2 * Copyright (C) 2006-2009 Brian Aker
5 * Use and distribution licensed under the BSD license. See
6 * the COPYING file in the parent directory for full text.
8 * Summary: Server IO, Not public!
14 #include <sys/select.h>
22 static ssize_t
io_flush(memcached_server_write_instance_st ptr
,
23 memcached_return_t
*error
);
24 static void increment_udp_message_id(memcached_server_write_instance_st ptr
);
26 static memcached_return_t
io_wait(memcached_server_write_instance_st ptr
,
27 memc_read_or_write read_or_write
)
35 unlikely (read_or_write
== MEM_WRITE
) /* write */
39 ** We are going to block on write, but at least on Solaris we might block
40 ** on write if we haven't read anything from our input buffer..
41 ** Try to purge the input buffer if we don't do any flow control in the
42 ** application layer (just sending a lot of data etc)
43 ** The test is moved down in the purge function to avoid duplication of
46 if (read_or_write
== MEM_WRITE
)
48 memcached_return_t rc
= memcached_purge(ptr
);
49 if (rc
!= MEMCACHED_SUCCESS
&& rc
!= MEMCACHED_STORED
)
50 return MEMCACHED_FAILURE
;
53 int timeout
= ptr
->root
->poll_timeout
;
54 if (ptr
->root
->flags
.no_block
== false)
60 error
= poll(&fds
, 1, timeout
);
65 return MEMCACHED_SUCCESS
;
67 return MEMCACHED_TIMEOUT
;
68 #ifdef TARGET_OS_LINUX
74 ptr
->cached_errno
= error
;
75 memcached_quit_server(ptr
, true);
77 return MEMCACHED_FAILURE
;
81 /* Imposssible for anything other then -1 */
82 WATCHPOINT_ASSERT(error
== -1);
83 ptr
->cached_errno
= error
;
84 memcached_quit_server(ptr
, true);
86 return MEMCACHED_FAILURE
;
90 * Try to fill the input buffer for a server with as much
93 * @param ptr the server to pack
95 static bool repack_input_buffer(memcached_server_write_instance_st ptr
)
97 if (ptr
->read_ptr
!= ptr
->read_buffer
)
99 /* Move all of the data to the beginning of the buffer so
100 ** that we can fit more data into the buffer...
102 memmove(ptr
->read_buffer
, ptr
->read_ptr
, ptr
->read_buffer_length
);
103 ptr
->read_ptr
= ptr
->read_buffer
;
104 ptr
->read_data_length
= ptr
->read_buffer_length
;
107 /* There is room in the buffer, try to fill it! */
108 if (ptr
->read_buffer_length
!= MEMCACHED_MAX_BUFFER
)
110 /* Just try a single read to grab what's available */
111 ssize_t nr
= read(ptr
->fd
,
112 ptr
->read_ptr
+ ptr
->read_data_length
,
113 MEMCACHED_MAX_BUFFER
- ptr
->read_data_length
);
117 ptr
->read_data_length
+= (size_t)nr
;
118 ptr
->read_buffer_length
+= (size_t)nr
;
126 * If the we have callbacks connected to this server structure
127 * we may start process the input queue and fire the callbacks
128 * for the incomming messages. This function is _only_ called
129 * when the input buffer is full, so that we _know_ that we have
130 * at least _one_ message to process.
