[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
[PATCH 16/16] block/io_uring: adds fd registration
From: |
Stefan Hajnoczi |
Subject: |
[PATCH 16/16] block/io_uring: adds fd registration |
Date: |
Mon, 7 Oct 2019 13:09:37 +0100 |
From: Aarushi Mehta <address@hidden>
File descriptor registration eliminates the need to get file descriptors
for each request on the host kernel side. The host kernel keeps
references to a set of files that were registered by the application.
These files can be quickly accessed by index instead of a traditional
file descriptor number.
Signed-off-by: Aarushi Mehta <address@hidden>
Signed-off-by: Stefan Hajnoczi <address@hidden>
---
v10:
* Add io_uring_register_files() return value to trace_luring_fd_register()
* Fix indentation in luring_fd_unregister()
* Set s->fd_reg.fd_array to NULL after g_free() to avoid dangling pointers
* Simplify fd registration code
* Add luring_fd_unregister() and call it from file-posix.c to prevent
fd leaks
* Add trace_luring_fd_unregister() trace event
---
include/block/raw-aio.h | 1 +
block/file-posix.c | 18 +++++
block/io_uring.c | 150 +++++++++++++++++++++++++++++++++++++++-
block/trace-events | 2 +
4 files changed, 170 insertions(+), 1 deletion(-)
diff --git a/include/block/raw-aio.h b/include/block/raw-aio.h
index 251b10d273..5fe6481bc4 100644
--- a/include/block/raw-aio.h
+++ b/include/block/raw-aio.h
@@ -68,6 +68,7 @@ void luring_detach_aio_context(LuringState *s, AioContext
*old_context);
void luring_attach_aio_context(LuringState *s, AioContext *new_context);
void luring_io_plug(BlockDriverState *bs, LuringState *s);
void luring_io_unplug(BlockDriverState *bs, LuringState *s);
+void luring_fd_unregister(LuringState *s, int fd);
#endif
#ifdef _WIN32
diff --git a/block/file-posix.c b/block/file-posix.c
index 264c44fc3f..604b5427c3 100644
--- a/block/file-posix.c
+++ b/block/file-posix.c
@@ -2010,6 +2010,20 @@ static int raw_co_flush_to_disk(BlockDriverState *bs)
return raw_thread_pool_submit(bs, handle_aiocb_flush, &acb);
}
+static void raw_aio_detach_aio_context(BlockDriverState *bs)
+{
+#ifdef CONFIG_LINUX_IO_URING
+ BDRVRawState *s = bs->opaque;
+
+ if (s->fd >= 0 && s->use_linux_io_uring) {
+ LuringState *luring;
+
+ luring = aio_get_linux_io_uring(bdrv_get_aio_context(bs));
+ luring_fd_unregister(luring, s->fd);
+ }
+#endif
+}
+
static void raw_aio_attach_aio_context(BlockDriverState *bs,
AioContext *new_context)
{
@@ -2994,6 +3008,7 @@ BlockDriver bdrv_file = {
.bdrv_refresh_limits = raw_refresh_limits,
.bdrv_io_plug = raw_aio_plug,
.bdrv_io_unplug = raw_aio_unplug,
+ .bdrv_detach_aio_context = raw_aio_detach_aio_context,
.bdrv_attach_aio_context = raw_aio_attach_aio_context,
.bdrv_co_truncate = raw_co_truncate,
@@ -3470,6 +3485,7 @@ static BlockDriver bdrv_host_device = {
.bdrv_refresh_limits = raw_refresh_limits,
.bdrv_io_plug = raw_aio_plug,
.bdrv_io_unplug = raw_aio_unplug,
+ .bdrv_detach_aio_context = raw_aio_detach_aio_context,
.bdrv_attach_aio_context = raw_aio_attach_aio_context,
.bdrv_co_truncate = raw_co_truncate,
@@ -3594,6 +3610,7 @@ static BlockDriver bdrv_host_cdrom = {
.bdrv_refresh_limits = raw_refresh_limits,
.bdrv_io_plug = raw_aio_plug,
.bdrv_io_unplug = raw_aio_unplug,
+ .bdrv_detach_aio_context = raw_aio_detach_aio_context,
.bdrv_attach_aio_context = raw_aio_attach_aio_context,
.bdrv_co_truncate = raw_co_truncate,
@@ -3726,6 +3743,7 @@ static BlockDriver bdrv_host_cdrom = {
.bdrv_refresh_limits = raw_refresh_limits,
.bdrv_io_plug = raw_aio_plug,
.bdrv_io_unplug = raw_aio_unplug,
+ .bdrv_detach_aio_context = raw_aio_detach_aio_context,
.bdrv_attach_aio_context = raw_aio_attach_aio_context,
.bdrv_co_truncate = raw_co_truncate,
diff --git a/block/io_uring.c b/block/io_uring.c
index 56892fd1ab..4d6e50ce19 100644
--- a/block/io_uring.c
+++ b/block/io_uring.c
@@ -46,10 +46,16 @@ typedef struct LuringQueue {
QSIMPLEQ_HEAD(, LuringAIOCB) submit_queue;
} LuringQueue;
+typedef struct LuringFd {
+ int *fd_array;
+ GHashTable *fd_lookup;
+} LuringFd;
+
typedef struct LuringState {
AioContext *aio_context;
struct io_uring ring;
+ LuringFd fd_reg;
/* io queue for submit at batch. Protected by AioContext lock. */
LuringQueue io_q;
@@ -298,6 +304,136 @@ static void ioq_init(LuringQueue *io_q)
io_q->blocked = false;
}
+/**
+ * luring_fd_unregister:
+ *
+ * Remove a file descriptor from the registered fds list. This is a slow
+ * operation because all registered fds are refreshed, but this function should
+ * not be called often anyway.
