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
;
69 WATCHPOINT_ERRNO(errno
);
74 ptr
->cached_errno
= error
;
75 memcached_quit_server(ptr
, true);
77 return MEMCACHED_FAILURE
;
83 if (loop_max
== 0 && error
== 0)
84 return MEMCACHED_TIMEOUT
;
86 /* Imposssible for anything other then -1 */
87 WATCHPOINT_ASSERT(error
== -1);
88 ptr
->cached_errno
= error
;
89 memcached_quit_server(ptr
, true);
91 return MEMCACHED_FAILURE
;
95 * Try to fill the input buffer for a server with as much
98 * @param ptr the server to pack
100 static bool repack_input_buffer(memcached_server_write_instance_st ptr
)
102 if (ptr
->read_ptr
!= ptr
->read_buffer
)
104 /* Move all of the data to the beginning of the buffer so
105 ** that we can fit more data into the buffer...
107 memmove(ptr
->read_buffer
, ptr
->read_ptr
, ptr
->read_buffer_length
);
108 ptr
->read_ptr
= ptr
->read_buffer
;
109 ptr
->read_data_length
= ptr
->read_buffer_length
;
112 /* There is room in the buffer, try to fill it! */
113 if (ptr
->read_buffer_length
!= MEMCACHED_MAX_BUFFER
)
115 /* Just try a single read to grab what's available */
116 ssize_t nr
= read(ptr
->fd
,
117 ptr
->read_ptr
+ ptr
->read_data_length
,
118 MEMCACHED_MAX_BUFFER
- ptr
->read_data_length
);
122 ptr
->read_data_length
+= (size_t)nr
;
123 ptr
->read_buffer_length
+= (size_t)nr
;
131 * If the we have callbacks connected to this server structure
132 * we may start process the input queue and fire the callbacks
133 * for the incomming messages. This function is _only_ called
134 * when the input buffer is full, so that we _know_ that we have
135 * at least _one_ message to process.
137 * @param ptr the server to star processing iput messages for
138 * @return true if we processed anything, false otherwise
140 static bool process_input_buffer(memcached_server_write_instance_st ptr
)
143 ** We might be able to process some of the response messages if we
144 ** have a callback set up
146 if (ptr
->root
->callbacks
!= NULL
&& ptr
->root
->flags
.use_udp
== false)
149 * We might have responses... try to read them out and fire
152 memcached_callback_st cb
= *ptr
->root
->callbacks
;
154 memcached_set_processing_input((memcached_st
*)ptr
->root
, true);
156 char buffer
[MEMCACHED_DEFAULT_COMMAND_SIZE
];
157 memcached_return_t error
;
158 memcached_st
*root
= (memcached_st
*)ptr
->root
;
159 error
= memcached_response(ptr
, buffer
, sizeof(buffer
),
162 memcached_set_processing_input(root
, false);
164 if (error
== MEMCACHED_SUCCESS
)
166 for (unsigned int x
= 0; x
< cb
.number_of_callback
; x
++)
168 error
= (*cb
.callback
[x
])(ptr
->root
, &root
->result
, cb
.context
);
169 if (error
!= MEMCACHED_SUCCESS
)
173 /* @todo what should I do with the error message??? */
175 /* @todo what should I do with other error messages?? */
182 static inline void memcached_io_cork_push(memcached_server_st
*ptr
)
186 if (ptr
->root
->flags
.cork
== false || ptr
->state
.is_corked
)
190 int err
= setsockopt(ptr
->fd
, IPPROTO_TCP
, CORK
,
191 &enable
, (socklen_t
)sizeof(int));
193 ptr
->state
.is_corked
= true;
195 WATCHPOINT_ASSERT(ptr
->state
.is_corked
== true);
199 static inline void memcached_io_cork_pop(memcached_server_st
*ptr
)
203 if (ptr
->root
->flags
.cork
== false || ptr
->state
.is_corked
== false)
207 int err
= setsockopt(ptr
->fd
, IPPROTO_TCP
, CORK
,
208 &enable
, (socklen_t
)sizeof(int));
210 ptr
->state
.is_corked
= false;
212 WATCHPOINT_ASSERT(ptr
->state
.is_corked
== false);
216 #if 0 // Dead code, this should be removed.
