_plugin_register_vcpu_mem_inline_per_vcpu( |
| 255 | insn, rw, QEMU_PLUGIN_INLINE_STORE_U64, |
| 256 | source_pc, pc); |
| 257 | } |
| 258 | qemu_plugin_register_vcpu_mem_cb(insn, vcpu_haddr, |
| 259 | QEMU_PLUGIN_CB_NO_REGS, rw, NULL); |
| 260 | } |
| 261 | } |
| 262 | |
| 263 | QEMU_PLUGIN_EXPORT |
| 264 | int qemu_plugin_install(qemu_plugin_id_t id, const qemu_info_t *info, |
| 265 | int argc, char **argv) |
| 266 | { |
| 267 | int i; |
| 268 | g_autoptr(GString) matches_raw = g_string_new(""); |
| 269 | |
| 270 | for (i = 0; i < argc; i++) { |
| 271 | char *opt = argv[i]; |
| 272 | g_auto(GStrv) tokens = g_strsplit(opt, "=", 2); |
| 273 | |
| 274 | if (g_strcmp0(tokens[0], "track") == 0) { |
| 275 | if (g_strcmp0(tokens[1], "read") == 0) { |
| 276 | rw = QEMU_PLUGIN_MEM_R; |
| 277 | } else if (g_strcmp0(tokens[1], "write") == 0) { |
| 278 | rw = QEMU_PLUGIN_MEM_W; |
| 279 | } else { |
| 280 | fprintf(stderr, "invalid value for track: %s\n", tokens[1]); |
| 281 | return -1; |
| 282 | } |
| 283 | } else if (g_strcmp0(tokens[0], "pattern") == 0) { |
| 284 | if (!qemu_plugin_bool_parse(tokens[0], tokens[1], &pattern)) { |
| 285 | fprintf(stderr, "boolean argument parsing failed: %s\n", opt); |
| 286 | return -1; |
| 287 | } |
| 288 | } else if (g_strcmp0(tokens[0], "source") == 0) { |
| 289 | if (!qemu_plugin_bool_parse(tokens[0], tokens[1], &source)) { |
| 290 | fprintf(stderr, "boolean argument parsing failed: %s\n", opt); |
| 291 | return -1; |
| 292 | } |
| 293 | } else if (g_strcmp0(tokens[0], "match") == 0) { |
| 294 | check_match = true; |
| 295 | g_string_append_printf(matches_raw, "%s,", tokens[1]); |
| 296 | } else { |
| 297 | fprintf(stderr, "option parsing failed: %s\n", opt); |
| 298 | return -1; |
| 299 | } |
| 300 | } |
| 301 | if (check_match) { |
| 302 | matches = g_strsplit(matches_raw->str, ",", -1); |
| 303 | } |
| 304 | |
| 305 | if (source && pattern) { |
| 306 | fprintf(stderr, "can only currently track either source or pattern.\n"); |
| 307 | return -1; |
| 308 | } |
| 309 | |
| 310 | if (!info->system_emulation) { |
| 311 | fprintf(stderr, "hwprofile: plugin only useful for system emulation\n"); |
| 312 | return -1; |
| 313 | } |
| 314 | |
| 315 | if (source) { |
| 316 | source_pc_scoreboard = qemu_plugin_scoreboard_new(sizeof(uint64_t)); |
| 317 | source_pc = qemu_plugin_scoreboard_u64(source_pc_scoreboard); |
| 318 | } |
| 319 | |
| 320 | plugin_init(); |
| 321 | |
| 322 | qemu_plugin_register_vcpu_tb_trans_cb(id, vcpu_tb_trans, NULL); |
| 323 | qemu_plugin_register_atexit_cb(id, plugin_exit, NULL); |
| 324 | return 0; |
| 325 | } |