1 /*
2 * mmap engine
3 *
4 * IO engine that reads/writes from files by doing memcpy to/from
5 * a memory mapped region of the file.
6 *
7 */
8 #include <stdio.h>
9 #include <stdlib.h>
10 #include <unistd.h>
11 #include <errno.h>
12 #include <sys/mman.h>
13
14 #include "../fio.h"
15 #include "../verify.h"
16
17 /*
18 * Limits us to 1GiB of mapped files in total
19 */
20 #define MMAP_TOTAL_SZ (1 * 1024 * 1024 * 1024UL)
21
22 static unsigned long mmap_map_size;
23
24 struct fio_mmap_data {
25 void *mmap_ptr;
26 size_t mmap_sz;
27 off_t mmap_off;
28 };
29
fio_mmap_file(struct thread_data * td,struct fio_file * f,size_t length,off_t off)30 static int fio_mmap_file(struct thread_data *td, struct fio_file *f,
31 size_t length, off_t off)
32 {
33 struct fio_mmap_data *fmd = FILE_ENG_DATA(f);
34 int flags = 0;
35
36 if (td_rw(td))
37 flags = PROT_READ | PROT_WRITE;
38 else if (td_write(td)) {
39 flags = PROT_WRITE;
40
41 if (td->o.verify != VERIFY_NONE)
42 flags |= PROT_READ;
43 } else
44 flags = PROT_READ;
45
46 fmd->mmap_ptr = mmap(NULL, length, flags, MAP_SHARED, f->fd, off);
47 if (fmd->mmap_ptr == MAP_FAILED) {
48 fmd->mmap_ptr = NULL;
49 td_verror(td, errno, "mmap");
50 goto err;
51 }
52
53 if (!td_random(td)) {
54 if (posix_madvise(fmd->mmap_ptr, length, POSIX_MADV_SEQUENTIAL) < 0) {
55 td_verror(td, errno, "madvise");
56 goto err;
57 }
58 } else {
59 if (posix_madvise(fmd->mmap_ptr, length, POSIX_MADV_RANDOM) < 0) {
60 td_verror(td, errno, "madvise");
61 goto err;
62 }
63 }
64 if (posix_madvise(fmd->mmap_ptr, length, POSIX_MADV_DONTNEED) < 0) {
65 td_verror(td, errno, "madvise");
66 goto err;
67 }
68
69 #ifdef FIO_MADV_FREE
70 if (f->filetype == FIO_TYPE_BLOCK)
71 (void) posix_madvise(fmd->mmap_ptr, fmd->mmap_sz, FIO_MADV_FREE);
72 #endif
73
74 err:
75 if (td->error && fmd->mmap_ptr)
76 munmap(fmd->mmap_ptr, length);
77
78 return td->error;
79 }
80
81 /*
82 * Just mmap an appropriate portion, we cannot mmap the full extent
83 */
fio_mmapio_prep_limited(struct thread_data * td,struct io_u * io_u)84 static int fio_mmapio_prep_limited(struct thread_data *td, struct io_u *io_u)
85 {
86 struct fio_file *f = io_u->file;
87 struct fio_mmap_data *fmd = FILE_ENG_DATA(f);
88
89 if (io_u->buflen > mmap_map_size) {
90 log_err("fio: bs too big for mmap engine\n");
91 return EIO;
92 }
93
94 fmd->mmap_sz = mmap_map_size;
95 if (fmd->mmap_sz > f->io_size)
96 fmd->mmap_sz = f->io_size;
97
98 fmd->mmap_off = io_u->offset;
99
100 return fio_mmap_file(td, f, fmd->mmap_sz, fmd->mmap_off);
101 }
102
103 /*
104 * Attempt to mmap the entire file
105 */
fio_mmapio_prep_full(struct thread_data * td,struct io_u * io_u)106 static int fio_mmapio_prep_full(struct thread_data *td, struct io_u *io_u)
107 {
108 struct fio_file *f = io_u->file;
109 struct fio_mmap_data *fmd = FILE_ENG_DATA(f);
110 int ret;
111
112 if (fio_file_partial_mmap(f))
113 return EINVAL;
114 if (io_u->offset != (size_t) io_u->offset ||
115 f->io_size != (size_t) f->io_size) {
116 fio_file_set_partial_mmap(f);
117 return EINVAL;
118 }
119
120 fmd->mmap_sz = f->io_size;
121 fmd->mmap_off = 0;
122
123 ret = fio_mmap_file(td, f, fmd->mmap_sz, fmd->mmap_off);
124 if (ret)
125 fio_file_set_partial_mmap(f);
126
127 return ret;
128 }
129
fio_mmapio_prep(struct thread_data * td,struct io_u * io_u)130 static int fio_mmapio_prep(struct thread_data *td, struct io_u *io_u)
131 {
132 struct fio_file *f = io_u->file;
133 struct fio_mmap_data *fmd = FILE_ENG_DATA(f);
134 int ret;
135
136 /*
137 * It fits within existing mapping, use it
138 */
139 if (io_u->offset >= fmd->mmap_off &&
140 io_u->offset + io_u->buflen < fmd->mmap_off + fmd->mmap_sz)