132 * @param ptr the server to star processing iput messages for
133 * @return true if we processed anything, false otherwise
135 static bool process_input_buffer(memcached_server_write_instance_st ptr
)
138 ** We might be able to process some of the response messages if we
139 ** have a callback set up
141 if (ptr
->root
->callbacks
!= NULL
&& ptr
->root
->flags
.use_udp
== false)
144 * We might have responses... try to read them out and fire
147 memcached_callback_st cb
= *ptr
->root
->callbacks
;
149 memcached_set_processing_input((memcached_st
*)ptr
->root
, true);
151 char buffer
[MEMCACHED_DEFAULT_COMMAND_SIZE
];
152 memcached_return_t error
;
153 memcached_st
*root
= (memcached_st
*)ptr
->root
;
154 error
= memcached_response(ptr
, buffer
, sizeof(buffer
),
157 memcached_set_processing_input(root
, false);
159 if (error
== MEMCACHED_SUCCESS
)
161 for (unsigned int x
= 0; x
< cb
.number_of_callback
; x
++)
163 error
= (*cb
.callback
[x
])(ptr
->root
, &root
->result
, cb
.context
);
164 if (error
!= MEMCACHED_SUCCESS
)
168 /* @todo what should I do with the error message??? */
170 /* @todo what should I do with other error messages?? */
177 static inline void memcached_io_cork_push(memcached_server_st
*ptr
)
181 if (ptr
->root
->flags
.cork
== false || ptr
->state
.is_corked
)
185 int err
= setsockopt(ptr
->fd
, IPPROTO_TCP
, CORK
,
186 &enable
, (socklen_t
)sizeof(int));
188 ptr
->state
.is_corked
= true;
190 WATCHPOINT_ASSERT(ptr
->state
.is_corked
== true);
194 static inline void memcached_io_cork_pop(memcached_server_st
*ptr
)
198 if (ptr
->root
->flags
.cork
== false || ptr
->state
.is_corked
== false)
202 int err
= setsockopt(ptr
->fd
, IPPROTO_TCP
, CORK
,
203 &enable
, (socklen_t
)sizeof(int));
205 ptr
->state
.is_corked
= false;
207 WATCHPOINT_ASSERT(ptr
->state
.is_corked
== false);
211 #if 0 // Dead code, this should be removed.
212 void memcached_io_preread(memcached_st
*ptr
)
218 for (x
= 0; x
< memcached_server_count(ptr
); x
++)
220 if (memcached_server_response_count(ptr
, x
) &&
221 ptr
->hosts
[x
].read_data_length
< MEMCACHED_MAX_BUFFER
)
225 data_read
= read(ptr
->hosts
[x
].fd
,
226 ptr
->hosts
[x
].read_ptr
+ ptr
->hosts
[x
].read_data_length
,
227 MEMCACHED_MAX_BUFFER
- ptr
->hosts
[x
].read_data_length
);
231 ptr
->hosts
[x
].read_buffer_length
+= data_read
;
232 ptr
->hosts
[x
].read_data_length
+= data_read
;
238 memcached_return_t
memcached_io_read(memcached_server_write_instance_st ptr
,
239 void *buffer
, size_t length
, ssize_t
*nread
)
247 if (!ptr
->read_buffer_length
)
253 data_read
= read(ptr
->fd
, ptr
->read_buffer
, MEMCACHED_MAX_BUFFER
);
256 else if (data_read
== -1)
258 ptr
->cached_errno
= errno
;
259 memcached_return_t rc
= MEMCACHED_UNKNOWN_READ_FAILURE
;
264 #ifdef TARGET_OS_LINUX
267 if ((rc
= io_wait(ptr
, MEM_READ
)) == MEMCACHED_SUCCESS
)
273 memcached_quit_server(ptr
, true);
282 EOF. Any data received so far is incomplete
283 so discard it. This always reads by byte in case of TCP
284 and protocol enforcement happens at memcached_response()
285 looking for '\n'. We do not care for UDB which requests 8 bytes
286 at once. Generally, this means that connection went away. Since
287 for blocking I/O we do not return 0 and for non-blocking case
288 it will return EGAIN if data is not immediatly available.