217 void memcached_io_preread(memcached_st
*ptr
)
223 for (x
= 0; x
< memcached_server_count(ptr
); x
++)
225 if (memcached_server_response_count(ptr
, x
) &&
226 ptr
->hosts
[x
].read_data_length
< MEMCACHED_MAX_BUFFER
)
230 data_read
= read(ptr
->hosts
[x
].fd
,
231 ptr
->hosts
[x
].read_ptr
+ ptr
->hosts
[x
].read_data_length
,
232 MEMCACHED_MAX_BUFFER
- ptr
->hosts
[x
].read_data_length
);
236 ptr
->hosts
[x
].read_buffer_length
+= data_read
;
237 ptr
->hosts
[x
].read_data_length
+= data_read
;
243 memcached_return_t
memcached_io_read(memcached_server_write_instance_st ptr
,
244 void *buffer
, size_t length
, ssize_t
*nread
)
252 if (!ptr
->read_buffer_length
)
258 data_read
= read(ptr
->fd
, ptr
->read_buffer
, MEMCACHED_MAX_BUFFER
);
261 else if (data_read
== -1)
263 ptr
->cached_errno
= errno
;
264 memcached_return_t rc
= MEMCACHED_UNKNOWN_READ_FAILURE
;
269 #ifdef TARGET_OS_LINUX
272 if ((rc
= io_wait(ptr
, MEM_READ
)) == MEMCACHED_SUCCESS
)
278 memcached_quit_server(ptr
, true);
287 EOF. Any data received so far is incomplete
288 so discard it. This always reads by byte in case of TCP
289 and protocol enforcement happens at memcached_response()
290 looking for '\n'. We do not care for UDB which requests 8 bytes
291 at once. Generally, this means that connection went away. Since
292 for blocking I/O we do not return 0 and for non-blocking case
293 it will return EGAIN if data is not immediatly available.
295 memcached_quit_server(ptr
, true);
297 return MEMCACHED_UNKNOWN_READ_FAILURE
;
301 ptr
->io_bytes_sent
= 0;
302 ptr
->read_data_length
= (size_t) data_read
;
303 ptr
->read_buffer_length
= (size_t) data_read
;
304 ptr
->read_ptr
= ptr
->read_buffer
;
311 difference
= (length
> ptr
->read_buffer_length
) ? ptr
->read_buffer_length
: length
;
313 memcpy(buffer_ptr
, ptr
->read_ptr
, difference
);
314 length
-= difference
;
315 ptr
->read_ptr
+= difference
;
316 ptr
->read_buffer_length
-= difference
;
317 buffer_ptr
+= difference
;
321 *buffer_ptr
= *ptr
->read_ptr
;
323 ptr
->read_buffer_length
--;
329 ptr
->server_failure_counter
= 0;
330 *nread
= (ssize_t
)(buffer_ptr
- (char*)buffer
);
331 return MEMCACHED_SUCCESS
;
334 static ssize_t
_io_write(memcached_server_write_instance_st ptr
,
335 const void *buffer
, size_t length
, bool with_flush
)
337 size_t original_length
;
338 const char* buffer_ptr
;
340 WATCHPOINT_ASSERT(ptr
->fd
!= -1);
342 original_length
= length
;
345 /* more writable data is coming if a flush isn't required, so delay send */
348 memcached_io_cork_push(ptr
);
357 if (ptr
->type
== MEMCACHED_CONNECTION_UDP
)
359 //UDP does not support partial writes
360 buffer_end
= MAX_UDP_DATAGRAM_LENGTH
;
361 should_write
= length
;
362 if (ptr
->write_buffer_offset
+ should_write
> buffer_end
)
367 buffer_end
= MEMCACHED_MAX_BUFFER
;
368 should_write
= buffer_end
- ptr
->write_buffer_offset
;
369 should_write
= (should_write
< length
) ? should_write
: length
;
372 write_ptr
= ptr
->write_buffer
+ ptr
->write_buffer_offset
;
373 memcpy(write_ptr
, buffer_ptr
, should_write
);
374 ptr
->write_buffer_offset
+= should_write
;
375 buffer_ptr
+= should_write
;
376 length
-= should_write
;
378 if (ptr
->write_buffer_offset
== buffer_end
&& ptr
->type
!= MEMCACHED_CONNECTION_UDP
)
380 memcached_return_t rc
;
383 WATCHPOINT_ASSERT(ptr
->fd
!= -1);
384 sent_length
= io_flush(ptr
, &rc
);
385 if (sent_length
== -1)
388 /* If io_flush calls memcached_purge, sent_length may be 0 */
389 unlikely (sent_length
!= 0)
391 WATCHPOINT_ASSERT(sent_length
== (ssize_t
)buffer_end
);
398 memcached_return_t rc
;
399 WATCHPOINT_ASSERT(ptr
->fd
!= -1);
400 if (io_flush(ptr
, &rc
) == -1)
405 memcached_io_cork_pop(ptr
);
408 return (ssize_t
) original_length
;
411 ssize_t
memcached_io_write(memcached_server_write_instance_st ptr
,
412 const void *buffer
, size_t length
, bool with_flush
)
