|
|
4a2fec |
From 16e6d181f5e881082a42dbae18c8f002fbbe689a Mon Sep 17 00:00:00 2001
|
|
|
4a2fec |
From: "Daniel P. Berrange" <berrange@redhat.com>
|
|
|
4a2fec |
Date: Fri, 1 Dec 2017 10:44:46 +0100
|
|
|
4a2fec |
Subject: [PATCH 10/36] block: use 1 MB bounce buffers for crypto instead of
|
|
|
4a2fec |
16KB
|
|
|
4a2fec |
|
|
|
4a2fec |
RH-Author: Daniel P. Berrange <berrange@redhat.com>
|
|
|
4a2fec |
Message-id: <20171201104446.7973-1-berrange@redhat.com>
|
|
|
4a2fec |
Patchwork-id: 78063
|
|
|
4a2fec |
O-Subject: [RHV-7.5 qemu-kvm-rhev PATCH] block: use 1 MB bounce buffers for crypto instead of 16KB
|
|
|
4a2fec |
Bugzilla: 1500334
|
|
|
4a2fec |
RH-Acked-by: Max Reitz <mreitz@redhat.com>
|
|
|
4a2fec |
RH-Acked-by: Jeffrey Cody <jcody@redhat.com>
|
|
|
4a2fec |
RH-Acked-by: Stefan Hajnoczi <stefanha@redhat.com>
|
|
|
4a2fec |
|
|
|
4a2fec |
Using 16KB bounce buffers creates a significant performance
|
|
|
4a2fec |
penalty for I/O to encrypted volumes on storage which high
|
|
|
4a2fec |
I/O latency (rotating rust & network drives), because it
|
|
|
4a2fec |
triggers lots of fairly small I/O operations.
|
|
|
4a2fec |
|
|
|
4a2fec |
On tests with rotating rust, and cache=none|directsync,
|
|
|
4a2fec |
write speed increased from 2MiB/s to 32MiB/s, on a par
|
|
|
4a2fec |
with that achieved by the in-kernel luks driver. With
|
|
|
4a2fec |
other cache modes the in-kernel driver is still notably
|
|
|
4a2fec |
faster because it is able to report completion of the
|
|
|
4a2fec |
I/O request before any encryption is done, while the
|
|
|
4a2fec |
in-QEMU driver must encrypt the data before completion.
|
|
|
4a2fec |
|
|
|
4a2fec |
Signed-off-by: Daniel P. Berrange <berrange@redhat.com>
|
|
|
4a2fec |
Message-id: 20170927125340.12360-2-berrange@redhat.com
|
|
|
4a2fec |
Reviewed-by: Eric Blake <eblake@redhat.com>
|
|
|
4a2fec |
Reviewed-by: Max Reitz <mreitz@redhat.com>
|
|
|
4a2fec |
Signed-off-by: Max Reitz <mreitz@redhat.com>
|
|
|
4a2fec |
(cherry picked from commit 161253e2d0a83a1b33bca019c6e926013e1a03db)
|
|
|
4a2fec |
Signed-off-by: Miroslav Rezanina <mrezanin@redhat.com>
|
|
|
4a2fec |
---
|
|
|
4a2fec |
block/crypto.c | 28 +++++++++++++++-------------
|
|
|
4a2fec |
1 file changed, 15 insertions(+), 13 deletions(-)
|
|
|
4a2fec |
|
|
|
4a2fec |
diff --git a/block/crypto.c b/block/crypto.c
|
|
|
4a2fec |
index 58ef6f2..684cabe 100644
|
|
|
4a2fec |
--- a/block/crypto.c
|
|
|
4a2fec |
+++ b/block/crypto.c
|
|
|
4a2fec |
@@ -379,7 +379,11 @@ static void block_crypto_close(BlockDriverState *bs)
|
|
|
4a2fec |
}
|
|
|
4a2fec |
|
|
|
4a2fec |
|
|
|
4a2fec |
-#define BLOCK_CRYPTO_MAX_SECTORS 32
|
|
|
4a2fec |
+/*
|
|
|
4a2fec |
+ * 1 MB bounce buffer gives good performance / memory tradeoff
|
|
|
4a2fec |
+ * when using cache=none|directsync.