290 memcached_quit_server(ptr
, true);
292 return MEMCACHED_UNKNOWN_READ_FAILURE
;
296 ptr
->io_bytes_sent
= 0;
297 ptr
->read_data_length
= (size_t) data_read
;
298 ptr
->read_buffer_length
= (size_t) data_read
;
299 ptr
->read_ptr
= ptr
->read_buffer
;
306 difference
= (length
> ptr
->read_buffer_length
) ? ptr
->read_buffer_length
: length
;
308 memcpy(buffer_ptr
, ptr
->read_ptr
, difference
);
309 length
-= difference
;
310 ptr
->read_ptr
+= difference
;
311 ptr
->read_buffer_length
-= difference
;
312 buffer_ptr
+= difference
;
316 *buffer_ptr
= *ptr
->read_ptr
;
318 ptr
->read_buffer_length
--;
324 ptr
->server_failure_counter
= 0;
325 *nread
= (ssize_t
)(buffer_ptr
- (char*)buffer
);
326 return MEMCACHED_SUCCESS
;
329 ssize_t
memcached_io_writev(memcached_server_write_instance_st ptr
,
330 struct __write_vector_st
*vector
,
331 size_t number_of
, bool with_flush
)
335 for (size_t x
= 0; x
< number_of
; x
++, vector
++)
339 if ((returnable
= memcached_io_write(ptr
, vector
->buffer
, vector
->length
, false)) == -1)
348 if (memcached_io_write(ptr
, NULL
, 0, true) == -1)
357 ssize_t
memcached_io_write(memcached_server_write_instance_st ptr
,
358 const void *buffer
, size_t length
, bool with_flush
)
360 size_t original_length
;
361 const char* buffer_ptr
;
363 WATCHPOINT_ASSERT(ptr
->fd
!= -1);
365 original_length
= length
;
368 /* more writable data is coming if a flush isn't required, so delay send */
371 memcached_io_cork_push(ptr
);
380 if (ptr
->type
== MEMCACHED_CONNECTION_UDP
)
382 //UDP does not support partial writes
383 buffer_end
= MAX_UDP_DATAGRAM_LENGTH
;
384 should_write
= length
;
385 if (ptr
->write_buffer_offset
+ should_write
> buffer_end
)
390 buffer_end
= MEMCACHED_MAX_BUFFER
;
391 should_write
= buffer_end
- ptr
->write_buffer_offset
;
392 should_write
= (should_write
< length
) ? should_write
: length
;
395 write_ptr
= ptr
->write_buffer
+ ptr
->write_buffer_offset
;
396 memcpy(write_ptr
, buffer_ptr
, should_write
);
397 ptr
->write_buffer_offset
+= should_write
;
398 buffer_ptr
+= should_write
;
399 length
-= should_write
;
401 if (ptr
->write_buffer_offset
== buffer_end
&& ptr
->type
!= MEMCACHED_CONNECTION_UDP
)
403 memcached_return_t rc
;
406 WATCHPOINT_ASSERT(ptr
->fd
!= -1);
407 sent_length
= io_flush(ptr
, &rc
);
408 if (sent_length
== -1)
411 /* If io_flush calls memcached_purge, sent_length may be 0 */
412 unlikely (sent_length
!= 0)
414 WATCHPOINT_ASSERT(sent_length
== (ssize_t
)buffer_end
);
421 memcached_return_t rc
;
422 WATCHPOINT_ASSERT(ptr
->fd
!= -1);
423 if (io_flush(ptr
, &rc
) == -1)
428 memcached_io_cork_pop(ptr
);
431 return (ssize_t
) original_length
;
434 memcached_return_t
memcached_io_close(memcached_server_write_instance_st ptr
)
438 return MEMCACHED_SUCCESS
;
441 /* in case of death shutdown to avoid blocking at close() */
442 if (shutdown(ptr
->fd
, SHUT_RDWR
) == -1 && errno
!= ENOTCONN
)
444 WATCHPOINT_NUMBER(ptr
->fd
);
445 WATCHPOINT_ERRNO(errno
);
446 WATCHPOINT_ASSERT(errno
);
449 if (close(ptr
->fd
) == -1)
451 WATCHPOINT_ERRNO(errno
);
454 return MEMCACHED_SUCCESS
;
457 memcached_server_write_instance_st
memcached_io_get_readable_server(memcached_st
*memc
)
459 #define MAX_SERVERS_TO_POLL 100
460 struct pollfd fds
[MAX_SERVERS_TO_POLL
];
461 unsigned int host_index
= 0;
464 x
< memcached_server_count(memc
) && host_index
< MAX_SERVERS_TO_POLL
;
467 memcached_server_write_instance_st instance
=
468 memcached_server_instance_fetch(memc
, x
);
470 if (instance
->read_buffer_length
> 0) /* I have data in the buffer */
