#include "common.h"
-#include <sys/select.h>
-#include <poll.h>
typedef enum {
MEM_READ,
MEM_WRITE
} memc_read_or_write;
-static ssize_t io_flush(memcached_server_st *ptr, memcached_return_t *error);
-static void increment_udp_message_id(memcached_server_st *ptr);
+static ssize_t io_flush(memcached_server_write_instance_st ptr,
+ memcached_return_t *error);
+static void increment_udp_message_id(memcached_server_write_instance_st ptr);
-static memcached_return_t io_wait(memcached_server_st *ptr,
+static memcached_return_t io_wait(memcached_server_write_instance_st ptr,
memc_read_or_write read_or_write)
{
struct pollfd fds= {
};
int error;
- unlikely (read_or_write == MEM_WRITE) /* write */
+ if (read_or_write == MEM_WRITE) /* write */
+ {
fds.events= POLLOUT;
+ WATCHPOINT_SET(ptr->io_wait_count.write++);
+ }
+ else
+ {
+ WATCHPOINT_SET(ptr->io_wait_count.read++);
+ }
/*
** We are going to block on write, but at least on Solaris we might block
return MEMCACHED_FAILURE;
}
- int timeout= ptr->root->poll_timeout;
- if (ptr->root->flags.no_block == false)
- timeout= -1;
+ size_t loop_max= 5;
+ while (--loop_max) // While loop is for ERESTART or EINTR
+ {
+ error= poll(&fds, 1, ptr->root->poll_timeout);
- error= poll(&fds, 1, timeout);
+ switch (error)
+ {
+ case 1: // Success!
+ WATCHPOINT_IF_LABELED_NUMBER(read_or_write && loop_max < 4, "read() times we had to loop, decremented down from 5", loop_max);
+ WATCHPOINT_IF_LABELED_NUMBER(!read_or_write && loop_max < 4, "write() times we had to loop, decremented down from 5", loop_max);
+
+ return MEMCACHED_SUCCESS;
+ case 0: // Timeout occured, we let the while() loop do its thing.
+ return MEMCACHED_TIMEOUT;
+ default:
+ WATCHPOINT_ERRNO(get_socket_errno());
+ switch (get_socket_errno())
+ {
+#ifdef TARGET_OS_LINUX
+ case ERESTART:
+#endif
+ case EINTR:
+ break;
+ default:
+ if (fds.revents & POLLERR)
+ {
+ int err;
+ socklen_t len= sizeof (err);
+ (void)getsockopt(ptr->fd, SOL_SOCKET, SO_ERROR, &err, &len);
+ ptr->cached_errno= (err == 0) ? get_socket_errno() : err;
+ }
+ else
+ {
+ ptr->cached_errno= get_socket_errno();
+ }
+ memcached_quit_server(ptr, true);
- if (error == 1)
- return MEMCACHED_SUCCESS;
- else if (error == 0)
- return MEMCACHED_TIMEOUT;
+ return MEMCACHED_FAILURE;
+ }
+ }
+ }
/* Imposssible for anything other then -1 */
WATCHPOINT_ASSERT(error == -1);
- memcached_quit_server(ptr, 1);
+ ptr->cached_errno= get_socket_errno();
+ memcached_quit_server(ptr, true);
return MEMCACHED_FAILURE;
}
+memcached_return_t memcached_io_wait_for_write(memcached_server_write_instance_st ptr)
+{
+ return io_wait(ptr, MEM_WRITE);
+}
+
/**
* Try to fill the input buffer for a server with as much
* data as possible.
