]>
Commit | Line | Data |
---|---|---|
04155bda | 1 | #include "cache.h" |
36bf1958 | 2 | #include "alloc.h" |
7531e4b6 | 3 | #include "config.h" |
04155bda MT |
4 | #include "entry.h" |
5 | #include "parallel-checkout.h" | |
e9e8adf1 | 6 | #include "pkt-line.h" |
1c4d6f46 | 7 | #include "progress.h" |
e9e8adf1 MT |
8 | #include "run-command.h" |
9 | #include "sigchain.h" | |
04155bda | 10 | #include "streaming.h" |
7531e4b6 | 11 | #include "thread-utils.h" |
6a7bc9d1 | 12 | #include "trace2.h" |
04155bda | 13 | |
e9e8adf1 MT |
14 | struct pc_worker { |
15 | struct child_process cp; | |
16 | size_t next_item_to_complete, nr_items_to_complete; | |
04155bda MT |
17 | }; |
18 | ||
19 | struct parallel_checkout { | |
20 | enum pc_status status; | |
21 | struct parallel_checkout_item *items; /* The parallel checkout queue. */ | |
22 | size_t nr, alloc; | |
1c4d6f46 MT |
23 | struct progress *progress; |
24 | unsigned int *progress_cnt; | |
04155bda MT |
25 | }; |
26 | ||
27 | static struct parallel_checkout parallel_checkout; | |
28 | ||
29 | enum pc_status parallel_checkout_status(void) | |
30 | { | |
31 | return parallel_checkout.status; | |
32 | } | |
33 | ||
7531e4b6 MT |
34 | static const int DEFAULT_THRESHOLD_FOR_PARALLELISM = 100; |
35 | static const int DEFAULT_NUM_WORKERS = 1; | |
36 | ||
37 | void get_parallel_checkout_configs(int *num_workers, int *threshold) | |
38 | { | |
87094fc2 MT |
39 | char *env_workers = getenv("GIT_TEST_CHECKOUT_WORKERS"); |
40 | ||
41 | if (env_workers && *env_workers) { | |
42 | if (strtol_i(env_workers, 10, num_workers)) { | |
1a8aea85 JNA |
43 | die(_("invalid value for '%s': '%s'"), |
44 | "GIT_TEST_CHECKOUT_WORKERS", env_workers); | |
87094fc2 MT |
45 | } |
46 | if (*num_workers < 1) | |
47 | *num_workers = online_cpus(); | |
48 | ||
49 | *threshold = 0; | |
50 | return; | |
51 | } | |
52 | ||
7531e4b6 MT |
53 | if (git_config_get_int("checkout.workers", num_workers)) |
54 | *num_workers = DEFAULT_NUM_WORKERS; | |
55 | else if (*num_workers < 1) | |
56 | *num_workers = online_cpus(); | |
57 | ||
58 | if (git_config_get_int("checkout.thresholdForParallelism", threshold)) | |
59 | *threshold = DEFAULT_THRESHOLD_FOR_PARALLELISM; | |
60 | } | |
61 | ||
04155bda MT |
62 | void init_parallel_checkout(void) |
63 | { | |
64 | if (parallel_checkout.status != PC_UNINITIALIZED) | |
65 | BUG("parallel checkout already initialized"); | |
66 | ||
67 | parallel_checkout.status = PC_ACCEPTING_ENTRIES; | |
68 | } | |
69 | ||
70 | static void finish_parallel_checkout(void) | |
71 | { | |
72 | if (parallel_checkout.status == PC_UNINITIALIZED) | |
73 | BUG("cannot finish parallel checkout: not initialized yet"); | |
74 | ||
75 | free(parallel_checkout.items); | |
76 | memset(¶llel_checkout, 0, sizeof(parallel_checkout)); | |
77 | } | |
78 | ||
79 | static int is_eligible_for_parallel_checkout(const struct cache_entry *ce, | |
80 | const struct conv_attrs *ca) | |
81 | { | |
82 | enum conv_attrs_classification c; | |
e9e8adf1 | 83 | size_t packed_item_size; |
04155bda MT |
84 | |
85 | /* | |
86 | * Symlinks cannot be checked out in parallel as, in case of path | |
87 | * collision, they could racily replace leading directories of other | |
88 | * entries being checked out. Submodules are checked out in child | |
89 | * processes, which have their own parallel checkout queues. | |
90 | */ | |
91 | if (!S_ISREG(ce->ce_mode)) | |
92 | return 0; | |
93 | ||
e9e8adf1 MT |
94 | packed_item_size = sizeof(struct pc_item_fixed_portion) + ce->ce_namelen + |
95 | (ca->working_tree_encoding ? strlen(ca->working_tree_encoding) : 0); | |
