1 /*
2 * Memory helpers
3 */
4 #include <fcntl.h>
5 #include <unistd.h>
6 #include <sys/mman.h>
7 #include <sys/stat.h>
8
9 #include "fio.h"
10 #ifndef FIO_NO_HAVE_SHM_H
11 #include <sys/shm.h>
12 #endif
13
fio_unpin_memory(struct thread_data * td)14 void fio_unpin_memory(struct thread_data *td)
15 {
16 if (td->pinned_mem) {
17 dprint(FD_MEM, "unpinning %llu bytes\n", td->o.lockmem);
18 if (munlock(td->pinned_mem, td->o.lockmem) < 0)
19 perror("munlock");
20 munmap(td->pinned_mem, td->o.lockmem);
21 td->pinned_mem = NULL;
22 }
23 }
24
fio_pin_memory(struct thread_data * td)25 int fio_pin_memory(struct thread_data *td)
26 {
27 unsigned long long phys_mem;
28
29 if (!td->o.lockmem)
30 return 0;
31
32 dprint(FD_MEM, "pinning %llu bytes\n", td->o.lockmem);
33
34 /*
35 * Don't allow mlock of more than real_mem-128MiB
36 */
37 phys_mem = os_phys_mem();
38 if (phys_mem) {
39 if ((td->o.lockmem + 128 * 1024 * 1024) > phys_mem) {
40 td->o.lockmem = phys_mem - 128 * 1024 * 1024;
41 log_info("fio: limiting mlocked memory to %lluMiB\n",
42 td->o.lockmem >> 20);
43 }
44 }
45
46 td->pinned_mem = mmap(NULL, td->o.lockmem, PROT_READ | PROT_WRITE,
47 MAP_PRIVATE | OS_MAP_ANON, -1, 0);
48 if (td->pinned_mem == MAP_FAILED) {
49 perror("malloc locked mem");
50 td->pinned_mem = NULL;
51 return 1;
52 }
53 if (mlock(td->pinned_mem, td->o.lockmem) < 0) {
54 perror("mlock");
55 munmap(td->pinned_mem, td->o.lockmem);
56 td->pinned_mem = NULL;
57 return 1;
58 }
59
60 return 0;
61 }
62
alloc_mem_shm(struct thread_data * td,unsigned int total_mem)63 static int alloc_mem_shm(struct thread_data *td, unsigned int total_mem)
64 {
65 #ifndef CONFIG_NO_SHM
66 int flags = IPC_CREAT | S_IRUSR | S_IWUSR;
67
68 if (td->o.mem_type == MEM_SHMHUGE) {
69 unsigned long mask = td->o.hugepage_size - 1;
70
71 flags |= SHM_HUGETLB;
72 total_mem = (total_mem + mask) & ~mask;
73 }
74
75 td->shm_id = shmget(IPC_PRIVATE, total_mem, flags);
76 dprint(FD_MEM, "shmget %u, %d\n", total_mem, td->shm_id);
77 if (td->shm_id < 0) {
78 td_verror(td, errno, "shmget");
79 if (geteuid() != 0 && (errno == ENOMEM || errno == EPERM))
80 log_err("fio: you may need to run this job as root\n");
81 if (td->o.mem_type == MEM_SHMHUGE) {
82 if (errno == EINVAL) {
83 log_err("fio: check that you have free huge"
84 " pages and that hugepage-size is"
85 " correct.\n");
86 } else if (errno == ENOSYS) {
87 log_err("fio: your system does not appear to"
88 " support huge pages.\n");
89 } else if (errno == ENOMEM) {
90 log_err("fio: no huge pages available, do you"
91 " need to allocate some? See HOWTO.\n");
92 }
93 }
94
95 return 1;
96 }
97
98 td->orig_buffer = shmat(td->shm_id, NULL, 0);
99 dprint(FD_MEM, "shmat %d, %p\n", td->shm_id, td->orig_buffer);
100 if (td->orig_buffer == (void *) -1) {
101 td_verror(td, errno, "shmat");
102 td->orig_buffer = NULL;
103 return 1;
104 }
105
106 return 0;
107 #else
108 log_err("fio: shm not supported\n");
109 return 1;
110 #endif
111 }
112
free_mem_shm(struct thread_data * td)113 static void free_mem_shm(struct thread_data *td)
114 {
115 #ifndef CONFIG_NO_SHM
116 struct shmid_ds sbuf;
117
118 dprint(FD_MEM, "shmdt/ctl %d %p\n", td->shm_id, td->orig_buffer);
119 shmdt(td->orig_buffer);
120 shmctl(td->shm_id, IPC_RMID, &sbuf);
121 #endif
122 }
123
alloc_mem_mmap(struct thread_data * td,size_t total_mem)124 static int alloc_mem_mmap(struct thread_data *td, size_t total_mem)
125 {
126 int flags = 0;
127
128 td->mmapfd = -1;
129
130 if (td->o.mem_type == MEM_MMAPHUGE) {
