Merge Monty
[m6w6/libmemcached] / libmemcached / io.c
1 /* LibMemcached
2 * Copyright (C) 2006-2009 Brian Aker
3 * All rights reserved.
4 *
5 * Use and distribution licensed under the BSD license. See
6 * the COPYING file in the parent directory for full text.
7 *
8 * Summary: Server IO, Not public!
9 *
10 */
11
12
13 #include "common.h"
14 #include <sys/select.h>
15 #include <poll.h>
16
17 typedef enum {
18 MEM_READ,
19 MEM_WRITE
20 } memc_read_or_write;
21
22 static ssize_t io_flush(memcached_server_instance_st *ptr, memcached_return_t *error);
23 static void increment_udp_message_id(memcached_server_instance_st *ptr);
24
25 static memcached_return_t io_wait(memcached_server_instance_st *ptr,
26 memc_read_or_write read_or_write)
27 {
28 struct pollfd fds= {
29 .fd= ptr->fd,
30 .events = POLLIN
31 };
32 int error;
33
34 unlikely (read_or_write == MEM_WRITE) /* write */
35 fds.events= POLLOUT;
36
37 /*
38 ** We are going to block on write, but at least on Solaris we might block
39 ** on write if we haven't read anything from our input buffer..
40 ** Try to purge the input buffer if we don't do any flow control in the
41 ** application layer (just sending a lot of data etc)
42 ** The test is moved down in the purge function to avoid duplication of
43 ** the test.
44 */
45 if (read_or_write == MEM_WRITE)
46 {
47 memcached_return_t rc= memcached_purge(ptr);
48 if (rc != MEMCACHED_SUCCESS && rc != MEMCACHED_STORED)
49 return MEMCACHED_FAILURE;
50 }
51
52 int timeout= ptr->root->poll_timeout;
53 if (ptr->root->flags.no_block == false)
54 timeout= -1;
55
56 error= poll(&fds, 1, timeout);
57
58 if (error == 1)
59 return MEMCACHED_SUCCESS;
60 else if (error == 0)
61 return MEMCACHED_TIMEOUT;
62
63 /* Imposssible for anything other then -1 */
64 WATCHPOINT_ASSERT(error == -1);
65 memcached_quit_server(ptr, 1);
66
67 return MEMCACHED_FAILURE;
68 }
69
70 /**
71 * Try to fill the input buffer for a server with as much
72 * data as possible.
73 *
74 * @param ptr the server to pack
75 */
76 static bool repack_input_buffer(memcached_server_instance_st *ptr)
77 {
78 if (ptr->read_ptr != ptr->read_buffer)
79 {
80 /* Move all of the data to the beginning of the buffer so
81 ** that we can fit more data into the buffer...
82 */
83 memmove(ptr->read_buffer, ptr->read_ptr, ptr->read_buffer_length);
84 ptr->read_ptr= ptr->read_buffer;
85 ptr->read_data_length= ptr->read_buffer_length;
86 }
87
88 /* There is room in the buffer, try to fill it! */
89 if (ptr->read_buffer_length != MEMCACHED_MAX_BUFFER)
90 {
91 /* Just try a single read to grab what's available */
92 ssize_t nr= read(ptr->fd,
93 ptr->read_ptr + ptr->read_data_length,
94 MEMCACHED_MAX_BUFFER - ptr->read_data_length);
95
96 if (nr > 0)
97 {
98 ptr->read_data_length+= (size_t)nr;
99 ptr->read_buffer_length+= (size_t)nr;
100 return true;
101 }
102 }
103 return false;
104 }
105
106 /**
107 * If the we have callbacks connected to this server structure
108 * we may start process the input queue and fire the callbacks
109 * for the incomming messages. This function is _only_ called
110 * when the input buffer is full, so that we _know_ that we have
111 * at least _one_ message to process.
