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

Write an essay explaining why ffs_write is one huge WAPBL transaction.

Browse Source
This commit is contained in:
riastradh 2015-03-31 00:22:50 +00:00
parent c6a78cfe85
commit e9f7453834
1 changed files with 24 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

26
sys/ufs/ufs/ufs_readwrite.c
View File

@ -1,4 +1,4 @@
/* $NetBSD: ufs_readwrite.c,v 1.116 2015/03/29 14:39:41 riastradh Exp $ */
/* $NetBSD: ufs_readwrite.c,v 1.117 2015/03/31 00:22:50 riastradh Exp $ */
/*-
* Copyright (c) 1993
@ -32,7 +32,7 @@
*/
#include <sys/cdefs.h>
__KERNEL_RCSID(1, "$NetBSD: ufs_readwrite.c,v 1.116 2015/03/29 14:39:41 riastradh Exp $");
__KERNEL_RCSID(1, "$NetBSD: ufs_readwrite.c,v 1.117 2015/03/31 00:22:50 riastradh Exp $");
#ifdef LFS_READWRITE
#define FS struct lfs
@ -328,6 +328,28 @@ WRITE(void *v)
KASSERT(vp->v_type == VREG);
/*
* XXX The entire write operation must occur in a single WAPBL
* transaction because it may allocate disk blocks, if
* appending or filling holes, which is allowed to happen only
* if the write fully succeeds.
*
* If ubc_uiomove fails in the middle with EFAULT, we can clean
* up at the end with UFS_TRUNCATE. But if the power fails in
* the middle, there would be nobody to deallocate the blocks,
* without an fsck to globally analyze the file system.
*
* If the increasingly inaccurately named WAPBL were augmented
* with rollback records for block allocations, then we could
* split this into multiple transactions and commit the
* allocations in the last one.
*
* But WAPBL doesn't have that notion now, so we'll have to
* live with gigantic transactions and WAPBL tentacles in
* genfs_getpages/putpages to cope with the possibility that
* the transaction may or may not be locked on entry to the
* page cache.
*/
error = UFS_WAPBL_BEGIN(vp->v_mount);
if (error) {
fstrans_done(vp->v_mount);