*
* @param ptr the server to pack
*/
-static bool repack_input_buffer(memcached_server_st *ptr)
+static bool repack_input_buffer(memcached_server_write_instance_st ptr)
{
if (ptr->read_ptr != ptr->read_buffer)
{
if (ptr->read_buffer_length != MEMCACHED_MAX_BUFFER)
{
/* Just try a single read to grab what's available */
- ssize_t nr= read(ptr->fd,
+ ssize_t nr= recv(ptr->fd,
ptr->read_ptr + ptr->read_data_length,
- MEMCACHED_MAX_BUFFER - ptr->read_data_length);
+ MEMCACHED_MAX_BUFFER - ptr->read_data_length,
+ 0);
if (nr > 0)
{
* @param ptr the server to star processing iput messages for
* @return true if we processed anything, false otherwise
*/
-static bool process_input_buffer(memcached_server_st *ptr)
+static bool process_input_buffer(memcached_server_write_instance_st ptr)
{
/*
** We might be able to process some of the response messages if we
*/
memcached_callback_st cb= *ptr->root->callbacks;
+ memcached_set_processing_input((memcached_st *)ptr->root, true);
+
char buffer[MEMCACHED_DEFAULT_COMMAND_SIZE];
memcached_return_t error;
+ memcached_st *root= (memcached_st *)ptr->root;
error= memcached_response(ptr, buffer, sizeof(buffer),
- &ptr->root->result);
+ &root->result);
+
+ memcached_set_processing_input(root, false);
+
if (error == MEMCACHED_SUCCESS)
{
for (unsigned int x= 0; x < cb.number_of_callback; x++)
{
- error= (*cb.callback[x])(ptr->root, &ptr->root->result, cb.context);
+ error= (*cb.callback[x])(ptr->root, &root->result, cb.context);
if (error != MEMCACHED_SUCCESS)
break;
}
return false;
}
-#ifdef UNUSED
+static inline void memcached_io_cork_push(memcached_server_st *ptr)
+{
+ (void)ptr;
+#ifdef CORK
+ if (ptr->root->flags.cork == false || ptr->state.is_corked)
+ return;
+
+ int enable= 1;
+ int err= setsockopt(ptr->fd, IPPROTO_TCP, CORK,
+ &enable, (socklen_t)sizeof(int));
+ if (! err)
+ ptr->state.is_corked= true;
+
+ WATCHPOINT_ASSERT(ptr->state.is_corked == true);
+#endif
+}
+
+static inline void memcached_io_cork_pop(memcached_server_st *ptr)
+{
+ (void)ptr;
+#ifdef CORK
+ if (ptr->root->flags.cork == false || ptr->state.is_corked == false)
+ return;
+
+ int enable= 0;
+ int err= setsockopt(ptr->fd, IPPROTO_TCP, CORK,
+ &enable, (socklen_t)sizeof(int));
+ if (! err)
+ ptr->state.is_corked= false;
+
+ WATCHPOINT_ASSERT(ptr->state.is_corked == false);
+#endif
+}
+
+#if 0 // Dead code, this should be removed.
void memcached_io_preread(memcached_st *ptr)
{
unsigned int x;
return;
- for (x= 0; x < ptr->number_of_hosts; x++)
+ for (x= 0; x < memcached_server_count(ptr); x++)
{
if (memcached_server_response_count(ptr, x) &&
ptr->hosts[x].read_data_length < MEMCACHED_MAX_BUFFER )
{
size_t data_read;
- data_read= read(ptr->hosts[x].fd,
+ data_read= recv(ptr->hosts[x].fd,
ptr->hosts[x].read_ptr + ptr->hosts[x].read_data_length,
- MEMCACHED_MAX_BUFFER - ptr->hosts[x].read_data_length);
- if (data_read == -1)
+ MEMCACHED_MAX_BUFFER - ptr->hosts[x].read_data_length, 0);
+ if (data_read == SOCKET_ERROR)
continue;
ptr->hosts[x].read_buffer_length+= data_read;
}
#endif
-memcached_return_t memcached_io_read(memcached_server_st *ptr,
+memcached_return_t memcached_io_read(memcached_server_write_instance_st ptr,
void *buffer, size_t length, ssize_t *nread)
{
char *buffer_ptr;
while (1)
{
- data_read= read(ptr->fd, ptr->read_buffer, MEMCACHED_MAX_BUFFER);
+ data_read= recv(ptr->fd, ptr->read_buffer, MEMCACHED_MAX_BUFFER, 0);
if (data_read > 0)
+ {
break;
- else if (data_read == -1)
+ }
+ else if (data_read == SOCKET_ERROR)
{
- ptr->cached_errno= errno;
- memcached_return_t rc= MEMCACHED_UNKNOWN_READ_FAILURE;
- switch (errno)
+ ptr->cached_errno= get_socket_errno();
+ memcached_return_t rc= MEMCACHED_ERRNO;
+ switch (get_socket_errno())
{
+ case EWOULDBLOCK:
+#ifdef USE_EAGAIN
case EAGAIN:
+#endif
case EINTR:
+#ifdef TARGET_OS_LINUX
+ case ERESTART:
+#endif
if ((rc= io_wait(ptr, MEM_READ)) == MEMCACHED_SUCCESS)
continue;
/* fall through */
default:
{
- memcached_quit_server(ptr, 1);
+ memcached_quit_server(ptr, true);
*nread= -1;
return rc;
}
for blocking I/O we do not return 0 and for non-blocking case
it will return EGAIN if data is not immediatly available.
