From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: from mga02.intel.com (mga02.intel.com [134.134.136.20]) by dpdk.org (Postfix) with ESMTP id 1DACC4C74 for ; Thu, 31 May 2018 12:58:05 +0200 (CEST) X-Amp-Result: SKIPPED(no attachment in message) X-Amp-File-Uploaded: False Received: from fmsmga004.fm.intel.com ([10.253.24.48]) by orsmga101.jf.intel.com with ESMTP/TLS/DHE-RSA-AES256-GCM-SHA384; 31 May 2018 03:58:03 -0700 X-ExtLoop1: 1 X-IronPort-AV: E=Sophos;i="5.49,463,1520924400"; d="scan'208";a="59202811" Received: from irvmail001.ir.intel.com ([163.33.26.43]) by fmsmga004.fm.intel.com with ESMTP; 31 May 2018 03:58:00 -0700 Received: from sivswdev01.ir.intel.com (sivswdev01.ir.intel.com [10.237.217.45]) by irvmail001.ir.intel.com (8.14.3/8.13.6/MailSET/Hub) with ESMTP id w4VAw0on004132; Thu, 31 May 2018 11:58:00 +0100 Received: from sivswdev01.ir.intel.com (localhost [127.0.0.1]) by sivswdev01.ir.intel.com with ESMTP id w4VAw0SJ002509; Thu, 31 May 2018 11:58:00 +0100 Received: (from aburakov@localhost) by sivswdev01.ir.intel.com with LOCAL id w4VAw0j2002505; Thu, 31 May 2018 11:58:00 +0100 From: Anatoly Burakov To: dev@dpdk.org Cc: thomas@monjalon.net, hemant.agrawal@nxp.com, bruce.richardson@intel.com, ferruh.yigit@intel.com, konstantin.ananyev@intel.com, jerin.jacob@caviumnetworks.com, olivier.matz@6wind.com, stephen@networkplumber.org, nhorman@tuxdriver.com, david.marchand@6wind.com, gowrishankar.m@linux.vnet.ibm.com Date: Thu, 31 May 2018 11:57:49 +0100 Message-Id: <5281a08a4d08db8692b38905ae1c6035343bca99.1527764061.git.anatoly.burakov@intel.com> X-Mailer: git-send-email 1.7.0.7 In-Reply-To: References: In-Reply-To: References: Subject: [dpdk-dev] [RFC 2/3] mem: add device memory reserve/free API X-BeenThere: dev@dpdk.org X-Mailman-Version: 2.1.15 Precedence: list List-Id: DPDK patches and discussions List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , X-List-Received-Date: Thu, 31 May 2018 10:58:06 -0000 In order for hotplug in multiprocess to work reliably, we will need a common shared memory area that is guaranteed to be accessible to all processes at all times. This is accomplished by pre-reserving memory that will be used for device mappings at startup, and managing it at runtime. Two new API calls are added: alloc and free of device memory. Once allocation is requested, memory is considered to be reserved until it is freed back using the same API. Usage of which blocks are occupied is tracked using shared fbarray. This allows us to give out device memory piecemeal and lessen fragmentation. Naturally, this adds a limitation of how much device memory DPDK can use. This is currently set to 2 gigabytes, but will be adjustable in later revisions. Signed-off-by: Anatoly Burakov --- lib/librte_eal/common/eal_common_memory.c | 270 ++++++++++++++++-- .../common/include/rte_eal_memconfig.h | 18 ++ lib/librte_eal/common/include/rte_memory.h | 40 +++ 3 files changed, 312 insertions(+), 16 deletions(-) diff --git a/lib/librte_eal/common/eal_common_memory.c b/lib/librte_eal/common/eal_common_memory.c index 4f0688f9d..8cae9b354 100644 --- a/lib/librte_eal/common/eal_common_memory.c +++ b/lib/librte_eal/common/eal_common_memory.c @@ -33,6 +33,7 @@ */ #define MEMSEG_LIST_FMT "memseg-%" PRIu64 "k-%i-%i" +#define DEVICE_MEMORY_NAME "device_memory" static uint64_t baseaddr_offset; static