->getevents() should take unsigned args
[fio.git] / engines / syslet-rw.c
... / ...
CommitLineData
1/*
2 * syslet engine
3 *
4 * IO engine that does regular pread(2)/pwrite(2) to transfer data, but
5 * with syslets to make the execution async.
6 *
7 */
8#include <stdio.h>
9#include <stdlib.h>
10#include <unistd.h>
11#include <errno.h>
12#include <assert.h>
13#include <malloc.h>
14#include <asm/unistd.h>
15
16#include "../fio.h"
17#include "../indirect.h"
18#include "../syslet.h"
19
20#ifdef FIO_HAVE_SYSLET
21
22#ifdef __NR_pread64
23#define __NR_fio_pread __NR_pread64
24#define __NR_fio_pwrite __NR_pwrite64
25#else
26#define __NR_fio_pread __NR_pread
27#define __NR_fio_pwrite __NR_pwrite
28#endif
29
30struct syslet_data {
31 struct io_u **events;
32 unsigned int nr_events;
33
34 struct syslet_ring *ring;
35 unsigned int ring_mask;
36 void *stack;
37};
38
39static void fio_syslet_add_event(struct thread_data *td, struct io_u *io_u)
40{
41 struct syslet_data *sd = td->io_ops->data;
42
43 assert(sd->nr_events < td->o.iodepth);
44 sd->events[sd->nr_events++] = io_u;
45}
46
47static void fio_syslet_add_events(struct thread_data *td, unsigned int nr)
48{
49 struct syslet_data *sd = td->io_ops->data;
50 unsigned int i, uidx;
51
52 uidx = sd->ring->user_tail;
53 read_barrier();
54
55 for (i = 0; i < nr; i++) {
56 unsigned int idx = (i + uidx) & sd->ring_mask;
57 struct syslet_completion *comp = &sd->ring->comp[idx];
58 struct io_u *io_u = (struct io_u *) (long) comp->caller_data;
59 long ret;
60
61 ret = comp->status;
62 if (ret <= 0) {
63 io_u->resid = io_u->xfer_buflen;
64 io_u->error = -ret;
65 } else {
66 io_u->resid = io_u->xfer_buflen - ret;
67 io_u->error = 0;
68 }
69
70 fio_syslet_add_event(td, io_u);
71 }
72}
73
74static void fio_syslet_wait_for_events(struct thread_data *td)
75{
76 struct syslet_data *sd = td->io_ops->data;
77 struct syslet_ring *ring = sd->ring;
78
79 do {
80 unsigned int kh = ring->kernel_head;
81 int ret;
82
83 /*
84 * first reap events that are already completed
85 */
86 if (ring->user_tail != kh) {
87 unsigned int nr = kh - ring->user_tail;
88
89 fio_syslet_add_events(td, nr);
90 ring->user_tail = kh;
91 break;
92 }
93
94 /*
95 * block waiting for at least one event
96 */
97 ret = syscall(__NR_syslet_ring_wait, ring, ring->user_tail);
98 assert(!ret);
99 } while (1);
100}
101
102static int fio_syslet_getevents(struct thread_data *td, unsigned int min,
103 unsigned int fio_unused max,
104 struct timespec fio_unused *t)
105{
106 struct syslet_data *sd = td->io_ops->data;
107 long ret;
108
109 /*
110 * While we have less events than requested, block waiting for them
111 */
112 while (sd->nr_events < (unsigned int) min)
113 fio_syslet_wait_for_events(td);
114
115 ret = sd->nr_events;
116 sd->nr_events = 0;
117 return ret;
118}
119
120static struct io_u *fio_syslet_event(struct thread_data *td, int event)
121{
122 struct syslet_data *sd = td->io_ops->data;
123
124 return sd->events[event];
125}
126
127static void fio_syslet_prep_sync(struct fio_file *f,
128 struct indirect_registers *regs)
129{
130 FILL_IN(*regs, __NR_fsync, (long) f->fd);
131}
132
133static void fio_syslet_prep_rw(struct io_u *io_u, struct fio_file *f,
134 struct indirect_registers *regs)
135{
136 long nr;
137
138 /*
139 * prepare rw
140 */
141 if (io_u->ddir == DDIR_READ)
142 nr = __NR_fio_pread;
143 else
144 nr = __NR_fio_pwrite;
145
146 FILL_IN(*regs, nr, (long) f->fd, (long) io_u->xfer_buf,
147 (long) io_u->xfer_buflen, (long) io_u->offset);
148}
149
150static void fio_syslet_prep(struct io_u *io_u, struct indirect_registers *regs)
151{
152 struct fio_file *f = io_u->file;
153
154 if (io_u->ddir == DDIR_SYNC)
155 fio_syslet_prep_sync(f, regs);
156 else
157 fio_syslet_prep_rw(io_u, f, regs);
158}
159
160static void ret_func(void)
161{
162 syscall(__NR_exit);
163}
164
165static int fio_syslet_queue(struct thread_data *td, struct io_u *io_u)
166{
167 struct syslet_data *sd = td->io_ops->data;
168 union indirect_params params;
169 struct indirect_registers regs;
170 int ret;
171
172 fio_ro_check(td, io_u);
173
174 memset(&params, 0, sizeof(params));
175 fill_syslet_args(&params.syslet, sd->ring, (long)io_u, ret_func, sd->stack);
176
177 fio_syslet_prep(io_u, &regs);
178
179 ret = syscall(__NR_indirect, &regs, &params, sizeof(params), 0);
180 if (ret == (int) io_u->xfer_buflen) {
181 /*
182 * completed sync, account. this also catches fsync().