96 | ||
97 | /* | |
98 | * The amount of data we send to the workers per checkout item is | |
99 | * typically small (75~300B). So unless we find an insanely huge path | |
100 | * of 64KB, we should never reach the 65KB limit of one pkt-line. If | |
101 | * that does happen, we let the sequential code handle the item. | |
102 | */ | |
103 | if (packed_item_size > LARGE_PACKET_DATA_MAX) | |
104 | return 0; | |
105 | ||
04155bda MT |
106 | c = classify_conv_attrs(ca); |
107 | switch (c) { | |
108 | case CA_CLASS_INCORE: | |
109 | return 1; | |
110 | ||
111 | case CA_CLASS_INCORE_FILTER: | |
112 | /* | |
113 | * It would be safe to allow concurrent instances of | |
114 | * single-file smudge filters, like rot13, but we should not | |
115 | * assume that all filters are parallel-process safe. So we | |
116 | * don't allow this. | |
117 | */ | |
118 | return 0; | |
119 | ||
120 | case CA_CLASS_INCORE_PROCESS: | |
121 | /* | |
122 | * The parallel queue and the delayed queue are not compatible, | |
123 | * so they must be kept completely separated. And we can't tell | |
124 | * if a long-running process will delay its response without | |
125 | * actually asking it to perform the filtering. Therefore, this | |
126 | * type of filter is not allowed in parallel checkout. | |
127 | * | |
128 | * Furthermore, there should only be one instance of the | |
129 | * long-running process filter as we don't know how it is | |
130 | * managing its own concurrency. So, spreading the entries that | |
131 | * requisite such a filter among the parallel workers would | |
132 | * require a lot more inter-process communication. We would | |
133 | * probably have to designate a single process to interact with | |
134 | * the filter and send all the necessary data to it, for each | |
135 | * entry. | |
136 | */ | |
137 | return 0; | |
138 | ||
139 | case CA_CLASS_STREAMABLE: | |
140 | return 1; | |
141 | ||
142 | default: | |
143 | BUG("unsupported conv_attrs classification '%d'", c); | |
144 | } | |
145 | } | |
146 | ||
611c7785 MT |
147 | int enqueue_checkout(struct cache_entry *ce, struct conv_attrs *ca, |
148 | int *checkout_counter) | |
04155bda MT |
149 | { |
150 | struct parallel_checkout_item *pc_item; | |
151 | ||
152 | if (parallel_checkout.status != PC_ACCEPTING_ENTRIES || | |
153 | !is_eligible_for_parallel_checkout(ce, ca)) | |
154 | return -1; | |
155 | ||
156 | ALLOC_GROW(parallel_checkout.items, parallel_checkout.nr + 1, | |
157 | parallel_checkout.alloc); | |
158 | ||
e9e8adf1 | 159 | pc_item = ¶llel_checkout.items[parallel_checkout.nr]; |
04155bda MT |
160 | pc_item->ce = ce; |
161 | memcpy(&pc_item->ca, ca, sizeof(pc_item->ca)); | |
162 | pc_item->status = PC_ITEM_PENDING; | |
e9e8adf1 | 163 | pc_item->id = parallel_checkout.nr; |
611c7785 | 164 | pc_item->checkout_counter = checkout_counter; |
e9e8adf1 | 165 | parallel_checkout.nr++; |
04155bda MT |
166 | |
167 | return 0; | |
168 | } | |
169 | ||
1c4d6f46 MT |
170 | size_t pc_queue_size(void) |
171 | { | |
172 | return parallel_checkout.nr; | |
173 | } | |
174 | ||
175 | static void advance_progress_meter(void) | |
176 | { | |
177 | if (parallel_checkout.progress) { | |
178 | (*parallel_checkout.progress_cnt)++; | |
179 | display_progress(parallel_checkout.progress, | |
180 | *parallel_checkout.progress_cnt); | |
181 | } | |
182 | } | |
183 | ||
04155bda MT |
184 | static int handle_results(struct checkout *state) |
185 | { | |
186 | int ret = 0; | |
187 | size_t i; | |
188 | int have_pending = 0; | |
189 | ||
190 | /* | |