|
|
|
4a2fec |
+ */
|
|
|
4a2fec |
+#define BLOCK_CRYPTO_MAX_IO_SIZE (1024 * 1024)
|
|
|
4a2fec |
|
|
|
4a2fec |
static coroutine_fn int
|
|
|
4a2fec |
block_crypto_co_readv(BlockDriverState *bs, int64_t sector_num,
|
|
|
4a2fec |
@@ -396,12 +400,11 @@ block_crypto_co_readv(BlockDriverState *bs, int64_t sector_num,
|
|
|
4a2fec |
|
|
|
4a2fec |
qemu_iovec_init(&hd_qiov, qiov->niov);
|
|
|
4a2fec |
|
|
|
4a2fec |
- /* Bounce buffer so we have a linear mem region for
|
|
|
4a2fec |
- * entire sector. XXX optimize so we avoid bounce
|
|
|
4a2fec |
- * buffer in case that qiov->niov == 1
|
|
|
4a2fec |
+ /* Bounce buffer because we don't wish to expose cipher text
|
|
|
4a2fec |
+ * in qiov which points to guest memory.
|
|
|
4a2fec |
*/
|
|
|
4a2fec |
cipher_data =
|
|
|
4a2fec |
- qemu_try_blockalign(bs->file->bs, MIN(BLOCK_CRYPTO_MAX_SECTORS * 512,
|
|
|
4a2fec |
+ qemu_try_blockalign(bs->file->bs, MIN(BLOCK_CRYPTO_MAX_IO_SIZE,
|
|
|
4a2fec |
qiov->size));
|
|
|
4a2fec |
if (cipher_data == NULL) {
|
|
|
4a2fec |
ret = -ENOMEM;
|
|
|
4a2fec |
@@ -411,8 +414,8 @@ block_crypto_co_readv(BlockDriverState *bs, int64_t sector_num,
|
|
|
4a2fec |
while (remaining_sectors) {
|
|
|
4a2fec |
cur_nr_sectors = remaining_sectors;
|
|
|
4a2fec |
|
|
|
4a2fec |
- if (cur_nr_sectors > BLOCK_CRYPTO_MAX_SECTORS) {
|
|
|
4a2fec |
- cur_nr_sectors = BLOCK_CRYPTO_MAX_SECTORS;
|
|
|
4a2fec |
+ if (cur_nr_sectors > (BLOCK_CRYPTO_MAX_IO_SIZE / 512)) {
|
|
|
4a2fec |
+ cur_nr_sectors = (BLOCK_CRYPTO_MAX_IO_SIZE / 512);
|
|
|
4a2fec |
}
|
|
|
4a2fec |
|
|
|
4a2fec |
qemu_iovec_reset(&hd_qiov);
|
|
|
4a2fec |
@@ -464,12 +467,11 @@ block_crypto_co_writev(BlockDriverState *bs, int64_t sector_num,
|
|
|
4a2fec |
|
|
|
4a2fec |
qemu_iovec_init(&hd_qiov, qiov->niov);
|
|
|
4a2fec |
|
|
|
4a2fec |
- /* Bounce buffer so we have a linear mem region for
|
|
|
4a2fec |
- * entire sector. XXX optimize so we avoid bounce
|
|
|
4a2fec |
- * buffer in case that qiov->niov == 1
|
|
|
4a2fec |
+ /* Bounce buffer because we're not permitted to touch
|
|
|
4a2fec |
+ * contents of qiov - it points to guest memory.
|
|
|
4a2fec |
*/
|
|
|
4a2fec |
cipher_data =
|
|
|
4a2fec |
- qemu_try_blockalign(bs->file->bs, MIN(BLOCK_CRYPTO_MAX_SECTORS * 512,
|
|
|
4a2fec |
+ qemu_try_blockalign(bs->file->bs, MIN(BLOCK_CRYPTO_MAX_IO_SIZE,
|
|
|
4a2fec |
qiov->size));
|
|
|
4a2fec |
if (cipher_data == NULL) {
|
|
|
4a2fec |
ret = -ENOMEM;
|
|
|
4a2fec |
@@ -479,8 +481,8 @@ block_crypto_co_writev(BlockDriverState *bs, int64_t sector_num,
|
|
|
4a2fec |
while (remaining_sectors) {
|
|
|
4a2fec |
cur_nr_sectors = remaining_sectors;
|
|
|
4a2fec |
|
|
|
4a2fec |
- if (cur_nr_sectors > BLOCK_CRYPTO_MAX_SECTORS) {
|
|
|
4a2fec |
- cur_nr_sectors = BLOCK_CRYPTO_MAX_SECTORS;
|
|
|
4a2fec |
+ if (cur_nr_sectors > (BLOCK_CRYPTO_MAX_IO_SIZE / 512)) {
|
|
|
4a2fec |
+ cur_nr_sectors = (BLOCK_CRYPTO_MAX_IO_SIZE / 512);
|
|
|
4a2fec |
}
|
|
|
4a2fec |
|
|
|
4a2fec |
qemu_iovec_to_buf(qiov, bytes_done,
|
|
|
4a2fec |
--
|
|
|
4a2fec |
1.8.3.1
|
|
|
4a2fec |
|