#include "jemalloc/internal/jemalloc_preamble.h" #include "jemalloc/internal/assert.h" #include "jemalloc/internal/bit_util.h" #include "jemalloc/internal/bitmap.h" #include "jemalloc/internal/pages.h" #include "jemalloc/internal/sc.h" /* * This module computes the size classes used to satisfy allocations. The logic * here was ported more or less line-by-line from a shell script, and because of * that is not the most idiomatic C. Eventually we should fix this, but for now * at least the damage is compartmentalized to this file. */ sc_data_t sc_data_global; static size_t reg_size_compute(int lg_base, int lg_delta, int ndelta) { return (ZU(1) << lg_base) + (ZU(ndelta) << lg_delta); } /* Returns the number of pages in the slab. */ static int slab_size(int lg_page, int lg_base, int lg_delta, int ndelta) { size_t page = (ZU(1) << lg_page); size_t reg_size = reg_size_compute(lg_base, lg_delta, ndelta); size_t try_slab_size = page; size_t try_nregs = try_slab_size / reg_size; size_t perfect_slab_size = 0; bool perfect = false; /* * This loop continues until we find the least common multiple of the * page size and size class size. Size classes are all of the form * base + ndelta * delta == (ndelta + base/ndelta) * delta, which is * (ndelta + ngroup) * delta. The way we choose slabbing strategies * means that delta is at most the page size and ndelta < ngroup. So * the loop executes for at most 2 * ngroup - 1 iterations, which is * also the bound on the number of pages in a slab chosen by default. * With the current default settings, this is at most 7. */ while (!perfect) { perfect_slab_size = try_slab_size; size_t perfect_nregs = try_nregs; try_slab_size += page; try_nregs = try_slab_size / reg_size; if (perfect_slab_size == perfect_nregs * reg_size) { perfect = true; } } return (int)(perfect_slab_size / page); } static void size_class( /* Output. */ sc_t *sc, /* Configuration decisions. */ int lg_max_lookup, int lg_page, int lg_ngroup, /* Inputs specific to the size class. */ int index, int lg_base, int lg_delta, int ndelta) { sc->index = index; sc->lg_base = lg_base; sc->lg_delta = lg_delta; sc->ndelta = ndelta; sc->psz = (reg_size_compute(lg_base, lg_delta, ndelta) % (ZU(1) << lg_page) == 0); size_t size = (ZU(1) << lg_base) + (ZU(ndelta) << lg_delta); if (index == 0) { assert(!sc->psz); } if (size < (ZU(1) << (lg_page + lg_ngroup))) { sc->bin = true; sc->pgs = slab_size(lg_page, lg_base, lg_delta, ndelta); } else { sc->bin = false; sc->pgs = 0; } if (size <= (ZU(1) << lg_max_lookup)) { sc->lg_delta_lookup = lg_delta; } else { sc->lg_delta_lookup = 0; } } static void size_classes( /* Output. */ sc_data_t *sc_data, /* Determined by the system. */ size_t lg_ptr_size, int lg_quantum, /* Configuration decisions. */ int lg_tiny_min, int lg_max_lookup, int lg_page, int lg_ngroup) { int ptr_bits = (1 << lg_ptr_size) * 8; int ngroup = (1 << lg_ngroup); int ntiny = 0; int nlbins = 0; int lg_tiny_maxclass = (unsigned)-1; int nbins = 0; int npsizes = 0; int index = 0; int ndelta = 0; int lg_base = lg_tiny_min; int lg_delta = lg_base; /* Outputs that we update as we go. */ size_t lookup_maxclass = 0; size_t small_maxclass = 0; int lg_large_minclass = 0; size_t large_maxclass = 0; /* Tiny size classes. */ while (lg_base < lg_quantum) { sc_t *sc = &sc_data->sc[index]; size_class(sc, lg_max_lookup, lg_page, lg_ngroup, index, lg_base, lg_delta, ndelta); if (sc->lg_delta_lookup != 0) { nlbins = index + 1; } if (sc->psz) { npsizes++; } if (sc->bin) { nbins++; } ntiny++; /* Final written value is correct. */ lg_tiny_maxclass = lg_base; index++; lg_delta = lg_base; lg_base++; } /* First non-tiny (pseudo) group. */ if (ntiny != 0) { sc_t *sc = &sc_data->sc[index]; /* * See the note in sc.h; the first non-tiny size class has an * unusual encoding. */ lg_base--; ndelta = 1; size_class(sc, lg_max_lookup, lg_page, lg_ngroup, index, lg_base, lg_delta, ndelta); index++; lg_base++; lg_delta++; if (sc->psz) { npsizes++; } if (sc->bin) { nbins++; } } while (ndelta < ngroup) { sc_t *sc = &sc_data->sc[index]; size_class(sc, lg_max_lookup, lg_page, lg_ngroup, index, lg_base, lg_delta, ndelta); index++; ndelta++; if (sc->psz) { npsizes++; } if (sc->bin) { nbins++; } } /* All remaining groups. */ lg_base = lg_base + lg_ngroup; while (lg_base < ptr_bits - 1) { ndelta = 1; int ndelta_limit; if (lg_base == ptr_bits - 2) { ndelta_limit = ngroup - 1; } else { ndelta_limit = ngroup; } while (ndelta <= ndelta_limit) { sc_t *sc = &sc_data->sc[index]; size_class(sc, lg_max_lookup, lg_page, lg_ngroup, index, lg_base, lg_delta, ndelta); if (sc->lg_delta_lookup != 0) { nlbins = index + 1; /* Final written value is correct. */ lookup_maxclass = (ZU(1) << lg_base) + (ZU(ndelta) << lg_delta); } if (sc->psz) { npsizes++; } if (sc->bin) { nbins++; /* Final written value is correct. */ small_maxclass = (ZU(1) << lg_base) + (ZU(ndelta) << lg_delta); if (lg_ngroup > 0) { lg_large_minclass = lg_base + 1; } else { lg_large_minclass = lg_base + 2; } } large_maxclass = (ZU(1) << lg_base) + (ZU(ndelta) << lg_delta); index++; ndelta++; } lg_base++; lg_delta++; } /* Additional outputs. */ int nsizes = index; unsigned lg_ceil_nsizes = lg_ceil(nsizes); /* Fill in the output data. */ sc_data->ntiny = ntiny; sc_data->nlbins = nlbins; sc_data->nbins = nbins; sc_data->nsizes = nsizes; sc_data->lg_ceil_nsizes = lg_ceil_nsizes; sc_data->npsizes = npsizes; sc_data->lg_tiny_maxclass = lg_tiny_maxclass; sc_data->lookup_maxclass = lookup_maxclass; sc_data->small_maxclass = small_maxclass; sc_data->lg_large_minclass = lg_large_minclass; sc_data->large_minclass = (ZU(1) << lg_large_minclass); sc_data->large_maxclass = large_maxclass; /* * We compute these values in two ways: * - Incrementally, as above. * - In macros, in sc.h. * The computation is easier when done incrementally, but putting it in * a constant makes it available to the fast paths without having to * touch the extra global cacheline. We assert, however, that the two * computations are equivalent. */ assert(sc_data->npsizes == SC_NPSIZES); assert(sc_data->lg_tiny_maxclass == SC_LG_TINY_MAXCLASS); assert(sc_data->small_maxclass == SC_SMALL_MAXCLASS); assert(sc_data->large_minclass == SC_LARGE_MINCLASS); assert(sc_data->lg_large_minclass == SC_LG_LARGE_MINCLASS); assert(sc_data->large_maxclass == SC_LARGE_MAXCLASS); /* * In the allocation fastpath, we want to assume that we can * unconditionally subtract the requested allocation size from * a ssize_t, and detect passing through 0 correctly. This * results in optimal generated code. For this to work, the * maximum allocation size must be less than SSIZE_MAX. */ assert(SC_LARGE_MAXCLASS < SSIZE_MAX); } void sc_data_init(sc_data_t *sc_data) { assert(!sc_data->initialized); int lg_max_lookup = 12; size_classes(sc_data, LG_SIZEOF_PTR, LG_QUANTUM, SC_LG_TINY_MIN, lg_max_lookup, LG_PAGE, 2); sc_data->initialized = true; } static void sc_data_update_sc_slab_size(sc_t *sc, size_t reg_size, size_t pgs_guess) { size_t min_pgs = reg_size / PAGE; if (reg_size % PAGE != 0) { min_pgs++; } /* * BITMAP_MAXBITS is actually determined by putting the smallest * possible size-class on one page, so this can never be 0. */ size_t max_pgs = BITMAP_MAXBITS * reg_size / PAGE; assert(min_pgs <= max_pgs); assert(min_pgs > 0); assert(max_pgs >= 1); if (pgs_guess < min_pgs) { sc->pgs = (int)min_pgs; } else if (pgs_guess > max_pgs) { sc->pgs = (int)max_pgs; } else { sc->pgs = (int)pgs_guess; } } void sc_data_update_slab_size(sc_data_t *data, size_t begin, size_t end, int pgs) { assert(data->initialized); for (int i = 0; i < data->nsizes; i++) { sc_t *sc = &data->sc[i]; if (!sc->bin) { break; } size_t reg_size = reg_size_compute(sc->lg_base, sc->lg_delta, sc->ndelta); if (begin <= reg_size && reg_size <= end) { sc_data_update_sc_slab_size(sc, reg_size, pgs); } } } void sc_boot(sc_data_t *data) { sc_data_init(data); }