191 | * We first update the successfully written entries with the collected | |
192 | * stat() data, so that they can be found by mark_colliding_entries(), | |
193 | * in the next loop, when necessary. | |
194 | */ | |
195 | for (i = 0; i < parallel_checkout.nr; i++) { | |
196 | struct parallel_checkout_item *pc_item = ¶llel_checkout.items[i]; | |
197 | if (pc_item->status == PC_ITEM_WRITTEN) | |
198 | update_ce_after_write(state, pc_item->ce, &pc_item->st); | |
199 | } | |
200 | ||
201 | for (i = 0; i < parallel_checkout.nr; i++) { | |
202 | struct parallel_checkout_item *pc_item = ¶llel_checkout.items[i]; | |
203 | ||
204 | switch(pc_item->status) { | |
205 | case PC_ITEM_WRITTEN: | |
611c7785 MT |
206 | if (pc_item->checkout_counter) |
207 | (*pc_item->checkout_counter)++; | |
04155bda MT |
208 | break; |
209 | case PC_ITEM_COLLIDED: | |
210 | /* | |
211 | * The entry could not be checked out due to a path | |
212 | * collision with another entry. Since there can only | |
213 | * be one entry of each colliding group on the disk, we | |
214 | * could skip trying to check out this one and move on. | |
215 | * However, this would leave the unwritten entries with | |
216 | * null stat() fields on the index, which could | |
217 | * potentially slow down subsequent operations that | |
218 | * require refreshing it: git would not be able to | |
219 | * trust st_size and would have to go to the filesystem | |
220 | * to see if the contents match (see ie_modified()). | |
221 | * | |
222 | * Instead, let's pay the overhead only once, now, and | |
223 | * call checkout_entry_ca() again for this file, to | |
224 | * have its stat() data stored in the index. This also | |
225 | * has the benefit of adding this entry and its | |
226 | * colliding pair to the collision report message. | |
227 | * Additionally, this overwriting behavior is consistent | |
228 | * with what the sequential checkout does, so it doesn't | |
229 | * add any extra overhead. | |
230 | */ | |
231 | ret |= checkout_entry_ca(pc_item->ce, &pc_item->ca, | |
611c7785 MT |
232 | state, NULL, |
233 | pc_item->checkout_counter); | |
1c4d6f46 | 234 | advance_progress_meter(); |
04155bda MT |
235 | break; |
236 | case PC_ITEM_PENDING: | |
237 | have_pending = 1; | |
238 | /* fall through */ | |
239 | case PC_ITEM_FAILED: | |
240 | ret = -1; | |
241 | break; | |
242 | default: | |
243 | BUG("unknown checkout item status in parallel checkout"); | |
244 | } | |
245 | } | |
246 | ||
247 | if (have_pending) | |
248 | error("parallel checkout finished with pending entries"); | |
249 | ||
250 | return ret; | |
251 | } | |
252 | ||
253 | static int reset_fd(int fd, const char *path) | |
254 | { | |
255 | if (lseek(fd, 0, SEEK_SET) != 0) | |
256 | return error_errno("failed to rewind descriptor of '%s'", path); | |
257 | if (ftruncate(fd, 0)) | |
258 | return error_errno("failed to truncate file '%s'", path); | |
259 | return 0; | |
260 | } | |
261 | ||
262 | static int write_pc_item_to_fd(struct parallel_checkout_item *pc_item, int fd, | |
263 | const char *path) | |
264 | { | |
265 | int ret; | |
266 | struct stream_filter *filter; | |
267 | struct strbuf buf = STRBUF_INIT; | |
268 | char *blob; | |
e9aa762c | 269 | size_t size; |
04155bda MT |
270 | ssize_t wrote; |
271 | ||
272 | /* Sanity check */ | |
273 | assert(is_eligible_for_parallel_checkout(pc_item->ce, &pc_item->ca)); | |
274 | ||
275 | filter = get_stream_filter_ca(&pc_item->ca, &pc_item->ce->oid); | |
276 | if (filter) { | |
277 | if (stream_blob_to_fd(fd, &pc_item->ce->oid, filter, 1)) { | |
