On Wed, Nov 07, 2012 at 11:25:55AM -0700, Bob Beck wrote:
> Hello tech@.
>
> I just committed some significant fixes to the buffer cache in -current.
> (you want to look for version 1.139 of vfs_bio.c)
>
> If you have such a current kernel, you may wish to try the following
> diff - it is a backout of a previous commit that emasculated softdep
> in order to make is safe with our buffer cache at the time. With
> version 1.139 of vfs_bio.c I believe this is no longer needed.
>
> I have done extensive testing with the diff below and beating up
> softdep filesystems - it both works, and, actually can provide
> significantly increased performance in some situations.
>
> You too may want to try what is below. If you do, please drop me a
> note indicating what you were doing with it, and what you noticed,
> sucesses, failures, performance changes, etc. I would like to know
> about all.
>
> WARNING - if you apply this to something without version 1.139 of
> vfs_bio.c, you may hang your machine.
>
> Thanks
>
> -Bob
>
Been running on my main dev machine for the past day or so, no regressions
seen. (Note I'm referring to the previous version of the diff circulated).
-ml
>
> ---------------8<----------------
> Index: ufs/ffs/ffs_inode.c
> ===================================================================
> RCS file: /cvs/src/sys/ufs/ffs/ffs_inode.c,v
> retrieving revision 1.61
> diff -u -p -r1.61 ffs_inode.c
> --- ufs/ffs/ffs_inode.c 23 Mar 2012 15:51:26 -0000 1.61
> +++ ufs/ffs/ffs_inode.c 20 Sep 2012 13:27:05 -0000
> @@ -164,10 +164,8 @@ ffs_truncate(struct inode *oip, off_t le
> struct buf *bp;
> int offset, size, level;
> long count, nblocks, vflags, blocksreleased = 0;
> - int i, aflags, error, allerror, indirect = 0;
> + int i, aflags, error, allerror;
> off_t osize;
> - extern int num_indirdep;
> - extern int max_indirdep;
>
> if (length < 0)
> return (EINVAL);
> @@ -243,8 +241,6 @@ ffs_truncate(struct inode *oip, off_t le
> cred, aflags, &bp);
> if (error)
> return (error);
> - if (bp->b_lblkno >= NDADDR)
> - indirect = 1;
> DIP_ASSIGN(oip, size, length);
> uvm_vnp_setsize(ovp, length);
> (void) uvm_vnp_uncache(ovp);
> @@ -253,20 +249,7 @@ ffs_truncate(struct inode *oip, off_t le
> else
> bawrite(bp);
> oip->i_flag |= IN_CHANGE | IN_UPDATE;
> - error = UFS_UPDATE(oip, MNT_WAIT);
> - if (DOINGSOFTDEP(ovp) && num_indirdep > max_indirdep)
> - if (indirect) {
> - /*
> - * If the number of pending indirect block
> - * dependencies is sufficiently close to the
> - * maximum number of simultaneously mappable
> - * buffers force a sync on the vnode to prevent
> - * buffer cache exhaustion.
> - */
> - VOP_FSYNC(ovp, curproc->p_ucred, MNT_WAIT,
> - curproc);
> - }
> - return (error);
> + return (UFS_UPDATE(oip, MNT_WAIT));
> }
> uvm_vnp_setsize(ovp, length);
>
> Index: ufs/ffs/ffs_softdep.c
> ===================================================================
> RCS file: /cvs/src/sys/ufs/ffs/ffs_softdep.c,v
> retrieving revision 1.112
> diff -u -p -r1.112 ffs_softdep.c
> --- ufs/ffs/ffs_softdep.c 18 Sep 2011 23:20:28 -0000 1.112
> +++ ufs/ffs/ffs_softdep.c 20 Sep 2012 13:27:05 -0000
> @@ -526,8 +526,6 @@ workitem_free(struct worklist *item)
> STATIC struct workhead softdep_workitem_pending;
> STATIC struct worklist *worklist_tail;
> STATIC int num_on_worklist; /* number of worklist items to be processed */