+ *
+ * Only call this function while there are no requests in flight.
+ */
+void luring_fd_unregister(LuringState *s, int fd)
+{
+ LuringFd *fd_reg = &s->fd_reg;
+ void *value;
+ int idx;
+ int ret;
+ int nr;
+ int i;
+
+ if (!g_hash_table_lookup_extended(fd_reg->fd_lookup, GINT_TO_POINTER(fd),
+ NULL, &value)) {
+ return;
+ }
+
+ ret = io_uring_unregister_files(&s->ring);
+ if (ret < 0) {
+ return;
+ }
+
+ idx = GPOINTER_TO_INT(value);
+ nr = g_hash_table_size(fd_reg->fd_lookup) - 1; /* minus this fd */
+
+ trace_luring_fd_unregister(s, fd, idx);
+
+ memmove(&fd_reg->fd_array[idx], &fd_reg->fd_array[idx + 1], nr - idx);
+
+ /* Rebuild hash table */
+ g_hash_table_remove_all(fd_reg->fd_lookup);
+ for (i = 0; i < nr; i++) {
+ g_hash_table_insert(fd_reg->fd_lookup,
+ GINT_TO_POINTER(fd_reg->fd_array[i]),
+ GINT_TO_POINTER(i));
+ }
+
+ io_uring_register_files(&s->ring, fd_reg->fd_array, nr);
+}
+
+/**
+ * luring_fd_register:
+ *
+ * Register file descriptors, see luring_fd_lookup
+ */
+static int luring_fd_register(struct io_uring *ring, LuringFd *fd_reg, int fd)
+{
+ int ret, nr;
+ GHashTable *lookup = fd_reg->fd_lookup;
+ nr = g_hash_table_size(lookup);
+
+ /* If adding new, API requires older registrations to be removed */
+ if (nr) {
+ /*
+ * See linux b19062a56726, register needs the ring mutex, any
+ * submission in progress will complete before unregistering begins
+ * and new ones will have to wait.
+ */
+ ret = io_uring_unregister_files(ring);
+ if (ret < 0) {
+ return ret;
+ }
+ }
+
+ fd_reg->fd_array = g_realloc_n(fd_reg->fd_array, nr + 1, sizeof(int));
+ fd_reg->fd_array[nr] = fd;
+
+ g_hash_table_insert(lookup, GINT_TO_POINTER(fd), GINT_TO_POINTER(nr));
+ ret = io_uring_register_files(ring, fd_reg->fd_array, nr + 1);
+ trace_luring_fd_register(fd, nr, ret);
+ if (ret < 0) {
+ /* Leave fd_array[] alone, fd will be overwritten next time anyway */
+ g_hash_table_remove(lookup, GINT_TO_POINTER(fd));
+ return ret;
+ }
+ return nr;
+}
+
+/**
+ * luring_fd_init:
+ *
+ * Initialize file descriptors
+ */
+static void luring_fd_init(LuringState *s)
+{
+ s->fd_reg.fd_lookup = g_hash_table_new(g_direct_hash, g_direct_equal);
+}
+
+/**
+ * luring_fd_cleanup:
+ *
+ * Unregisters file descriptors, TODO: error handling
+ */
+static void luring_fd_cleanup(LuringState *s)
+{
+ io_uring_unregister_files(&s->ring);
+ g_hash_table_unref(s->fd_reg.fd_lookup);
+ g_free(s->fd_reg.fd_array);
+ s->fd_reg.fd_array = NULL;
+}
+
+/**
+ * luring_fd_lookup:
+ *
+ * Used to lookup fd index in registered array at submission time
+ * If the lookup table has not been created or the fd is not in the table,
+ * the fd is registered.
+ *
+ * If registration errors, this function returns -errno.
+ *
+ * Unregistering is done in luring_detach_aio_context().