278 | /* On error, reset fd to try writing without streaming */ | |
279 | if (reset_fd(fd, path)) | |
280 | return -1; | |
281 | } else { | |
282 | return 0; | |
283 | } | |
284 | } | |
285 | ||
286 | blob = read_blob_entry(pc_item->ce, &size); | |
287 | if (!blob) | |
288 | return error("cannot read object %s '%s'", | |
289 | oid_to_hex(&pc_item->ce->oid), pc_item->ce->name); | |
290 | ||
291 | /* | |
292 | * checkout metadata is used to give context for external process | |
293 | * filters. Files requiring such filters are not eligible for parallel | |
e9e8adf1 MT |
294 | * checkout, so pass NULL. Note: if that changes, the metadata must also |
295 | * be passed from the main process to the workers. | |
04155bda MT |
296 | */ |
297 | ret = convert_to_working_tree_ca(&pc_item->ca, pc_item->ce->name, | |
298 | blob, size, &buf, NULL); | |
299 | ||
300 | if (ret) { | |
301 | size_t newsize; | |
302 | free(blob); | |
303 | blob = strbuf_detach(&buf, &newsize); | |
304 | size = newsize; | |
305 | } | |
306 | ||
307 | wrote = write_in_full(fd, blob, size); | |
308 | free(blob); | |
309 | if (wrote < 0) | |
310 | return error("unable to write file '%s'", path); | |
311 | ||
312 | return 0; | |
313 | } | |
314 | ||
315 | static int close_and_clear(int *fd) | |
316 | { | |
317 | int ret = 0; | |
318 | ||
319 | if (*fd >= 0) { | |
320 | ret = close(*fd); | |
321 | *fd = -1; | |
322 | } | |
323 | ||
324 | return ret; | |
325 | } | |
326 | ||
e9e8adf1 MT |
327 | void write_pc_item(struct parallel_checkout_item *pc_item, |
328 | struct checkout *state) | |
04155bda MT |
329 | { |
330 | unsigned int mode = (pc_item->ce->ce_mode & 0100) ? 0777 : 0666; | |
331 | int fd = -1, fstat_done = 0; | |
332 | struct strbuf path = STRBUF_INIT; | |
333 | const char *dir_sep; | |
334 | ||
335 | strbuf_add(&path, state->base_dir, state->base_dir_len); | |
336 | strbuf_add(&path, pc_item->ce->name, pc_item->ce->ce_namelen); | |
337 | ||
338 | dir_sep = find_last_dir_sep(path.buf); | |
339 | ||
340 | /* | |
341 | * The leading dirs should have been already created by now. But, in | |
342 | * case of path collisions, one of the dirs could have been replaced by | |
343 | * a symlink (checked out after we enqueued this entry for parallel | |
344 | * checkout). Thus, we must check the leading dirs again. | |
345 | */ | |
346 | if (dir_sep && !has_dirs_only_path(path.buf, dir_sep - path.buf, | |
347 | state->base_dir_len)) { | |
348 | pc_item->status = PC_ITEM_COLLIDED; | |
6a7bc9d1 | 349 | trace2_data_string("pcheckout", NULL, "collision/dirname", path.buf); |
04155bda MT |
350 | goto out; |
351 | } | |
352 | ||
353 | fd = open(path.buf, O_WRONLY | O_CREAT | O_EXCL, mode); | |
354 | ||
355 | if (fd < 0) { | |
356 | if (errno == EEXIST || errno == EISDIR) { | |
357 | /* | |
358 | * Errors which probably represent a path collision. | |
359 | * Suppress the error message and mark the item to be | |
360 | * retried later, sequentially. ENOTDIR and ENOENT are | |
361 | * also interesting, but the above has_dirs_only_path() | |
362 | * call should have already caught these cases. | |
363 | */ | |
364 | pc_item->status = PC_ITEM_COLLIDED; | |
6a7bc9d1 MT |
365 | trace2_data_string("pcheckout", NULL, |
366 | "collision/basename", path.buf); | |
04155bda MT |
367 | } else { |
368 | error_errno("failed to open file '%s'", path.buf); | |
369 | pc_item->status = PC_ITEM_FAILED; | |
370 | } | |
371 | goto out; | |
372 | } | |
373 | ||
374 | if (write_pc_item_to_fd(pc_item, fd, path.buf)) { | |
375 | /* Error was already reported. */ | |
376 | pc_item->status = PC_ITEM_FAILED; | |
