diff options
Diffstat (limited to 'platform/linux-generic/odp_ishmpool.c')
| -rw-r--r-- | platform/linux-generic/odp_ishmpool.c | 657 |
1 files changed, 657 insertions, 0 deletions
diff --git a/platform/linux-generic/odp_ishmpool.c b/platform/linux-generic/odp_ishmpool.c new file mode 100644 index 000000000..fb2cf43c4 --- /dev/null +++ b/platform/linux-generic/odp_ishmpool.c @@ -0,0 +1,657 @@ +/* SPDX-License-Identifier: BSD-3-Clause + * Copyright (c) 2017-2018 Linaro Limited + */ + +/* This file gathers the buddy and slab allocation functionality provided + * by _ishm. + * _odp_ishmpool_create() can be used to create a pool for buddy/slab + * allocation. _odp_ishmpool_create() will allocate a memory area using + * ishm_reserve() for both the control part (needed for tracking + * allocation/free...) and the user memory itself (part of which will be given + * at each ishmpool_alloc()). + * The element size provided at pool creation time determines whether + * to pool will of type buddy or slab. + * For buddy, all allocations are rounded to the nearest power of 2. + * + * The implementation of the buddy allocator is very traditional: it + * maintains N lists of free buffers. + * The control part actually contains these N queue heads, (N-M are actually + * used), the free buffers themselves being used for chaining (the chaining info + * is in the buffers: as they are "free" they should not be touched by the + * user). The control part also contains a array of bytes for remembering + * the size (actually the order) of the allocated buffers: + * There are 2^(N-M) such bytes, this number being the maximum number of + * allocated buffers (when all allocation are <= 2^M bytes) + * Buddy allocators handle fragmentation by splitting or merging blocks by 2. + * They guarantee a minimum efficiency of 50%, at worse case fragmentation. + * + * Slab implementation is even simpler, all free elements being queued in + * one single queue at init, taken from this queue when allocated and + * returned to this same queue when freed. + * + * The reason for not using malloc() is that malloc does not guarantee + * memory sharability between ODP threads (regardless of their implementation) + * which ishm_reserve() can do. see the comments around + * _odp_ishmbud_pool_create() and ishm_reserve() for more details. + * + * This file is divided in 3 sections: the first one regroups functions + * needed by the buddy allocation. + * The second one regroups the functions needed by the slab allocator. + * The third section regroups the common functions exported externally. + */ + +#include <odp_posix_extensions.h> + +#include <odp/api/spinlock.h> +#include <odp/api/align.h> +#include <odp/api/debug.h> + +#include <odp_shm_internal.h> +#include <odp_debug_internal.h> +#include <odp_macros_internal.h> +#include <odp_shm_internal.h> +#include <odp_ishmpool_internal.h> + +#include <stdlib.h> +#include <stdio.h> +#include <unistd.h> +#include <string.h> +#include <inttypes.h> + +#define BUDDY_MIN_SIZE 32 /* minimal buddy allocation size */ + +typedef _odp_ishm_pool_t pool_t; /* for shorter writing */ + +/* array of ishm block index used for pools. only used for pool + * lookup by name */ +#define MAX_NB_POOL 100 +static int pool_blk_idx[MAX_NB_POOL]; + +/* section 1: functions for buddy allocation: */ + +/* free buddy blocks contains the following structure, used to link the + * free blocks together. + */ +typedef struct bblock_t { + struct bblock_t *next; + uint32_t order; +} bblock_t; + +/* value set in the 'order' table when the block is not allocated: */ +#define BBLOCK_FREE 0 + +/* compute ceil(log2(size)) */ +static uint8_t clog2(uint64_t size) +{ + uint64_t sz; + uint32_t bit; + uint8_t res; + + sz = size; /* we start by computing res = log2(sz)... */ + res = 0; + for (bit = 32; bit ; bit >>= 1) { + if (sz >= ((uint64_t)1 << bit)) { + sz >>= bit; + res += bit; + } + } + if (((uint64_t)1 << res) < size) /* ...and then ceil(x) */ + res++; + + return res; +} + +/* + * given a bblock address, and an order value, returns the address + * of the buddy bblock (the other "half") + */ +static inline bblock_t *get_bblock_buddy(pool_t *bpool, bblock_t *addr, + uint8_t order) +{ + uintptr_t b; + + b = ((uintptr_t)addr - (uintptr_t)bpool->ctrl.user_addr); + b ^= 1 << order; + return (void *)(b + (uintptr_t)bpool->ctrl.user_addr); +} + +/* + * given a buddy block address, return its number (used for busy flags): + */ +static inline uintptr_t get_bblock_nr(pool_t *bpool, void *addr) +{ + uintptr_t b; + uint8_t min_order; + + min_order = bpool->ctrl.min_order; + b = ((uintptr_t)addr - (uintptr_t)bpool->ctrl.user_addr) >> min_order; + return b; +} + +/* remove bblock from the list for bblocks of rank order. The bblock to be + * removed is really expected to be on the list: not finding it is an error */ +static inline void remove_from_list(pool_t *bpool, uint8_t order, + bblock_t *bblock) +{ + bblock_t *curr; + bblock_t *prev = NULL; + + curr = bpool->ctrl.free_heads[order]; + if (!curr) + goto remove_from_list_error; + + if (curr == bblock) { + bpool->ctrl.free_heads[order] = curr->next; + return; + } + + while (curr) { + if (curr == bblock) { + prev->next = curr->next; + return; + } + prev = curr; + curr = curr->next; + } + +remove_from_list_error: + _ODP_ERR("List corrupted\n"); +} + +/* + * create a buddy memory pool of given size (actually nearest power of 2), + * where allocation will never be smaller than min_alloc. + * returns a pointer to the created buddy_pool + * The allocated area contains: + * - The _odp_ishm_pool_ctrl_t structure + * - The array of ((order - min_order) of free list heads + * - The array of 'order' values, remembering sizes of allocated bblocks + * - alignment to cache line + * - The user memory + */ +static pool_t *_odp_ishmbud_pool_create(const char *pool_name, int store_idx, + uint64_t size, + uint64_t min_alloc, int flags) +{ + uint8_t order; /* pool order = ceil(log2(size)) */ + uint8_t min_order; /* pool min_order = ceil(log2(min_alloc))*/ + uint32_t max_nb_bblock; /* max number of bblock, when smallest */ + uint32_t control_sz; /* size of control area */ + uint32_t free_head_sz; /* mem area needed for list heads */ + uint32_t saved_order_sz; /* mem area to remember given sizes */ + uint64_t user_sz; /* 2^order bytes */ + uint64_t total_sz; /* total size to request */ + int blk_idx; /* as returned by _ishm_resrve() */ + pool_t *bpool; + int i; + bblock_t *first_block; + + /* a bblock_t must fit in the buffers for linked chain! */ + if (min_alloc < sizeof(bblock_t)) + min_alloc = sizeof(bblock_t); + + /* pool order is such that 2^order = size. same for min_order */ + order = clog2(size); + min_order = clog2(min_alloc); + + /* check parameters obvious wishes: */ + if (order >= 64) + return NULL; + if (order < min_order) + return NULL; + + /* at worst case, all bblocks have smallest (2^min_order) size */ + max_nb_bblock = (1 << (order - min_order)); + + /* space needed for the control area (padded to cache line size)*/ + control_sz = _ODP_ROUNDUP_CACHE_LINE(sizeof(_odp_ishm_pool_ctrl_t)); + + /* space needed for 'order' free bblock list heads: */ + /* Note that only lists from min_order to order are really used.