remove v_number
PATCH 1/1 remove v_number
v_number is unused except for the naming some locks (which is a functionality totally unused by Linux), so remove it and assorted crap. Besides saving two words in struct vnode this also gets rid of a spinlock per inode allocation.
fs/xfs/linux-2.6/xfs_vnode.c +0 -8
@@ -17,8 +17,6 @@*/ #include "xfs.h" -uint64_t vn_generation;/* vnode generation number */ -DEFINE_SPINLOCK (vnumber_lock);/* * Dedicated vnode inactive/reclaim sync semaphores.
@@ -82,12 +80,6 @@ vp->v_flag =VMODIFIED ;spinlock_init (&vp->v_lock,"v_lock" ); -spin_lock (&vnumber_lock); -if (!++vn_generation)/* v_number shouldn't be zero */ - vn_generation++; - vp->v_number = vn_generation; -spin_unlock (&vnumber_lock); -ASSERT (VN_CACHED (vp) ==0 );/* Initialize the first behavior and the behavior chain head. */
fs/xfs/linux-2.6/xfs_vnode.h +0 -1
@@ -41,7 +41,6 @@ typedef struct bhv_vnode { bhv_vflags_t v_flag; /* vnode flags (see above) */ bhv_vfs_t *v_vfsp; /* ptr to containing VFS */ - bhv_vnumber_t v_number; /* in-core vnode number */ bhv_head_t v_bh; /* behavior head */ spinlock_t v_lock; /* VN_LOCK/VN_UNLOCK */ atomic_t v_iocount; /* outstanding I/O count */
fs/xfs/xfs_iget.c +2 -2
@@ -570,8 +570,8 @@ bhv_vnode_t *vp) {mrlock_init (&ip->i_lock,MRLOCK_ALLOW_EQUAL_PRI |MRLOCK_BARRIER , -"xfsino" , (long )vp->v_number); -mrlock_init (&ip->i_iolock,MRLOCK_BARRIER ,"xfsio" , vp->v_number); +"xfsino" , ip->i_ino); +mrlock_init (&ip->i_iolock,MRLOCK_BARRIER ,"xfsio" , ip->i_ino);init_waitqueue_head (&ip->i_ipin_wait);atomic_set (&ip->i_pincount,0 );initnsema (&ip->i_flock,1 ,"xfsfino" );
Quoted 4 lines · Christoph Hellwig
v_numberis unused except for the naming some locks (which is a functionality totally unused by Linux), so remove it and assorted crap. Besides saving two words in struct vnode this also gets rid of a spinlock per inode allocation.
Hmm - given that I've just used the v_number in post-mortem analysis of a nasty bug to correlate the sequence of events during a series of mkdir operations (i.e. transactions in the incore log buffers, the resulting xfs_inodes and some screwed up dentries) that lead to a BUG_ON being tripped in d_instantiate.
So, while it appears to be unused, it is _very_ useful for determining the SOE that has occurred in certain types of problems.
FWIW, while analysing this crash dump a couple of days ago I was wishing that dentries had an equivalent sequence number because there is no way to tell what dentry was supposed to be related to what inode after it got screwed up...
Cheers,
Dave.
Signature · 3 lines
Dave Chinner Principal Engineer SGI Australian Software Group
Quoted 12 lines · David Chinner
Quoted 4 lines · Christoph Hellwig
v_numberis unused except for the naming some locks (which is a functionality totally unused by Linux), so remove it and assorted crap. Besides saving two words in struct vnode this also gets rid of a spinlock per inode allocation.Hmm - given that I've just used the
v_numberin post-mortem analysis of a nasty bug to correlate the sequence of events during a series of mkdir operations (i.e. transactions in the incore log buffers, the resultingxfs_inodesand some screwed up dentries) that lead to aBUG_ONbeing tripped ind_instantiate.
Quoted 7 lines
So, while it appears to be unused, it is _very_ useful for determining the SOE that has occurred in certain types of problems.
FWIW, while analysing this crash dump a couple of days ago I was wishing that dentries had an equivalent sequence number because there is no way to tell what dentry was supposed to be related to what inode after it got screwed up...
Putting in sequence counting is trivial using kprobes. Will you put in this patch after I write you a kprobes modules to do the sequence numbering?
Quoted 16 lines · Christoph Hellwig
Quoted 15 lines · David Chinner
Quoted 4 lines · Christoph Hellwig
v_numberis unused except for the naming some locks (which is a functionality totally unused by Linux), so remove it and assorted crap. Besides saving two words in struct vnode this also gets rid of a spinlock per inode allocation.Hmm - given that I've just used the
v_numberin post-mortem analysis of a nasty bug to correlate the sequence of events during a series of mkdir operations (i.e. transactions in the incore log buffers, the resultingxfs_inodesand some screwed up dentries) that lead to aBUG_ONbeing tripped ind_instantiate.So, while it appears to be unused, it is _very_ useful for determining the SOE that has occurred in certain types of problems.
Quoted 6 lines
Quoted 4 lines
FWIW, while analysing this crash dump a couple of days ago I was wishing that dentries had an equivalent sequence number because there is no way to tell what dentry was supposed to be related to what inode after it got screwed up...
Putting in sequence counting is trivial using kprobes.
Never used kprobes - I'll take your word that it's easy....
Quoted 3 lines
Will you put
in this patch after I write you a kprobes modules to do the sequence numbering?
I wasn't NACKing your patch - I was making the point that this apparently unused code does have some useful, non-obvious, redeeming features.
Pardon my lack of kprobes knowledge, but what is the point of removing this code if we have to then go and add kprobes instrumentation to every XFS installation we support to provide the same information? Where does the kprobes data get kept? In the kernel, or partially in kernel and in userspace? How would you correlate the sequence number and the vnode from the crash dump? How much extra runtime memory does it use?
I'm not saying no, Christoph - I'm just trying to understand the implications of what you are suggesting....
Cheers,
Dave.
Signature · 3 lines
Dave Chinner Principal Engineer SGI Australian Software Group