|
|
1bdc94 |
From 6a6e25b0dff81b5a6dc4272a176f5987b7474a93 Mon Sep 17 00:00:00 2001
|
|
|
1bdc94 |
From: "Dr. David Alan Gilbert" <dgilbert@redhat.com>
|
|
|
1bdc94 |
Date: Fri, 22 Jun 2018 18:59:49 +0200
|
|
|
1bdc94 |
Subject: [PATCH 10/57] migration: stop compression to allocate and free memory
|
|
|
1bdc94 |
frequently
|
|
|
1bdc94 |
|
|
|
1bdc94 |
RH-Author: Dr. David Alan Gilbert <dgilbert@redhat.com>
|
|
|
1bdc94 |
Message-id: <20180622190005.21297-3-dgilbert@redhat.com>
|
|
|
1bdc94 |
Patchwork-id: 80998
|
|
|
1bdc94 |
O-Subject: [RHEL7.6 qemu-kvm-rhev PATCH 02/18] migration: stop compression to allocate and free memory frequently
|
|
|
1bdc94 |
Bugzilla: 1584139
|
|
|
1bdc94 |
RH-Acked-by: Juan Quintela <quintela@redhat.com>
|
|
|
1bdc94 |
RH-Acked-by: Peter Xu <peterx@redhat.com>
|
|
|
1bdc94 |
RH-Acked-by: Laurent Vivier <lvivier@redhat.com>
|
|
|
1bdc94 |
|
|
|
1bdc94 |
From: Xiao Guangrong <xiaoguangrong@tencent.com>
|
|
|
1bdc94 |
|
|
|
1bdc94 |
Current code uses compress2() to compress memory which manages memory
|
|
|
1bdc94 |
internally, that causes huge memory is allocated and freed very
|
|
|
1bdc94 |
frequently
|
|
|
1bdc94 |
|
|
|
1bdc94 |
More worse, frequently returning memory to kernel will flush TLBs
|
|
|
1bdc94 |
and trigger invalidation callbacks on mmu-notification which
|
|
|
1bdc94 |
interacts with KVM MMU, that dramatically reduce the performance
|
|
|
1bdc94 |
of VM
|
|
|
1bdc94 |
|
|
|
1bdc94 |
So, we maintain the memory by ourselves and reuse it for each
|
|
|
1bdc94 |
compression
|
|
|
1bdc94 |
|
|
|
1bdc94 |
Reviewed-by: Peter Xu <peterx@redhat.com>
|
|
|
1bdc94 |
Reviewed-by: Jiang Biao <jiang.biao2@zte.com.cn>
|
|
|
1bdc94 |
Signed-off-by: Xiao Guangrong <xiaoguangrong@tencent.com>
|
|
|
1bdc94 |
Message-Id: <20180330075128.26919-3-xiaoguangrong@tencent.com>
|
|
|
1bdc94 |
Signed-off-by: Dr. David Alan Gilbert <dgilbert@redhat.com>
|
|
|
1bdc94 |
(cherry picked from commit dcaf446ebda5d87e05eb41cdbafb7ae4a7cc4a62)
|
|
|
1bdc94 |
Signed-off-by: Miroslav Rezanina <mrezanin@redhat.com>
|
|
|
1bdc94 |
---
|
|
|
1bdc94 |
migration/qemu-file.c | 39 ++++++++++++++++++++++++++++++++-------
|
|
|
1bdc94 |
migration/qemu-file.h | 6 ++++--
|
|
|
1bdc94 |
migration/ram.c | 41 ++++++++++++++++++++++++++++++++---------
|
|
|
1bdc94 |
3 files changed, 68 insertions(+), 18 deletions(-)
|
|
|
1bdc94 |
|
|
|
1bdc94 |
diff --git a/migration/qemu-file.c b/migration/qemu-file.c
|
|
|
1bdc94 |
index bb63c77..bafe3a0 100644
|
|
|
1bdc94 |
--- a/migration/qemu-file.c
|
|
|
1bdc94 |
+++ b/migration/qemu-file.c
|
|
|
1bdc94 |
@@ -658,8 +658,32 @@ uint64_t qemu_get_be64(QEMUFile *f)
|
|
|
1bdc94 |
return v;
|
|
|
1bdc94 |
}
|
|
|
1bdc94 |
|
|
|
1bdc94 |
-/* Compress size bytes of data start at p with specific compression
|
|
|
1bdc94 |
- * level and store the compressed data to the buffer of f.