*/
- memcached_quit_server(ptr, 1);
+ WATCHPOINT_STRING("We had a zero length recv()");
+ memcached_quit_server(ptr, true);
*nread= -1;
return MEMCACHED_UNKNOWN_READ_FAILURE;
}
return MEMCACHED_SUCCESS;
}
-ssize_t memcached_io_write(memcached_server_st *ptr,
- const void *buffer, size_t length, char with_flush)
+static ssize_t _io_write(memcached_server_write_instance_st ptr,
+ const void *buffer, size_t length, bool with_flush)
{
size_t original_length;
const char* buffer_ptr;
- WATCHPOINT_ASSERT(ptr->fd != -1);
+ WATCHPOINT_ASSERT(ptr->fd != INVALID_SOCKET);
original_length= length;
buffer_ptr= buffer;
+ /* more writable data is coming if a flush isn't required, so delay send */
+ if (! with_flush)
+ {
+ memcached_io_cork_push(ptr);
+ }
+
while (length)
{
char *write_ptr;
buffer_end= MAX_UDP_DATAGRAM_LENGTH;
should_write= length;
if (ptr->write_buffer_offset + should_write > buffer_end)
+ {
return -1;
+ }
}
else
{
memcached_return_t rc;
ssize_t sent_length;
- WATCHPOINT_ASSERT(ptr->fd != -1);
+ WATCHPOINT_ASSERT(ptr->fd != INVALID_SOCKET);
sent_length= io_flush(ptr, &rc);
if (sent_length == -1)
+ {
return -1;
+ }
/* If io_flush calls memcached_purge, sent_length may be 0 */
unlikely (sent_length != 0)
if (with_flush)
{
memcached_return_t rc;
- WATCHPOINT_ASSERT(ptr->fd != -1);
+ WATCHPOINT_ASSERT(ptr->fd != INVALID_SOCKET);
if (io_flush(ptr, &rc) == -1)
+ {
return -1;
+ }
+
+ memcached_io_cork_pop(ptr);
}
return (ssize_t) original_length;
}
-memcached_return_t memcached_io_close(memcached_server_st *ptr)
+ssize_t memcached_io_write(memcached_server_write_instance_st ptr,
+ const void *buffer, size_t length, bool with_flush)
{
- int r;
+ return _io_write(ptr, buffer, length, with_flush);
+}
- if (ptr->fd == -1)
- return MEMCACHED_SUCCESS;
+ssize_t memcached_io_writev(memcached_server_write_instance_st ptr,
+ const struct libmemcached_io_vector_st *vector,
+ size_t number_of, bool with_flush)
+{
+ ssize_t total= 0;
- /* in case of death shutdown to avoid blocking at close() */
- if (1)
+ for (size_t x= 0; x < number_of; x++, vector++)
{
- r= shutdown(ptr->fd, SHUT_RDWR);
+ ssize_t returnable;
-#ifdef DEBUG
- if (r && errno != ENOTCONN)
+ if ((returnable= _io_write(ptr, vector->buffer, vector->length, false)) == -1)
{
- WATCHPOINT_NUMBER(ptr->fd);
- WATCHPOINT_ERRNO(errno);
- WATCHPOINT_ASSERT(errno);
+ return -1;
}
-#endif
+ total+= returnable;
}
- r= close(ptr->fd);
-#ifdef DEBUG
- if (r != 0)
- WATCHPOINT_ERRNO(errno);
-#endif
+ if (with_flush)
+ {
+ if (memcached_io_write(ptr, NULL, 0, true) == -1)
+ {
+ return -1;
+ }
+ }
+
+ return total;
+}
+
+
+memcached_return_t memcached_io_close(memcached_server_write_instance_st ptr)
+{
+ if (ptr->fd == INVALID_SOCKET)
+ {
+ return MEMCACHED_SUCCESS;
+ }
+
+ /* in case of death shutdown to avoid blocking at close() */
+ if (shutdown(ptr->fd, SHUT_RDWR) == SOCKET_ERROR && get_socket_errno() != ENOTCONN)
+ {
+ WATCHPOINT_NUMBER(ptr->fd);
+ WATCHPOINT_ERRNO(get_socket_errno());
+ WATCHPOINT_ASSERT(get_socket_errno());
+ }
+
