Re: [PATCH 1/2 linux-next] Revert "ufs: fix deadlocks introduced by sb mutex merge"

[Fabian Frederick <fabf@skynet.be>]

[-- Attachment #1: Type: text/plain, Size: 2821 bytes --]



> On 04 June 2015 at 07:01 Al Viro <viro@ZenIV.linux.org.uk> wrote:
>
>
> On Wed, May 27, 2015 at 02:57:35PM -0700, Andrew Morton wrote:
> > On Wed, 27 May 2015 21:15:30 +0200 Fabian Frederick <fabf@skynet.be> wrote:
> >
> > > This reverts commit 9ef7db7f38d0
> > > ("ufs: fix deadlocks introduced by sb mutex merge")
> > > That patch tried to solve
> > >     Commit 0244756edc4b98c
> > >     ("ufs: sb mutex merge + mutex_destroy")
> > > which is itself partially reverted due to multiple deadlocks
> >
> > This is all very vague.  The changelogs are missing any description of
> > the deadlocks: how they are triggered, why they occur.  And there's no
> > description of how the patches fix these deadlocks.  And as we're
> > reverting a bunch of things one wonders whether the problems which the
> > now-reverted patches fixed are being reintroduced.
> >
> > Has anyone (Ian?) confirmed that the fs works OK with these patches?
>
> Folks, how about we figure out what's really being protected by that
> mutex?  IIRC, the main irregularity about ufs is the need to deal with
> growing the partial final block - unlike e.g ext2, ufs has differently-sized
> blocks and fragments.  Basically, it's tail-packing - short files have
> the last used direct pointer refer to a group of adjacent fragments that
> doesn't have to be block-aligned or fill the entire block.  They can't
> cross the disk block boundary and write might have to reallocate the partial
> block.
>
> So we need
>       * per-page exclusion for reallocation time (normal page locks are
> doing that)
>       * per-fs exclusion for block and fragment allocations (->s_lock?)
>       * per-fs exclusion for inode allocations (->s_lock?)
>       * per-inode exclusion for mapping changes (a-la ext2 truncate_mutex)
>       * per-directory exclusion for contents access (->i_mutex gives that)
>
> Looks like we ought to add ->truncate_mutex and shove lock_ufs() calls
> all way down into balloc.c (and ialloc.c for inode allocations)...

If we look at linux-next with the original mutex behavior restored,
mutex/spinlocks are the following:

struct ufs_sb_info{

        struct mutex mutex; (lock_ufs()/unlock_ufs())
        struct mutex s_lock; (mutex_lock()/mutex_unlock())
        spinlock_t work_lock; /* protects sync_work and work_queued */

}

You're asking to remove lock_ufs() in allocation and replace it by
truncate_mutex. I guess you're talking about doing that on current rc
(without s_lock restored).

I tried a quick patch on rc trying to convert lock_ufs()/unlock_ufs()
with per inode truncate_mutex (see attachment).
Is it going the right direction ? That would involve dropping the two linux-next
reverts in ufs.

Regards,
Fabian

[-- Attachment #2: ufsmutex2 --]
[-- Type: application/octet-stream, Size: 7541 bytes --]

diff --git a/fs/ufs/balloc.c b/fs/ufs/balloc.c
index 2c10360..23f9daf 100644
--- a/fs/ufs/balloc.c
+++ b/fs/ufs/balloc.c
@@ -40,6 +40,7 @@ void ufs_free_fragments(struct inode *inode, u64 fragment, unsigned count)
 	struct ufs_sb_private_info * uspi;
 	struct ufs_cg_private_info * ucpi;
 	struct ufs_cylinder_group * ucg;
+	struct ufs_inode_info *ei = UFS_I(inode);
 	unsigned cgno, bit, end_bit, bbase, blkmap, i;
 	u64 blkno;
 	
@@ -52,7 +53,7 @@ void ufs_free_fragments(struct inode *inode, u64 fragment, unsigned count)
 	if (ufs_fragnum(fragment) + count > uspi->s_fpg)
 		ufs_error (sb, "ufs_free_fragments", "internal error");
 	
-	lock_ufs(sb);
+	mutex_lock(&ei->truncate_mutex);
 	
 	cgno = ufs_dtog(uspi, fragment);
 	bit = ufs_dtogd(uspi, fragment);
@@ -116,12 +117,12 @@ void ufs_free_fragments(struct inode *inode, u64 fragment, unsigned count)
 		ubh_sync_block(UCPI_UBH(ucpi));
 	ufs_mark_sb_dirty(sb);
 	
-	unlock_ufs(sb);
+	mutex_unlock(&ei->truncate_mutex);
 	UFSD("EXIT\n");
 	return;
 
 failed:
-	unlock_ufs(sb);
+	mutex_unlock(&ei->truncate_mutex);
 	UFSD("EXIT (FAILED)\n");
 	return;
 }
@@ -135,6 +136,7 @@ void ufs_free_blocks(struct inode *inode, u64 fragment, unsigned count)
 	struct ufs_sb_private_info * uspi;
 	struct ufs_cg_private_info * ucpi;
 	struct ufs_cylinder_group * ucg;
+	struct ufs_inode_info *ei = UFS_I(inode);
 	unsigned overflow, cgno, bit, end_bit, i;
 	u64 blkno;
 	
