Aren
/
NetBSD
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

according to benchmark extracting pkgsrc.tar, using FUA and hence waiting 

for each transfer to write through to the medium is way slower than just
letting the drive use a cached write and doing DIOCCACHESYNC on the end

Results were (fs block 32KB / frag 4KB, partition aligned on 32KB boundary):
HDD at siisata(4):  no-FUA: 108 sec w/FUA: 294 sec
SSD at ahcisata(4): no-FUA:  73 sec w/FUA: 502 sec

change the flag so that FUA is only used for the commit block write;
for journal data write, only pass DPO, rely on the cache flush to get them
to media
This commit is contained in:
jdolecek 2017-12-02 17:29:55 +00:00
parent 7398143904
commit 508f8978c8
1 changed files with 27 additions and 24 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

51
sys/kern/vfs_wapbl.c
View File

@ -1,4 +1,4 @@
/* $NetBSD: vfs_wapbl.c,v 1.100 2017/10/27 12:25:15 joerg Exp $ */
/* $NetBSD: vfs_wapbl.c,v 1.101 2017/12/02 17:29:55 jdolecek Exp $ */
/*-
* Copyright (c) 2003, 2008, 2009 The NetBSD Foundation, Inc.
@ -36,7 +36,7 @@
#define WAPBL_INTERNAL
#include <sys/cdefs.h>
__KERNEL_RCSID(0, "$NetBSD: vfs_wapbl.c,v 1.100 2017/10/27 12:25:15 joerg Exp $");
__KERNEL_RCSID(0, "$NetBSD: vfs_wapbl.c,v 1.101 2017/12/02 17:29:55 jdolecek Exp $");
#include <sys/param.h>
#include <sys/bitops.h>
@ -239,10 +239,9 @@ struct wapbl {
(wapbl_allow_dpofua && ISSET((wl)->wl_dkcache, DKCACHE_FUA))
#define WAPBL_JFLAGS(wl) \
(WAPBL_USE_FUA(wl) ? (wl)->wl_jwrite_flags : 0)
#define WAPBL_MFLAGS(wl) \
(WAPBL_USE_FUA(wl) ? (wl)->wl_mwrite_flags : 0)
#define WAPBL_JDATA_FLAGS(wl) \
(WAPBL_JFLAGS(wl) & B_MEDIA_DPO) /* only DPO */
int wl_jwrite_flags; /* r: journal write flags */
int wl_mwrite_flags; /* r: metadata write flags */
};
#ifdef WAPBL_DEBUG_PRINT
@ -444,10 +443,8 @@ wapbl_dkcache_init(struct wapbl *wl)
}
/* Use FUA instead of cache flush if available */
if (ISSET(wl->wl_dkcache, DKCACHE_FUA)) {
if (ISSET(wl->wl_dkcache, DKCACHE_FUA))
wl->wl_jwrite_flags |= B_MEDIA_FUA;
wl->wl_mwrite_flags |= B_MEDIA_FUA;
}
/* Use DPO for journal writes if available */
if (ISSET(wl->wl_dkcache, DKCACHE_DPO))
@ -998,7 +995,7 @@ wapbl_buffered_write_async(struct wapbl *wl, struct buf *bp)
KASSERT(TAILQ_FIRST(&wl->wl_iobufs) == bp);
TAILQ_REMOVE(&wl->wl_iobufs, bp, b_wapbllist);
bp->b_flags = B_WRITE | WAPBL_JFLAGS(wl);
bp->b_flags |= B_WRITE;
bp->b_cflags = BC_BUSY; /* mandatory, asserted by biowait() */
bp->b_oflags = 0;
bp->b_bcount = bp->b_resid;
@ -1043,7 +1040,7 @@ again:
error = biowait(bp);
/* reset for reuse */
bp->b_blkno = bp->b_resid = 0;
bp->b_blkno = bp->b_resid = bp->b_flags = 0;
TAILQ_INSERT_TAIL(&wl->wl_iobufs, bp, b_wapbllist);
found = true;
@ -1067,7 +1064,8 @@ again:
* wapbl_buffered_flush.
*/
static int
wapbl_buffered_write(void *data, size_t len, struct wapbl *wl, daddr_t pbn)
wapbl_buffered_write(void *data, size_t len, struct wapbl *wl, daddr_t pbn,