112 *
113 * @param ptr the server to star processing iput messages for
114 * @return true if we processed anything, false otherwise
115 */
116 static bool process_input_buffer(memcached_server_instance_st *ptr)
117 {
118 /*
119 ** We might be able to process some of the response messages if we
120 ** have a callback set up
121 */
122 if (ptr->root->callbacks != NULL && ptr->root->flags.use_udp == false)
123 {
124 /*
125 * We might have responses... try to read them out and fire
126 * callbacks
127 */
128 memcached_callback_st cb= *ptr->root->callbacks;
129
130 ptr->root->options.is_processing_input= true;
131
132 char buffer[MEMCACHED_DEFAULT_COMMAND_SIZE];
133 memcached_return_t error;
134 error= memcached_response(ptr, buffer, sizeof(buffer),
135 &ptr->root->result);
136
137 ptr->root->options.is_processing_input = false;
138
139 if (error == MEMCACHED_SUCCESS)
140 {
141 for (unsigned int x= 0; x < cb.number_of_callback; x++)
142 {
143 error= (*cb.callback[x])(ptr->root, &ptr->root->result, cb.context);
144 if (error != MEMCACHED_SUCCESS)
145 break;
146 }
147
148 /* @todo what should I do with the error message??? */
149 }
150 /* @todo what should I do with other error messages?? */
151 return true;
152 }
153
154 return false;
155 }
156
157 #ifdef UNUSED
158 void memcached_io_preread(memcached_st *ptr)
159 {
160 unsigned int x;
161
162 return;
163
164 for (x= 0; x < memcached_server_count(ptr); x++)
165 {
166 if (memcached_server_response_count(ptr, x) &&
167 ptr->hosts[x].read_data_length < MEMCACHED_MAX_BUFFER )
168 {
169 size_t data_read;
170
171 data_read= read(ptr->hosts[x].fd,
172 ptr->hosts[x].read_ptr + ptr->hosts[x].read_data_length,
173 MEMCACHED_MAX_BUFFER - ptr->hosts[x].read_data_length);
174 if (data_read == -1)
175 continue;
176
177 ptr->hosts[x].read_buffer_length+= data_read;
178 ptr->hosts[x].read_data_length+= data_read;
179 }
180 }
181 }
182 #endif
183
184 memcached_return_t memcached_io_read(memcached_server_instance_st *ptr,
185 void *buffer, size_t length, ssize_t *nread)
186 {
187 char *buffer_ptr;
188
189 buffer_ptr= buffer;
190
191 while (length)
192 {
193 if (!ptr->read_buffer_length)
194 {
195 ssize_t data_read;
196
197 while (1)
198 {
199 data_read= read(ptr->fd, ptr->read_buffer, MEMCACHED_MAX_BUFFER);
200 if (data_read > 0)
201 break;
202 else if (data_read == -1)
203 {
204 ptr->cached_errno= errno;
205 memcached_return_t rc= MEMCACHED_UNKNOWN_READ_FAILURE;
206 switch (errno)
207 {
208 case EAGAIN:
209 case EINTR:
210 if ((rc= io_wait(ptr, MEM_READ)) == MEMCACHED_SUCCESS)
211 continue;
212 /* fall through */
213
214 default:
215 {
216 memcached_quit_server(ptr, 1);
217 *nread= -1;
218 return rc;
219 }
220 }
221 }
222 else
223 {
224 /*
225 EOF. Any data received so far is incomplete
226 so discard it. This always reads by byte in case of TCP
227 and protocol enforcement happens at memcached_response()
228 looking for '\n'. We do not care for UDB which requests 8 bytes
229 at once. Generally, this means that connection went away. Since
230 for blocking I/O we do not return 0 and for non-blocking case
231 it will return EGAIN if data is not immediatly available.