|
|
|
1bdc94 |
+/* return the size after compression, or negative value on error */
|
|
|
1bdc94 |
+static int qemu_compress_data(z_stream *stream, uint8_t *dest, size_t dest_len,
|
|
|
1bdc94 |
+ const uint8_t *source, size_t source_len)
|
|
|
1bdc94 |
+{
|
|
|
1bdc94 |
+ int err;
|
|
|
1bdc94 |
+
|
|
|
1bdc94 |
+ err = deflateReset(stream);
|
|
|
1bdc94 |
+ if (err != Z_OK) {
|
|
|
1bdc94 |
+ return -1;
|
|
|
1bdc94 |
+ }
|
|
|
1bdc94 |
+
|
|
|
1bdc94 |
+ stream->avail_in = source_len;
|
|
|
1bdc94 |
+ stream->next_in = (uint8_t *)source;
|
|
|
1bdc94 |
+ stream->avail_out = dest_len;
|
|
|
1bdc94 |
+ stream->next_out = dest;
|
|
|
1bdc94 |
+
|
|
|
1bdc94 |
+ err = deflate(stream, Z_FINISH);
|
|
|
1bdc94 |
+ if (err != Z_STREAM_END) {
|
|
|
1bdc94 |
+ return -1;
|
|
|
1bdc94 |
+ }
|
|
|
1bdc94 |
+
|
|
|
1bdc94 |
+ return stream->next_out - dest;
|
|
|
1bdc94 |
+}
|
|
|
1bdc94 |
+
|
|
|
1bdc94 |
+/* Compress size bytes of data start at p and store the compressed
|
|
|
1bdc94 |
+ * data to the buffer of f.
|
|
|
1bdc94 |
*
|
|
|
1bdc94 |
* When f is not writable, return -1 if f has no space to save the
|
|
|
1bdc94 |
* compressed data.
|
|
|
1bdc94 |
@@ -667,9 +691,8 @@ uint64_t qemu_get_be64(QEMUFile *f)
|
|
|
1bdc94 |
* do fflush first, if f still has no space to save the compressed
|
|
|
1bdc94 |
* data, return -1.
|
|
|
1bdc94 |
*/
|
|
|
1bdc94 |
-
|
|
|
1bdc94 |
-ssize_t qemu_put_compression_data(QEMUFile *f, const uint8_t *p, size_t size,
|
|
|
1bdc94 |
- int level)
|
|
|
1bdc94 |
+ssize_t qemu_put_compression_data(QEMUFile *f, z_stream *stream,
|
|
|
1bdc94 |
+ const uint8_t *p, size_t size)
|
|
|
1bdc94 |
{
|
|
|
1bdc94 |
ssize_t blen = IO_BUF_SIZE - f->buf_index - sizeof(int32_t);
|
|
|
1bdc94 |
|
|
|
1bdc94 |
@@ -683,8 +706,10 @@ ssize_t qemu_put_compression_data(QEMUFile *f, const uint8_t *p, size_t size,
|
|
|
1bdc94 |
return -1;
|
|
|
1bdc94 |
}
|
|
|
1bdc94 |
}
|
|
|
1bdc94 |
- if (compress2(f->buf + f->buf_index + sizeof(int32_t), (uLongf *)&blen,
|
|
|
1bdc94 |
- (Bytef *)p, size, level) != Z_OK) {
|
|
|
1bdc94 |
+
|
|
|
1bdc94 |
+ blen = qemu_compress_data(stream, f->buf + f->buf_index + sizeof(int32_t),
|
|
|
1bdc94 |
+ blen, p, size);
|
|
|
1bdc94 |
+ if (blen < 0) {
|
|
|
1bdc94 |
error_report("Compress Failed!");
|
|
|
1bdc94 |
return 0;
|
|
|
1bdc94 |
}
|
|
|
1bdc94 |
