11 #ifndef FIO_NO_HAVE_SHM_H
15 void fio_unpin_memory(struct thread_data *td)
18 dprint(FD_MEM, "unpinning %llu bytes\n", td->o.lockmem);
19 if (munlock(td->pinned_mem, td->o.lockmem) < 0)
21 munmap(td->pinned_mem, td->o.lockmem);
22 td->pinned_mem = NULL;
26 int fio_pin_memory(struct thread_data *td)
28 unsigned long long phys_mem;
33 dprint(FD_MEM, "pinning %llu bytes\n", td->o.lockmem);
36 * Don't allow mlock of more than real_mem-128MiB
38 phys_mem = os_phys_mem();
40 if ((td->o.lockmem + 128 * 1024 * 1024) > phys_mem) {
41 td->o.lockmem = phys_mem - 128 * 1024 * 1024;
42 log_info("fio: limiting mlocked memory to %lluMiB\n",
47 td->pinned_mem = mmap(NULL, td->o.lockmem, PROT_READ | PROT_WRITE,
48 MAP_PRIVATE | OS_MAP_ANON, -1, 0);
49 if (td->pinned_mem == MAP_FAILED) {
50 perror("malloc locked mem");
51 td->pinned_mem = NULL;
54 if (mlock(td->pinned_mem, td->o.lockmem) < 0) {
56 munmap(td->pinned_mem, td->o.lockmem);
57 td->pinned_mem = NULL;
64 static int alloc_mem_shm(struct thread_data *td, unsigned int total_mem)
67 int flags = IPC_CREAT | S_IRUSR | S_IWUSR;
69 if (td->o.mem_type == MEM_SHMHUGE) {
70 unsigned long mask = td->o.hugepage_size - 1;
73 total_mem = (total_mem + mask) & ~mask;
76 td->shm_id = shmget(IPC_PRIVATE, total_mem, flags);
77 dprint(FD_MEM, "shmget %u, %d\n", total_mem, td->shm_id);
79 td_verror(td, errno, "shmget");
80 if (geteuid() != 0 && (errno == ENOMEM || errno == EPERM))
81 log_err("fio: you may need to run this job as root\n");
82 if (td->o.mem_type == MEM_SHMHUGE) {
83 if (errno == EINVAL) {
84 log_err("fio: check that you have free huge"
85 " pages and that hugepage-size is"
87 } else if (errno == ENOSYS) {
88 log_err("fio: your system does not appear to"
89 " support huge pages.\n");
90 } else if (errno == ENOMEM) {
91 log_err("fio: no huge pages available, do you"
92 " need to allocate some? See HOWTO.\n");
99 td->orig_buffer = shmat(td->shm_id, NULL, 0);
100 dprint(FD_MEM, "shmat %d, %p\n", td->shm_id, td->orig_buffer);
101 if (td->orig_buffer == (void *) -1) {
102 td_verror(td, errno, "shmat");
103 td->orig_buffer = NULL;
109 log_err("fio: shm not supported\n");
114 static void free_mem_shm(struct thread_data *td)
116 #ifndef CONFIG_NO_SHM
117 struct shmid_ds sbuf;
119 dprint(FD_MEM, "shmdt/ctl %d %p\n", td->shm_id, td->orig_buffer);
120 shmdt(td->orig_buffer);
121 shmctl(td->shm_id, IPC_RMID, &sbuf);
125 static int alloc_mem_mmap(struct thread_data *td, size_t total_mem)
131 if (td->o.mem_type == MEM_MMAPHUGE) {
132 unsigned long mask = td->o.hugepage_size - 1;
134 /* TODO: make sure the file is a real hugetlbfs file */
136 flags |= MAP_HUGETLB;
137 total_mem = (total_mem + mask) & ~mask;
140 if (td->o.mmapfile) {
141 td->mmapfd = open(td->o.mmapfile, O_RDWR|O_CREAT, 0644);
143 if (td->mmapfd < 0) {
144 td_verror(td, errno, "open mmap file");
145 td->orig_buffer = NULL;
148 if (td->o.mem_type != MEM_MMAPHUGE &&
149 td->o.mem_type != MEM_MMAPSHARED &&
150 ftruncate(td->mmapfd, total_mem) < 0) {
151 td_verror(td, errno, "truncate mmap file");
152 td->orig_buffer = NULL;
155 if (td->o.mem_type == MEM_MMAPHUGE ||
156 td->o.mem_type == MEM_MMAPSHARED)
159 flags |= MAP_PRIVATE;
161 flags |= OS_MAP_ANON | MAP_PRIVATE;
163 td->orig_buffer = mmap(NULL, total_mem, PROT_READ | PROT_WRITE, flags,
165 dprint(FD_MEM, "mmap %llu/%d %p\n", (unsigned long long) total_mem,
166 td->mmapfd, td->orig_buffer);
167 if (td->orig_buffer == MAP_FAILED) {
168 td_verror(td, errno, "mmap");
169 td->orig_buffer = NULL;
170 if (td->mmapfd != 1 && td->mmapfd != -1) {
173 unlink(td->o.mmapfile);
182 static void free_mem_mmap(struct thread_data *td, size_t total_mem)
184 dprint(FD_MEM, "munmap %llu %p\n", (unsigned long long) total_mem,
186 munmap(td->orig_buffer, td->orig_buffer_size);
187 if (td->o.mmapfile) {
188 if (td->mmapfd != -1)
190 unlink(td->o.mmapfile);
191 free(td->o.mmapfile);
195 static int alloc_mem_malloc(struct thread_data *td, size_t total_mem)