414 return _io_write(ptr
, buffer
, length
, with_flush
);
417 ssize_t
memcached_io_writev(memcached_server_write_instance_st ptr
,
418 const struct __write_vector_st
*vector
,
419 size_t number_of
, bool with_flush
)
423 for (size_t x
= 0; x
< number_of
; x
++, vector
++)
427 if ((returnable
= _io_write(ptr
, vector
->buffer
, vector
->length
, false)) == -1)
436 if (memcached_io_write(ptr
, NULL
, 0, true) == -1)
446 memcached_return_t
memcached_io_close(memcached_server_write_instance_st ptr
)
450 return MEMCACHED_SUCCESS
;
453 /* in case of death shutdown to avoid blocking at close() */
454 if (shutdown(ptr
->fd
, SHUT_RDWR
) == -1 && errno
!= ENOTCONN
)
456 WATCHPOINT_NUMBER(ptr
->fd
);
457 WATCHPOINT_ERRNO(errno
);
458 WATCHPOINT_ASSERT(errno
);
461 if (close(ptr
->fd
) == -1)
463 WATCHPOINT_ERRNO(errno
);
466 return MEMCACHED_SUCCESS
;
469 memcached_server_write_instance_st
memcached_io_get_readable_server(memcached_st
*memc
)
471 #define MAX_SERVERS_TO_POLL 100
472 struct pollfd fds
[MAX_SERVERS_TO_POLL
];
473 unsigned int host_index
= 0;
476 x
< memcached_server_count(memc
) && host_index
< MAX_SERVERS_TO_POLL
;
479 memcached_server_write_instance_st instance
=
480 memcached_server_instance_fetch(memc
, x
);
482 if (instance
->read_buffer_length
> 0) /* I have data in the buffer */
485 if (memcached_server_response_count(instance
) > 0)
487 fds
[host_index
].events
= POLLIN
;
488 fds
[host_index
].revents
= 0;
489 fds
[host_index
].fd
= instance
->fd
;
496 /* We have 0 or 1 server with pending events.. */
497 for (uint32_t x
= 0; x
< memcached_server_count(memc
); ++x
)
499 memcached_server_write_instance_st instance
=
500 memcached_server_instance_fetch(memc
, x
);
502 if (memcached_server_response_count(instance
) > 0)
511 int err
= poll(fds
, host_index
, memc
->poll_timeout
);
514 memc
->cached_errno
= errno
;
519 for (size_t x
= 0; x
< host_index
; ++x
)
521 if (fds
[x
].revents
& POLLIN
)
523 for (uint32_t y
= 0; y
< memcached_server_count(memc
); ++y
)
525 memcached_server_write_instance_st instance
=
526 memcached_server_instance_fetch(memc
, y
);
528 if (instance
->fd
== fds
[x
].fd
)
538 static ssize_t
io_flush(memcached_server_write_instance_st ptr
,
539 memcached_return_t
*error
)
542 ** We might want to purge the input buffer if we haven't consumed
543 ** any output yet... The test for the limits is the purge is inline
544 ** in the purge function to avoid duplicating the logic..
547 memcached_return_t rc
;
548 WATCHPOINT_ASSERT(ptr
->fd
!= -1);
549 rc
= memcached_purge(ptr
);
551 if (rc
!= MEMCACHED_SUCCESS
&& rc
!= MEMCACHED_STORED
)
555 size_t return_length
;
556 char *local_write_ptr
= ptr
->write_buffer
;
557 size_t write_length
= ptr
->write_buffer_offset
;
559 *error
= MEMCACHED_SUCCESS
;
561 WATCHPOINT_ASSERT(ptr
->fd
!= -1);
563 // UDP Sanity check, make sure that we are not sending somthing too big
564 if (ptr
->type
== MEMCACHED_CONNECTION_UDP
&& write_length
> MAX_UDP_DATAGRAM_LENGTH
)
567 if (ptr
->write_buffer_offset
== 0 || (ptr
->type
== MEMCACHED_CONNECTION_UDP
568 && ptr
->write_buffer_offset
== UDP_DATAGRAM_HEADER_LENGTH
))
571 /* Looking for memory overflows */
573 if (write_length
== MEMCACHED_MAX_BUFFER
)
574 WATCHPOINT_ASSERT(ptr
->write_buffer
== local_write_ptr
);
575 WATCHPOINT_ASSERT((ptr
->write_buffer
+ MEMCACHED_MAX_BUFFER
) >= (local_write_ptr
+ write_length
));
581 WATCHPOINT_ASSERT(ptr
->fd
!= -1);
582 WATCHPOINT_ASSERT(write_length
> 0);
584 if (ptr
->type
== MEMCACHED_CONNECTION_UDP
)
585 increment_udp_message_id(ptr
);
586 sent_length
= write(ptr
->fd
, local_write_ptr
, write_length
);
588 if (sent_length
== -1)
590 ptr
->cached_errno
= errno
;
591 WATCHPOINT_ERRNO(errno
);
599 * We may be blocked on write because the input buffer
600 * is full. Let's check if we have room in our input
601 * buffer for more data and retry the write before
