From 402a47411bff5e849dc880dd08ba7e6564e6e4f4 Mon Sep 17 00:00:00 2001 From: Paolo Bonzini Date: Tue, 22 Jan 2013 09:03:14 +0100 Subject: [PATCH] mirror: support more than one in-flight AIO operation With AIO support in place, we can start copying more than one chunk in parallel. This patch introduces the required infrastructure for this: the buffer is split into multiple granularity-sized chunks, and there is a free list to access them. Because of copy-on-write, a single operation may already require multiple chunks to be available on the free list. In addition, two different iterations on the HBitmap may want to copy the same cluster. We avoid this by keeping a bitmap of in-flight I/O operations, and blocking until the previous iteration completes. This should be a pretty rare occurrence, though; as long as there is no overlap the next iteration can start before the previous one finishes. Signed-off-by: Paolo Bonzini Signed-off-by: Kevin Wolf --- block/mirror.c | 102 +++++++++++++++++++++++++++++++++++++++++++------ trace-events | 4 +- 2 files changed, 94 insertions(+), 12 deletions(-) diff --git a/block/mirror.c b/block/mirror.c index 896972c297..4696dc84ea 100644 --- a/block/mirror.c +++ b/block/mirror.c @@ -17,7 +17,15 @@ #include "qemu/ratelimit.h" #include "qemu/bitmap.h" -#define SLICE_TIME 100000000ULL /* ns */ +#define SLICE_TIME 100000000ULL /* ns */ +#define MAX_IN_FLIGHT 16 + +/* The mirroring buffer is a list of granularity-sized chunks. + * Free chunks are organized in a list. + */ +typedef struct MirrorBuffer { + QSIMPLEQ_ENTRY(MirrorBuffer) next; +} MirrorBuffer; typedef struct MirrorBlockJob { BlockJob common; @@ -33,7 +41,10 @@ typedef struct MirrorBlockJob { unsigned long *cow_bitmap; HBitmapIter hbi; uint8_t *buf; + QSIMPLEQ_HEAD(, MirrorBuffer) buf_free; + int buf_free_count; + unsigned long *in_flight_bitmap; int in_flight; int ret; } MirrorBlockJob; @@ -41,7 +52,6 @@ typedef struct MirrorBlockJob { typedef struct MirrorOp { MirrorBlockJob *s; QEMUIOVector qiov; - struct iovec iov; int64_t sector_num; int nb_sectors; } MirrorOp; @@ -62,15 +72,24 @@ static BlockErrorAction mirror_error_action(MirrorBlockJob *s, bool read, static void mirror_iteration_done(MirrorOp *op, int ret) { MirrorBlockJob *s = op->s; + struct iovec *iov; int64_t chunk_num; - int nb_chunks, sectors_per_chunk; + int i, nb_chunks, sectors_per_chunk; trace_mirror_iteration_done(s, op->sector_num, op->nb_sectors, ret); s->in_flight--; + iov = op->qiov.iov; + for (i = 0; i < op->qiov.niov; i++) { + MirrorBuffer *buf = (MirrorBuffer *) iov[i].iov_base; + QSIMPLEQ_INSERT_TAIL(&s->buf_free, buf, next); + s->buf_free_count++; + } + sectors_per_chunk = s->granularity >> BDRV_SECTOR_BITS; chunk_num = op->sector_num / sectors_per_chunk; nb_chunks = op->nb_sectors / sectors_per_chunk; + bitmap_clear(s->in_flight_bitmap, chunk_num, nb_chunks); if (s->cow_bitmap && ret >= 0) { bitmap_set(s->cow_bitmap, chunk_num, nb_chunks); } @@ -120,8 +139,8 @@ static void mirror_read_complete(void *opaque, int ret) static void coroutine_fn mirror_iteration(MirrorBlockJob *s) { BlockDriverState *source = s->common.bs; - int nb_sectors, sectors_per_chunk; - int64_t end, sector_num, chunk_num; + int nb_sectors, sectors_per_chunk, nb_chunks; + int64_t end, sector_num, chunk_num, next_sector, hbitmap_next_sector; MirrorOp *op; s->sector_num = hbitmap_iter_next(&s->hbi); @@ -132,6 +151,8 @@ static void coroutine_fn mirror_iteration(MirrorBlockJob *s) assert(s->sector_num >= 0); } + hbitmap_next_sector = s->sector_num; + /* If we have no backing file yet in the destination, and the cluster size * is very large, we need to do COW ourselves. The first time a cluster is * copied, copy it entirely. @@ -147,19 +168,56 @@ static void coroutine_fn mirror_iteration(MirrorBlockJob *s) bdrv_round_to_clusters(s->target, sector_num, sectors_per_chunk, §or_num, &nb_sectors); + + /* The rounding may make us copy sectors before the + * first dirty one. + */ + chunk_num = sector_num / sectors_per_chunk; + } + + /* Wait for I/O to this cluster (from a previous iteration) to be done. */ + while (test_bit(chunk_num, s->in_flight_bitmap)) { + trace_mirror_yield_in_flight(s, sector_num, s->in_flight); + qemu_coroutine_yield(); } end = s->common.len >> BDRV_SECTOR_BITS; nb_sectors = MIN(nb_sectors, end - sector_num); + nb_chunks = (nb_sectors + sectors_per_chunk - 1) / sectors_per_chunk; + while (s->buf_free_count < nb_chunks) { + trace_mirror_yield_buf_busy(s, nb_chunks, s->in_flight); + qemu_coroutine_yield(); + } + + /* We have enough free space to copy these sectors. */ + bitmap_set(s->in_flight_bitmap, chunk_num, nb_chunks); /* Allocate a MirrorOp that is used as an AIO callback. */ op = g_slice_new(MirrorOp); op->s = s; - op->iov.iov_base = s->buf; - op->iov.iov_len = nb_sectors * 512; op->sector_num = sector_num; op->nb_sectors = nb_sectors; - qemu_iovec_init_external(&op->qiov, &op->iov, 1); + + /* Now make a QEMUIOVector taking enough granularity-sized chunks + * from s->buf_free. + */ + qemu_iovec_init(&op->qiov, nb_chunks); + next_sector = sector_num; + while (nb_chunks-- > 0) { + MirrorBuffer *buf = QSIMPLEQ_FIRST(&s->buf_free); + QSIMPLEQ_REMOVE_HEAD(&s->buf_free, next); + s->buf_free_count--; + qemu_iovec_add(&op->qiov, buf, s->granularity); + + /* Advance the HBitmapIter in parallel, so that we do not examine + * the same sector twice. + */ + if (next_sector > hbitmap_next_sector && bdrv_get_dirty(source, next_sector)) { + hbitmap_next_sector = hbitmap_iter_next(&s->hbi); + } + + next_sector += sectors_per_chunk; + } bdrv_reset_dirty(source, sector_num, nb_sectors); @@ -170,6 +228,23 @@ static void coroutine_fn mirror_iteration(MirrorBlockJob *s) mirror_read_complete, op); } +static void mirror_free_init(MirrorBlockJob *s) +{ + int granularity = s->granularity; + size_t buf_size = s->buf_size; + uint8_t *buf = s->buf; + + assert(s->buf_free_count == 0); + QSIMPLEQ_INIT(&s->buf_free); + while (buf_size != 0) { + MirrorBuffer *cur = (MirrorBuffer *)buf; + QSIMPLEQ_INSERT_TAIL(&s->buf_free, cur, next); + s->buf_free_count++; + buf_size -= granularity; + buf += granularity; + } +} + static void mirror_drain(MirrorBlockJob *s) { while (s->in_flight > 0) { @@ -198,6 +273,9 @@ static void coroutine_fn mirror_run(void *opaque) return; } + length = (bdrv_getlength(bs) + s->granularity - 1) / s->granularity; + s->in_flight_bitmap = bitmap_new(length); + /* If we have no backing file yet in the destination, we cannot let * the destination do COW. Instead, we copy sectors around the * dirty data if needed. We need a bitmap to do that. @@ -208,7 +286,6 @@ static void coroutine_fn mirror_run(void *opaque) bdrv_get_info(s->target, &bdi); if (s->granularity < bdi.cluster_size) { s->buf_size = MAX(s->buf_size, bdi.cluster_size); - length = (bdrv_getlength(bs) + s->granularity - 1) / s->granularity; s->cow_bitmap = bitmap_new(length); } } @@ -216,6 +293,7 @@ static void coroutine_fn mirror_run(void *opaque) end = s->common.len >> BDRV_SECTOR_BITS; s->buf = qemu_blockalign(bs, s->buf_size); sectors_per_chunk = s->granularity >> BDRV_SECTOR_BITS; + mirror_free_init(s); if (s->mode != MIRROR_SYNC_MODE_NONE) { /* First part, loop on the sectors and initialize the dirty bitmap. */ @@ -261,8 +339,9 @@ static void coroutine_fn mirror_run(void *opaque) */ if (qemu_get_clock_ns(rt_clock) - last_pause_ns < SLICE_TIME && s->common.iostatus == BLOCK_DEVICE_IO_STATUS_OK) { - if (s->in_flight > 0) { - trace_mirror_yield(s, s->in_flight, cnt); + if (s->in_flight == MAX_IN_FLIGHT || s->buf_free_count == 0 || + (cnt == 0 && s->in_flight > 0)) { + trace_mirror_yield(s, s->in_flight, s->buf_free_count, cnt); qemu_coroutine_yield(); continue; } else if (cnt != 0) { @@ -354,6 +433,7 @@ immediate_exit: assert(s->in_flight == 0); qemu_vfree(s->buf); g_free(s->cow_bitmap); + g_free(s->in_flight_bitmap); bdrv_set_dirty_tracking(bs, 0); bdrv_iostatus_disable(s->target); if (s->should_complete && ret == 0) { diff --git a/trace-events b/trace-events index 43e6b73fa5..b4a23cdc2e 100644 --- a/trace-events +++ b/trace-events @@ -86,7 +86,9 @@ mirror_before_sleep(void *s, int64_t cnt, int synced) "s %p dirty count %"PRId64 mirror_one_iteration(void *s, int64_t sector_num, int nb_sectors) "s %p sector_num %"PRId64" nb_sectors %d" mirror_cow(void *s, int64_t sector_num) "s %p sector_num %"PRId64 mirror_iteration_done(void *s, int64_t sector_num, int nb_sectors, int ret) "s %p sector_num %"PRId64" nb_sectors %d ret %d" -mirror_yield(void *s, int64_t cnt, int in_flight) "s %p dirty count %"PRId64" in_flight %d" +mirror_yield(void *s, int64_t cnt, int buf_free_count, int in_flight) "s %p dirty count %"PRId64" free buffers %d in_flight %d" +mirror_yield_in_flight(void *s, int64_t sector_num, int in_flight) "s %p sector_num %"PRId64" in_flight %d" +mirror_yield_buf_busy(void *s, int nb_chunks, int in_flight) "s %p requested chunks %d in_flight %d" # blockdev.c qmp_block_job_cancel(void *job) "job %p"