+ if (closesocket(ptr->fd) == SOCKET_ERROR)
+ {
+ WATCHPOINT_ERRNO(get_socket_errno());
+ }
return MEMCACHED_SUCCESS;
}
-memcached_server_st *memcached_io_get_readable_server(memcached_st *memc)
+memcached_server_write_instance_st memcached_io_get_readable_server(memcached_st *memc)
{
#define MAX_SERVERS_TO_POLL 100
struct pollfd fds[MAX_SERVERS_TO_POLL];
unsigned int host_index= 0;
- for (unsigned int x= 0;
- x< memc->number_of_hosts && host_index < MAX_SERVERS_TO_POLL;
+ for (uint32_t x= 0;
+ x< memcached_server_count(memc) && host_index < MAX_SERVERS_TO_POLL;
++x)
{
- if (memc->hosts[x].read_buffer_length > 0) /* I have data in the buffer */
- return &memc->hosts[x];
+ memcached_server_write_instance_st instance=
+ memcached_server_instance_fetch(memc, x);
+
+ if (instance->read_buffer_length > 0) /* I have data in the buffer */
+ return instance;
- if (memcached_server_response_count(&memc->hosts[x]) > 0)
+ if (memcached_server_response_count(instance) > 0)
{
fds[host_index].events = POLLIN;
fds[host_index].revents = 0;
- fds[host_index].fd = memc->hosts[x].fd;
+ fds[host_index].fd = instance->fd;
++host_index;
}
}
if (host_index < 2)
{
/* We have 0 or 1 server with pending events.. */
- for (unsigned int x= 0; x< memc->number_of_hosts; ++x)
- if (memcached_server_response_count(&memc->hosts[x]) > 0)
- return &memc->hosts[x];
+ for (uint32_t x= 0; x< memcached_server_count(memc); ++x)
+ {
+ memcached_server_write_instance_st instance=
+ memcached_server_instance_fetch(memc, x);
+
+ if (memcached_server_response_count(instance) > 0)
+ {
+ return instance;
+ }
+ }
return NULL;
}
int err= poll(fds, host_index, memc->poll_timeout);
switch (err) {
case -1:
- memc->cached_errno = errno;
+ memc->cached_errno = get_socket_errno();
/* FALLTHROUGH */
case 0:
break;
default:
- for (unsigned int x= 0; x < host_index; ++x)
+ for (size_t x= 0; x < host_index; ++x)
+ {
if (fds[x].revents & POLLIN)
- for (unsigned int y= 0; y < memc->number_of_hosts; ++y)
- if (memc->hosts[y].fd == fds[x].fd)
- return &memc->hosts[y];
+ {
+ for (uint32_t y= 0; y < memcached_server_count(memc); ++y)
+ {
+ memcached_server_write_instance_st instance=
+ memcached_server_instance_fetch(memc, y);
+
+ if (instance->fd == fds[x].fd)
+ return instance;
+ }
+ }
+ }
}
return NULL;
}
-static ssize_t io_flush(memcached_server_st *ptr,
+static ssize_t io_flush(memcached_server_write_instance_st ptr,
memcached_return_t *error)
{
/*
*/
{
memcached_return_t rc;
- WATCHPOINT_ASSERT(ptr->fd != -1);
+ WATCHPOINT_ASSERT(ptr->fd != INVALID_SOCKET);
rc= memcached_purge(ptr);
if (rc != MEMCACHED_SUCCESS && rc != MEMCACHED_STORED)
+ {
return -1;
+ }
}
ssize_t sent_length;
size_t return_length;
*error= MEMCACHED_SUCCESS;
- WATCHPOINT_ASSERT(ptr->fd != -1);
+ WATCHPOINT_ASSERT(ptr->fd != INVALID_SOCKET);
// UDP Sanity check, make sure that we are not sending somthing too big
if (ptr->type == MEMCACHED_CONNECTION_UDP && write_length > MAX_UDP_DATAGRAM_LENGTH)
+ {
return -1;
+ }
if (ptr->write_buffer_offset == 0 || (ptr->type == MEMCACHED_CONNECTION_UDP