183 */
184 return FIO_Q_COMPLETED;
185 } else if (ret < 0) {
186 /*
187 * queued for async execution
188 */
189 if (errno == ESYSLETPENDING)
190 return FIO_Q_QUEUED;
191 }
192
193 io_u->error = errno;
194 td_verror(td, io_u->error, "xfer");
195 return FIO_Q_COMPLETED;
196}
197
198static int check_syslet_support(struct syslet_data *sd)
199{
200 union indirect_params params;
201 struct indirect_registers regs;
202 pid_t pid, my_pid = getpid();
203
204 memset(&params, 0, sizeof(params));
205 fill_syslet_args(&params.syslet, sd->ring, 0, ret_func, sd->stack);
206
207 FILL_IN(regs, __NR_getpid);
208
209 pid = syscall(__NR_indirect, &regs, &params, sizeof(params), 0);
210 if (pid == my_pid)
211 return 0;
212
213 return 1;
214}
215
216static void fio_syslet_cleanup(struct thread_data *td)
217{
218 struct syslet_data *sd = td->io_ops->data;
219
220 if (sd) {
221 free(sd->events);
222 free(sd->ring);
223 free(sd->stack);
224 free(sd);
225 td->io_ops->data = NULL;
226 }
227}
228
229static int fio_syslet_init(struct thread_data *td)
230{
231 struct syslet_data *sd;
232 void *ring = NULL, *stack = NULL;
233 unsigned int ring_size, ring_nr;
234
235 sd = malloc(sizeof(*sd));
236 memset(sd, 0, sizeof(*sd));
237
238 sd->events = malloc(sizeof(struct io_u *) * td->o.iodepth);
239 memset(sd->events, 0, sizeof(struct io_u *) * td->o.iodepth);
240
241 /*
242 * The ring needs to be a power-of-2, so round it up if we have to
243 */
244 ring_nr = td->o.iodepth;
245 if (ring_nr & (ring_nr - 1)) {
246 int bits = 1;
247
248 while (ring_nr >>= 1)
249 bits++;
250
251 ring_nr = 1 << bits;
252 }
253
254 ring_size = sizeof(struct syslet_ring) +
255 ring_nr * sizeof(struct syslet_completion);
256 if (posix_memalign(&ring, sizeof(uint64_t), ring_size))
257 goto err_mem;
258 if (posix_memalign(&stack, page_size, page_size))
259 goto err_mem;
260
261 sd->ring = ring;
262 sd->ring_mask = ring_nr - 1;
263 sd->stack = stack;
264
265 memset(sd->ring, 0, ring_size);
266 sd->ring->elements = ring_nr;
267
268 if (!check_syslet_support(sd)) {
269 td->io_ops->data = sd;
270 return 0;
271 }
272
273 log_err("fio: syslets do not appear to work\n");
274err_mem:
275 free(sd->events);
276 if (ring)
277 free(ring);
278 if (stack)
279 free(stack);
280 free(sd);
281 return 1;
282}
283
284static struct ioengine_ops ioengine = {
285 .name = "syslet-rw",
286 .version = FIO_IOOPS_VERSION,
287 .init = fio_syslet_init,
288 .queue = fio_syslet_queue,
289 .getevents = fio_syslet_getevents,
290 .event = fio_syslet_event,
291 .cleanup = fio_syslet_cleanup,
292 .open_file = generic_open_file,
293 .close_file = generic_close_file,
294};
295
296#else /* FIO_HAVE_SYSLET */
297
298/*
299 * When we have a proper configure system in place, we simply wont build
300 * and install this io engine. For now install a crippled version that
301 * just complains and fails to load.
302 */
303static int fio_syslet_init(struct thread_data fio_unused *td)
304{
305 fprintf(stderr, "fio: syslet not available\n");
306 return 1;
307}
308
309static struct ioengine_ops ioengine = {
310 .name = "syslet-rw",
311 .version = FIO_IOOPS_VERSION,
312 .init = fio_syslet_init,
313};
314
315#endif /* FIO_HAVE_SYSLET */
316
317static void fio_init fio_syslet_register(void)
318{
319 register_ioengine(&ioengine);
320}
321
322static void fio_exit fio_syslet_unregister(void)
323{
324 unregister_ioengine(&ioengine);
325}