*/ + free_head_sz = _ODP_ROUNDUP_CACHE_LINE(sizeof(void *) * (order + 1)); + + /* space needed for order -i.e. size- storage of alloc'd bblock:*/ + saved_order_sz = _ODP_ROUNDUP_CACHE_LINE(max_nb_bblock * sizeof(uint8_t)); + + /* space needed for user area is 2^order bytes: */ + user_sz = 1ULL << order; + + total_sz = control_sz + + free_head_sz + + saved_order_sz + + user_sz; + + /* allocate required memory: */ + blk_idx = _odp_ishm_reserve(pool_name, total_sz, -1, + ODP_CACHE_LINE_SIZE, 0, flags, 0); + if (blk_idx < 0) { + _ODP_ERR("_odp_ishm_reserve failed."); + return NULL; + } + + bpool = _odp_ishm_address(blk_idx); + if (bpool == NULL) { + _ODP_ERR("_odp_ishm_address failed."); + return NULL; + } + + /* store in pool array (needed for look up): */ + pool_blk_idx[store_idx] = blk_idx; + + /* remember block index, needed when pool is destroyed */ + bpool->ctrl.ishm_blk_idx = blk_idx; + + /* remember element size: 0 means unknown size, i.e. buddy alloation*/ + bpool->ctrl.element_sz = 0; + + /* prepare mutex: */ + odp_spinlock_init(&bpool->ctrl.lock); + + /* initialise pointers and things... */ + bpool->ctrl.order = order; + bpool->ctrl.min_order = min_order; + bpool->ctrl.free_heads = + (void *)((uintptr_t)bpool + control_sz); + bpool->ctrl.alloced_order = + (uint8_t *)((uintptr_t)bpool->ctrl.free_heads + free_head_sz); + bpool->ctrl.user_addr = + (void *)((uintptr_t)bpool->ctrl.alloced_order + saved_order_sz); + + /* initialize all free list to NULL, except the top biggest element:*/ + for (i = 0; i < (order - min_order); i++) + bpool->ctrl.free_heads[i] = NULL; + bpool->ctrl.free_heads[order] = bpool->ctrl.user_addr; + first_block = (bblock_t *)bpool->ctrl.user_addr; + first_block->next = NULL; + first_block->order = order; + + /* set all 'order' of allocated bblocks to free: */ + memset(bpool->ctrl.alloced_order, BBLOCK_FREE, saved_order_sz); + + return bpool; +} + +/* allocated memory from the given buddy pool */ +static void *_odp_ishmbud_alloc(pool_t *bpool, uint64_t size) +{ + uint32_t rq_order; /* requested order */ + uint32_t try_order; + bblock_t *bblock; + bblock_t *buddy; + uintptr_t nr; + + /* if size is zero or too big reject: */ + if ((!size) && (size > (1ULL << bpool->ctrl.order))) { + _ODP_ERR("Invalid alloc size (0 or larger than whole pool)\n"); + return NULL; + } + + /* compute ceil(log2(size)), to get the requested block order: */ + rq_order = clog2(size); + + /* make sure the requested order is bigger (or same) as minimum! */ + if (rq_order < bpool->ctrl.min_order) + rq_order = bpool->ctrl.min_order; + + /* mutex from here: */ + odp_spinlock_lock(&bpool->ctrl.lock); + + /* now, start trying to allocate a bblock of rq_order. If that + * fails keep trying larger orders until pool order is reached */ + bblock = NULL; + for (try_order = rq_order; try_order <= bpool->ctrl.order; + try_order++) { + if (bpool->ctrl.free_heads[try_order]) { + /* remove from list: */ + bblock = + (bblock_t *)(bpool->ctrl.free_heads[try_order]); + bpool->ctrl.free_heads[try_order] = bblock->next; + break; + } + } + + if (!bblock) { + odp_spinlock_unlock(&bpool->ctrl.lock); + _ODP_ERR("Out of memory. (Buddy pool full)\n"); + return NULL; + } + + /* OK: we got a block, but possibbly too large (if try_order>rq_order) + * return the extra halves to the pool hence splitting the bblock at + * each 'extra' order: */ + while (try_order-- > rq_order) { + /* split: */ + buddy = (bblock_t *)((uintptr_t)bblock + (1 << try_order)); + buddy->order = try_order; + /* add to list: */ + buddy->next = bpool->ctrl.free_heads[try_order]; + bpool->ctrl.free_heads[try_order] = buddy; + /* mark as free (non allocated block get size 0): */ + nr = get_bblock_nr(bpool, buddy); + bpool->ctrl.alloced_order[nr] = BBLOCK_FREE; + } + + /* remember the size if the allocated block: */ + nr = get_bblock_nr(bpool, bblock); + bpool->ctrl.alloced_order[nr] = rq_order; + + /* and return the allocated block! */ + odp_spinlock_unlock(&bpool->ctrl.lock); + return (void *)bblock; +} + +/* free a previously allocated buffer from a given buddy pool */ +static int _odp_ishmbud_free(pool_t *bpool, void *addr) +{ + uintptr_t user_start; /* start of user area */ + uintptr_t user_stop; /* stop of user area */ + uintptr_t mask; /* 2^min_order - 1 */ + bblock_t *bblock; /* bblock being freed */ + bblock_t *buddy; /* buddy bblock of bblock being freed */ + uint8_t order; /* order of block being freed */ + uintptr_t nr; /* block number */ + + /* freeing NULL is regarded as OK, though without any effect: */ + if (!addr) + return 0; + + user_start = (uintptr_t)bpool->ctrl.user_addr; + user_stop = user_start + ((uintptr_t)1 << bpool->ctrl.order); + mask = ((uintptr_t)1 << bpool->ctrl.min_order) - 1; + + /* some sanity checks: check that given address is within pool and + * that relative address has 2^min_order granularity: */ + if (((uintptr_t)addr < user_start) || + ((uintptr_t)addr > user_stop) || + (((uintptr_t)addr - user_start) & mask)) { + _ODP_ERR("Invalid address to be freed\n"); + return -1; + } + + /* mutex from here: */ + odp_spinlock_lock(&bpool->ctrl.lock); + + /* collect saved block order and make sure bblock was allocated */ + bblock = (bblock_t *)addr; + nr = get_bblock_nr(bpool, bblock); + order = bpool->ctrl.alloced_order[nr]; + if (order == BBLOCK_FREE) { + _ODP_ERR("Double free error\n"); + odp_spinlock_unlock(&bpool->ctrl.lock); + return -1; + } + + /* this looks like a valid free, mark at least this as free: */ + bpool->ctrl.alloced_order[nr] = BBLOCK_FREE; + + /* go up in orders, trying to merge buddies... */ + while (order < bpool->ctrl.order) { + buddy = get_bblock_buddy(bpool, bblock, order); + /*if buddy is not free: no further merge possible */ + nr = get_bblock_nr(bpool, buddy); + if (bpool->ctrl.alloced_order[nr] != BBLOCK_FREE) + break; + /*merge only bblock of same order:*/ + if (buddy->order != order) + break; + /*merge: remove buddy from free list: */ + remove_from_list(bpool, order, buddy); + /*merge: make sure we point at start of block: */ + if (bblock > buddy) + bblock = buddy; + /*merge: size of block has doubled: increase order: */ + order++; + } + + /* insert the bblock into its correct free block list: */ + bblock->next = bpool->ctrl.free_heads[order]; + bpool->ctrl.free_heads[order] = bblock; + + /* remember the (possibly now merged) block order: */ + bblock->order = order; + + odp_spinlock_unlock(&bpool->ctrl.lock); + return 0; +} + +/* section 2: functions for slab allocation: */ + +/* free slab blocks contains the following structure, used to link the + * free blocks together. + */ +typedef struct sblock_t { + struct sblock_t *next; +} sblock_t; + +/* + * create a slab memory pool of given size (rounded up to the nearest integer + * number of element, where each element has size 'elt_size'). + * returns a pointer to the created slab pool. + * The allocated area contains: + * - The _odp_ishm_pool_ctrl_t structure + * - alignment to cache line + * - The user memory + */ +static pool_t *_odp_ishmslab_pool_create(const char *pool_name, int store_idx, + uint64_t size, + uint64_t elt_size, int flags) +{ + uint32_t nb_sblock; /* number of elements in the pool */ + uint32_t control_sz; /* size of control area */ + uint64_t total_sz; /* total size to request */ + uint64_t user_sz; /* 2^order bytes */ + int blk_idx; /* as returned by _ishm_reserve() */ + pool_t *spool; + unsigned int i; + sblock_t *block; + + /* a sblock_t must fit in the buffers for linked chain! */ + if (elt_size < sizeof(bblock_t)) { + elt_size = sizeof(bblock_t); + size = size * (sizeof(bblock_t) / elt_size + + ((sizeof(bblock_t) % elt_size) ? 1 : 0)); + } + + /* nb of element fitting in the pool is just ceil(size/elt_size)*/ + nb_sblock = (size / elt_size) + ((size % elt_size) ? 1 : 0); + + /* space needed for the control area (padded to cache line size)*/ + control_sz = _ODP_ROUNDUP_CACHE_LINE(sizeof(_odp_ishm_pool_ctrl_t)); + + /* space needed for user area is : */ + user_sz = nb_sblock * elt_size; + + total_sz = control_sz + + user_sz; + + /* allocate required memory: */ + blk_idx = _odp_ishm_reserve(pool_name, total_sz, -1, + ODP_CACHE_LINE_SIZE, 0, flags, 0); + if (blk_idx < 0) { + _ODP_ERR("_odp_ishm_reserve failed."); + return NULL; + } + + spool = _odp_ishm_address(blk_idx); + if (spool == NULL) { + _ODP_ERR("_odp_ishm_address failed."); + return NULL; + } + + /* store in pool array (needed for look up): */ + pool_blk_idx[store_idx] = blk_idx; + + /* remember block index, needed when pool is destroyed */ + spool->ctrl.ishm_blk_idx = blk_idx; + + /* remember element (sblock) size and their number: */ + spool->ctrl.element_sz = elt_size; + spool->ctrl.nb_elem = nb_sblock; + + /* prepare mutex: */ + odp_spinlock_init(&spool->ctrl.lock); + + /* initialise pointers and things... */ + spool->ctrl.user_addr = + (void *)((uintptr_t)spool + control_sz); + + /* initialise the free list with the list of all elements:*/ + spool->ctrl.free_head = spool->ctrl.user_addr; + for (i = 0; i < nb_sblock - 1; i++) { + block = (sblock_t *)((uintptr_t)spool->ctrl.user_addr + + i * (uintptr_t)elt_size); + block->next = (sblock_t *)((uintptr_t)block + + (uintptr_t)elt_size); + } + block = (sblock_t *)((uintptr_t)spool->ctrl.user_addr + + (nb_sblock - 1) * (uintptr_t)elt_size); + block->next = NULL; + + return spool; +} + +/* allocated memory from the given slab pool */ +static void *_odp_ishmslab_alloc(pool_t *spool, uint64_t size) +{ + void *ret; + sblock_t *block; + + if (size > spool->ctrl.element_sz) + return NULL; + + odp_spinlock_lock(&spool->ctrl.lock); + ret = spool->ctrl.free_head; + if (!ret) { + odp_spinlock_unlock(&spool->ctrl.lock); + _ODP_ERR("Out of memory. (Slab pool full)\n"); + return NULL; + } + + block = (sblock_t *)ret; + spool->ctrl.free_head = block->next; + + odp_spinlock_unlock(&spool->ctrl.lock); + return ret; +} + +/* free a previously allocated buffer from a given