377 | close_and_clear(&fd); | |
378 | unlink(path.buf); | |
379 | goto out; | |
380 | } | |
381 | ||
382 | fstat_done = fstat_checkout_output(fd, state, &pc_item->st); | |
383 | ||
384 | if (close_and_clear(&fd)) { | |
385 | error_errno("unable to close file '%s'", path.buf); | |
386 | pc_item->status = PC_ITEM_FAILED; | |
387 | goto out; | |
388 | } | |
389 | ||
390 | if (state->refresh_cache && !fstat_done && lstat(path.buf, &pc_item->st) < 0) { | |
391 | error_errno("unable to stat just-written file '%s'", path.buf); | |
392 | pc_item->status = PC_ITEM_FAILED; | |
393 | goto out; | |
394 | } | |
395 | ||
396 | pc_item->status = PC_ITEM_WRITTEN; | |
397 | ||
398 | out: | |
399 | strbuf_release(&path); | |
400 | } | |
401 | ||
e9e8adf1 MT |
402 | static void send_one_item(int fd, struct parallel_checkout_item *pc_item) |
403 | { | |
404 | size_t len_data; | |
405 | char *data, *variant; | |
406 | struct pc_item_fixed_portion *fixed_portion; | |
407 | const char *working_tree_encoding = pc_item->ca.working_tree_encoding; | |
408 | size_t name_len = pc_item->ce->ce_namelen; | |
409 | size_t working_tree_encoding_len = working_tree_encoding ? | |
410 | strlen(working_tree_encoding) : 0; | |
411 | ||
412 | /* | |
413 | * Any changes in the calculation of the message size must also be made | |
414 | * in is_eligible_for_parallel_checkout(). | |
415 | */ | |
416 | len_data = sizeof(struct pc_item_fixed_portion) + name_len + | |
417 | working_tree_encoding_len; | |
418 | ||
3d20ed27 | 419 | data = xmalloc(len_data); |
e9e8adf1 MT |
420 | |
421 | fixed_portion = (struct pc_item_fixed_portion *)data; | |
422 | fixed_portion->id = pc_item->id; | |
423 | fixed_portion->ce_mode = pc_item->ce->ce_mode; | |
424 | fixed_portion->crlf_action = pc_item->ca.crlf_action; | |
425 | fixed_portion->ident = pc_item->ca.ident; | |
426 | fixed_portion->name_len = name_len; | |
427 | fixed_portion->working_tree_encoding_len = working_tree_encoding_len; | |
428 | /* | |
3d20ed27 MT |
429 | * We pad the unused bytes in the hash array because, otherwise, |
430 | * Valgrind would complain about passing uninitialized bytes to a | |
431 | * write() syscall. The warning doesn't represent any real risk here, | |
432 | * but it could hinder the detection of actual errors. | |
e9e8adf1 | 433 | */ |
3d20ed27 | 434 | oidcpy_with_padding(&fixed_portion->oid, &pc_item->ce->oid); |
e9e8adf1 MT |
435 | |
436 | variant = data + sizeof(*fixed_portion); | |
437 | if (working_tree_encoding_len) { | |
438 | memcpy(variant, working_tree_encoding, working_tree_encoding_len); | |
439 | variant += working_tree_encoding_len; | |
440 | } | |
441 | memcpy(variant, pc_item->ce->name, name_len); | |
442 | ||
443 | packet_write(fd, data, len_data); | |
444 | ||
445 | free(data); | |
446 | } | |
447 | ||
448 | static void send_batch(int fd, size_t start, size_t nr) | |
449 | { | |
450 | size_t i; | |
451 | sigchain_push(SIGPIPE, SIG_IGN); | |
452 | for (i = 0; i < nr; i++) | |
453 | send_one_item(fd, ¶llel_checkout.items[start + i]); | |
454 | packet_flush(fd); | |
455 | sigchain_pop(SIGPIPE); | |
456 | } | |
457 | ||
458 | static struct pc_worker *setup_workers(struct checkout *state, int num_workers) | |
459 | { | |
460 | struct pc_worker *workers; | |
461 | int i, workers_with_one_extra_item; | |
462 | size_t base_batch_size, batch_beginning = 0; | |
463 | ||
464 | ALLOC_ARRAY(workers, num_workers); | |
465 | ||
466 | for (i = 0; i < num_workers; i++) { | |
467 | struct child_process *cp = &workers[i].cp; | |
468 | ||
469 | child_process_init(cp); | |
470 | cp->git_cmd = 1; | |