141 goto done;
142
143 /*
144 * unmap any existing mapping
145 */
146 if (fmd->mmap_ptr) {
147 if (munmap(fmd->mmap_ptr, fmd->mmap_sz) < 0)
148 return errno;
149 fmd->mmap_ptr = NULL;
150 }
151
152 if (fio_mmapio_prep_full(td, io_u)) {
153 td_clear_error(td);
154 ret = fio_mmapio_prep_limited(td, io_u);
155 if (ret)
156 return ret;
157 }
158
159 done:
160 io_u->mmap_data = fmd->mmap_ptr + io_u->offset - fmd->mmap_off -
161 f->file_offset;
162 return 0;
163 }
164
fio_mmapio_queue(struct thread_data * td,struct io_u * io_u)165 static int fio_mmapio_queue(struct thread_data *td, struct io_u *io_u)
166 {
167 struct fio_file *f = io_u->file;
168 struct fio_mmap_data *fmd = FILE_ENG_DATA(f);
169
170 fio_ro_check(td, io_u);
171
172 if (io_u->ddir == DDIR_READ)
173 memcpy(io_u->xfer_buf, io_u->mmap_data, io_u->xfer_buflen);
174 else if (io_u->ddir == DDIR_WRITE)
175 memcpy(io_u->mmap_data, io_u->xfer_buf, io_u->xfer_buflen);
176 else if (ddir_sync(io_u->ddir)) {
177 if (msync(fmd->mmap_ptr, fmd->mmap_sz, MS_SYNC)) {
178 io_u->error = errno;
179 td_verror(td, io_u->error, "msync");
180 }
181 } else if (io_u->ddir == DDIR_TRIM) {
182 int ret = do_io_u_trim(td, io_u);
183
184 if (!ret)
185 td_verror(td, io_u->error, "trim");
186 }
187
188
189 /*
190 * not really direct, but should drop the pages from the cache
191 */
192 if (td->o.odirect && ddir_rw(io_u->ddir)) {
193 if (msync(io_u->mmap_data, io_u->xfer_buflen, MS_SYNC) < 0) {
194 io_u->error = errno;
195 td_verror(td, io_u->error, "msync");
196 }
197 if (posix_madvise(io_u->mmap_data, io_u->xfer_buflen, POSIX_MADV_DONTNEED) < 0) {
198 io_u->error = errno;
199 td_verror(td, io_u->error, "madvise");
200 }
201 }
202
203 return FIO_Q_COMPLETED;
204 }
205
fio_mmapio_init(struct thread_data * td)206 static int fio_mmapio_init(struct thread_data *td)
207 {
208 struct thread_options *o = &td->o;
209
210 if ((o->rw_min_bs & page_mask) &&
211 (o->odirect || o->fsync_blocks || o->fdatasync_blocks)) {
212 log_err("fio: mmap options dictate a minimum block size of "
213 "%llu bytes\n", (unsigned long long) page_size);
214 return 1;
215 }
216
217 mmap_map_size = MMAP_TOTAL_SZ / o->nr_files;
218 return 0;
219 }
220
fio_mmapio_open_file(struct thread_data * td,struct fio_file * f)221 static int fio_mmapio_open_file(struct thread_data *td, struct fio_file *f)
222 {
223 struct fio_mmap_data *fmd;
224 int ret;
225
226 ret = generic_open_file(td, f);
227 if (ret)
228 return ret;
229
230 fmd = calloc(1, sizeof(*fmd));
231 if (!fmd) {
232 int fio_unused __ret;
233 __ret = generic_close_file(td, f);
234 return 1;
235 }
236
237 FILE_SET_ENG_DATA(f, fmd);
238 return 0;
239 }
240
fio_mmapio_close_file(struct thread_data * td,struct fio_file * f)241 static int fio_mmapio_close_file(struct thread_data *td, struct fio_file *f)
242 {
243 struct fio_mmap_data *fmd = FILE_ENG_DATA(f);
244
245 FILE_SET_ENG_DATA(f, NULL);
246 free(fmd);
247 fio_file_clear_partial_mmap(f);
248
249 return generic_close_file(td, f);
250 }
251
252 static struct ioengine_ops ioengine = {
253 .name = "mmap",
254 .version = FIO_IOOPS_VERSION,
255 .init = fio_mmapio_init,
256 .prep = fio_mmapio_prep,
257 .queue = fio_mmapio_queue,
258 .open_file = fio_mmapio_open_file,
259 .close_file = fio_mmapio_close_file,
260 .get_file_size = generic_get_file_size,
261 .flags = FIO_SYNCIO | FIO_NOEXTEND,
262 };
263
fio_mmapio_register(void)264 static void fio_init fio_mmapio_register(void)
265 {
266 register_ioengine(&ioengine);
267 }
268
fio_mmapio_unregister(void)269 static void fio_exit fio_mmapio_unregister(void)
270 {
271 unregister_ioengine(&ioengine);
272 }
273