diff options
author | Yinan Zhang <zyn8950@gmail.com> | 2019-03-15 11:01:45 -0700 |
---|---|---|
committer | Qi Wang <interwq@gmail.com> | 2019-04-04 13:48:39 -0700 |
commit | 9aab3f2be041b09f42375d3bf173d1a8795a1ee9 (patch) | |
tree | 73c326e39d0c8158310e2c207e37ffe1d632b89f /src | |
parent | b0b3e49a54ec29e32636f4577d9d5a896d67fd20 (diff) |
Add memory utilization analytics to mallctl
The analytics tool is put under experimental.utilization namespace in
mallctl. Input is one pointer or an array of pointers and the output
is a list of memory utilization statistics.
Diffstat (limited to 'src')
-rw-r--r-- | src/ctl.c | 236 | ||||
-rw-r--r-- | src/extent.c | 69 |
2 files changed, 301 insertions, 4 deletions
@@ -216,6 +216,8 @@ CTL_PROTO(stats_mapped) CTL_PROTO(stats_retained) CTL_PROTO(experimental_hooks_install) CTL_PROTO(experimental_hooks_remove) +CTL_PROTO(experimental_utilization_query) +CTL_PROTO(experimental_utilization_batch_query) #define MUTEX_STATS_CTL_PROTO_GEN(n) \ CTL_PROTO(stats_##n##_num_ops) \ @@ -574,11 +576,17 @@ static const ctl_named_node_t stats_node[] = { static const ctl_named_node_t hooks_node[] = { {NAME("install"), CTL(experimental_hooks_install)}, - {NAME("remove"), CTL(experimental_hooks_remove)}, + {NAME("remove"), CTL(experimental_hooks_remove)} +}; + +static const ctl_named_node_t utilization_node[] = { + {NAME("query"), CTL(experimental_utilization_query)}, + {NAME("batch_query"), CTL(experimental_utilization_batch_query)} }; static const ctl_named_node_t experimental_node[] = { - {NAME("hooks"), CHILD(named, hooks)} + {NAME("hooks"), CHILD(named, hooks)}, + {NAME("utilization"), CHILD(named, utilization)} }; static const ctl_named_node_t root_node[] = { @@ -2714,7 +2722,7 @@ static int prof_log_start_ctl(tsd_t *tsd, const size_t *mib, size_t miblen, void *oldp, size_t *oldlenp, void *newp, size_t newlen) { int ret; - + const char *filename = NULL; if (!config_prof) { @@ -2726,7 +2734,7 @@ prof_log_start_ctl(tsd_t *tsd, const size_t *mib, size_t miblen, void *oldp, if (prof_log_start(tsd_tsdn(tsd), filename)) { ret = EFAULT; - goto label_return; + goto label_return; } ret = 0; @@ -3083,3 +3091,223 @@ experimental_hooks_remove_ctl(tsd_t *tsd, const size_t *mib, size_t miblen, label_return: return ret; } + +/* + * Output six memory utilization entries for an input pointer, the first one of + * type (void *) and the remaining five of type size_t, describing the following + * (in the same order): + * + * (a) memory address of the extent a potential reallocation would go into, + * == the five fields below describe about the extent the pointer resides in == + * (b) number of free regions in the extent, + * (c) number of regions in the extent, + * (d) size of the extent in terms of bytes, + * (e) total number of free regions in the bin the extent belongs to, and + * (f) total number of regions in the bin the extent belongs to. + * + * Note that "(e)" and "(f)" are only available when stats are enabled; + * otherwise both are set zero. + * + * This API is mainly intended for small class allocations, where extents are + * used as slab. + * + * In case of large class allocations, "(a)" will be NULL, and "(e)" and "(f)" + * will be zero. The other three fields will be properly set though the values + * are trivial: "(b)" will be 0, "(c)" will be 1, and "(d)" will be the usable + * size. + * + * The input pointer and size are respectively passed in by newp and newlen, + * and the output fields and size are respectively oldp and *oldlenp. + * + * It can be beneficial to define the following macros to make it easier to + * access the output: + * + * #define SLABCUR_READ(out) (*(void **)out) + * #define COUNTS(out) ((size_t *)((void **)out + 1)) + * #define NFREE_READ(out) COUNTS(out)[0] + * #define NREGS_READ(out) COUNTS(out)[1] + * #define SIZE_READ(out) COUNTS(out)[2] + * #define BIN_NFREE_READ(out) COUNTS(out)[3] + * #define BIN_NREGS_READ(out) COUNTS(out)[4] + * + * and then write e.g. NFREE_READ(oldp) to fetch the output. See the unit test + * test_utilization_query in test/unit/mallctl.c for an example. + * + * For a typical defragmentation workflow making use of this API for + * understanding the fragmentation level, please refer to the comment for + * experimental_utilization_batch_query_ctl. + * + * It's up to the application how to determine the significance of + * fragmentation relying on the outputs returned. Possible choices are: + * + * (a) if extent utilization ratio is below certain threshold, + * (b) if extent memory consumption