471 | cp->in = -1; | |
472 | cp->out = -1; | |
473 | cp->clean_on_exit = 1; | |
474 | strvec_push(&cp->args, "checkout--worker"); | |
475 | if (state->base_dir_len) | |
476 | strvec_pushf(&cp->args, "--prefix=%s", state->base_dir); | |
477 | if (start_command(cp)) | |
478 | die("failed to spawn checkout worker"); | |
479 | } | |
480 | ||
481 | base_batch_size = parallel_checkout.nr / num_workers; | |
482 | workers_with_one_extra_item = parallel_checkout.nr % num_workers; | |
483 | ||
484 | for (i = 0; i < num_workers; i++) { | |
485 | struct pc_worker *worker = &workers[i]; | |
486 | size_t batch_size = base_batch_size; | |
487 | ||
488 | /* distribute the extra work evenly */ | |
489 | if (i < workers_with_one_extra_item) | |
490 | batch_size++; | |
491 | ||
492 | send_batch(worker->cp.in, batch_beginning, batch_size); | |
493 | worker->next_item_to_complete = batch_beginning; | |
494 | worker->nr_items_to_complete = batch_size; | |
495 | ||
496 | batch_beginning += batch_size; | |
497 | } | |
498 | ||
499 | return workers; | |
500 | } | |
501 | ||
502 | static void finish_workers(struct pc_worker *workers, int num_workers) | |
503 | { | |
504 | int i; | |
505 | ||
506 | /* | |
507 | * Close pipes before calling finish_command() to let the workers | |
508 | * exit asynchronously and avoid spending extra time on wait(). | |
509 | */ | |
510 | for (i = 0; i < num_workers; i++) { | |
511 | struct child_process *cp = &workers[i].cp; | |
512 | if (cp->in >= 0) | |
513 | close(cp->in); | |
514 | if (cp->out >= 0) | |
515 | close(cp->out); | |
516 | } | |
517 | ||
518 | for (i = 0; i < num_workers; i++) { | |
519 | int rc = finish_command(&workers[i].cp); | |
520 | if (rc > 128) { | |
521 | /* | |
522 | * For a normal non-zero exit, the worker should have | |
523 | * already printed something useful to stderr. But a | |
524 | * death by signal should be mentioned to the user. | |
525 | */ | |
526 | error("checkout worker %d died of signal %d", i, rc - 128); | |
527 | } | |
528 | } | |
529 | ||
530 | free(workers); | |
531 | } | |
532 | ||
533 | static inline void assert_pc_item_result_size(int got, int exp) | |
534 | { | |
535 | if (got != exp) | |
536 | BUG("wrong result size from checkout worker (got %dB, exp %dB)", | |
537 | got, exp); | |
538 | } | |
539 | ||
540 | static void parse_and_save_result(const char *buffer, int len, | |
541 | struct pc_worker *worker) | |
542 | { | |
543 | struct pc_item_result *res; | |
544 | struct parallel_checkout_item *pc_item; | |
545 | struct stat *st = NULL; | |
546 | ||
547 | if (len < PC_ITEM_RESULT_BASE_SIZE) | |
548 | BUG("too short result from checkout worker (got %dB, exp >=%dB)", | |
549 | len, (int)PC_ITEM_RESULT_BASE_SIZE); | |
550 | ||
551 | res = (struct pc_item_result *)buffer; | |
552 | ||
553 | /* | |
554 | * Worker should send either the full result struct on success, or | |
555 | * just the base (i.e. no stat data), otherwise. | |
556 | */ | |
557 | if (res->status == PC_ITEM_WRITTEN) { | |
558 | assert_pc_item_result_size(len, (int)sizeof(struct pc_item_result)); | |
559 | st = &res->st; | |
560 | } else { | |
561 | assert_pc_item_result_size(len, (int)PC_ITEM_RESULT_BASE_SIZE); | |
562 | } | |
563 | ||
564 | if (!worker->nr_items_to_complete) | |
565 | BUG("received result from supposedly finished checkout worker"); | |
566 | if (res->id != worker->next_item_to_complete) | |
567 | BUG("unexpected item id from checkout worker (got %"PRIuMAX", exp %"PRIuMAX")", | |
568 | (uintmax_t)res->id, (uintmax_t)worker->next_item_to_complete); | |
569 | ||