diff --git a/migration/qemu-file.h b/migration/qemu-file.h
|
|
|
1bdc94 |
index f4f356a..2ccfcfb 100644
|
|
|
1bdc94 |
--- a/migration/qemu-file.h
|
|
|
1bdc94 |
+++ b/migration/qemu-file.h
|
|
|
1bdc94 |
@@ -25,6 +25,8 @@
|
|
|
1bdc94 |
#ifndef MIGRATION_QEMU_FILE_H
|
|
|
1bdc94 |
#define MIGRATION_QEMU_FILE_H
|
|
|
1bdc94 |
|
|
|
1bdc94 |
+#include <zlib.h>
|
|
|
1bdc94 |
+
|
|
|
1bdc94 |
/* Read a chunk of data from a file at the given position. The pos argument
|
|
|
1bdc94 |
* can be ignored if the file is only be used for streaming. The number of
|
|
|
1bdc94 |
* bytes actually read should be returned.
|
|
|
1bdc94 |
@@ -132,8 +134,8 @@ bool qemu_file_is_writable(QEMUFile *f);
|
|
|
1bdc94 |
|
|
|
1bdc94 |
size_t qemu_peek_buffer(QEMUFile *f, uint8_t **buf, size_t size, size_t offset);
|
|
|
1bdc94 |
size_t qemu_get_buffer_in_place(QEMUFile *f, uint8_t **buf, size_t size);
|
|
|
1bdc94 |
-ssize_t qemu_put_compression_data(QEMUFile *f, const uint8_t *p, size_t size,
|
|
|
1bdc94 |
- int level);
|
|
|
1bdc94 |
+ssize_t qemu_put_compression_data(QEMUFile *f, z_stream *stream,
|
|
|
1bdc94 |
+ const uint8_t *p, size_t size);
|
|
|
1bdc94 |
int qemu_put_qemu_file(QEMUFile *f_des, QEMUFile *f_src);
|
|
|
1bdc94 |
|
|
|
1bdc94 |
/*
|
|
|
1bdc94 |
diff --git a/migration/ram.c b/migration/ram.c
|
|
|
1bdc94 |
index f27038a..7d3b1da 100644
|
|
|
1bdc94 |
--- a/migration/ram.c
|
|
|
1bdc94 |
+++ b/migration/ram.c
|
|
|
1bdc94 |
@@ -269,6 +269,7 @@ struct CompressParam {
|
|
|
1bdc94 |
QemuCond cond;
|
|
|
1bdc94 |
RAMBlock *block;
|
|
|
1bdc94 |
ram_addr_t offset;
|
|
|
1bdc94 |
+ z_stream stream;
|
|
|
1bdc94 |
};
|
|
|
1bdc94 |
typedef struct CompressParam CompressParam;
|
|
|
1bdc94 |
|
|
|
1bdc94 |
@@ -299,7 +300,7 @@ static QemuThread *decompress_threads;
|
|
|
1bdc94 |
static QemuMutex decomp_done_lock;
|
|
|
1bdc94 |
static QemuCond decomp_done_cond;
|
|
|
1bdc94 |
|
|
|
1bdc94 |
-static int do_compress_ram_page(QEMUFile *f, RAMBlock *block,
|
|
|
1bdc94 |
+static int do_compress_ram_page(QEMUFile *f, z_stream *stream, RAMBlock *block,
|
|
|
1bdc94 |
ram_addr_t offset);
|
|
|
1bdc94 |
|
|
|
1bdc94 |
static void *do_data_compress(void *opaque)
|
|
|
1bdc94 |
@@ -316,7 +317,7 @@ static void *do_data_compress(void *opaque)
|
|
|
1bdc94 |
param->block = NULL;
|
|
|
1bdc94 |
qemu_mutex_unlock(¶m->mutex);
|