+ */
+static int luring_fd_lookup(LuringState *s, int fd)
+{
+ void *index;
+
+ if (g_hash_table_lookup_extended(s->fd_reg.fd_lookup, GINT_TO_POINTER(fd),
+ NULL, &index)) {
+ return GPOINTER_TO_INT(index);
+ }
+
+ return luring_fd_register(&s->ring, &s->fd_reg, fd);
+}
+
void luring_io_plug(BlockDriverState *bs, LuringState *s)
{
trace_luring_io_plug(s);
@@ -329,9 +465,14 @@ void luring_io_unplug(BlockDriverState *bs, LuringState *s)
static int luring_do_submit(int fd, LuringAIOCB *luringcb, LuringState *s,
uint64_t offset, int type)
{
- int ret;
+ int ret, fd_index;
struct io_uring_sqe *sqes = &luringcb->sqeq;
+ fd_index = luring_fd_lookup(s, fd);
+ if (fd_index >= 0) {
+ fd = fd_index;
+ }
+
switch (type) {
case QEMU_AIO_WRITE:
io_uring_prep_writev(sqes, fd, luringcb->qiov->iov,
@@ -349,7 +490,11 @@ static int luring_do_submit(int fd, LuringAIOCB *luringcb,
LuringState *s,
__func__, type);
abort();
}
+
io_uring_sqe_set_data(sqes, luringcb);
+ if (fd_index >= 0) {
+ io_uring_sqe_set_flags(sqes, IOSQE_FIXED_FILE);
+ }
QSIMPLEQ_INSERT_TAIL(&s->io_q.submit_queue, luringcb, next);
s->io_q.in_queue++;
@@ -375,6 +520,7 @@ int coroutine_fn luring_co_submit(BlockDriverState *bs,
LuringState *s, int fd,
.qiov = qiov,
.is_read = (type == QEMU_AIO_READ),
};
+
trace_luring_co_submit(bs, s, &luringcb, fd, offset, qiov ? qiov->size : 0,
type);
ret = luring_do_submit(fd, &luringcb, s, offset, type);
@@ -391,6 +537,7 @@ int coroutine_fn luring_co_submit(BlockDriverState *bs,
LuringState *s, int fd,
void luring_detach_aio_context(LuringState *s, AioContext *old_context)
{
+ luring_fd_cleanup(s);
aio_set_fd_handler(old_context, s->ring.ring_fd, false, NULL, NULL, NULL,
s);
qemu_bh_delete(s->completion_bh);
@@ -403,6 +550,7 @@ void luring_attach_aio_context(LuringState *s, AioContext
*new_context)
s->completion_bh = aio_bh_new(new_context, qemu_luring_completion_bh, s);
aio_set_fd_handler(s->aio_context, s->ring.ring_fd, false,
qemu_luring_completion_cb, NULL, qemu_luring_poll_cb,
s);
+ luring_fd_init(s);
}
LuringState *luring_init(Error **errp)
diff --git a/block/trace-events b/block/trace-events
index 2e3a9851b6..6a19cc00b7 100644
--- a/block/trace-events
+++ b/block/trace-events
@@ -72,6 +72,8 @@ luring_co_submit(void *bs, void *s, void *luringcb, int fd,
uint64_t offset, siz
luring_process_completion(void *s, void *aiocb, int ret) "LuringState %p
luringcb %p ret %d"
luring_io_uring_submit(void *s, int ret) "LuringState %p ret %d"
luring_resubmit_short_read(void *s, void *luringcb, int nread) "LuringState %p
luringcb %p nread %d"
+luring_fd_register(int fd, int index, int ret) "fd %d index %d ret %d"
+luring_fd_unregister(void *s, int fd, int index) "LuringState %p fd %d index
%d"
# qcow2.c
qcow2_writev_start_req(void *co, int64_t offset, int bytes) "co %p offset 0x%"
PRIx64 " bytes %d"
--
2.21.0
- [PATCH 06/16] util/async: add aio interfaces for io_uring, (continued)
- [PATCH 06/16] util/async: add aio interfaces for io_uring, Stefan Hajnoczi, 2019/10/07
- [PATCH 07/16] blockdev: adds bdrv_parse_aio to use io_uring, Stefan Hajnoczi, 2019/10/07
- [PATCH 08/16] block/file-posix.c: extend to use io_uring, Stefan Hajnoczi, 2019/10/07
- [PATCH 09/16] block: add trace events for io_uring, Stefan Hajnoczi, 2019/10/07
- [PATCH 10/16] block/io_uring: adds userspace completion polling, Stefan Hajnoczi, 2019/10/07
- [PATCH 11/16] qemu-io: adds option to use aio engine, Stefan Hajnoczi, 2019/10/07
- [PATCH 12/16] qemu-img: adds option to use aio engine for benchmarking, Stefan Hajnoczi, 2019/10/07
- [PATCH 13/16] qemu-nbd: adds option for aio engines, Stefan Hajnoczi, 2019/10/07
- [PATCH 14/16] tests/qemu-iotests: enable testing with aio options, Stefan Hajnoczi, 2019/10/07
- [PATCH 15/16] tests/qemu-iotests: use AIOMODE with various tests, Stefan Hajnoczi, 2019/10/07
- [PATCH 16/16] block/io_uring: adds fd registration,
Stefan Hajnoczi <=