473 if (memcached_server_response_count(instance
) > 0)
475 fds
[host_index
].events
= POLLIN
;
476 fds
[host_index
].revents
= 0;
477 fds
[host_index
].fd
= instance
->fd
;
484 /* We have 0 or 1 server with pending events.. */
485 for (uint32_t x
= 0; x
< memcached_server_count(memc
); ++x
)
487 memcached_server_write_instance_st instance
=
488 memcached_server_instance_fetch(memc
, x
);
490 if (memcached_server_response_count(instance
) > 0)
499 int err
= poll(fds
, host_index
, memc
->poll_timeout
);
502 memc
->cached_errno
= errno
;
507 for (size_t x
= 0; x
< host_index
; ++x
)
509 if (fds
[x
].revents
& POLLIN
)
511 for (uint32_t y
= 0; y
< memcached_server_count(memc
); ++y
)
513 memcached_server_write_instance_st instance
=
514 memcached_server_instance_fetch(memc
, y
);
516 if (instance
->fd
== fds
[x
].fd
)
526 static ssize_t
io_flush(memcached_server_write_instance_st ptr
,
527 memcached_return_t
*error
)
530 ** We might want to purge the input buffer if we haven't consumed
531 ** any output yet... The test for the limits is the purge is inline
532 ** in the purge function to avoid duplicating the logic..
535 memcached_return_t rc
;
536 WATCHPOINT_ASSERT(ptr
->fd
!= -1);
537 rc
= memcached_purge(ptr
);
539 if (rc
!= MEMCACHED_SUCCESS
&& rc
!= MEMCACHED_STORED
)
543 size_t return_length
;
544 char *local_write_ptr
= ptr
->write_buffer
;
545 size_t write_length
= ptr
->write_buffer_offset
;
547 *error
= MEMCACHED_SUCCESS
;
549 WATCHPOINT_ASSERT(ptr
->fd
!= -1);
551 // UDP Sanity check, make sure that we are not sending somthing too big
552 if (ptr
->type
== MEMCACHED_CONNECTION_UDP
&& write_length
> MAX_UDP_DATAGRAM_LENGTH
)
555 if (ptr
->write_buffer_offset
== 0 || (ptr
->type
== MEMCACHED_CONNECTION_UDP
556 && ptr
->write_buffer_offset
== UDP_DATAGRAM_HEADER_LENGTH
))
559 /* Looking for memory overflows */
561 if (write_length
== MEMCACHED_MAX_BUFFER
)
562 WATCHPOINT_ASSERT(ptr
->write_buffer
== local_write_ptr
);
563 WATCHPOINT_ASSERT((ptr
->write_buffer
+ MEMCACHED_MAX_BUFFER
) >= (local_write_ptr
+ write_length
));
569 WATCHPOINT_ASSERT(ptr
->fd
!= -1);
570 WATCHPOINT_ASSERT(write_length
> 0);
572 if (ptr
->type
== MEMCACHED_CONNECTION_UDP
)
573 increment_udp_message_id(ptr
);
574 sent_length
= write(ptr
->fd
, local_write_ptr
, write_length
);
576 if (sent_length
== -1)
578 ptr
->cached_errno
= errno
;
586 * We may be blocked on write because the input buffer
587 * is full. Let's check if we have room in our input
588 * buffer for more data and retry the write before
591 if (repack_input_buffer(ptr
) ||
592 process_input_buffer(ptr
))
595 memcached_return_t rc
;
596 rc
= io_wait(ptr
, MEM_WRITE
);
598 if (rc
== MEMCACHED_SUCCESS
|| rc
== MEMCACHED_TIMEOUT
)
601 memcached_quit_server(ptr
, true);
605 memcached_quit_server(ptr
, true);
606 *error
= MEMCACHED_ERRNO
;
611 if (ptr
->type
== MEMCACHED_CONNECTION_UDP
&&
612 (size_t)sent_length
!= write_length
)
614 memcached_quit_server(ptr
, true);
618 ptr
->io_bytes_sent
+= (uint32_t) sent_length
;
620 local_write_ptr
+= sent_length
;
621 write_length
-= (uint32_t) sent_length
;
622 return_length
+= (uint32_t) sent_length
;
625 WATCHPOINT_ASSERT(write_length
== 0);
626 // Need to study this assert() WATCHPOINT_ASSERT(return_length ==
627 // ptr->write_buffer_offset);
629 // if we are a udp server, the begining of the buffer is reserverd for
630 // the upd frame header
631 if (ptr
->type
== MEMCACHED_CONNECTION_UDP
)
632 ptr
->write_buffer_offset
= UDP_DATAGRAM_HEADER_LENGTH
;
634 ptr
->write_buffer_offset
= 0;
636 return (ssize_t
) return_length
;