@@ -151,7 +153,7 @@ void ufs_free_blocks(struct inode *inode, u64 fragment, unsigned count)
 		goto failed;
 	}
 
-	lock_ufs(sb);
+	mutex_lock(&ei->truncate_mutex);
 	
 do_more:
 	overflow = 0;
@@ -211,12 +213,12 @@ do_more:
 	}
 
 	ufs_mark_sb_dirty(sb);
-	unlock_ufs(sb);
+	mutex_unlock(&ei->truncate_mutex);
 	UFSD("EXIT\n");
 	return;
 
 failed_unlock:
-	unlock_ufs(sb);
+	mutex_unlock(&ei->truncate_mutex);
 failed:
 	UFSD("EXIT (FAILED)\n");
 	return;
@@ -345,6 +347,7 @@ u64 ufs_new_fragments(struct inode *inode, void *p, u64 fragment,
 	struct super_block * sb;
 	struct ufs_sb_private_info * uspi;
 	struct ufs_super_block_first * usb1;
+	struct ufs_inode_info *ei = UFS_I(inode);
 	unsigned cgno, oldcount, newcount;
 	u64 tmp, request, result;
 	
@@ -357,7 +360,7 @@ u64 ufs_new_fragments(struct inode *inode, void *p, u64 fragment,
 	usb1 = ubh_get_usb_first(uspi);
 	*err = -ENOSPC;
 
-	lock_ufs(sb);
+	mutex_lock(&ei->truncate_mutex);
 	tmp = ufs_data_ptr_to_cpu(sb, p);
 
 	if (count + ufs_fragnum(fragment) > uspi->s_fpb) {
@@ -378,19 +381,19 @@ u64 ufs_new_fragments(struct inode *inode, void *p, u64 fragment,
 				  "fragment %llu, tmp %llu\n",
 				  (unsigned long long)fragment,
 				  (unsigned long long)tmp);
-			unlock_ufs(sb);
+			mutex_unlock(&ei->truncate_mutex);
 			return INVBLOCK;
 		}
 		if (fragment < UFS_I(inode)->i_lastfrag) {
 			UFSD("EXIT (ALREADY ALLOCATED)\n");
-			unlock_ufs(sb);
+			mutex_unlock(&ei->truncate_mutex);
 			return 0;
 		}
 	}
 	else {
 		if (tmp) {
 			UFSD("EXIT (ALREADY ALLOCATED)\n");
-			unlock_ufs(sb);
+			mutex_unlock(&ei->truncate_mutex);
 			return 0;
 		}
 	}
@@ -399,7 +402,7 @@ u64 ufs_new_fragments(struct inode *inode, void *p, u64 fragment,
 	 * There is not enough space for user on the device
 	 */
 	if (!capable(CAP_SYS_RESOURCE) && ufs_freespace(uspi, UFS_MINFREE) <= 0) {
-		unlock_ufs(sb);
+		mutex_unlock(&ei->truncate_mutex);
 		UFSD("EXIT (FAILED)\n");
 		return 0;
 	}
@@ -424,7 +427,7 @@ u64 ufs_new_fragments(struct inode *inode, void *p, u64 fragment,
 			ufs_clear_frags(inode, result + oldcount,
 					newcount - oldcount, locked_page != NULL);
 		}
-		unlock_ufs(sb);
+		mutex_lock(&ei->truncate_mutex);
 		UFSD("EXIT, result %llu\n", (unsigned long long)result);
 		return result;
 	}
@@ -439,7 +442,7 @@ u64 ufs_new_fragments(struct inode *inode, void *p, u64 fragment,
 						fragment + count);
 		ufs_clear_frags(inode, result + oldcount, newcount - oldcount,
 				locked_page != NULL);
-		unlock_ufs(sb);
+		mutex_unlock(&ei->truncate_mutex);
 		UFSD("EXIT, result %llu\n", (unsigned long long)result);
 		return result;
 	}
@@ -477,7 +480,7 @@ u64 ufs_new_fragments(struct inode *inode, void *p, u64 fragment,
 		*err = 0;
 		UFS_I(inode)->i_lastfrag = max(UFS_I(inode)->i_lastfrag,
 						fragment + count);
-		unlock_ufs(sb);
+		mutex_unlock(&ei->truncate_mutex);
 		if (newcount < request)
 			ufs_free_fragments (inode, result + newcount, request - newcount);
 		ufs_free_fragments (inode, tmp, oldcount);
@@ -485,7 +488,7 @@ u64 ufs_new_fragments(struct inode *inode, void *p, u64 fragment,
 		return result;
 	}
 