|
|
1bdc94 |
|
|
|
1bdc94 |
- do_compress_ram_page(param->file, block, offset);
|
|
|
1bdc94 |
+ do_compress_ram_page(param->file, ¶m->stream, block, offset);
|
|
|
1bdc94 |
|
|
|
1bdc94 |
qemu_mutex_lock(&comp_done_lock);
|
|
|
1bdc94 |
param->done = true;
|
|
|
1bdc94 |
@@ -357,10 +358,19 @@ static void compress_threads_save_cleanup(void)
|
|
|
1bdc94 |
terminate_compression_threads();
|
|
|
1bdc94 |
thread_count = migrate_compress_threads();
|
|
|
1bdc94 |
for (i = 0; i < thread_count; i++) {
|
|
|
1bdc94 |
+ /*
|
|
|
1bdc94 |
+ * we use it as a indicator which shows if the thread is
|
|
|
1bdc94 |
+ * properly init'd or not
|
|
|
1bdc94 |
+ */
|
|
|
1bdc94 |
+ if (!comp_param[i].file) {
|
|
|
1bdc94 |
+ break;
|
|
|
1bdc94 |
+ }
|
|
|
1bdc94 |
qemu_thread_join(compress_threads + i);
|
|
|
1bdc94 |
- qemu_fclose(comp_param[i].file);
|
|
|
1bdc94 |
qemu_mutex_destroy(&comp_param[i].mutex);
|
|
|
1bdc94 |
qemu_cond_destroy(&comp_param[i].cond);
|
|
|
1bdc94 |
+ deflateEnd(&comp_param[i].stream);
|
|
|
1bdc94 |
+ qemu_fclose(comp_param[i].file);
|
|
|
1bdc94 |
+ comp_param[i].file = NULL;
|
|
|
1bdc94 |
}
|
|
|
1bdc94 |
qemu_mutex_destroy(&comp_done_lock);
|
|
|
1bdc94 |
qemu_cond_destroy(&comp_done_cond);
|
|
|
1bdc94 |
@@ -370,12 +380,12 @@ static void compress_threads_save_cleanup(void)
|
|
|
1bdc94 |
comp_param = NULL;
|
|
|
1bdc94 |
}
|
|
|
1bdc94 |
|
|
|
1bdc94 |
-static void compress_threads_save_setup(void)
|
|
|
1bdc94 |
+static int compress_threads_save_setup(void)
|
|
|
1bdc94 |
{
|
|
|
1bdc94 |
int i, thread_count;
|
|
|
1bdc94 |
|
|
|
1bdc94 |
if (!migrate_use_compression()) {
|
|
|
1bdc94 |
- return;
|
|
|
1bdc94 |
+ return 0;
|
|
|
1bdc94 |
}
|
|
|
1bdc94 |
thread_count = migrate_compress_threads();
|
|
|
1bdc94 |
compress_threads = g_new0(QemuThread, thread_count);
|
|
|
1bdc94 |
@@ -383,6 +393,11 @@ static void compress_threads_save_setup(void)
|
|
|
1bdc94 |
qemu_cond_init(&comp_done_cond);
|
|
|
1bdc94 |
qemu_mutex_init(&comp_done_lock);
|
|
|
1bdc94 |
for (i = 0; i < thread_count; i++) {
|
|
|
1bdc94 |
+ if (deflateInit(&comp_param[i].stream,
|
|
|
1bdc94 |
+ migrate_compress_level()) != Z_OK) {
|
|
|
1bdc94 |
+ goto exit;
|
|
|
1bdc94 |
+ }
|
|
|
1bdc94 |
+
|
|
|
1bdc94 |
/* comp_param[i].file is just used as a dummy buffer to save data,
|
|
|
1bdc94 |
* set its ops to empty.
