xfs: don't nest transactions when scanning for eofblocks

Brian Foster reported a lockdep warning on xfs/167:

============================================
WARNING: possible recursive locking detected
5.11.0-rc4 #35 Tainted: G        W I
--------------------------------------------
fsstress/17733 is trying to acquire lock:
ffff8e0fd1d90650 (sb_internal){++++}-{0:0}, at: xfs_free_eofblocks+0x104/0x1d0 [xfs]

but task is already holding lock:
ffff8e0fd1d90650 (sb_internal){++++}-{0:0}, at: xfs_trans_alloc_inode+0x5f/0x160 [xfs]

stack backtrace:
CPU: 38 PID: 17733 Comm: fsstress Tainted: G        W I       5.11.0-rc4 #35
Hardware name: Dell Inc. PowerEdge R740/01KPX8, BIOS 1.6.11 11/20/2018
Call Trace:
 dump_stack+0x8b/0xb0
 __lock_acquire.cold+0x159/0x2ab
 lock_acquire+0x116/0x370
 xfs_trans_alloc+0x1ad/0x310 [xfs]
 xfs_free_eofblocks+0x104/0x1d0 [xfs]
 xfs_blockgc_scan_inode+0x24/0x60 [xfs]
 xfs_inode_walk_ag+0x202/0x4b0 [xfs]
 xfs_inode_walk+0x66/0xc0 [xfs]
 xfs_trans_alloc+0x160/0x310 [xfs]
 xfs_trans_alloc_inode+0x5f/0x160 [xfs]
 xfs_alloc_file_space+0x105/0x300 [xfs]
 xfs_file_fallocate+0x270/0x460 [xfs]
 vfs_fallocate+0x14d/0x3d0
 __x64_sys_fallocate+0x3e/0x70
 do_syscall_64+0x33/0x40
 entry_SYSCALL_64_after_hwframe+0x44/0xa9

The cause of this is the new code that spurs a scan to garbage collect
speculative preallocations if we fail to reserve enough blocks while
allocating a transaction.  While the warning itself is a fairly benign
lockdep complaint, it does expose a potential livelock if the rwsem
behavior ever changes with regards to nesting read locks when someone's
waiting for a write lock.

Fix this by freeing the transaction and jumping back to xfs_trans_alloc
like this patch in the V4 submission[1].

[1] https://lore.kernel.org/linux-xfs/161142798066.2171939.9311024588681972086.stgit@magnolia/

Fixes: a1a7d05 ("xfs: flush speculative space allocations when we run out of space")
Reported-by: Brian Foster <bfoster@redhat.com>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Allison Henderson <allison.henderson@oracle.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>

fs/xfs/xfs_trans.c

@@ -260,13 +260,15 @@ xfs_trans_alloc(
 	struct xfs_trans	**tpp)
 {
 	struct xfs_trans	*tp;
+	bool			want_retry = true;
 	int			error;
 
 	/*
 	 * Allocate the handle before we do our freeze accounting and setting up
 	 * GFP_NOFS allocation context so that we avoid lockdep false positives
 	 * by doing GFP_KERNEL allocations inside sb_start_intwrite().
 	 */
+retry:
 	tp = kmem_cache_zalloc(xfs_trans_zone, GFP_KERNEL | __GFP_NOFAIL);
 	if (!(flags & XFS_TRANS_NO_WRITECOUNT))
 		sb_start_intwrite(mp->m_super);
@@ -289,16 +291,21 @@ xfs_trans_alloc(
 	tp->t_firstblock = NULLFSBLOCK;
 
 	error = xfs_trans_reserve(tp, resp, blocks, rtextents);
-	if (error == -ENOSPC) {
+	if (error == -ENOSPC && want_retry) {
+		xfs_trans_cancel(tp);
+
 		/*
 		 * We weren't able to reserve enough space for the transaction.
 		 * Flush the other speculative space allocations to free space.
 		 * Do not perform a synchronous scan because callers can hold
 		 * other locks.
 		 */
 		error = xfs_blockgc_free_space(mp, NULL);
-		if (!error)
-			error = xfs_trans_reserve(tp, resp, blocks, rtextents);
+		if (error)
+			return error;
+
+		want_retry = false;
+		goto retry;
 	}
 	if (error) {
 		xfs_trans_cancel(tp);