slab pool */ +static int _odp_ishmslab_free(pool_t *spool, void *addr) +{ + uintptr_t user_start; /* start of user area */ + uintptr_t user_stop; /* stop of user area */ + sblock_t *block; + + /* freeing NULL is regarded as OK, though without any effect: */ + if (!addr) + return 0; + + user_start = (uintptr_t)spool->ctrl.user_addr; + user_stop = user_start + spool->ctrl.element_sz * spool->ctrl.nb_elem; + + /* some sanity checks: check that given address is within pool and + * that relative address has element_sz granularity: */ + if (((uintptr_t)addr < user_start) || + ((uintptr_t)addr > user_stop) || + (((uintptr_t)addr - user_start) % spool->ctrl.element_sz)) { + _ODP_ERR("Invalid address to be freed\n"); + return -1; + } + + odp_spinlock_lock(&spool->ctrl.lock); + block = (sblock_t *)addr; + block->next = (sblock_t *)spool->ctrl.free_head; + spool->ctrl.free_head = addr; + odp_spinlock_unlock(&spool->ctrl.lock); + + return 0; +} + +/* section 3: common, external functions: */ + +/* create a pool: either with fixed alloc size (if max_alloc/min_alloc<2) or + * of variable block size (if max_alloc == 0) */ +pool_t *_odp_ishm_pool_create(const char *pool_name, uint64_t size, + uint64_t min_alloc, uint64_t max_alloc, int flags) +{ + int store_idx; + uint64_t real_pool_sz; + + if (min_alloc > max_alloc) { + _ODP_ERR("invalid parameter: min_alloc > max_alloc"); + return NULL; + } + + /* search for a free index in pool_blk_idx for the pool */ + for (store_idx = 0; store_idx < MAX_NB_POOL; store_idx++) { + if (pool_blk_idx[store_idx] < 0) + break; + } + if (store_idx == MAX_NB_POOL) { + _ODP_ERR("Max number of pool reached (MAX_NB_POOL)"); + return NULL; + } + + if ((min_alloc == 0) || ((max_alloc / min_alloc) > 2)) { + /* alloc variation is not constant enough: we go for a buddy + * allocator. The pool efficiency may go as low as 50% + * so we double the required size to make sure we can satisfy + * the user request */ + real_pool_sz = 2 * size; + return _odp_ishmbud_pool_create(pool_name, store_idx, + real_pool_sz, + BUDDY_MIN_SIZE, flags); + } else { + /* min and max are close enough so we go for constant size + * allocator: + * make sure the pool can fit the required size, even when + * only min_alloc allocation are performed: */ + real_pool_sz = ((size / min_alloc) + + ((size % min_alloc) ? 1 : 0)) + * max_alloc; + return _odp_ishmslab_pool_create(pool_name, store_idx, + real_pool_sz, + max_alloc, flags); + } +} + +/* destroy a pool. everything goes away. no operation on the pool should + * follow. */ +int _odp_ishm_pool_destroy(pool_t *pool) +{ + int store_idx; + + for (store_idx = 0; store_idx < MAX_NB_POOL; store_idx++) { + if (pool_blk_idx[store_idx] == pool->ctrl.ishm_blk_idx) { + pool_blk_idx[store_idx] = -1; + break; + } + } + + return _odp_ishm_free_by_index(pool->ctrl.ishm_blk_idx); +} + +/* allocated a buffer from a pool */ +void *_odp_ishm_pool_alloc(_odp_ishm_pool_t *pool, uint64_t size) +{ + if (!pool->ctrl.element_sz) + return _odp_ishmbud_alloc(pool, size); + else + return _odp_ishmslab_alloc(pool, size); +} + +/* free a previously allocated buffer from a pool */ +int _odp_ishm_pool_free(_odp_ishm_pool_t *pool, void *addr) +{ + if (!pool->ctrl.element_sz) + return _odp_ishmbud_free(pool, addr); + else + return _odp_ishmslab_free(pool, addr); +} + +void _odp_ishm_pool_init(void) +{ + int i; + + for (i = 0; i < MAX_NB_POOL; i++) + pool_blk_idx[i] = -1; +} |