From ceb029cd6feccf9f7607833b71dd609d149421a1 Mon Sep 17 00:00:00 2001 From: Vladimir Sementsov-Ogievskiy Date: Wed, 20 Jun 2018 17:48:37 +0300 Subject: [PATCH] qcow2: add compress threads Do data compression in separate threads. This significantly improve performance for qemu-img convert with -W (allow async writes) and -c (compressed) options. Signed-off-by: Vladimir Sementsov-Ogievskiy Signed-off-by: Kevin Wolf --- block/qcow2.c | 62 ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++- block/qcow2.h | 3 +++ 2 files changed, 64 insertions(+), 1 deletion(-) diff --git a/block/qcow2.c b/block/qcow2.c index 9cee653d96..33b61b7480 100644 --- a/block/qcow2.c +++ b/block/qcow2.c @@ -44,6 +44,7 @@ #include "qapi/qobject-input-visitor.h" #include "qapi/qapi-visit-block-core.h" #include "crypto.h" +#include "block/thread-pool.h" /* Differences with QCOW: @@ -1544,6 +1545,9 @@ static int coroutine_fn qcow2_do_open(BlockDriverState *bs, QDict *options, qcow2_check_refcounts(bs, &result, 0); } #endif + + qemu_co_queue_init(&s->compress_wait_queue); + return ret; fail: @@ -3695,6 +3699,62 @@ static ssize_t qcow2_compress(void *dest, const void *src, size_t size) return ret; } +#define MAX_COMPRESS_THREADS 4 + +typedef struct Qcow2CompressData { + void *dest; + const void *src; + size_t size; + ssize_t ret; +} Qcow2CompressData; + +static int qcow2_compress_pool_func(void *opaque) +{ + Qcow2CompressData *data = opaque; + + data->ret = qcow2_compress(data->dest, data->src, data->size); + + return 0; +} + +static void qcow2_compress_complete(void *opaque, int ret) +{ + qemu_coroutine_enter(opaque); +} + +/* See qcow2_compress definition for parameters description */ +static ssize_t qcow2_co_compress(BlockDriverState *bs, + void *dest, const void *src, size_t size) +{ + BDRVQcow2State *s = bs->opaque; + BlockAIOCB *acb; + ThreadPool *pool = aio_get_thread_pool(bdrv_get_aio_context(bs)); + Qcow2CompressData arg = { + .dest = dest, + .src = src, + .size = size, + }; + + while (s->nb_compress_threads >= MAX_COMPRESS_THREADS) { + qemu_co_queue_wait(&s->compress_wait_queue, NULL); + } + + s->nb_compress_threads++; + acb = thread_pool_submit_aio(pool, qcow2_compress_pool_func, &arg, + qcow2_compress_complete, + qemu_coroutine_self()); + + if (!acb) { + s->nb_compress_threads--; + return -EINVAL; + } + qemu_coroutine_yield(); + s->nb_compress_threads--; + qemu_co_queue_next(&s->compress_wait_queue); + + return arg.ret; +} + /* XXX: put compressed sectors first, then all the cluster aligned tables to avoid losing bytes in alignment */ static coroutine_fn int @@ -3739,7 +3799,7 @@ qcow2_co_pwritev_compressed(BlockDriverState *bs, uint64_t offset, out_buf = g_malloc(s->cluster_size); - out_len = qcow2_compress(out_buf, buf, s->cluster_size); + out_len = qcow2_co_compress(bs, out_buf, buf, s->cluster_size); if (out_len == -2) { ret = -EINVAL; goto fail; diff --git a/block/qcow2.h b/block/qcow2.h index 1c9c0d3631..d6aca687d6 100644 --- a/block/qcow2.h +++ b/block/qcow2.h @@ -326,6 +326,9 @@ typedef struct BDRVQcow2State { * override) */ char *image_backing_file; char *image_backing_format; + + CoQueue compress_wait_queue; + int nb_compress_threads; } BDRVQcow2State; typedef struct Qcow2COWRegion { -- 2.11.4.GIT