570 | worker->next_item_to_complete++; | |
571 | worker->nr_items_to_complete--; | |
572 | ||
573 | pc_item = ¶llel_checkout.items[res->id]; | |
574 | pc_item->status = res->status; | |
575 | if (st) | |
576 | pc_item->st = *st; | |
1c4d6f46 MT |
577 | |
578 | if (res->status != PC_ITEM_COLLIDED) | |
579 | advance_progress_meter(); | |
e9e8adf1 MT |
580 | } |
581 | ||
582 | static void gather_results_from_workers(struct pc_worker *workers, | |
583 | int num_workers) | |
584 | { | |
585 | int i, active_workers = num_workers; | |
586 | struct pollfd *pfds; | |
587 | ||
588 | CALLOC_ARRAY(pfds, num_workers); | |
589 | for (i = 0; i < num_workers; i++) { | |
590 | pfds[i].fd = workers[i].cp.out; | |
591 | pfds[i].events = POLLIN; | |
592 | } | |
593 | ||
594 | while (active_workers) { | |
595 | int nr = poll(pfds, num_workers, -1); | |
596 | ||
597 | if (nr < 0) { | |
598 | if (errno == EINTR) | |
599 | continue; | |
600 | die_errno("failed to poll checkout workers"); | |
601 | } | |
602 | ||
603 | for (i = 0; i < num_workers && nr > 0; i++) { | |
604 | struct pc_worker *worker = &workers[i]; | |
605 | struct pollfd *pfd = &pfds[i]; | |
606 | ||
607 | if (!pfd->revents) | |
608 | continue; | |
609 | ||
610 | if (pfd->revents & POLLIN) { | |
ec9a37d6 | 611 | int len = packet_read(pfd->fd, packet_buffer, |
e9e8adf1 MT |
612 | sizeof(packet_buffer), 0); |
613 | ||
614 | if (len < 0) { | |
615 | BUG("packet_read() returned negative value"); | |
616 | } else if (!len) { | |
617 | pfd->fd = -1; | |
618 | active_workers--; | |
619 | } else { | |
620 | parse_and_save_result(packet_buffer, | |
621 | len, worker); | |
622 | } | |
623 | } else if (pfd->revents & POLLHUP) { | |
624 | pfd->fd = -1; | |
625 | active_workers--; | |
626 | } else if (pfd->revents & (POLLNVAL | POLLERR)) { | |
627 | die("error polling from checkout worker"); | |
628 | } | |
629 | ||
630 | nr--; | |
631 | } | |
632 | } | |
633 | ||
634 | free(pfds); | |
635 | } | |
636 | ||
04155bda MT |
637 | static void write_items_sequentially(struct checkout *state) |
638 | { | |
639 | size_t i; | |
640 | ||
1c4d6f46 MT |
641 | for (i = 0; i < parallel_checkout.nr; i++) { |
642 | struct parallel_checkout_item *pc_item = ¶llel_checkout.items[i]; | |
643 | write_pc_item(pc_item, state); | |
644 | if (pc_item->status != PC_ITEM_COLLIDED) | |
645 | advance_progress_meter(); | |
646 | } | |
04155bda MT |
647 | } |
648 | ||
1c4d6f46 MT |
649 | int run_parallel_checkout(struct checkout *state, int num_workers, int threshold, |
650 | struct progress *progress, unsigned int *progress_cnt) | |
04155bda | 651 | { |
7531e4b6 | 652 | int ret; |
04155bda MT |
653 | |
654 | if (parallel_checkout.status != PC_ACCEPTING_ENTRIES) | |
655 | BUG("cannot run parallel checkout: uninitialized or already running"); | |
656 | ||
657 | parallel_checkout.status = PC_RUNNING; | |
1c4d6f46 MT |
658 | parallel_checkout.progress = progress; |
659 | parallel_checkout.progress_cnt = progress_cnt; | |
04155bda | 660 | |
e9e8adf1 MT |
661 | if (parallel_checkout.nr < num_workers) |
662 | num_workers = parallel_checkout.nr; | |
663 | ||
7531e4b6 | 664 | if (num_workers <= 1 || parallel_checkout.nr < threshold) { |
e9e8adf1 MT |
665 | write_items_sequentially(state); |
666 | } else { | |
667 | struct pc_worker *workers = setup_workers(state, num_workers); | |
668 | gather_results_from_workers(workers, num_workers); | |
669 | finish_workers(workers, num_workers); | |
670 | } | |
671 | ||
04155bda MT |
672 | ret = handle_results(state); |
673 | ||
674 | finish_parallel_checkout(); | |
675 | return ret; | |
676 | } |