diff -urN lvm-ref/drivers/md/lvm.c lvm/drivers/md/lvm.c
--- lvm-ref/drivers/md/lvm.c	Wed Jul 18 17:11:13 2001
+++ lvm/drivers/md/lvm.c	Wed Jul 18 17:15:00 2001
@@ -1111,11 +1111,33 @@
  * block device support function for /usr/src/linux/drivers/block/ll_rw_blk.c
  * (see init_module/lvm_init)
  */
-static inline void __remap_snapshot(kdev_t rdev, ulong rsector,
-				    ulong pe_start, lv_t *lv, vg_t *vg) {
+static void __remap_snapshot(kdev_t rdev, ulong rsector,
+			     ulong pe_start, lv_t *lv, vg_t *vg) {
+
+	/* copy a chunk from the origin to a snapshot device */
+	down_write(&lv->lv_lock);
+
+	/* we must redo lvm_snapshot_remap_block in order to avoid a
+	   race condition in the gap where no lock was held */
 	if (!lvm_snapshot_remap_block(&rdev, &rsector, pe_start, lv) &&
 	    !lvm_snapshot_COW(rdev, rsector, pe_start, rsector, vg, lv))
 		lvm_write_COW_table_block(vg, lv);
+
+	up_write(&lv->lv_lock);
+}
+
+static inline void _remap_snapshot(kdev_t rdev, ulong rsector,
+				   ulong pe_start, lv_t *lv, vg_t *vg) {
+	int r;
+
+	/* check to see if this chunk is already in the snapshot */
+	down_read(&lv->lv_lock);
+	r = lvm_snapshot_remap_block(&rdev, &rsector, pe_start, lv);
+	up_read(&lv->lv_lock);
+
+	if (!r)
+		/* we haven't yet copied this block to the snapshot */
+		__remap_snapshot(rdev, rsector, pe_start, lv, vg);
 }
 
 static int lvm_map(struct buffer_head *bh, int rw)
@@ -1131,7 +1153,7 @@
 	lv_t *lv = vg_this->lv[LV_BLK(minor)];
 
 
-	down(&lv->lv_snapshot_sem);
+	down_read(&lv->lv_lock);
 	if (!(lv->lv_status & LV_ACTIVE)) {
 		printk(KERN_ALERT
 		       "%s - lvm_map: ll_rw_blk for inactive LV %s\n",
@@ -1197,8 +1219,16 @@
 		      rsector_map, stripe_length, stripe_index);
 	}
 
+#if 0 /*
+       * drop the semaphore to avoid hurting I/O performance, pv_move
+       * is racy anyways in beta7. You are required to avoid pv_moves
+       * with writes going on the LV in beta7 anyways. So this doesn't make
+       * any reliability difference and it fixes the performance
+       * showstopper during production.
+       */
 	/* handle physical extents on the move */
 	down(&_pe_lock);
+#endif
 	if((pe_lock_req.lock == LOCK_PE) &&
 	   (rdev_map == pe_lock_req.data.pv_dev) &&
 	   (rsector_map >= pe_lock_req.data.pv_offset) &&
@@ -1209,17 +1239,23 @@
 			printk(KERN_ERR 
 			       "%s -- bh uses low 2 bits of pointer\n",
 			       lvm_name);
+#if 0
 			up(&_pe_lock);
+#endif
 			goto bad;
 		}
 
 		bh->b_reqnext = _pe_requests;
 		_pe_requests = (struct buffer_head *) ((int) bh | rw);
+#if 0
 		up(&_pe_lock);
-		up(&lv->lv_snapshot_sem);
+#endif
+		up_read(&lv->lv_lock);
 		return 0;
 	}
+#if 0
 	up(&_pe_lock);
+#endif
 
 	/* statistic */
 	if (rw == WRITE || rw == WRITEA)
@@ -1252,22 +1288,20 @@
 
 			/* Serializes the COW with the accesses to the
 			   snapshot device */
-			down(&snap->lv_snapshot_sem);
-			__remap_snapshot(rdev_map, rsector_map,
-					 pe_start, snap, vg_this);
-			up(&snap->lv_snapshot_sem);
+			_remap_snapshot(rdev_map, rsector_map,
+					pe_start, snap, vg_this);
 		}
 	}
 
  out:
 	bh->b_rdev = rdev_map;
 	bh->b_rsector = rsector_map;
-	up(&lv->lv_snapshot_sem);
+	up_read(&lv->lv_lock);
 	return 1;
 
  bad:
 	buffer_IO_error(bh);
-	up(&lv->lv_snapshot_sem);
+	up_read(&lv->lv_lock);
 	return -1;
 } /* lvm_map() */
 