&& ptr->write_buffer_offset == UDP_DATAGRAM_HEADER_LENGTH))
return_length= 0;
while (write_length)
{
- WATCHPOINT_ASSERT(ptr->fd != -1);
+ WATCHPOINT_ASSERT(ptr->fd != INVALID_SOCKET);
WATCHPOINT_ASSERT(write_length > 0);
sent_length= 0;
if (ptr->type == MEMCACHED_CONNECTION_UDP)
increment_udp_message_id(ptr);
- sent_length= write(ptr->fd, local_write_ptr, write_length);
- if (sent_length == -1)
+ WATCHPOINT_ASSERT(ptr->fd != -1);
+ sent_length= send(ptr->fd, local_write_ptr, write_length, MSG_NOSIGNAL|MSG_DONTWAIT);
+ if (sent_length == SOCKET_ERROR)
{
- ptr->cached_errno= errno;
- switch (errno)
+ ptr->cached_errno= get_socket_errno();
+#if 0 // @todo I should look at why we hit this bit of code hard frequently
+ WATCHPOINT_ERRNO(get_socket_errno());
+ WATCHPOINT_NUMBER(get_socket_errno());
+#endif
+ switch (get_socket_errno())
{
case ENOBUFS:
continue;
+ case EWOULDBLOCK:
+#ifdef USE_EAGAIN
case EAGAIN:
+#endif
{
/*
* We may be blocked on write because the input buffer
if (rc == MEMCACHED_SUCCESS || rc == MEMCACHED_TIMEOUT)
continue;
- memcached_quit_server(ptr, 1);
+ memcached_quit_server(ptr, true);
return -1;
}
+ case ENOTCONN:
+ case EPIPE:
default:
- memcached_quit_server(ptr, 1);
+ memcached_quit_server(ptr, true);
*error= MEMCACHED_ERRNO;
+ WATCHPOINT_ASSERT(ptr->fd == -1);
return -1;
}
}
if (ptr->type == MEMCACHED_CONNECTION_UDP &&
(size_t)sent_length != write_length)
{
- memcached_quit_server(ptr, 1);
+ memcached_quit_server(ptr, true);
return -1;
}
/*
Eventually we will just kill off the server with the problem.
*/
-void memcached_io_reset(memcached_server_st *ptr)
+void memcached_io_reset(memcached_server_write_instance_st ptr)
{
- memcached_quit_server(ptr, 1);
+ memcached_quit_server(ptr, true);
}
/**
* Read a given number of bytes from the server and place it into a specific
* buffer. Reset the IO channel on this server if an error occurs.
*/
-memcached_return_t memcached_safe_read(memcached_server_st *ptr,
+memcached_return_t memcached_safe_read(memcached_server_write_instance_st ptr,
void *dta,
size_t size)
{
return MEMCACHED_SUCCESS;
}
-memcached_return_t memcached_io_readline(memcached_server_st *ptr,
+memcached_return_t memcached_io_readline(memcached_server_write_instance_st ptr,
char *buffer_ptr,
size_t size)
{
* extracts the message number from message_id, increments it and then
* writes the new value back into the header
*/
-static void increment_udp_message_id(memcached_server_st *ptr)
+static void increment_udp_message_id(memcached_server_write_instance_st ptr)
{
struct udp_datagram_header_st *header= (struct udp_datagram_header_st *)ptr->write_buffer;
uint16_t cur_req= get_udp_datagram_request_id(header);
header->request_id= htons((uint16_t) (thread_id | msg_num));
}
-memcached_return_t memcached_io_init_udp_header(memcached_server_st *ptr, uint16_t thread_id)
+memcached_return_t memcached_io_init_udp_header(memcached_server_write_instance_st ptr, uint16_t thread_id)
{
if (thread_id > UDP_REQUEST_ID_MAX_THREAD_ID)
return MEMCACHED_FAILURE;