| 1 | /* |
| 2 | * SPDX-License-Identifier: LGPL-2.1-or-later |
| 3 | * |
| 4 | * QEMU TCG statistics |
| 5 | * |
| 6 | * Copyright (c) 2003-2005 Fabrice Bellard |
| 7 | */ |
| 8 | |
| 9 | #include "qemu/osdep.h" |
| 10 | #include "qemu/accel.h" |
| 11 | #include "qemu/qht.h" |
| 12 | #include "qapi/error.h" |
| 13 | #include "system/cpu-timers.h" |
| 14 | #include "exec/icount.h" |
| 15 | #include "hw/core/cpu.h" |
| 16 | #include "tcg/tcg.h" |
| 17 | #include "internal-common.h" |
| 18 | #include "tb-context.h" |
| 19 | #include <math.h> |
| 20 | |
| 21 | static void dump_drift_info(GString *buf) |
| 22 | { |
| 23 | if (!icount_enabled()) { |
| 24 | return; |
| 25 | } |
| 26 | |
| 27 | g_string_append_printf(buf, "Host - Guest clock %"PRIi64" ms\n", |
| 28 | (cpu_get_clock() - icount_get()) / SCALE_MS); |
| 29 | if (icount_align_option) { |
| 30 | g_string_append_printf(buf, "Max guest delay %"PRIi64" ms\n", |
| 31 | -max_delay / SCALE_MS); |
| 32 | g_string_append_printf(buf, "Max guest advance %"PRIi64" ms\n", |
| 33 | max_advance / SCALE_MS); |
| 34 | } else { |
| 35 | g_string_append_printf(buf, "Max guest delay NA\n"); |
| 36 | g_string_append_printf(buf, "Max guest advance NA\n"); |
| 37 | } |
| 38 | } |
| 39 | |
| 40 | static void dump_accel_info(AccelState *accel, GString *buf) |
| 41 | { |
| 42 | bool one_insn_per_tb = object_property_get_bool(OBJECT(accel), |
| 43 | "one-insn-per-tb", |
| 44 | &error_fatal); |
| 45 | |
| 46 | g_string_append_printf(buf, "Accelerator settings:\n"); |
| 47 | g_string_append_printf(buf, "one-insn-per-tb: %s\n\n", |
| 48 | one_insn_per_tb ? "on" : "off"); |
| 49 | } |
| 50 | |
| 51 | static void print_qht_statistics(struct qht_stats hst, GString *buf) |
| 52 | { |
| 53 | uint32_t hgram_opts; |
| 54 | size_t hgram_bins; |
| 55 | char *hgram; |
| 56 | double avg; |
| 57 | |
| 58 | if (!hst.head_buckets) { |
| 59 | return; |
| 60 | } |
| 61 | g_string_append_printf(buf, "TB hash buckets %zu/%zu " |
| 62 | "(%0.2f%% head buckets used)\n", |
| 63 | hst.used_head_buckets, hst.head_buckets, |
| 64 | (double)hst.used_head_buckets / |
| 65 | hst.head_buckets * 100); |
| 66 | |
| 67 | hgram_opts = QDIST_PR_BORDER | QDIST_PR_LABELS; |
| 68 | hgram_opts |= QDIST_PR_100X | QDIST_PR_PERCENT; |
| 69 | if (qdist_xmax(&hst.occupancy) - qdist_xmin(&hst.occupancy) == 1) { |
| 70 | hgram_opts |= QDIST_PR_NODECIMAL; |
| 71 | } |
| 72 | hgram = qdist_pr(&hst.occupancy, 10, hgram_opts); |
| 73 | avg = qdist_avg(&hst.occupancy); |
| 74 | if (!isnan(avg)) { |
| 75 | g_string_append_printf(buf, "TB hash occupancy " |
| 76 | "%0.2f%% avg chain occ. " |
| 77 | "Histogram: %s\n", |
| 78 | avg * 100, hgram); |
| 79 | } |
| 80 | g_free(hgram); |
| 81 | |
| 82 | hgram_opts = QDIST_PR_BORDER | QDIST_PR_LABELS; |
| 83 | hgram_bins = qdist_xmax(&hst.chain) - qdist_xmin(&hst.chain); |
| 84 | if (hgram_bins > 10) { |
| 85 | hgram_bins = 10; |
| 86 | } else { |
| 87 | hgram_bins = 0; |
| 88 | hgram_opts |= QDIST_PR_NODECIMAL | QDIST_PR_NOBINRANGE; |
| 89 | } |
| 90 | hgram = qdist_pr(&hst.chain, hgram_bins, hgram_opts); |
| 91 | avg = qdist_avg(&hst.chain); |
| 92 | if (!isnan(avg)) { |
| 93 | g_string_append_printf(buf, "TB hash avg chain %0.3f buckets. " |
| 94 | "Histogram: %s\n", |
| 95 | avg, hgram); |
| 96 | } |
| 97 | g_free(hgram); |
| 98 | } |
| 99 | |
| 100 | struct tb_tree_stats { |
| 101 | size_t nb_tbs; |
| 102 | size_t host_size; |
| 103 | size_t target_size; |
| 104 | size_t max_target_size; |
| 105 | size_t direct_jmp_count; |
| 106 | size_t direct_jmp2_count; |
| 107 | size_t cross_page; |
| 108 | }; |
| 109 | |
| 110 | static gboolean tb_tree_stats_iter(gpointer key, gpointer value, gpointer data) |
| 111 | { |
| 112 | const TranslationBlock *tb = value; |
| 113 | struct tb_tree_stats *tst = data; |
| 114 | |
| 115 | tst->nb_tbs++; |
| 116 | tst->host_size += tb->tc.size; |
| 117 | tst->target_size += tb->size; |
| 118 | if (tb->size > tst->max_target_size) { |