> -STATIC int num_indirdep; /* number of indirdep items to be processed */
> -STATIC int max_indirdep; /* maximum number of indirdep items allowed */
> STATIC int softdep_worklist_busy; /* 1 => trying to do unmount */
> STATIC int softdep_worklist_req; /* serialized waiters */
> STATIC int max_softdeps; /* maximum number of structs before slowdown */
> @@ -1142,8 +1140,6 @@ top:
> void
> softdep_initialize(void)
> {
> - extern vsize_t bufkvm;
> - max_indirdep = (int)bufkvm / MAXPHYS * 80 / 100;
>
> bioops.io_start = softdep_disk_io_initiation;
> bioops.io_complete = softdep_disk_write_complete;
> @@ -1819,14 +1815,12 @@ setup_allocindir_phase2(struct buf *bp,
> if (indirdep->ir_savebp != NULL)
> brelse(newindirdep->ir_savebp);
> WORKITEM_FREE(newindirdep, D_INDIRDEP);
> - num_indirdep--;
> }
> if (indirdep)
> break;
> newindirdep = pool_get(&indirdep_pool, PR_WAITOK);
> newindirdep->ir_list.wk_type = D_INDIRDEP;
> newindirdep->ir_state = ATTACHED;
> - num_indirdep++;
> if (ip->i_ump->um_fstype == UM_UFS1)
> newindirdep->ir_state |= UFS1FMT;
> LIST_INIT(&newindirdep->ir_deplisthd);
> @@ -2446,7 +2440,6 @@ indir_trunc(struct inode *ip, daddr64_t
> }
> WORKLIST_REMOVE(wk);
> WORKITEM_FREE(indirdep, D_INDIRDEP);
> - num_indirdep--;
> if (LIST_FIRST(&bp->b_dep) != NULL) {
> FREE_LOCK(&lk);
> panic("indir_trunc: dangling dep");
> @@ -3288,7 +3281,6 @@ softdep_disk_io_initiation(struct buf *b
> wk->wk_state &= ~ONWORKLIST;
> LIST_REMOVE(wk, wk_list);
> WORKITEM_FREE(indirdep, D_INDIRDEP);
> - num_indirdep--;
> FREE_LOCK(&lk);
> brelse(sbp);
> ACQUIRE_LOCK(&lk);
> Index: ufs/ffs/ffs_vnops.c
> ===================================================================
> RCS file: /cvs/src/sys/ufs/ffs/ffs_vnops.c,v
> retrieving revision 1.69
> diff -u -p -r1.69 ffs_vnops.c
> --- ufs/ffs/ffs_vnops.c 17 Sep 2012 14:22:23 -0000 1.69
> +++ ufs/ffs/ffs_vnops.c 20 Sep 2012 13:37:00 -0000
> @@ -295,12 +295,10 @@ ffs_write(void *v)
> struct inode *ip;
> struct fs *fs;
> struct buf *bp;
> - struct proc *p;
> daddr64_t lbn;
> off_t osize;
> int blkoffset, error, extended, flags, ioflag, resid, size, xfersize;
> int overrun;
> - extern int num_indirdep, max_indirdep;
>
> extended = 0;
> ioflag = ap->a_ioflag;
> @@ -367,11 +365,7 @@ ffs_write(void *v)
> if (uio->uio_offset + xfersize > DIP(ip, size)) {
> DIP_ASSIGN(ip, size, uio->uio_offset + xfersize);
> uvm_vnp_setsize(vp, DIP(ip, size));
> - /* Are we extending into an indirect block? */
> - if (bp->b_lblkno < NDADDR)
> - extended = 1;
> - else
> - extended = 2;
> + extended = 1;
> }
> (void)uvm_vnp_uncache(vp);
>
> @@ -415,21 +409,8 @@ ffs_write(void *v)
> uio->uio_offset -= resid - uio->uio_resid;
> uio->uio_resid = resid;
> }
> - } else if (resid > uio->uio_resid) {
> - if (ioflag & IO_SYNC)
> - error = UFS_UPDATE(ip, MNT_WAIT);
> - if (DOINGSOFTDEP(vp) && num_indirdep > max_indirdep)
> - if (extended > 1) {
> - /*
> - * If the number of pending indirect block
> - * dependencies is sufficiently close to the
> - * maximum number of simultaneously mappable
> - * buffers force a sync on the vnode to prevent
> - * buffer cache exhaustion.
> - */
> - p = uio->uio_procp;
> - VOP_FSYNC(vp, NULL, MNT_WAIT, p);
> - }
> + } else if (resid > uio->uio_resid && (ioflag & IO_SYNC)) {
> + error = UFS_UPDATE(ip, MNT_WAIT);
> }
> /* correct the result for writes clamped by vn_fsizechk() */
> uio->uio_resid += overrun;