640 Eventually we will just kill off the server with the problem.
642 void memcached_io_reset(memcached_server_write_instance_st ptr
)
644 memcached_quit_server(ptr
, true);
648 * Read a given number of bytes from the server and place it into a specific
649 * buffer. Reset the IO channel on this server if an error occurs.
651 memcached_return_t
memcached_safe_read(memcached_server_write_instance_st ptr
,
658 while (offset
< size
)
661 memcached_return_t rc
= memcached_io_read(ptr
, data
+ offset
, size
- offset
,
663 if (rc
!= MEMCACHED_SUCCESS
)
666 offset
+= (size_t) nread
;
669 return MEMCACHED_SUCCESS
;
672 memcached_return_t
memcached_io_readline(memcached_server_write_instance_st ptr
,
676 bool line_complete
= false;
679 while (!line_complete
)
681 if (ptr
->read_buffer_length
== 0)
684 * We don't have any data in the buffer, so let's fill the read
685 * buffer. Call the standard read function to avoid duplicating
689 memcached_return_t rc
= memcached_io_read(ptr
, buffer_ptr
, 1, &nread
);
690 if (rc
!= MEMCACHED_SUCCESS
)
693 if (*buffer_ptr
== '\n')
700 /* Now let's look in the buffer and copy as we go! */
701 while (ptr
->read_buffer_length
&& total_nr
< size
&& !line_complete
)
703 *buffer_ptr
= *ptr
->read_ptr
;
704 if (*buffer_ptr
== '\n')
705 line_complete
= true;
706 --ptr
->read_buffer_length
;
712 if (total_nr
== size
)
713 return MEMCACHED_PROTOCOL_ERROR
;
716 return MEMCACHED_SUCCESS
;
720 * The udp request id consists of two seperate sections
722 * 2) The message number
723 * The thread id should only be set when the memcached_st struct is created
724 * and should not be changed.
726 * The message num is incremented for each new message we send, this function
727 * extracts the message number from message_id, increments it and then
728 * writes the new value back into the header
730 static void increment_udp_message_id(memcached_server_write_instance_st ptr
)
732 struct udp_datagram_header_st
*header
= (struct udp_datagram_header_st
*)ptr
->write_buffer
;
733 uint16_t cur_req
= get_udp_datagram_request_id(header
);
734 int msg_num
= get_msg_num_from_request_id(cur_req
);
735 int thread_id
= get_thread_id_from_request_id(cur_req
);
737 if (((++msg_num
) & UDP_REQUEST_ID_THREAD_MASK
) != 0)
740 header
->request_id
= htons((uint16_t) (thread_id
| msg_num
));
743 memcached_return_t
memcached_io_init_udp_header(memcached_server_write_instance_st ptr
, uint16_t thread_id
)
745 if (thread_id
> UDP_REQUEST_ID_MAX_THREAD_ID
)
746 return MEMCACHED_FAILURE
;
748 struct udp_datagram_header_st
*header
= (struct udp_datagram_header_st
*)ptr
->write_buffer
;
749 header
->request_id
= htons((uint16_t) (generate_udp_request_thread_id(thread_id
)));
750 header
->num_datagrams
= htons(1);
751 header
->sequence_number
= htons(0);
753 return MEMCACHED_SUCCESS
;