232 */
233 memcached_quit_server(ptr, 1);
234 *nread= -1;
235 return MEMCACHED_UNKNOWN_READ_FAILURE;
236 }
237 }
238
239 ptr->io_bytes_sent = 0;
240 ptr->read_data_length= (size_t) data_read;
241 ptr->read_buffer_length= (size_t) data_read;
242 ptr->read_ptr= ptr->read_buffer;
243 }
244
245 if (length > 1)
246 {
247 size_t difference;
248
249 difference= (length > ptr->read_buffer_length) ? ptr->read_buffer_length : length;
250
251 memcpy(buffer_ptr, ptr->read_ptr, difference);
252 length -= difference;
253 ptr->read_ptr+= difference;
254 ptr->read_buffer_length-= difference;
255 buffer_ptr+= difference;
256 }
257 else
258 {
259 *buffer_ptr= *ptr->read_ptr;
260 ptr->read_ptr++;
261 ptr->read_buffer_length--;
262 buffer_ptr++;
263 break;
264 }
265 }
266
267 ptr->server_failure_counter= 0;
268 *nread = (ssize_t)(buffer_ptr - (char*)buffer);
269 return MEMCACHED_SUCCESS;
270 }
271
272 ssize_t memcached_io_write(memcached_server_instance_st *ptr,
273 const void *buffer, size_t length, char with_flush)
274 {
275 size_t original_length;
276 const char* buffer_ptr;
277
278 WATCHPOINT_ASSERT(ptr->fd != -1);
279
280 original_length= length;
281 buffer_ptr= buffer;
282
283 while (length)
284 {
285 char *write_ptr;
286 size_t should_write;
287 size_t buffer_end;
288
289 if (ptr->type == MEMCACHED_CONNECTION_UDP)
290 {
291 //UDP does not support partial writes
292 buffer_end= MAX_UDP_DATAGRAM_LENGTH;
293 should_write= length;
294 if (ptr->write_buffer_offset + should_write > buffer_end)
295 return -1;
296 }
297 else
298 {
299 buffer_end= MEMCACHED_MAX_BUFFER;
300 should_write= buffer_end - ptr->write_buffer_offset;
301 should_write= (should_write < length) ? should_write : length;
302 }
303
304 write_ptr= ptr->write_buffer + ptr->write_buffer_offset;
305 memcpy(write_ptr, buffer_ptr, should_write);
306 ptr->write_buffer_offset+= should_write;
307 buffer_ptr+= should_write;
308 length-= should_write;
309
310 if (ptr->write_buffer_offset == buffer_end && ptr->type != MEMCACHED_CONNECTION_UDP)
311 {
312 memcached_return_t rc;
313 ssize_t sent_length;
314
315 WATCHPOINT_ASSERT(ptr->fd != -1);
316 sent_length= io_flush(ptr, &rc);
317 if (sent_length == -1)
318 return -1;
319
320 /* If io_flush calls memcached_purge, sent_length may be 0 */
321 unlikely (sent_length != 0)
322 {
323 WATCHPOINT_ASSERT(sent_length == (ssize_t)buffer_end);
324 }
325 }
326 }
327
328 if (with_flush)
329 {
330 memcached_return_t rc;
331 WATCHPOINT_ASSERT(ptr->fd != -1);
332 if (io_flush(ptr, &rc) == -1)
333 return -1;
334 }
335
336 return (ssize_t) original_length;
337 }
338
339 memcached_return_t memcached_io_close(memcached_server_instance_st *ptr)
340 {
341 if (ptr->fd == -1)
342 {
343 return MEMCACHED_SUCCESS;
344 }
345
346 /* in case of death shutdown to avoid blocking at close() */
347 if (shutdown(ptr->fd, SHUT_RDWR) == -1 && errno != ENOTCONN)
348 {
349 WATCHPOINT_NUMBER(ptr->fd);
350 WATCHPOINT_ERRNO(errno);
351 WATCHPOINT_ASSERT(errno);
352 }
353
354 if (close(ptr->fd) == -1)
355 {
356 WATCHPOINT_ERRNO(errno);
357 }
358
359 return MEMCACHED_SUCCESS;
360 }
361
362 memcached_server_instance_st *memcached_io_get_readable_server(memcached_st *memc)
363 {
364 #define MAX_SERVERS_TO_POLL 100
365 struct pollfd fds[MAX_SERVERS_TO_POLL];
366 unsigned int host_index= 0;
367
368 for (uint32_t x= 0;
369 x< memcached_server_count(memc) && host_index < MAX_SERVERS_TO_POLL;
370 ++x)
371 {
372 memcached_server_instance_st *instance=
373 memcached_server_instance_fetch(memc, x);
374
375 if (instance->read_buffer_length > 0) /* I have data in the buffer */
376 return instance;
377
378 if (memcached_server_response_count(instance) > 0)
379 {
380 fds[host_index].events = POLLIN;
381 fds[host_index].revents = 0;
382 fds[host_index].fd = instance->fd;
383 ++host_index;
384 }
385 }
386
387 if (host_index < 2)
388 {
389 /* We have 0 or 1 server with pending events.. */
390 for (uint32_t x= 0; x< memcached_server_count(memc); ++x)
391 {
392 memcached_server_instance_st *instance=
393 memcached_server_instance_fetch(memc, x);
394
395 if (memcached_server_response_count(instance) > 0)
396 {
397 return instance;
398 }
399 }
400
401 return NULL;
402 }
403
404 int err= poll(fds, host_index, memc->poll_timeout);
405 switch (err) {
406 case -1:
407 memc->cached_errno = errno;
408 /* FALLTHROUGH */
409 case 0:
410 break;
411 default:
412 for (size_t x= 0; x < host_index; ++x)
413 {
414 if (fds[x].revents & POLLIN)
415 {
416 for (uint32_t y= 0; y < memcached_server_count(memc); ++y)
417 {
418 memcached_server_instance_st *instance=
419 memcached_server_instance_fetch(memc, y);
420
421 if (instance->fd == fds[x].fd)
422 return instance;
423 }
424 }
425 }
426 }
427
428 return NULL;
429 }
430
431 static ssize_t io_flush(memcached_server_instance_st *ptr,
432 memcached_return_t *error)
433 {
434 /*
435 ** We might want to purge the input buffer if we haven't consumed
436 ** any output yet... The test for the limits is the purge is inline
437 ** in the purge function to avoid duplicating the logic..