is above certain threshold, + * (c) if extent utilization ratio is significantly below bin utilization ratio, + * (d) if input pointer deviates a lot from potential reallocation address, or + * (e) some selection/combination of the above. + * + * The caller needs to make sure that the input/output arguments are valid, + * in particular, that the size of the output is correct, i.e.: + * + * *oldlenp = sizeof(void *) + sizeof(size_t) * 5 + * + * Otherwise, the function immediately returns EINVAL without touching anything. + * + * In the rare case where there's no associated extent found for the input + * pointer, the function zeros out all output fields and return. Please refer + * to the comment for experimental_utilization_batch_query_ctl to understand the + * motivation from C++. + */ +static int +experimental_utilization_query_ctl(tsd_t *tsd, const size_t *mib, + size_t miblen, void *oldp, size_t *oldlenp, void *newp, size_t newlen) { + int ret; + + assert(sizeof(extent_util_stats_verbose_t) + == sizeof(void *) + sizeof(size_t) * 5); + + if (oldp == NULL || oldlenp == NULL + || *oldlenp != sizeof(extent_util_stats_verbose_t) + || newp == NULL) { + ret = EINVAL; + goto label_return; + } + + void *ptr = NULL; + WRITE(ptr, void *); + extent_util_stats_verbose_t *util_stats + = (extent_util_stats_verbose_t *)oldp; + extent_util_stats_verbose_get(tsd_tsdn(tsd), ptr, + &util_stats->nfree, &util_stats->nregs, &util_stats->size, + &util_stats->bin_nfree, &util_stats->bin_nregs, + &util_stats->slabcur_addr); + ret = 0; + +label_return: + return ret; +} + +/* + * Given an input array of pointers, output three memory utilization entries of + * type size_t for each input pointer about the extent it resides in: + * + * (a) number of free regions in the extent, + * (b) number of regions in the extent, and + * (c) size of the extent in terms of bytes. + * + * This API is mainly intended for small class allocations, where extents are + * used as slab. In case of large class allocations, the outputs are trivial: + * "(a)" will be 0, "(b)" will be 1, and "(c)" will be the usable size. + * + * Note that multiple input pointers may reside on a same extent so the output + * fields may contain duplicates. + * + * The format of the input/output looks like: + * + * input[0]: 1st_pointer_to_query | output[0]: 1st_extent_n_free_regions + * | output[1]: 1st_extent_n_regions + * | output[2]: 1st_extent_size + * input[1]: 2nd_pointer_to_query | output[3]: 2nd_extent_n_free_regions + * | output[4]: 2nd_extent_n_regions + * | output[5]: 2nd_extent_size + * ... | ... + * + * The input array and size are respectively passed in by newp and newlen, and + * the output array and size are respectively oldp and *oldlenp. + * + * It can be beneficial to define the following macros to make it easier to + * access the output: + * + * #define NFREE_READ(out, i) out[(i) * 3] + * #define NREGS_READ(out, i) out[(i) * 3 + 1] + * #define SIZE_READ(out, i) out[(i) * 3 + 2] + * + * and then write e.g. NFREE_READ(oldp, i) to fetch the output. See the unit + * test test_utilization_batch in test/unit/mallctl.c for a concrete example. + * + * A typical workflow would be composed of the following steps: + * + * (1) flush tcache: mallctl("thread.tcache.flush", ...) + * (2) initialize input array of pointers to query fragmentation + * (3) allocate output array to hold utilization statistics + * (4) query utilization: mallctl("experimental.utilization.batch_query", ...) + * (5) (optional) decide if it's worthwhile to defragment; otherwise stop here + * (6) disable tcache: mallctl("thread.tcache.enabled", ...) + * (7) defragment allocations with significant fragmentation, e.g.: + * for each allocation { + * if it's fragmented { + * malloc(...); + * memcpy(...); + * free(...); + * } + * } + * (8) enable tcache: mallctl("thread.tcache.enabled", ...) + * + * The application can determine the significance of fragmentation themselves + * relying on the statistics returned, both at the overall level i.e. step "(5)" + * and at individual allocation level i.e. within step "(7)". Possible choices + * are: + * + * (a) whether memory utilization ratio is below certain threshold, + * (b) whether memory consumption is above certain threshold, or + * (c) some combination of the two. + * + * The caller needs to make sure that the input/output arrays are valid and + * their sizes are proper as well as matched, meaning: + * + * (a) newlen = n_pointers * sizeof(const void *) + * (b) *oldlenp = n_pointers * sizeof(size_t) * 3 + * (c) n_pointers > 0 + * + * Otherwise, the function immediately returns EINVAL without touching anything. + * + * In the rare case where there's no associated extent found for some pointers, + * rather than immediately terminating the computation and raising an error, + * the function simply zeros out the corresponding output fields and continues + * the computation until all input pointers are handled. The motivations of + * such a design are as follows: + * + * (a) The function always either processes nothing or processes everything, and + * never leaves the output half touched and half untouched. + * + * (b) It facilitates usage needs especially common in C++. A vast variety of + * C++ objects are instantiated with multiple dynamic memory allocations. For + * example, std::string and std::vector typically use at least two allocations, + * one for the metadata and one for the actual content. Other types may use + * even more allocations. When inquiring about utilization statistics, the + * caller often wants to examine into all such allocations, especially internal + * one(s), rather than just the topmost one. The issue comes when some + * implementations do certain optimizations to reduce/aggregate some internal + * allocations, e.g. putting short strings directly into the metadata, and such + * decisions are not known to the caller. Therefore, we permit pointers to + * memory usages that may not be returned by previous malloc calls, and we + * provide the caller a convenient way to identify such cases. + */ +static int +experimental_utilization_batch_query_ctl(tsd_t *tsd, const size_t *mib, + size_t miblen, void *oldp, size_t *oldlenp, void *newp, size_t newlen) { + int ret; + + assert(sizeof(extent_util_stats_t) == sizeof(size_t) * 3); + + const size_t len = newlen / sizeof(const void *); + if (oldp == NULL || oldlenp == NULL || newp == NULL || newlen == 0 + || newlen != len * sizeof(const void *) + || *oldlenp != len * sizeof(extent_util_stats_t)) { + ret = EINVAL; + goto label_return; + } + + void **ptrs = (void **)newp; + extent_util_stats_t *util_stats = (extent_util_stats_t *)oldp; + size_t i; + for (i = 0; i < len; ++i) { + extent_util_stats_get(tsd_tsdn(tsd), ptrs[i], + &util_stats[i].nfree, &util_stats[i].nregs, + &util_stats[i].size); + } + ret = 0; + +label_return: + return ret; +} diff --git a/src/extent.c b/src/extent.c index 62086c7d..814f0a32 100644 --- a/src/extent.c +++ b/src/extent.c @@ -2280,3 +2280,72 @@ extent_boot(void) { return false; } + +void +extent_util_stats_get(tsdn_t *tsdn, const void *ptr, + size_t *nfree, size_t *nregs, size_t *size) { + assert(ptr != NULL && nfree != NULL && nregs != NULL && size != NULL); + + const extent_t *extent = iealloc(tsdn, ptr); + if (unlikely(extent == NULL)) { + *nfree = *nregs = *size = 0; + return; + } + + *size = extent_size_get(extent); + if (!extent_slab_get(extent)) { + *nfree = 0; + *nregs = 1; + } else { + *nfree = extent_nfree_get(extent); + *nregs = bin_infos[extent_szind_get(extent)].nregs; + assert(*nfree <= *nregs); + assert(*nfree * extent_usize_get(extent) <= *size); + } +} + +void +extent_util_stats_verbose_get(tsdn_t *tsdn, const void *ptr, + size_t *nfree, size_t *nregs, size_t *size, + size_t *bin_nfree, size_t *bin_nregs, void **slabcur_addr) { + assert(ptr != NULL && nfree != NULL && nregs != NULL && size != NULL + && bin_nfree != NULL && bin_nregs != NULL && slabcur_addr != NULL); + + const extent_t *extent = iealloc(tsdn, ptr); + if (unlikely(extent == NULL)) { + *nfree = *nregs = *size = *bin_nfree = *bin_nregs = 0; + *slabcur_addr = NULL; + return; + } + + *size = extent_size_get(extent); + if (!extent_slab_get(extent)) { + *nfree = *bin_nfree = *bin_nregs = 0; + *nregs = 1; + *slabcur_addr = NULL; + return; + } + + *nfree = extent_nfree_get(extent); + const szind_t szind = extent_szind_get(extent); + *nregs = bin_infos[szind].nregs; + assert(*nfree <= *nregs); + assert(*nfree * extent_usize_get(extent) <= *size); + + const arena_t *arena = extent_arena_get(extent); + assert(arena != NULL); + const unsigned binshard = extent_binshard_get(extent); + bin_t *bin = &arena->bins[szind].bin_shards[binshard]; + + malloc_mutex_lock(tsdn, &bin->lock); + if (config_stats) { + *bin_nregs = *nregs * bin->stats.curslabs; + assert(*bin_nregs >= bin->stats.curregs); + *bin_nfree = *bin_nregs - bin->stats.curregs; + } else { + *bin_nfree = *bin_nregs = 0; + } + *slabcur_addr = extent_addr_get(bin->slabcur); + assert(*slabcur_addr != NULL); + malloc_mutex_unlock(tsdn, &bin->lock); +} |