@@ -1364,9 +1398,8 @@
 		}
 		if (p == vg_ptr->pv_max) return -ENXIO;
 
-		pe_lock_req = new_lock;
-
 		down(&_pe_lock);
+		pe_lock_req = new_lock;
 		pe_lock_req.lock = UNLOCK_PE;
 		up(&_pe_lock);
 
@@ -1865,7 +1898,7 @@
 	lv_ptr->lv_snapshot_hash_table = NULL;
 	lv_ptr->lv_snapshot_hash_table_size = 0;
 	lv_ptr->lv_snapshot_hash_mask = 0;
-	init_MUTEX(&lv_ptr->lv_snapshot_sem);
+	init_rwsem(&lv_ptr->lv_lock);
 	lv_ptr->lv_snapshot_use_rate = 0;
 
 	vg_ptr->lv[l] = lv_ptr;
@@ -2007,7 +2040,7 @@
 		fsync_dev_lockfs(org->lv_dev);
 #endif
 
-		down(&org->lv_snapshot_sem);
+		down_write(&org->lv_lock);
 		org->lv_access |= LV_SNAPSHOT_ORG;
 		lv_ptr->lv_access &= ~LV_SNAPSHOT_ORG; /* this can only hide an userspace bug */
 
@@ -2016,7 +2049,7 @@
 		for (last = org; last->lv_snapshot_next; last = last->lv_snapshot_next);
 		lv_ptr->lv_snapshot_prev = last;
 		last->lv_snapshot_next = lv_ptr;
-		up(&org->lv_snapshot_sem);
+		up_write(&org->lv_lock);
 	}
 
 	/* activate the logical volume */
@@ -2085,7 +2118,7 @@
 		 * to the original lv before playing with it.
 		 */
 		lv_t * org = lv_ptr->lv_snapshot_org;
-		down(&org->lv_snapshot_sem);
+		down_write(&org->lv_lock);
 
 		/* remove this snapshot logical volume from the chain */
 		lv_ptr->lv_snapshot_prev->lv_snapshot_next = lv_ptr->lv_snapshot_next;
@@ -2098,7 +2131,7 @@
 		if (!org->lv_snapshot_next) {
 			org->lv_access &= ~LV_SNAPSHOT_ORG;
 		}
-		up(&org->lv_snapshot_sem);
+		up_write(&org->lv_lock);
 
 		lvm_snapshot_release(lv_ptr);
 
@@ -2295,7 +2328,7 @@
 		return r;
 
 	/* copy relevent fields */
-	down(&old_lv->lv_snapshot_sem);
+	down_write(&old_lv->lv_lock);
 
 	if(new_lv->lv_access & LV_SNAPSHOT) {
 
@@ -2337,7 +2370,7 @@
 			lv_t *snap;
 			for(snap = old_lv->lv_snapshot_next; snap;
 			    snap = snap->lv_snapshot_next) {
-				down(&snap->lv_snapshot_sem);
+				down_write(&snap->lv_lock);
 				snap->lv_current_pe = old_lv->lv_current_pe;
 				snap->lv_allocated_le =
 					old_lv->lv_allocated_le;
@@ -2349,13 +2382,13 @@
 				lvm_size[MINOR(snap->lv_dev)] =
 					old_lv->lv_size >> 1;
 				__update_hardblocksize(snap);
-				up(&snap->lv_snapshot_sem);
+				up_write(&snap->lv_lock);
 			}
 		}
 	}
 
 	__update_hardblocksize(old_lv);
-	up(&old_lv->lv_snapshot_sem);
+	up_write(&old_lv->lv_lock);
 
 	return 0;
 } /* lvm_do_lv_extend_reduce() */
diff -urN lvm-ref/include/linux/lvm.h lvm/include/linux/lvm.h
--- lvm-ref/include/linux/lvm.h	Wed Jul 18 17:11:13 2001
+++ lvm/include/linux/lvm.h	Wed Jul 18 17:11:45 2001
@@ -650,7 +650,7 @@
 #ifdef __KERNEL__
 	struct kiobuf *lv_iobuf;
 	struct kiobuf *lv_COW_table_iobuf;
-	struct semaphore lv_snapshot_sem;
+	struct rw_semaphore lv_lock;
 	struct list_head *lv_snapshot_hash_table;
 	uint32_t lv_snapshot_hash_table_size;
 	uint32_t lv_snapshot_hash_mask;