438 */
439 {
440 memcached_return_t rc;
441 WATCHPOINT_ASSERT(ptr->fd != -1);
442 rc= memcached_purge(ptr);
443
444 if (rc != MEMCACHED_SUCCESS && rc != MEMCACHED_STORED)
445 return -1;
446 }
447 ssize_t sent_length;
448 size_t return_length;
449 char *local_write_ptr= ptr->write_buffer;
450 size_t write_length= ptr->write_buffer_offset;
451
452 *error= MEMCACHED_SUCCESS;
453
454 WATCHPOINT_ASSERT(ptr->fd != -1);
455
456 // UDP Sanity check, make sure that we are not sending somthing too big
457 if (ptr->type == MEMCACHED_CONNECTION_UDP && write_length > MAX_UDP_DATAGRAM_LENGTH)
458 return -1;
459
460 if (ptr->write_buffer_offset == 0 || (ptr->type == MEMCACHED_CONNECTION_UDP
461 && ptr->write_buffer_offset == UDP_DATAGRAM_HEADER_LENGTH))
462 return 0;
463
464 /* Looking for memory overflows */
465 #if defined(DEBUG)
466 if (write_length == MEMCACHED_MAX_BUFFER)
467 WATCHPOINT_ASSERT(ptr->write_buffer == local_write_ptr);
468 WATCHPOINT_ASSERT((ptr->write_buffer + MEMCACHED_MAX_BUFFER) >= (local_write_ptr + write_length));
469 #endif
470
471 return_length= 0;
472 while (write_length)
473 {
474 WATCHPOINT_ASSERT(ptr->fd != -1);
475 WATCHPOINT_ASSERT(write_length > 0);
476 sent_length= 0;
477 if (ptr->type == MEMCACHED_CONNECTION_UDP)
478 increment_udp_message_id(ptr);
479 sent_length= write(ptr->fd, local_write_ptr, write_length);
480
481 if (sent_length == -1)
482 {
483 ptr->cached_errno= errno;
484 switch (errno)
485 {
486 case ENOBUFS:
487 continue;
488 case EAGAIN:
489 {
490 /*
491 * We may be blocked on write because the input buffer
492 * is full. Let's check if we have room in our input
493 * buffer for more data and retry the write before
494 * waiting..
495 */
496 if (repack_input_buffer(ptr) ||
497 process_input_buffer(ptr))
498 continue;
499
500 memcached_return_t rc;
501 rc= io_wait(ptr, MEM_WRITE);
502
503 if (rc == MEMCACHED_SUCCESS || rc == MEMCACHED_TIMEOUT)
504 continue;
505
506 memcached_quit_server(ptr, 1);
507 return -1;
508 }
509 default:
510 memcached_quit_server(ptr, 1);
511 *error= MEMCACHED_ERRNO;
512 return -1;
513 }
514 }
515
516 if (ptr->type == MEMCACHED_CONNECTION_UDP &&
517 (size_t)sent_length != write_length)
518 {
519 memcached_quit_server(ptr, 1);
520 return -1;
521 }
522
523 ptr->io_bytes_sent += (uint32_t) sent_length;
524
525 local_write_ptr+= sent_length;
526 write_length-= (uint32_t) sent_length;
527 return_length+= (uint32_t) sent_length;
528 }
529
530 WATCHPOINT_ASSERT(write_length == 0);
531 // Need to study this assert() WATCHPOINT_ASSERT(return_length ==
532 // ptr->write_buffer_offset);
533
534 // if we are a udp server, the begining of the buffer is reserverd for
535 // the upd frame header
536 if (ptr->type == MEMCACHED_CONNECTION_UDP)
537 ptr->write_buffer_offset= UDP_DATAGRAM_HEADER_LENGTH;
538 else
539 ptr->write_buffer_offset= 0;
540
541 return (ssize_t) return_length;