131 unsigned long mask = td->o.hugepage_size - 1;
132
133 /* TODO: make sure the file is a real hugetlbfs file */
134 if (!td->o.mmapfile)
135 flags |= MAP_HUGETLB;
136 total_mem = (total_mem + mask) & ~mask;
137 }
138
139 if (td->o.mmapfile) {
140 if (access(td->o.mmapfile, F_OK) == 0)
141 td->flags |= TD_F_MMAP_KEEP;
142
143 td->mmapfd = open(td->o.mmapfile, O_RDWR|O_CREAT, 0644);
144
145 if (td->mmapfd < 0) {
146 td_verror(td, errno, "open mmap file");
147 td->orig_buffer = NULL;
148 return 1;
149 }
150 if (td->o.mem_type != MEM_MMAPHUGE &&
151 td->o.mem_type != MEM_MMAPSHARED &&
152 ftruncate(td->mmapfd, total_mem) < 0) {
153 td_verror(td, errno, "truncate mmap file");
154 td->orig_buffer = NULL;
155 return 1;
156 }
157 if (td->o.mem_type == MEM_MMAPHUGE ||
158 td->o.mem_type == MEM_MMAPSHARED)
159 flags |= MAP_SHARED;
160 else
161 flags |= MAP_PRIVATE;
162 } else
163 flags |= OS_MAP_ANON | MAP_PRIVATE;
164
165 td->orig_buffer = mmap(NULL, total_mem, PROT_READ | PROT_WRITE, flags,
166 td->mmapfd, 0);
167 dprint(FD_MEM, "mmap %llu/%d %p\n", (unsigned long long) total_mem,
168 td->mmapfd, td->orig_buffer);
169 if (td->orig_buffer == MAP_FAILED) {
170 td_verror(td, errno, "mmap");
171 td->orig_buffer = NULL;
172 if (td->mmapfd != 1 && td->mmapfd != -1) {
173 close(td->mmapfd);
174 if (td->o.mmapfile && !(td->flags & TD_F_MMAP_KEEP))
175 unlink(td->o.mmapfile);
176 }
177
178 return 1;
179 }
180
181 return 0;
182 }
183
free_mem_mmap(struct thread_data * td,size_t total_mem)184 static void free_mem_mmap(struct thread_data *td, size_t total_mem)
185 {
186 dprint(FD_MEM, "munmap %llu %p\n", (unsigned long long) total_mem,
187 td->orig_buffer);
188 munmap(td->orig_buffer, td->orig_buffer_size);
189 if (td->o.mmapfile) {
190 if (td->mmapfd != -1)
191 close(td->mmapfd);
192 if (!(td->flags & TD_F_MMAP_KEEP))
193 unlink(td->o.mmapfile);
194 free(td->o.mmapfile);
195 }
196 }
197
alloc_mem_malloc(struct thread_data * td,size_t total_mem)198 static int alloc_mem_malloc(struct thread_data *td, size_t total_mem)
199 {
200 td->orig_buffer = malloc(total_mem);
201 dprint(FD_MEM, "malloc %llu %p\n", (unsigned long long) total_mem,
202 td->orig_buffer);
203
204 return td->orig_buffer == NULL;
205 }
206
free_mem_malloc(struct thread_data * td)207 static void free_mem_malloc(struct thread_data *td)
208 {
209 dprint(FD_MEM, "free malloc mem %p\n", td->orig_buffer);
210 free(td->orig_buffer);
211 }
212
alloc_mem_cudamalloc(struct thread_data * td,size_t total_mem)213 static int alloc_mem_cudamalloc(struct thread_data *td, size_t total_mem)
214 {
215 #ifdef CONFIG_CUDA
216 CUresult ret;
217 char name[128];
218
219 ret = cuInit(0);
220 if (ret != CUDA_SUCCESS) {
221 log_err("fio: failed initialize cuda driver api\n");
222 return 1;
223 }
224
225 ret = cuDeviceGetCount(&td->gpu_dev_cnt);
226 if (ret != CUDA_SUCCESS) {
227 log_err("fio: failed get device count\n");
228 return 1;
229 }
230 dprint(FD_MEM, "found %d GPU devices\n", td->gpu_dev_cnt);
231
232 if (td->gpu_dev_cnt == 0) {
233 log_err("fio: no GPU device found. "
234 "Can not perform GPUDirect RDMA.\n");
235 return 1;
236 }
237
238 td->gpu_dev_id = td->o.gpu_dev_id;
239 ret = cuDeviceGet(&td->cu_dev, td->gpu_dev_id);
240 if (ret != CUDA_SUCCESS) {
241 log_err("fio: failed get GPU device\n");
242 return 1;
243 }
244
245 ret = cuDeviceGetName(name, sizeof(name), td->gpu_dev_id);
246 if (ret != CUDA_SUCCESS) {
247 log_err("fio: failed get device name\n");
248 return 1;
