diff -rupN xen-unstable.hg-3.0.5-pre-14797.orig/tools/blktap/drivers/block-aio.c xen-unstable.hg-3.0.5-pre-14797.new/tools/blktap/drivers/block-aio.c
--- xen-unstable.hg-3.0.5-pre-14797.orig/tools/blktap/drivers/block-aio.c 2007-04-11 19:10:30.000000000 -0400
+++ xen-unstable.hg-3.0.5-pre-14797.new/tools/blktap/drivers/block-aio.c 2007-04-13 11:17:08.000000000 -0400
@@ -44,14 +44,6 @@
#include <linux/fs.h>
#include "tapdisk.h"
-
-/**
- * We used a kernel patch to return an fd associated with the AIO context
- * so that we can concurrently poll on synchronous and async descriptors.
- * This is signalled by passing 1 as the io context to io_setup.
- */
-#define REQUEST_ASYNC_FD 1
-
#define MAX_AIO_REQS (MAX_REQUESTS * MAX_SEGMENTS_PER_REQ)
struct pending_aio {
@@ -72,12 +64,18 @@ struct tdaio_state {
int iocb_free_count;
struct iocb *iocb_queue[MAX_AIO_REQS];
int iocb_queued;
- int poll_fd; /* NB: we require aio_poll support */
struct io_event aio_events[MAX_AIO_REQS];
+
+ pthread_t aio_thread;
+ /* pipe fds for communication with the aio completion thread */
+ int command_fd[2];
+ int completion_fd[2];
};
#define IOCB_IDX(_s, _io) ((_io) - (_s)->iocb_list)
+static void *tdaio_completion_thread(void *);
+
/*Get Image size, secsize*/
static int get_image_info(struct td_state *s, int fd)
{
@@ -109,7 +107,6 @@ static int get_image_info(struct td_stat
/*Get the sector size*/
#if defined(BLKSSZGET)
{
- int arg;
s->sector_size = DEFAULT_SECTOR_SIZE;
ioctl(fd, BLKSSZGET, &s->sector_size);
@@ -148,7 +145,7 @@ static inline void init_fds(struct disk_
for(i = 0; i < MAX_IOFD; i++)
dd->io_fd[i] = 0;
- dd->io_fd[0] = prv->poll_fd;
+ dd->io_fd[0] = prv->completion_fd[0];
}
/* Open the disk file and initialize aio state. */
@@ -163,11 +160,10 @@ int tdaio_open (struct disk_driver *dd,
prv->iocb_free_count = MAX_AIO_REQS;
prv->iocb_queued = 0;
- prv->aio_ctx = (io_context_t) REQUEST_ASYNC_FD;
- prv->poll_fd = io_setup(MAX_AIO_REQS, &prv->aio_ctx);
+ prv->aio_ctx = (io_context_t) 0;
+ ret = io_setup(MAX_AIO_REQS, &prv->aio_ctx);
- if (prv->poll_fd < 0) {
- ret = prv->poll_fd;
+ if (ret < 0) {
if (ret == -EAGAIN) {
DPRINTF("Couldn't setup AIO context. If you are "
"trying to concurrently use a large number "
@@ -176,9 +172,7 @@ int tdaio_open (struct disk_driver *dd,
"(e.g. 'echo echo 1048576 > /proc/sys/fs/"
"aio-max-nr')\n");
} else {
- DPRINTF("Couldn't get fd for AIO poll support. This "
- "is probably because your kernel does not "
- "have the aio-poll patch applied.\n");
+ DPRINTF("Couldn't setup AIO context.\n");
}
goto done;
}
@@ -209,6 +203,14 @@ int tdaio_open (struct disk_driver *dd,
prv->fd = fd;
+ pipe(prv->command_fd);
+ pipe(prv->completion_fd);
+
+ ret = pthread_create(&prv->aio_thread, NULL,
+ tdaio_completion_thread, prv);
+ ret = 0;
+ write(prv->command_fd[1], &ret, sizeof(ret));
+
init_fds(dd);
ret = get_image_info(s, fd);
@@ -306,17 +308,62 @@ int tdaio_close(struct disk_driver *dd)
return 0;
}
+/*
+ * We don't have any way to do epoll on aio events in a normal kernel, so
+ * wait for aio events in a separate thread and return completion status
+ * that via a pipe that can be waited on normally.
+ *
+ * To keep locking problems between the completion thread and the submit
+ * thread to a minimum, there's a handshake which allows only one thread
+ * to be doing work on the completion queue at a time:
+ *
+ * 1) main thread sends completion thread a command via the command pipe;
+ * 2) completion thread waits for aio events and returns the number
+ * received on the completion pipe
+ * 3) main thread processes the received prv->aio_events events
+ * 4) loop back to 1) to let the completion thread refill the aio_events
+ * buffer.
+ *
+ * This workaround needs to disappear once the kernel provides a single
+ * mechanism for waiting on both aio and normal fd wakeups.
+ */
+
+static void *tdaio_completion_thread(void *arg)
+{
+ struct tdaio_state *prv = (struct tdaio_state *) arg;
+ int command;
+ int nr_events;
+ int rc;
+
+ while (1) {
+ rc = read(prv->command_fd[0], &command, sizeof(command));
+
+ do {
+ /* Non-blocking test for completed io. */
+ rc = io_getevents(prv->aio_ctx, 1,
+ MAX_AIO_REQS, prv->aio_events,
+ NULL);
+ if (rc) {
+ nr_events = rc;
+ rc = write(prv->completion_fd[1], &nr_events,
+ sizeof(nr_events));
+ }
+ } while (!rc);
+ }
+}
+
+
int tdaio_do_callbacks(struct disk_driver *dd, int sid)
{
- int ret, i, rsp = 0;
+ int ret, i, nr_events, rsp = 0;
struct io_event *ep;
struct tdaio_state *prv = (struct tdaio_state *)dd->private;
/* Non-blocking test for completed io. */
- ret = io_getevents(prv->aio_ctx, 0, MAX_AIO_REQS, prv->aio_events,
- NULL);
-
- for (ep=prv->aio_events,i=ret; i-->0; ep++) {
+ ret = read(prv->completion_fd[0], &nr_events, sizeof(nr_events));
+
+repeat:
+ for (ep=prv->aio_events,i=nr_events; i-->0; ep++) {
struct iocb *io = ep->obj;
struct pending_aio *pio;
@@ -327,6 +374,16 @@ int tdaio_do_callbacks(struct disk_drive
prv->iocb_free[prv->iocb_free_count++] = io;
}
+
+ if (nr_events) {
+ nr_events = io_getevents(prv->aio_ctx, 0,
+ MAX_AIO_REQS, prv->aio_events,
+ NULL);
+ goto repeat;
+ }
+
+ write(prv->command_fd[1], &nr_events, sizeof(nr_events));
+
return rsp;
}