542 }
543
544 /*
545 Eventually we will just kill off the server with the problem.
546 */
547 void memcached_io_reset(memcached_server_instance_st *ptr)
548 {
549 memcached_quit_server(ptr, 1);
550 }
551
552 /**
553 * Read a given number of bytes from the server and place it into a specific
554 * buffer. Reset the IO channel on this server if an error occurs.
555 */
556 memcached_return_t memcached_safe_read(memcached_server_instance_st *ptr,
557 void *dta,
558 size_t size)
559 {
560 size_t offset= 0;
561 char *data= dta;
562
563 while (offset < size)
564 {
565 ssize_t nread;
566 memcached_return_t rc= memcached_io_read(ptr, data + offset, size - offset,
567 &nread);
568 if (rc != MEMCACHED_SUCCESS)
569 return rc;
570
571 offset+= (size_t) nread;
572 }
573
574 return MEMCACHED_SUCCESS;
575 }
576
577 memcached_return_t memcached_io_readline(memcached_server_instance_st *ptr,
578 char *buffer_ptr,
579 size_t size)
580 {
581 bool line_complete= false;
582 size_t total_nr= 0;
583
584 while (!line_complete)
585 {
586 if (ptr->read_buffer_length == 0)
587 {
588 /*
589 * We don't have any data in the buffer, so let's fill the read
590 * buffer. Call the standard read function to avoid duplicating
591 * the logic.
592 */
593 ssize_t nread;
594 memcached_return_t rc= memcached_io_read(ptr, buffer_ptr, 1, &nread);
595 if (rc != MEMCACHED_SUCCESS)
596 return rc;
597
598 if (*buffer_ptr == '\n')
599 line_complete= true;
600
601 ++buffer_ptr;
602 ++total_nr;
603 }
604
605 /* Now let's look in the buffer and copy as we go! */
606 while (ptr->read_buffer_length && total_nr < size && !line_complete)
607 {
608 *buffer_ptr = *ptr->read_ptr;
609 if (*buffer_ptr == '\n')
610 line_complete = true;
611 --ptr->read_buffer_length;
612 ++ptr->read_ptr;
613 ++total_nr;
614 ++buffer_ptr;
615 }
616
617 if (total_nr == size)
618 return MEMCACHED_PROTOCOL_ERROR;
619 }
620
621 return MEMCACHED_SUCCESS;
622 }
623
624 /*
625 * The udp request id consists of two seperate sections
626 * 1) The thread id
627 * 2) The message number
628 * The thread id should only be set when the memcached_st struct is created
629 * and should not be changed.
630 *
631 * The message num is incremented for each new message we send, this function
632 * extracts the message number from message_id, increments it and then
633 * writes the new value back into the header
634 */
635 static void increment_udp_message_id(memcached_server_instance_st *ptr)
636 {
637 struct udp_datagram_header_st *header= (struct udp_datagram_header_st *)ptr->write_buffer;
638 uint16_t cur_req= get_udp_datagram_request_id(header);
639 int msg_num= get_msg_num_from_request_id(cur_req);
640 int thread_id= get_thread_id_from_request_id(cur_req);
641
642 if (((++msg_num) & UDP_REQUEST_ID_THREAD_MASK) != 0)
643 msg_num= 0;
644
645 header->request_id= htons((uint16_t) (thread_id | msg_num));
646 }
647
648 memcached_return_t memcached_io_init_udp_header(memcached_server_instance_st *ptr, uint16_t thread_id)
649 {
650 if (thread_id > UDP_REQUEST_ID_MAX_THREAD_ID)
651 return MEMCACHED_FAILURE;
652
653 struct udp_datagram_header_st *header= (struct udp_datagram_header_st *)ptr->write_buffer;
654 header->request_id= htons((uint16_t) (generate_udp_request_thread_id(thread_id)));
655 header->num_datagrams= htons(1);
656 header->sequence_number= htons(0);
657
658 return MEMCACHED_SUCCESS;
659 }