| 119 | tst->max_target_size = tb->size; |
| 120 | } |
| 121 | #ifndef CONFIG_USER_ONLY |
| 122 | if (tb->page_addr[1] != -1) { |
| 123 | tst->cross_page++; |
| 124 | } |
| 125 | #endif |
| 126 | if (tb->jmp_reset_offset[0] != TB_JMP_OFFSET_INVALID) { |
| 127 | tst->direct_jmp_count++; |
| 128 | if (tb->jmp_reset_offset[1] != TB_JMP_OFFSET_INVALID) { |
| 129 | tst->direct_jmp2_count++; |
| 130 | } |
| 131 | } |
| 132 | return false; |
| 133 | } |
| 134 | |
| 135 | static void tlb_flush_counts(size_t *pfull, size_t *ppart, size_t *pelide) |
| 136 | { |
| 137 | CPUState *cpu; |
| 138 | size_t full = 0, part = 0, elide = 0; |
| 139 | |
| 140 | CPU_FOREACH(cpu) { |
| 141 | full += qatomic_read(&cpu->neg.tlb.c.full_flush_count); |
| 142 | part += qatomic_read(&cpu->neg.tlb.c.part_flush_count); |
| 143 | elide += qatomic_read(&cpu->neg.tlb.c.elide_flush_count); |
| 144 | } |
| 145 | *pfull = full; |
| 146 | *ppart = part; |
| 147 | *pelide = elide; |
| 148 | } |
| 149 | |
| 150 | static void tcg_dump_flush_info(GString *buf) |
| 151 | { |
| 152 | size_t flush_full, flush_part, flush_elide; |
| 153 | |
| 154 | g_string_append_printf(buf, "TB flush count %u\n", |
| 155 | qatomic_read(&tb_ctx.tb_flush_count)); |
| 156 | g_string_append_printf(buf, "TB invalidate count %u\n", |
| 157 | qatomic_read(&tb_ctx.tb_phys_invalidate_count)); |
| 158 | |
| 159 | tlb_flush_counts(&flush_full, &flush_part, &flush_elide); |
| 160 | g_string_append_printf(buf, "TLB full flushes %zu\n", flush_full); |
| 161 | g_string_append_printf(buf, "TLB partial flushes %zu\n", flush_part); |
| 162 | g_string_append_printf(buf, "TLB elided flushes %zu\n", flush_elide); |
| 163 | } |
| 164 | |
| 165 | static void dump_exec_info(GString *buf) |
| 166 | { |
| 167 | struct tb_tree_stats tst = {}; |
| 168 | struct qht_stats hst; |
| 169 | size_t nb_tbs; |
| 170 | |
| 171 | tcg_tb_foreach(tb_tree_stats_iter, &tst); |
| 172 | nb_tbs = tst.nb_tbs; |
| 173 | /* XXX: avoid using doubles ? */ |
| 174 | g_string_append_printf(buf, "Translation buffer state:\n"); |
| 175 | /* |
| 176 | * Report total code size including the padding and TB structs; |
| 177 | * otherwise users might think "-accel tcg,tb-size" is not honoured. |
| 178 | * For avg host size we use the precise numbers from tb_tree_stats though. |
| 179 | */ |
| 180 | g_string_append_printf(buf, "gen code size %zu/%zu\n", |
| 181 | tcg_code_size(), tcg_code_capacity()); |
| 182 | g_string_append_printf(buf, "TB count %zu\n", nb_tbs); |
| 183 | g_string_append_printf(buf, "TB avg target size %zu max=%zu bytes\n", |
| 184 | nb_tbs ? tst.target_size / nb_tbs : 0, |
| 185 | tst.max_target_size); |
| 186 | g_string_append_printf(buf, "TB avg host size %zu bytes " |
| 187 | "(expansion ratio: %0.1f)\n", |
| 188 | nb_tbs ? tst.host_size / nb_tbs : 0, |
| 189 | tst.target_size ? |
| 190 | (double)tst.host_size / tst.target_size : 0); |
| 191 | g_string_append_printf(buf, "cross page TB count %zu (%zu%%)\n", |
| 192 | tst.cross_page, |
| 193 | nb_tbs ? (tst.cross_page * 100) / nb_tbs : 0); |
| 194 | g_string_append_printf(buf, "direct jump count %zu (%zu%%) " |
| 195 | "(2 jumps=%zu %zu%%)\n", |
| 196 | tst.direct_jmp_count, |
| 197 | nb_tbs ? (tst.direct_jmp_count * 100) / nb_tbs : 0, |
| 198 | tst.direct_jmp2_count, |
| 199 | nb_tbs ? (tst.direct_jmp2_count * 100) / nb_tbs : 0); |
| 200 | |
| 201 | qht_statistics_init(&tb_ctx.htable, &hst); |
| 202 | print_qht_statistics(hst, buf); |
| 203 | qht_statistics_destroy(&hst); |
| 204 | |
| 205 | g_string_append_printf(buf, "\nStatistics:\n"); |
| 206 | tcg_dump_flush_info(buf); |
| 207 | } |
| 208 | |
| 209 | void tcg_get_stats(AccelState *accel, GString *buf) |
| 210 | { |
| 211 | dump_accel_info(accel, buf); |
| 212 | dump_exec_info(buf); |
| 213 | dump_drift_info(buf); |
| 214 | } |
| 215 | |
| 216 | void tcg_dump_stats(GString *buf) |
| 217 | { |
| 218 | tcg_get_stats(current_accel(), buf); |
| 219 | } |