249 }
250 dprint(FD_MEM, "dev_id = [%d], device name = [%s]\n", \
251 td->gpu_dev_id, name);
252
253 ret = cuCtxCreate(&td->cu_ctx, CU_CTX_MAP_HOST, td->cu_dev);
254 if (ret != CUDA_SUCCESS) {
255 log_err("fio: failed to create cuda context: %d\n", ret);
256 return 1;
257 }
258
259 ret = cuMemAlloc(&td->dev_mem_ptr, total_mem);
260 if (ret != CUDA_SUCCESS) {
261 log_err("fio: cuMemAlloc %zu bytes failed\n", total_mem);
262 return 1;
263 }
264 td->orig_buffer = (void *) td->dev_mem_ptr;
265
266 dprint(FD_MEM, "cudaMalloc %llu %p\n", \
267 (unsigned long long) total_mem, td->orig_buffer);
268 return 0;
269 #else
270 return -EINVAL;
271 #endif
272 }
273
free_mem_cudamalloc(struct thread_data * td)274 static void free_mem_cudamalloc(struct thread_data *td)
275 {
276 #ifdef CONFIG_CUDA
277 if (td->dev_mem_ptr)
278 cuMemFree(td->dev_mem_ptr);
279
280 if (cuCtxDestroy(td->cu_ctx) != CUDA_SUCCESS)
281 log_err("fio: failed to destroy cuda context\n");
282 #endif
283 }
284
285 /*
286 * Set up the buffer area we need for io.
287 */
allocate_io_mem(struct thread_data * td)288 int allocate_io_mem(struct thread_data *td)
289 {
290 size_t total_mem;
291 int ret = 0;
292
293 if (td_ioengine_flagged(td, FIO_NOIO))
294 return 0;
295
296 total_mem = td->orig_buffer_size;
297
298 if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
299 td_ioengine_flagged(td, FIO_MEMALIGN)) {
300 total_mem += page_mask;
301 if (td->o.mem_align && td->o.mem_align > page_size)
302 total_mem += td->o.mem_align - page_size;
303 }
304
305 dprint(FD_MEM, "Alloc %llu for buffers\n", (unsigned long long) total_mem);
306
307 /*
308 * If the IO engine has hooks to allocate/free memory, use those. But
309 * error out if the user explicitly asked for something else.
310 */
311 if (td->io_ops->iomem_alloc) {
312 if (fio_option_is_set(&td->o, mem_type)) {
313 log_err("fio: option 'mem/iomem' conflicts with specified IO engine\n");
314 ret = 1;
315 } else
316 ret = td->io_ops->iomem_alloc(td, total_mem);
317 } else if (td->o.mem_type == MEM_MALLOC)
318 ret = alloc_mem_malloc(td, total_mem);
319 else if (td->o.mem_type == MEM_SHM || td->o.mem_type == MEM_SHMHUGE)
320 ret = alloc_mem_shm(td, total_mem);
321 else if (td->o.mem_type == MEM_MMAP || td->o.mem_type == MEM_MMAPHUGE ||
322 td->o.mem_type == MEM_MMAPSHARED)
323 ret = alloc_mem_mmap(td, total_mem);
324 else if (td->o.mem_type == MEM_CUDA_MALLOC)
325 ret = alloc_mem_cudamalloc(td, total_mem);
326 else {
327 log_err("fio: bad mem type: %d\n", td->o.mem_type);
328 ret = 1;
329 }
330
331 if (ret)
332 td_verror(td, ENOMEM, "iomem allocation");
333
334 return ret;
335 }
336
free_io_mem(struct thread_data * td)337 void free_io_mem(struct thread_data *td)
338 {
339 unsigned int total_mem;
340
341 total_mem = td->orig_buffer_size;
342 if (td->o.odirect || td->o.oatomic)
343 total_mem += page_mask;
344
345 if (td->io_ops->iomem_alloc) {
346 if (td->io_ops->iomem_free)
347 td->io_ops->iomem_free(td);
348 } else if (td->o.mem_type == MEM_MALLOC)
349 free_mem_malloc(td);
350 else if (td->o.mem_type == MEM_SHM || td->o.mem_type == MEM_SHMHUGE)
351 free_mem_shm(td);
352 else if (td->o.mem_type == MEM_MMAP || td->o.mem_type == MEM_MMAPHUGE ||
353 td->o.mem_type == MEM_MMAPSHARED)
354 free_mem_mmap(td, total_mem);
355 else if (td->o.mem_type == MEM_CUDA_MALLOC)
356 free_mem_cudamalloc(td);
357 else
358 log_err("Bad memory type %u\n", td->o.mem_type);
359
360 td->orig_buffer = NULL;
361 td->orig_buffer_size = 0;
362 }
363