int bflags)
{
size_t resid;
struct buf *bp;
@ -1096,8 +1094,10 @@ again:
* If this write goes to an empty buffer we have to
* save the disk block address first.
*/
if (bp->b_blkno == 0)
if (bp->b_blkno == 0) {
bp->b_blkno = pbn;
bp->b_flags |= bflags;
}
/*
* Remaining space so this buffer ends on a buffer size boundary.
@ -1164,7 +1164,8 @@ wapbl_circ_write(struct wapbl *wl, void *data, size_t len, off_t *offp)
#ifdef _KERNEL
pbn = btodb(pbn << wl->wl_log_dev_bshift);
#endif
error = wapbl_buffered_write(data, slen, wl, pbn);
error = wapbl_buffered_write(data, slen, wl, pbn,
WAPBL_JDATA_FLAGS(wl));
if (error)
return error;
data = (uint8_t *)data + slen;
@ -1175,7 +1176,8 @@ wapbl_circ_write(struct wapbl *wl, void *data, size_t len, off_t *offp)
#ifdef _KERNEL
pbn = btodb(pbn << wl->wl_log_dev_bshift);
#endif
error = wapbl_buffered_write(data, len, wl, pbn);
error = wapbl_buffered_write(data, len, wl, pbn,
WAPBL_JDATA_FLAGS(wl));
if (error)
return error;
off += len;
@ -1925,9 +1927,6 @@ wapbl_flush(struct wapbl *wl, int waitfor)
bp->b_iodone = wapbl_biodone;
bp->b_private = we;
/* make sure the block is saved sync when FUA in use */
bp->b_flags |= WAPBL_MFLAGS(wl);
bremfree(bp);
wapbl_remove_buf_locked(wl, bp);
mutex_exit(&wl->wl_mtx);
@ -2399,8 +2398,8 @@ wapbl_cache_sync(struct wapbl *wl, const char *msg)
int force = 1;
int error;
/* Skip full cache sync if disabled, or when using FUA */
if (!wapbl_flush_disk_cache || WAPBL_USE_FUA(wl)) {
/* Skip full cache sync if disabled */
if (!wapbl_flush_disk_cache) {
return 0;
}
if (verbose) {
@ -2459,8 +2458,10 @@ wapbl_write_commit(struct wapbl *wl, off_t head, off_t tail)
if (error)
return error;
/*
* flush disk cache to ensure that blocks we've written are actually
* Flush disk cache to ensure that blocks we've written are actually
* written to the stable storage before the commit header.
* This flushes to disk not only journal blocks, but also all
* metadata blocks, written asynchronously since previous commit.
*
* XXX Calc checksum here, instead we do this for now
*/
@ -2489,7 +2490,7 @@ wapbl_write_commit(struct wapbl *wl, off_t head, off_t tail)
#ifdef _KERNEL
pbn = btodb(pbn << wc->wc_log_dev_bshift);
#endif
error = wapbl_buffered_write(wc, wc->wc_len, wl, pbn);
error = wapbl_buffered_write(wc, wc->wc_len, wl, pbn, WAPBL_JFLAGS(wl));
if (error)
return error;
error = wapbl_buffered_flush(wl, true);
@ -2497,10 +2498,12 @@ wapbl_write_commit(struct wapbl *wl, off_t head, off_t tail)
return error;
/*
* flush disk cache to ensure that the commit header is actually
* written before meta data blocks.
* Flush disk cache to ensure that the commit header is actually
* written before meta data blocks. Commit block is written using
* FUA when enabled, in that case this flush is not needed.
*/
wapbl_cache_sync(wl, "2");
if (!WAPBL_USE_FUA(wl))
wapbl_cache_sync(wl, "2");
/*
* If the generation number was zero, write it out a second time.