197 td->orig_buffer = malloc(total_mem);
198 dprint(FD_MEM, "malloc %llu %p\n", (unsigned long long) total_mem,
201 return td->orig_buffer == NULL;
204 static void free_mem_malloc(struct thread_data *td)
206 dprint(FD_MEM, "free malloc mem %p\n", td->orig_buffer);
207 free(td->orig_buffer);
210 static int alloc_mem_cudamalloc(struct thread_data *td, size_t total_mem)
217 if (ret != CUDA_SUCCESS) {
218 log_err("fio: failed initialize cuda driver api\n");
222 ret = cuDeviceGetCount(&td->gpu_dev_cnt);
223 if (ret != CUDA_SUCCESS) {
224 log_err("fio: failed get device count\n");
227 dprint(FD_MEM, "found %d GPU devices\n", td->gpu_dev_cnt);
229 if (td->gpu_dev_cnt == 0) {
230 log_err("fio: no GPU device found. "
231 "Can not perform GPUDirect RDMA.\n");
235 td->gpu_dev_id = td->o.gpu_dev_id;
236 ret = cuDeviceGet(&td->cu_dev, td->gpu_dev_id);
237 if (ret != CUDA_SUCCESS) {
238 log_err("fio: failed get GPU device\n");
242 ret = cuDeviceGetName(name, sizeof(name), td->gpu_dev_id);
243 if (ret != CUDA_SUCCESS) {
244 log_err("fio: failed get device name\n");
247 dprint(FD_MEM, "dev_id = [%d], device name = [%s]\n", \
248 td->gpu_dev_id, name);
250 ret = cuCtxCreate(&td->cu_ctx, CU_CTX_MAP_HOST, td->cu_dev);
251 if (ret != CUDA_SUCCESS) {
252 log_err("fio: failed to create cuda context: %d\n", ret);
256 ret = cuMemAlloc(&td->dev_mem_ptr, total_mem);
257 if (ret != CUDA_SUCCESS) {
258 log_err("fio: cuMemAlloc %zu bytes failed\n", total_mem);
261 td->orig_buffer = (void *) td->dev_mem_ptr;
263 dprint(FD_MEM, "cudaMalloc %llu %p\n", \
264 (unsigned long long) total_mem, td->orig_buffer);
271 static void free_mem_cudamalloc(struct thread_data *td)
274 if (td->dev_mem_ptr != NULL)
275 cuMemFree(td->dev_mem_ptr);
277 if (cuCtxDestroy(td->cu_ctx) != CUDA_SUCCESS)
278 log_err("fio: failed to destroy cuda context\n");
283 * Set up the buffer area we need for io.
285 int allocate_io_mem(struct thread_data *td)
290 if (td_ioengine_flagged(td, FIO_NOIO))
293 total_mem = td->orig_buffer_size;
295 if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
296 td_ioengine_flagged(td, FIO_MEMALIGN)) {
297 total_mem += page_mask;
298 if (td->o.mem_align && td->o.mem_align > page_size)
299 total_mem += td->o.mem_align - page_size;
302 dprint(FD_MEM, "Alloc %llu for buffers\n", (unsigned long long) total_mem);
305 * If the IO engine has hooks to allocate/free memory, use those. But
306 * error out if the user explicitly asked for something else.
308 if (td->io_ops->iomem_alloc) {
309 if (fio_option_is_set(&td->o, mem_type)) {
310 log_err("fio: option 'mem/iomem' conflicts with specified IO engine\n");
313 ret = td->io_ops->iomem_alloc(td, total_mem);
314 } else if (td->o.mem_type == MEM_MALLOC)
315 ret = alloc_mem_malloc(td, total_mem);
316 else if (td->o.mem_type == MEM_SHM || td->o.mem_type == MEM_SHMHUGE)
317 ret = alloc_mem_shm(td, total_mem);
318 else if (td->o.mem_type == MEM_MMAP || td->o.mem_type == MEM_MMAPHUGE ||
319 td->o.mem_type == MEM_MMAPSHARED)
320 ret = alloc_mem_mmap(td, total_mem);
321 else if (td->o.mem_type == MEM_CUDA_MALLOC)
322 ret = alloc_mem_cudamalloc(td, total_mem);
324 log_err("fio: bad mem type: %d\n", td->o.mem_type);
329 td_verror(td, ENOMEM, "iomem allocation");
334 void free_io_mem(struct thread_data *td)
336 unsigned int total_mem;
338 total_mem = td->orig_buffer_size;
339 if (td->o.odirect || td->o.oatomic)
340 total_mem += page_mask;
342 if (td->io_ops->iomem_alloc) {
343 if (td->io_ops->iomem_free)
344 td->io_ops->iomem_free(td);
345 } else if (td->o.mem_type == MEM_MALLOC)
347 else if (td->o.mem_type == MEM_SHM || td->o.mem_type == MEM_SHMHUGE)
349 else if (td->o.mem_type == MEM_MMAP || td->o.mem_type == MEM_MMAPHUGE ||
350 td->o.mem_type == MEM_MMAPSHARED)
351 free_mem_mmap(td, total_mem);
352 else if (td->o.mem_type == MEM_CUDA_MALLOC)
353 free_mem_cudamalloc(td);
355 log_err("Bad memory type %u\n", td->o.mem_type);
357 td->orig_buffer = NULL;
358 td->orig_buffer_size = 0;