bdev/aio: Reap completions from userspace if supported

Change-Id: I30d9cc619df2fddb870ed7bf187f14cd44376d19
Signed-off-by: Ben Walker <benjamin.walker@intel.com>
Reviewed-on: https://review.gerrithub.io/c/443468
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Jim Harris <james.r.harris@intel.com>
Reviewed-by: Shuhei Matsumoto <shuhei.matsumoto.xt@hitachi.com>
This commit is contained in:
Ben Walker 2019-02-05 12:17:31 -07:00 committed by Jim Harris
parent 5e5acd3dbf
commit dadb948585

View File

@ -35,10 +35,12 @@
#include "spdk/stdinc.h"
#include "spdk/barrier.h"
#include "spdk/bdev.h"
#include "spdk/conf.h"
#include "spdk/env.h"
#include "spdk/fd.h"
#include "spdk/likely.h"
#include "spdk/thread.h"
#include "spdk/json.h"
#include "spdk/util.h"
@ -75,6 +77,21 @@ struct file_disk {
bool block_size_override;
};
/* For user space reaping of completions */
struct spdk_aio_ring {
uint32_t id;
uint32_t size;
uint32_t head;
uint32_t tail;
uint32_t version;
uint32_t compat_features;
uint32_t incompat_features;
uint32_t header_length;
};
#define SPDK_AIO_RING_VERSION 0xa10a10a1
static int bdev_aio_initialize(void);
static void bdev_aio_fini(void);
static void aio_free_disk(struct file_disk *fdisk);
@ -230,6 +247,69 @@ bdev_aio_destruct(void *ctx)
return rc;
}
static int
bdev_user_io_getevents(io_context_t io_ctx, unsigned int max, struct io_event *uevents)
{
uint32_t head, tail, count;
struct spdk_aio_ring *ring;
struct timespec timeout;
struct io_event *kevents;
ring = (struct spdk_aio_ring *)io_ctx;
if (spdk_unlikely(ring->version != SPDK_AIO_RING_VERSION || ring->incompat_features != 0)) {
timeout.tv_sec = 0;
timeout.tv_nsec = 0;
return io_getevents(io_ctx, 0, max, uevents, &timeout);
}
/* Read the current state out of the ring */
head = ring->head;
tail = ring->tail;
/* This memory barrier is required to prevent the loads above
* from being re-ordered with stores to the events array
* potentially occurring on other threads. */
spdk_smp_rmb();
/* Calculate how many items are in the circular ring */
count = tail - head;
if (tail < head) {
count += ring->size;
}
/* Reduce the count to the limit provided by the user */
count = spdk_min(max, count);
/* Grab the memory location of the event array */
kevents = (struct io_event *)((uintptr_t)ring + ring->header_length);
/* Copy the events out of the ring. */
if ((head + count) <= ring->size) {
/* Only one copy is required */
memcpy(uevents, &kevents[head], count * sizeof(struct io_event));
} else {
uint32_t first_part = ring->size - head;
/* Two copies are required */
memcpy(uevents, &kevents[head], first_part * sizeof(struct io_event));
memcpy(&uevents[first_part], &kevents[0], (count - first_part) * sizeof(struct io_event));
}
/* Update the head pointer. On x86, stores will not be reordered with older loads,
* so the copies out of the event array will always be complete prior to this
* update becoming visible. On other architectures this is not guaranteed, so
* add a barrier. */
#if defined(__i386__) || defined(__x86_64__)
spdk_compiler_barrier();
#else
spdk_mb();
#endif
ring->head = (head + count) % ring->size;
return count;
}
static int
bdev_aio_group_poll(void *arg)
{
@ -237,14 +317,9 @@ bdev_aio_group_poll(void *arg)
int nr, i = 0;
enum spdk_bdev_io_status status;
struct bdev_aio_task *aio_task;
struct timespec timeout;
struct io_event events[SPDK_AIO_QUEUE_DEPTH];
timeout.tv_sec = 0;
timeout.tv_nsec = 0;
nr = io_getevents(group_ch->io_ctx, 0, SPDK_AIO_QUEUE_DEPTH,
events, &timeout);
nr = bdev_user_io_getevents(group_ch->io_ctx, SPDK_AIO_QUEUE_DEPTH, events);
if (nr < 0) {
return -1;