uint64_t system_page_sz; @@ -904,6 +905,227 @@ rte_memseg_list_walk(rte_memseg_list_walk_t func, void *arg) return ret; } +void * __rte_experimental +rte_mem_dev_memory_alloc(size_t size, size_t align) +{ + struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config; + struct rte_fbarray *arr = &mcfg->device_memory.mem_map_arr; + unsigned int n_pages, page_align; + int start_idx, cur_idx; + void *addr = NULL; + + /* check parameters first */ + if (size == 0 || (size & (system_page_sz - 1)) != 0) { + RTE_LOG(ERR, EAL, "%s(): size is not page-aligned\n", + __func__); + rte_errno = EINVAL; + return NULL; + } + if ((align & (system_page_sz - 1)) != 0) { + RTE_LOG(ERR, EAL, "%s(): alignment is not page-aligned\n", + __func__); + rte_errno = EINVAL; + return NULL; + } + /* PCI BAR sizes can only be powers of two, but this memory may be used + * for more than just PCI BAR mappings, so only check if alignment is + * power of two. + */ + if (align != 0 && !rte_is_power_of_2(align)) { + RTE_LOG(ERR, EAL, "%s(): alignment is not a power of two\n", + __func__); + rte_errno = EINVAL; + return NULL; + } + /* check if device memory map is uninitialized. */ + if (mcfg->device_memory.base_va == NULL || arr->len == 0) { + RTE_LOG(ERR, EAL, "%s(): device memory map is not initialized\n", + __func__); + rte_errno = ENODEV; + return NULL; + } + + n_pages = size / system_page_sz; + page_align = align / system_page_sz; + + /* lock the device memory map */ + rte_spinlock_lock(&mcfg->device_memory.lock); + + start_idx = 0; + while (1) { + size_t offset; + int end; + + cur_idx = rte_fbarray_find_next_n_free(arr, start_idx, n_pages); + if (cur_idx < 0) + break; + + /* if there are alignment requirements, check if the offset we + * found is aligned, and if not, align it and check if we still + * have enough space. + */ + if (page_align != 0 && (cur_idx & (page_align - 1)) != 0) { + unsigned int aligned, len; + + aligned = RTE_ALIGN_CEIL(cur_idx, page_align); + len = rte_fbarray_find_contig_free(arr, aligned); + + /* if there's not enough space, keep looking */ + if (len < n_pages) { + start_idx = aligned + len; + continue; + } + + /* we've found space */ + cur_idx = aligned; + } + end = cur_idx + n_pages; + offset = cur_idx * system_page_sz; + addr = RTE_PTR_ADD(mcfg->device_memory.base_va, + offset); + + /* now, mark all space as occupied */ + for (; cur_idx < end; cur_idx++) + rte_fbarray_set_used(arr, cur_idx); + break; + } + rte_spinlock_unlock(&mcfg->device_memory.lock); + + if (addr != NULL) + RTE_LOG(DEBUG, EAL, "%s(): allocated %p-%p (%lu bytes) for hardware device usage\n", + __func__, addr, RTE_PTR_ADD(addr, size), size); + + return addr; +} + +int __rte_experimental +rte_mem_dev_memory_free(void *addr, size_t size) +{ + struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config; + struct rte_fbarray *arr = &mcfg->device_memory.mem_map_arr; + int cur_idx, end, ret, n_pages, len; + void *map_end, *mem_end; + + /* check parameters first */ + if (size == 0 || (size & (system_page_sz - 1)) != 0) { + RTE_LOG(ERR, EAL, "%s(): size is not page-aligned\n", + __func__); + rte_errno = EINVAL; + return -1; + } + /* check if device memory map is uninitialized. */ + if (mcfg->device_memory.base_va == NULL || arr->len == 0) { + RTE_LOG(ERR, EAL, "%s(): device memory map is not initialized\n", + __func__); + rte_errno = ENODEV; + return -1; + } + map_end = RTE_PTR_ADD(mcfg->device_memory.base_va, + arr->len * system_page_sz); + mem_end = RTE_PTR_ADD(addr, size); + + /* check if address is within the memory map */ + if (addr < mcfg->device_memory.base_va || addr >= map_end || + mem_end > map_end) { + RTE_LOG(ERR, EAL, "%s(): address is beyond device memory map range\n", + __func__); + rte_errno = EINVAL; + return -1; + } + + rte_spinlock_lock(&mcfg->device_memory.lock); + + n_pages = size / system_page_sz; + cur_idx = RTE_PTR_DIFF(addr, mcfg->device_memory.base_va) / + system_page_sz; + end = cur_idx + n_pages; + + /* check all space we will be marking as free is currently occupied */ + len = rte_fbarray_find_contig_used(arr, cur_idx); + if (len < n_pages) { + RTE_LOG(ERR, EAL, "%s(): attempting to free unoccupied space\n", + __func__); + rte_errno = EINVAL; + ret = -1; + goto unlock; + } + /* now, mark all space as free */ + for (; cur_idx < end; cur_idx++) + rte_fbarray_set_free(arr, cur_idx); + + /* success */ + ret = 0; + + RTE_LOG(DEBUG, EAL, "%s(): deallocated %p-%p (%lu bytes) for hardware device usage\n", + __func__, addr, RTE_PTR_ADD(addr, size), size); +unlock: + rte_spinlock_unlock(&mcfg->device_memory.lock); + return ret; +} + +static int +dev_memory_init(void) +{ + struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config; + size_t size; + void *addr; + int retval; + + if (system_page_sz == 0) + system_page_sz = sysconf(_SC_PAGESIZE); + + size = (size_t) 2 << 30; + + addr = eal_get_virtual_area(NULL, &size, system_page_sz, 0, 0); + if (addr == NULL) { + RTE_LOG(ERR, EAL, "Cannot reserve device memory\n"); + return -1; + } + + retval = rte_fbarray_init(&mcfg->device_memory.mem_map_arr, + DEVICE_MEMORY_NAME, size / system_page_sz, 0); + if (retval < 0) { + RTE_LOG(ERR, EAL, "Cannot initialize device memory map\n"); + return -1; + } + mcfg->device_memory.base_va = addr; + rte_spinlock_init(&mcfg->device_memory.lock); + return 0; +} + +static int +dev_memory_attach(void) +{ + struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config; + size_t size; + void *addr; + int retval; + + rte_spinlock_lock(&mcfg->device_memory.lock); + + if (system_page_sz == 0) + system_page_sz = sysconf(_SC_PAGESIZE); + + size = mcfg->device_memory.mem_map_arr.len * system_page_sz; + + addr = eal_get_virtual_area(mcfg->device_memory.base_va, &size, + system_page_sz, 0, 0); + if (addr == NULL) { + RTE_LOG(ERR, EAL, "Cannot reserve device memory\n"); + return -1; + } + + retval = rte_fbarray_attach(&mcfg->device_memory.mem_map_arr); + if (retval < 0) { + RTE_LOG(ERR, EAL, "Cannot attach to device memory map\n"); + return -1; + } + + rte_spinlock_unlock(&mcfg->device_memory.lock); + + return 0; +} + /* init memory subsystem */ int rte_eal_memory_init(void) @@ -918,25 +1140,41 @@ rte_eal_memory_init(void) /* lock mem hotplug here, to prevent races while we init */ rte_rwlock_read_lock(&mcfg->memory_hotplug_lock); - retval = rte_eal_process_type() == RTE_PROC_PRIMARY ? + if (rte_eal_process_type() == RTE_PROC_PRIMARY) { + retval = dev_memory_init(); + if (retval < 0) + goto fail; + #ifndef RTE_ARCH_64 - memseg_primary_init_32() : + retval = memseg_primary_init_32(); #else - memseg_primary_init() : + retval = memseg_primary_init(); #endif - memseg_secondary_init(); - - if (retval < 0) - goto fail; - - if (eal_memalloc_init() < 0) - goto fail; - - retval = rte_eal_process_type() == RTE_PROC_PRIMARY ? - rte_eal_hugepage_init() : - rte_eal_hugepage_attach(); - if (retval < 0) - goto fail; + if (retval < 0) + goto fail; + + if (eal_memalloc_init() < 0) + goto fail; + + retval = rte_eal_hugepage_init(); + if (retval < 0) + goto fail; + } else { + retval = dev_memory_attach(); + if (retval < 0) + goto fail; + + retval = memseg_secondary_init(); + if (retval < 0) + goto fail; + + if (eal_memalloc_init() < 0) + goto fail; + + retval = rte_eal_hugepage_attach(); + if (retval < 0) + goto fail; + } if (internal_config.no_shconf == 0 && rte_eal_memdevice_init() < 0) goto fail; diff --git a/lib/librte_eal/common/include/rte_eal_memconfig.h b/lib/librte_eal/common/include/rte_eal_memconfig.h index aff0688dd..a8e7c39ff 100644 --- a/lib/librte_eal/common/include/rte_eal_memconfig.h +++ b/lib/librte_eal/common/include/rte_eal_memconfig.h @@ -36,6 +36,22 @@ struct rte_memseg_list { struct rte_fbarray memseg_arr; }; +/** + * mem map is a special case because we need to store a bunch of other data + * together with the array itself. + */ +struct rte_mem_map { + RTE_STD_C11 + union { + void *base_va; + /**< Base virtual address for this mem map. */ + uint64_t addr_64; + /**< Makes sure addr is always 64-bits */ + }; + rte_spinlock_t lock; + struct rte_fbarray mem_map_arr; +}; + /** * the structure for the memory configuration for the RTE. * Used by the rte_config structure. It is separated out, as for multi-process @@ -68,6 +84,8 @@ struct rte_mem_config { struct rte_memseg_list memsegs[RTE_MAX_MEMSEG_LISTS]; /**< list of dynamic arrays holding memsegs */ + struct rte_mem_map device_memory; /**< Occupancy map of preallocated device memory */ + struct rte_tailq_head tailq_head[RTE_MAX_TAILQ]; /**< Tailqs for objects */ /* Heaps of Malloc per socket */ diff --git a/lib/librte_eal/common/include/rte_memory.h b/lib/librte_eal/common/include/rte_memory.h index aab9f6fe5..4cf58bd2a 100644 --- a/lib/librte_eal/common/include/rte_memory.h +++ b/lib/librte_eal/common/include/rte_memory.h @@ -445,6 +445,46 @@ rte_mem_alloc_validator_register(const char *name, int __rte_experimental rte_mem_alloc_validator_unregister(const char *name, int socket_id); +/** + * @brief Request memory for device mapping. + * + * @note after this call, reserved memory will be marked as unavailable in all + * processes until it is released, even if it goes unused. + * + * @param size + * Size of memory to request. + * + * @param align + * Alignment of memory to be returned. + * + * @return + * Valid pointer on successful fulfillment of request. + * NULL on unsuccessful fulfillment of request, with rte_errno indicating the + * case of error. + */ +void * __rte_experimental +rte_mem_dev_memory_alloc(size_t size, size_t align); + +/** + * @brief Release memory for device mapping. + * + * @note by the time this call is made, memory region being freed must not be in + * use. + * + * @param addr + * Address of previously requested block of memory. + * + * @param size + * Size of memory to request. + * + * @return + * 0 on successful memory release. + * -1 on unsuccessful memory release, with rte_errno indicating the cause of + * error. + */ +int __rte_experimental +rte_mem_dev_memory_free(void *addr, size_t size); + #ifdef __cplusplus } #endif -- 2.17.0