604 if (repack_input_buffer(ptr
) ||
605 process_input_buffer(ptr
))
608 memcached_return_t rc
;
609 rc
= io_wait(ptr
, MEM_WRITE
);
611 if (rc
== MEMCACHED_SUCCESS
|| rc
== MEMCACHED_TIMEOUT
)
614 memcached_quit_server(ptr
, true);
618 memcached_quit_server(ptr
, true);
619 *error
= MEMCACHED_ERRNO
;
624 if (ptr
->type
== MEMCACHED_CONNECTION_UDP
&&
625 (size_t)sent_length
!= write_length
)
627 memcached_quit_server(ptr
, true);
631 ptr
->io_bytes_sent
+= (uint32_t) sent_length
;
633 local_write_ptr
+= sent_length
;
634 write_length
-= (uint32_t) sent_length
;
635 return_length
+= (uint32_t) sent_length
;
638 WATCHPOINT_ASSERT(write_length
== 0);
639 // Need to study this assert() WATCHPOINT_ASSERT(return_length ==
640 // ptr->write_buffer_offset);
642 // if we are a udp server, the begining of the buffer is reserverd for
643 // the upd frame header
644 if (ptr
->type
== MEMCACHED_CONNECTION_UDP
)
645 ptr
->write_buffer_offset
= UDP_DATAGRAM_HEADER_LENGTH
;
647 ptr
->write_buffer_offset
= 0;
649 return (ssize_t
) return_length
;
653 Eventually we will just kill off the server with the problem.
655 void memcached_io_reset(memcached_server_write_instance_st ptr
)
657 memcached_quit_server(ptr
, true);
661 * Read a given number of bytes from the server and place it into a specific
662 * buffer. Reset the IO channel on this server if an error occurs.
664 memcached_return_t
memcached_safe_read(memcached_server_write_instance_st ptr
,
671 while (offset
< size
)
674 memcached_return_t rc
= memcached_io_read(ptr
, data
+ offset
, size
- offset
,
676 if (rc
!= MEMCACHED_SUCCESS
)
679 offset
+= (size_t) nread
;
682 return MEMCACHED_SUCCESS
;
685 memcached_return_t
memcached_io_readline(memcached_server_write_instance_st ptr
,
689 bool line_complete
= false;
692 while (!line_complete
)
694 if (ptr
->read_buffer_length
== 0)
697 * We don't have any data in the buffer, so let's fill the read
698 * buffer. Call the standard read function to avoid duplicating
702 memcached_return_t rc
= memcached_io_read(ptr
, buffer_ptr
, 1, &nread
);
703 if (rc
!= MEMCACHED_SUCCESS
)
706 if (*buffer_ptr
== '\n')
713 /* Now let's look in the buffer and copy as we go! */
714 while (ptr
->read_buffer_length
&& total_nr
< size
&& !line_complete
)
716 *buffer_ptr
= *ptr
->read_ptr
;
717 if (*buffer_ptr
== '\n')
718 line_complete
= true;
719 --ptr
->read_buffer_length
;
725 if (total_nr
== size
)
726 return MEMCACHED_PROTOCOL_ERROR
;
729 return MEMCACHED_SUCCESS
;
733 * The udp request id consists of two seperate sections
735 * 2) The message number
736 * The thread id should only be set when the memcached_st struct is created
737 * and should not be changed.
739 * The message num is incremented for each new message we send, this function
740 * extracts the message number from message_id, increments it and then
741 * writes the new value back into the header
743 static void increment_udp_message_id(memcached_server_write_instance_st ptr
)
745 struct udp_datagram_header_st
*header
= (struct udp_datagram_header_st
*)ptr
->write_buffer
;
746 uint16_t cur_req
= get_udp_datagram_request_id(header
);
747 int msg_num
= get_msg_num_from_request_id(cur_req
);
748 int thread_id
= get_thread_id_from_request_id(cur_req
);
750 if (((++msg_num
) & UDP_REQUEST_ID_THREAD_MASK
) != 0)
753 header
->request_id
= htons((uint16_t) (thread_id
| msg_num
));
756 memcached_return_t
memcached_io_init_udp_header(memcached_server_write_instance_st ptr
, uint16_t thread_id
)
758 if (thread_id
> UDP_REQUEST_ID_MAX_THREAD_ID
)
759 return MEMCACHED_FAILURE
;
761 struct udp_datagram_header_st
*header
= (struct udp_datagram_header_st
*)ptr
->write_buffer
;
762 header
->request_id
= htons((uint16_t) (generate_udp_request_thread_id(thread_id
)));
763 header
->num_datagrams
= htons(1);
764 header
->sequence_number
= htons(0);
766 return MEMCACHED_SUCCESS
;