-	unlock_ufs(sb);
+	mutex_unlock(&ei->truncate_mutex);
 	UFSD("EXIT (FAILED)\n");
 	return 0;
 }		
diff --git a/fs/ufs/ialloc.c b/fs/ufs/ialloc.c
index 7caa016..49bd9c4 100644
--- a/fs/ufs/ialloc.c
+++ b/fs/ufs/ialloc.c
@@ -59,6 +59,7 @@ void ufs_free_inode (struct inode * inode)
 	struct ufs_sb_private_info * uspi;
 	struct ufs_cg_private_info * ucpi;
 	struct ufs_cylinder_group * ucg;
+	struct ufs_inode_info *ei = UFS_I(inode);
 	int is_directory;
 	unsigned ino, cg, bit;
 	
@@ -69,11 +70,11 @@ void ufs_free_inode (struct inode * inode)
 	
 	ino = inode->i_ino;
 
-	lock_ufs(sb);
+	mutex_lock(&ei->truncate_mutex);
 
 	if (!((ino > 1) && (ino < (uspi->s_ncg * uspi->s_ipg )))) {
 		ufs_warning(sb, "ufs_free_inode", "reserved inode or nonexistent inode %u\n", ino);
-		unlock_ufs(sb);
+		mutex_unlock(&ei->truncate_mutex);
 		return;
 	}
 	
@@ -81,7 +82,7 @@ void ufs_free_inode (struct inode * inode)
 	bit = ufs_inotocgoff (ino);
 	ucpi = ufs_load_cylinder (sb, cg);
 	if (!ucpi) {
-		unlock_ufs(sb);
+		mutex_unlock(&ei->truncate_mutex);
 		return;
 	}
 	ucg = ubh_get_ucg(UCPI_UBH(ucpi));
@@ -115,7 +116,7 @@ void ufs_free_inode (struct inode * inode)
 		ubh_sync_block(UCPI_UBH(ucpi));
 	
 	ufs_mark_sb_dirty(sb);
-	unlock_ufs(sb);
+	mutex_unlock(&ei->truncate_mutex);
 	UFSD("EXIT\n");
 }
 
@@ -176,6 +177,7 @@ struct inode *ufs_new_inode(struct inode *dir, umode_t mode)
 	struct ufs_cg_private_info * ucpi;
 	struct ufs_cylinder_group * ucg;
 	struct inode * inode;
+	struct ufs_inode_info *ei = UFS_I(dir);
 	unsigned cg, bit, i, j, start;
 	struct ufs_inode_info *ufsi;
 	int err = -ENOSPC;
@@ -193,7 +195,7 @@ struct inode *ufs_new_inode(struct inode *dir, umode_t mode)
 	sbi = UFS_SB(sb);
 	uspi = sbi->s_uspi;
 
-	lock_ufs(sb);
+	mutex_lock(&ei->truncate_mutex);
 
 	/*
 	 * Try to place the inode in its parent directory
@@ -331,21 +333,21 @@ cg_found:
 			sync_dirty_buffer(bh);
 		brelse(bh);
 	}
-	unlock_ufs(sb);
+	mutex_unlock(&ei->truncate_mutex);
 
 	UFSD("allocating inode %lu\n", inode->i_ino);
 	UFSD("EXIT\n");
 	return inode;
 
 fail_remove_inode:
-	unlock_ufs(sb);
+	mutex_unlock(&ei->truncate_mutex);
 	clear_nlink(inode);
 	unlock_new_inode(inode);
 	iput(inode);
 	UFSD("EXIT (FAILED): err %d\n", err);
 	return ERR_PTR(err);
 failed:
-	unlock_ufs(sb);
+	mutex_unlock(&ei->truncate_mutex);
 	make_bad_inode(inode);
 	iput (inode);
 	UFSD("EXIT (FAILED): err %d\n", err);
diff --git a/fs/ufs/super.c b/fs/ufs/super.c
index b3bc3e7..1a695a4 100644
--- a/fs/ufs/super.c
+++ b/fs/ufs/super.c
@@ -1435,6 +1435,7 @@ static void init_once(void *foo)
 {
 	struct ufs_inode_info *ei = (struct ufs_inode_info *) foo;
 
+	mutex_init(&ei->truncate_mutex);
 	inode_init_once(&ei->vfs_inode);
 }
 
diff --git a/fs/ufs/ufs.h b/fs/ufs/ufs.h
index 2a07396..b530c2f 100644
--- a/fs/ufs/ufs.h
+++ b/fs/ufs/ufs.h
@@ -46,6 +46,7 @@ struct ufs_inode_info {
 	__u16	i_osync;
 	__u64	i_lastfrag;
 	__u32   i_dir_start_lookup;
+	struct mutex truncate_mutex;
 	struct inode vfs_inode;
 };
 

[-- Attachment #3: Type: text/plain, Size: 126 bytes --]

_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xen.org
http://lists.xen.org/xen-devel