|
|
|
1bdc94 |
*/
|
|
|
1bdc94 |
@@ -395,6 +410,11 @@ static void compress_threads_save_setup(void)
|
|
|
1bdc94 |
do_data_compress, comp_param + i,
|
|
|
1bdc94 |
QEMU_THREAD_JOINABLE);
|
|
|
1bdc94 |
}
|
|
|
1bdc94 |
+ return 0;
|
|
|
1bdc94 |
+
|
|
|
1bdc94 |
+exit:
|
|
|
1bdc94 |
+ compress_threads_save_cleanup();
|
|
|
1bdc94 |
+ return -1;
|
|
|
1bdc94 |
}
|
|
|
1bdc94 |
|
|
|
1bdc94 |
/* Multiple fd's */
|
|
|
1bdc94 |
@@ -1032,7 +1052,7 @@ static int ram_save_page(RAMState *rs, PageSearchStatus *pss, bool last_stage)
|
|
|
1bdc94 |
return pages;
|
|
|
1bdc94 |
}
|
|
|
1bdc94 |
|
|
|
1bdc94 |
-static int do_compress_ram_page(QEMUFile *f, RAMBlock *block,
|
|
|
1bdc94 |
+static int do_compress_ram_page(QEMUFile *f, z_stream *stream, RAMBlock *block,
|
|
|
1bdc94 |
ram_addr_t offset)
|
|
|
1bdc94 |
{
|
|
|
1bdc94 |
RAMState *rs = ram_state;
|
|
|
1bdc94 |
@@ -1041,8 +1061,7 @@ static int do_compress_ram_page(QEMUFile *f, RAMBlock *block,
|
|
|
1bdc94 |
|
|
|
1bdc94 |
bytes_sent = save_page_header(rs, f, block, offset |
|
|
|
1bdc94 |
RAM_SAVE_FLAG_COMPRESS_PAGE);
|
|
|
1bdc94 |
- blen = qemu_put_compression_data(f, p, TARGET_PAGE_SIZE,
|
|
|
1bdc94 |
- migrate_compress_level());
|
|
|
1bdc94 |
+ blen = qemu_put_compression_data(f, stream, p, TARGET_PAGE_SIZE);
|
|
|
1bdc94 |
if (blen < 0) {
|
|
|
1bdc94 |
bytes_sent = 0;
|
|
|
1bdc94 |
qemu_file_set_error(migrate_get_current()->to_dst_file, blen);
|
|
|
1bdc94 |
@@ -2215,9 +2234,14 @@ static int ram_save_setup(QEMUFile *f, void *opaque)
|
|
|
1bdc94 |
RAMState **rsp = opaque;
|
|
|
1bdc94 |
RAMBlock *block;
|
|
|
1bdc94 |
|
|
|
1bdc94 |
+ if (compress_threads_save_setup()) {
|
|
|
1bdc94 |
+ return -1;
|
|
|
1bdc94 |
+ }
|
|
|
1bdc94 |
+
|
|
|
1bdc94 |
/* migration has already setup the bitmap, reuse it. */
|
|
|
1bdc94 |
if (!migration_in_colo_state()) {
|
|
|
1bdc94 |
if (ram_init_all(rsp) != 0) {
|
|
|
1bdc94 |
+ compress_threads_save_cleanup();
|
|
|
1bdc94 |
return -1;
|
|
|
1bdc94 |
}
|
|
|
1bdc94 |
}
|
|
|
1bdc94 |
@@ -2237,7 +2261,6 @@ static int ram_save_setup(QEMUFile *f, void *opaque)
|
|
|
1bdc94 |
}
|
|
|
1bdc94 |
|
|
|
1bdc94 |
rcu_read_unlock();
|
|
|
1bdc94 |
- compress_threads_save_setup();
|
|
|
1bdc94 |
|
|
|
1bdc94 |
ram_control_before_iterate(f, RAM_CONTROL_SETUP);
|
|
|
1bdc94 |
ram_control_after_iterate(f, RAM_CONTROL_SETUP);
|
|
|
1bdc94 |
--
|
|
|
1bdc94 |
1.8.3.1
|
|
|
1bdc94 |
|