diff options
author | Andrew Tridgell <tridge@samba.org> | 2006-12-11 08:53:29 +0000 |
---|---|---|
committer | Gerald (Jerry) Carter <jerry@samba.org> | 2007-10-10 14:29:06 -0500 |
commit | 9ca93ced8bc719fbfc444c472cf0d058409cbce1 (patch) | |
tree | 59fbb331c5ca426383bb036a7a4732afe6352676 /source4 | |
parent | 47e03d96a263a7f9a5e3061c04147e559f2fe9e0 (diff) | |
download | samba-9ca93ced8bc719fbfc444c472cf0d058409cbce1.tar.gz samba-9ca93ced8bc719fbfc444c472cf0d058409cbce1.tar.bz2 samba-9ca93ced8bc719fbfc444c472cf0d058409cbce1.zip |
r20104: this is a alternative events backend, which uses a hybrid of aio and
epoll. It is not linked in anywhere yet - I'm committing it in case
anyone else wants to have a look at it.
The concept is quite strange really, but it seems to be the only way
that Linux 2.6.x can currently use a unified event model allowing for
AIO events and socket events to be waited for by a single unified
event wait function. You setup a epoll system, then setup a weird aio
event that points at the epoll system, then use io_getevents() to
actually do the waiting.
I'm hoping that kevents or a proper integration of epoll will allow us
to avoid ths rather hackish scheme, but meanwhile this is the only
path to proper AIO in Samba on Linux (without a horrible signals mess)
(NOTE: this code requires some kernel patches to work at the moment)
(This used to be commit 195051fdee341e8d8cb76e5c91dcc0f6c246a870)
Diffstat (limited to 'source4')
-rw-r--r-- | source4/lib/events/events_aio.c | 518 |
1 files changed, 518 insertions, 0 deletions
diff --git a/source4/lib/events/events_aio.c b/source4/lib/events/events_aio.c new file mode 100644 index 0000000000..97305fdeb3 --- /dev/null +++ b/source4/lib/events/events_aio.c @@ -0,0 +1,518 @@ +/* + Unix SMB/CIFS implementation. + + main select loop and event handling - aio/epoll hybrid implementation + + Copyright (C) Andrew Tridgell 2006 + + based on events_standard.c + + This program is free software; you can redistribute it and/or modify + it under the terms of the GNU General Public License as published by + the Free Software Foundation; either version 2 of the License, or + (at your option) any later version. + + This program is distributed in the hope that it will be useful, + but WITHOUT ANY WARRANTY; without even the implied warranty of + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the + GNU General Public License for more details. + + You should have received a copy of the GNU General Public License + along with this program; if not, write to the Free Software + Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. +*/ +/* + this is a very strange beast. The Linux AIO implementation doesn't + yet integrate properly with epoll, but there is a kernel patch that + allows the aio wait primitives to be used to wait for epoll events, + and this can be used to give us a unified event system incorporating + both aio events and epoll events + + this is _very_ experimental code +*/ + +#include "includes.h" +#include "system/filesys.h" +#include "system/select.h" /* needed for WITH_EPOLL */ +#include "lib/util/dlinklist.h" +#include "lib/events/events.h" +#include "lib/events/events_internal.h" +#include <libaio.h> + +#define MAX_AIO_QUEUE_DEPTH 10 +#define IOCB_CMD_EPOLL_WAIT 9 + +struct aio_event_context { + /* a pointer back to the generic event_context */ + struct event_context *ev; + + /* number of registered fd event handlers */ + int num_fd_events; + + /* list of timed events */ + struct timed_event *timed_events; + + uint32_t destruction_count; + + io_context_t ioctx; + + struct io_event events[MAX_AIO_QUEUE_DEPTH]; + struct epoll_event epevent; + + struct iocb *epoll_iocb; + + int epoll_fd; +}; + +static void aio_event_loop_timer(struct aio_event_context *aio_ev); + +/* + map from EVENT_FD_* to EPOLLIN/EPOLLOUT +*/ +static uint32_t epoll_map_flags(uint16_t flags) +{ + uint32_t ret = 0; + if (flags & EVENT_FD_READ) ret |= (EPOLLIN | EPOLLERR | EPOLLHUP); + if (flags & EVENT_FD_WRITE) ret |= (EPOLLOUT | EPOLLERR | EPOLLHUP); + return ret; +} + +/* + free the epoll fd +*/ +static int aio_ctx_destructor(struct aio_event_context *aio_ev) +{ + close(aio_ev->epoll_fd); + aio_ev->epoll_fd = -1; + return 0; +} + +#define EPOLL_ADDITIONAL_FD_FLAG_HAS_EVENT (1<<0) +#define EPOLL_ADDITIONAL_FD_FLAG_REPORT_ERROR (1<<1) +#define EPOLL_ADDITIONAL_FD_FLAG_GOT_ERROR (1<<2) + +/* + add the epoll event to the given fd_event +*/ +static void epoll_add_event(struct aio_event_context *aio_ev, struct fd_event *fde) +{ + struct epoll_event event; + if (aio_ev->epoll_fd == -1) return; + + fde->additional_flags &= ~EPOLL_ADDITIONAL_FD_FLAG_REPORT_ERROR; + + /* if we don't want events yet, don't add an aio_event */ + if (fde->flags == 0) return; + + ZERO_STRUCT(event); + event.events = epoll_map_flags(fde->flags); + event.data.ptr = fde; + epoll_ctl(aio_ev->epoll_fd, EPOLL_CTL_ADD, fde->fd, &event); + fde->additional_flags |= EPOLL_ADDITIONAL_FD_FLAG_HAS_EVENT; + + /* only if we want to read we want to tell the event handler about errors */ + if (fde->flags & EVENT_FD_READ) { + fde->additional_flags |= EPOLL_ADDITIONAL_FD_FLAG_REPORT_ERROR; + } +} + +/* + delete the epoll event for given fd_event +*/ +static void epoll_del_event(struct aio_event_context *aio_ev, struct fd_event *fde) +{ + struct epoll_event event; + if (aio_ev->epoll_fd == -1) return; + + fde->additional_flags &= ~EPOLL_ADDITIONAL_FD_FLAG_REPORT_ERROR; + + /* if there's no aio_event, we don't need to delete it */ + if (!(fde->additional_flags & EPOLL_ADDITIONAL_FD_FLAG_HAS_EVENT)) return; + + ZERO_STRUCT(event); + event.events = epoll_map_flags(fde->flags); + event.data.ptr = fde; + epoll_ctl(aio_ev->epoll_fd, EPOLL_CTL_DEL, fde->fd, &event); + + fde->additional_flags &= ~EPOLL_ADDITIONAL_FD_FLAG_HAS_EVENT; +} + +/* + change the epoll event to the given fd_event +*/ +static void epoll_mod_event(struct aio_event_context *aio_ev, struct fd_event *fde) +{ + struct epoll_event event; + if (aio_ev->epoll_fd == -1) return; + + fde->additional_flags &= ~EPOLL_ADDITIONAL_FD_FLAG_REPORT_ERROR; + + ZERO_STRUCT(event); + event.events = epoll_map_flags(fde->flags); + event.data.ptr = fde; + epoll_ctl(aio_ev->epoll_fd, EPOLL_CTL_MOD, fde->fd, &event); + + /* only if we want to read we want to tell the event handler about errors */ + if (fde->flags & EVENT_FD_READ) { + fde->additional_flags |= EPOLL_ADDITIONAL_FD_FLAG_REPORT_ERROR; + } +} + +static void epoll_change_event(struct aio_event_context *aio_ev, struct fd_event *fde) +{ + BOOL got_error = (fde->additional_flags & EPOLL_ADDITIONAL_FD_FLAG_GOT_ERROR); + BOOL want_read = (fde->flags & EVENT_FD_READ); + BOOL want_write= (fde->flags & EVENT_FD_WRITE); + + if (aio_ev->epoll_fd == -1) return; + + fde->additional_flags &= ~EPOLL_ADDITIONAL_FD_FLAG_REPORT_ERROR; + + /* there's already an event */ + if (fde->additional_flags & EPOLL_ADDITIONAL_FD_FLAG_HAS_EVENT) { + if (want_read || (want_write && !got_error)) { + epoll_mod_event(aio_ev, fde); + return; + } + epoll_del_event(aio_ev, fde); + return; + } + + /* there's no aio_event attached to the fde */ + if (want_read || (want_write && !got_error)) { + epoll_add_event(aio_ev, fde); + return; + } +} + +static int setup_epoll_wait(struct aio_event_context *aio_ev) +{ + struct io_event r; + + memset(aio_ev->epoll_iocb, 0, sizeof(*aio_ev->epoll_iocb)); + aio_ev->epoll_iocb->aio_fildes = aio_ev->epoll_fd; + aio_ev->epoll_iocb->aio_lio_opcode = IOCB_CMD_EPOLL_WAIT; + aio_ev->epoll_iocb->aio_reqprio = 0; + + aio_ev->epoll_iocb->u.c.nbytes = 1; + aio_ev->epoll_iocb->u.c.offset = -1; + aio_ev->epoll_iocb->u.c.buf = &aio_ev->epevent; + + if (io_submit(aio_ev->ioctx, 1, &aio_ev->epoll_iocb) != 1) { + return -1; + } +} + + +/* + event loop handling using aio/epoll hybrid +*/ +static int aio_event_loop(struct aio_event_context *aio_ev, struct timeval *tvalp) +{ + int ret, i; + uint32_t destruction_count = aio_ev->destruction_count; + struct timespec timeout; + + if (aio_ev->epoll_fd == -1) return -1; + + if (tvalp) { + timeout.tv_sec = tvalp->tv_sec; + timeout.tv_nsec = tvalp->tv_usec; + timeout.tv_nsec *= 1000; + } + + setup_epoll_wait(aio_ev); + + ret = io_getevents(aio_ev->ioctx, 1, MAX_AIO_QUEUE_DEPTH, + aio_ev->events, tvalp?&timeout:NULL); + if (ret == -EINTR) { + return 0; + } + + if (ret == 0 && tvalp) { + aio_event_loop_timer(aio_ev); + return 0; + } + + for (i=0;i<ret;i++) { + struct iocb *finished = aio_ev->events[i].obj; + switch (finished->aio_lio_opcode) { + case IOCB_CMD_EPOLL_WAIT: { + struct epoll_event *ep = (struct epoll_event *)finished->u.c.buf; + struct fd_event *fde = talloc_get_type(ep->data.ptr, + struct fd_event); + uint16_t flags = 0; + + if (fde == NULL) { + return -1; + } + if (ep->events & (EPOLLHUP|EPOLLERR)) { + fde->additional_flags |= EPOLL_ADDITIONAL_FD_FLAG_GOT_ERROR; + if (!(fde->additional_flags & EPOLL_ADDITIONAL_FD_FLAG_REPORT_ERROR)) { + epoll_del_event(aio_ev, fde); + continue; + } + flags |= EVENT_FD_READ; + } + if (ep->events & EPOLLIN) flags |= EVENT_FD_READ; + if (ep->events & EPOLLOUT) flags |= EVENT_FD_WRITE; + if (flags) { + fde->handler(aio_ev->ev, fde, flags, fde->private_data); + if (destruction_count != aio_ev->destruction_count) { + return 0; + } + } + break; + } + } + } + + return 0; +} + +/* + create a aio_event_context structure. +*/ +static int aio_event_context_init(struct event_context *ev, void *private_data) +{ + struct aio_event_context *aio_ev; + + aio_ev = talloc_zero(ev, struct aio_event_context); + if (!aio_ev) return -1; + + aio_ev->ev = ev; + aio_ev->epoll_iocb = talloc(aio_ev, struct iocb); + + if (io_queue_init(MAX_AIO_QUEUE_DEPTH, &aio_ev->ioctx) != 0) { + return -1; + } + + aio_ev->epoll_fd = epoll_create(MAX_AIO_QUEUE_DEPTH); + if (aio_ev->epoll_fd == -1) return -1; + + talloc_set_destructor(aio_ev, aio_ctx_destructor); + + ev->additional_data = aio_ev; + return 0; +} + +/* + destroy an fd_event +*/ +static int aio_event_fd_destructor(struct fd_event *fde) +{ + struct event_context *ev = fde->event_ctx; + struct aio_event_context *aio_ev = talloc_get_type(ev->additional_data, + struct aio_event_context); + + aio_ev->num_fd_events--; + aio_ev->destruction_count++; + + epoll_del_event(aio_ev, fde); + + return 0; +} + +/* + add a fd based event + return NULL on failure (memory allocation error) +*/ +static struct fd_event *aio_event_add_fd(struct event_context *ev, TALLOC_CTX *mem_ctx, + int fd, uint16_t flags, + event_fd_handler_t handler, + void *private_data) +{ + struct aio_event_context *aio_ev = talloc_get_type(ev->additional_data, + struct aio_event_context); + struct fd_event *fde; + + fde = talloc(mem_ctx?mem_ctx:ev, struct fd_event); + if (!fde) return NULL; + + fde->event_ctx = ev; + fde->fd = fd; + fde->flags = flags; + fde->handler = handler; + fde->private_data = private_data; + fde->additional_flags = 0; + fde->additional_data = NULL; + + aio_ev->num_fd_events++; + talloc_set_destructor(fde, aio_event_fd_destructor); + + epoll_add_event(aio_ev, fde); + + return fde; +} + + +/* + return the fd event flags +*/ +static uint16_t aio_event_get_fd_flags(struct fd_event *fde) +{ + return fde->flags; +} + +/* + set the fd event flags +*/ +static void aio_event_set_fd_flags(struct fd_event *fde, uint16_t flags) +{ + struct event_context *ev; + struct aio_event_context *aio_ev; + + if (fde->flags == flags) return; + + ev = fde->event_ctx; + aio_ev = talloc_get_type(ev->additional_data, struct aio_event_context); + + fde->flags = flags; + + epoll_change_event(aio_ev, fde); +} + +/* + destroy a timed event +*/ +static int aio_event_timed_destructor(struct timed_event *te) +{ + struct aio_event_context *aio_ev = talloc_get_type(te->event_ctx->additional_data, + struct aio_event_context); + DLIST_REMOVE(aio_ev->timed_events, te); + return 0; +} + +static int aio_event_timed_deny_destructor(struct timed_event *te) +{ + return -1; +} + +/* + add a timed event + return NULL on failure (memory allocation error) +*/ +static struct timed_event *aio_event_add_timed(struct event_context *ev, TALLOC_CTX *mem_ctx, + struct timeval next_event, + event_timed_handler_t handler, + void *private_data) +{ + struct aio_event_context *aio_ev = talloc_get_type(ev->additional_data, + struct aio_event_context); + struct timed_event *te, *last_te, *cur_te; + + te = talloc(mem_ctx?mem_ctx:ev, struct timed_event); + if (te == NULL) return NULL; + + te->event_ctx = ev; + te->next_event = next_event; + te->handler = handler; + te->private_data = private_data; + te->additional_data = NULL; + + /* keep the list ordered */ + last_te = NULL; + for (cur_te = aio_ev->timed_events; cur_te; cur_te = cur_te->next) { + /* if the new event comes before the current one break */ + if (!timeval_is_zero(&cur_te->next_event) && + timeval_compare(&te->next_event, + &cur_te->next_event) < 0) { + break; + } + + last_te = cur_te; + } + + DLIST_ADD_AFTER(aio_ev->timed_events, te, last_te); + + talloc_set_destructor(te, aio_event_timed_destructor); + + return te; +} + +/* + a timer has gone off - call it +*/ +static void aio_event_loop_timer(struct aio_event_context *aio_ev) +{ + struct timeval t = timeval_current(); + struct timed_event *te = aio_ev->timed_events; + + if (te == NULL) { + return; + } + + /* deny the handler to free the event */ + talloc_set_destructor(te, aio_event_timed_deny_destructor); + + /* We need to remove the timer from the list before calling the + * handler because in a semi-async inner event loop called from the + * handler we don't want to come across this event again -- vl */ + DLIST_REMOVE(aio_ev->timed_events, te); + + te->handler(aio_ev->ev, te, t, te->private_data); + + /* The destructor isn't necessary anymore, we've already removed the + * event from the list. */ + talloc_set_destructor(te, NULL); + + talloc_free(te); +} + +/* + do a single event loop using the events defined in ev +*/ +static int aio_event_loop_once(struct event_context *ev) +{ + struct aio_event_context *aio_ev = talloc_get_type(ev->additional_data, + struct aio_event_context); + struct timeval tval; + + /* work out the right timeout for all timed events */ + if (aio_ev->timed_events) { + struct timeval t = timeval_current(); + tval = timeval_until(&t, &aio_ev->timed_events->next_event); + if (timeval_is_zero(&tval)) { + aio_event_loop_timer(aio_ev); + return 0; + } + } else { + /* have a default tick time of 30 seconds. This guarantees + that code that uses its own timeout checking will be + able to proceeed eventually */ + tval = timeval_set(30, 0); + } + + return aio_event_loop(aio_ev, &tval); +} + +/* + return on failure or (with 0) if all fd events are removed +*/ +static int aio_event_loop_wait(struct event_context *ev) +{ + struct aio_event_context *aio_ev = talloc_get_type(ev->additional_data, + struct aio_event_context); + while (aio_ev->num_fd_events) { + if (aio_event_loop_once(ev) != 0) { + break; + } + } + + return 0; +} + +static const struct event_ops aio_event_ops = { + .context_init = aio_event_context_init, + .add_fd = aio_event_add_fd, + .get_fd_flags = aio_event_get_fd_flags, + .set_fd_flags = aio_event_set_fd_flags, + .add_timed = aio_event_add_timed, + .loop_once = aio_event_loop_once, + .loop_wait = aio_event_loop_wait, +}; + +const struct event_ops *event_aio_get_ops(void) +{ + return &aio_event_ops; +} |