Skip to contentSkip to conversation

libext2fs: revert "try to always use PUNCH_HOLE for unix_discard"

PATCH Applied Darrick J. Wong 3 participants 4 messages Jan 14, 2019 – Feb 15, 2019

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 libext2fs: revert "try to always use PUNCH_HOLE for unix_discard"

From: Darrick J. Wong <[email protected]>

Revert bcca9876a3428c10417c660b78933e6e70e8a5f5, because fallocate(PUNCH_HOLE) on block devices was changed to use zeroout instead of discard shortly after block device fallocate was merged. zeroout isn't necessarily a "drop storage" operation like discard is, so we prefer to use that on block devices.

---
 lib/ext2fs/unix_io.c |   61 ++++++++++++++++++++------------------------------
 1 file changed, 25 insertions(+), 36 deletions(-)
lib/ext2fs/unix_io.c +25 -36
index 16e2052cd..74fc8a75d 100644
@@ -1081,38 +1081,6 @@ static errcode_t unix_set_option(io_channel channel, const char *option,
 #define BLKDISCARD		_IO(0x12,119)
 #endif
 
-/*
- * Try a PUNCH_HOLE to unmap blocks, then BLKDISCARD if that doesn't work.
- * We prefer PUNCH_HOLE because it invalidates the page cache, even on block
- * devices.
- */
-static int __unix_discard(int fd, int is_bdev, off_t offset, off_t len)
-{
-#ifdef BLKDISCARD
-	__u64 range[2];
-#endif
-	int ret = -1;
-
-#if defined(HAVE_FALLOCATE) && defined(FALLOC_FL_PUNCH_HOLE) && defined(FALLOC_FL_KEEP_SIZE)
-	ret = fallocate(fd, FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
-			offset, len);
-	if (ret == 0)
-		return 0;
-#endif
-#ifdef BLKDISCARD
-	if (is_bdev) {
-		range[0] = (__u64)offset;
-		range[1] = (__u64)len;
-
-		ret = ioctl(fd, BLKDISCARD, &range);
-		if (ret == 0)
-			return 0;
-	}
-#endif
-	errno = EOPNOTSUPP;
-	return ret;
-}
-
 static errcode_t unix_discard(io_channel channel, unsigned long long block,
 			      unsigned long long count)
 {
@@ -1123,10 +1091,31 @@ static errcode_t unix_discard(io_channel channel, unsigned long long block,
 	data = (struct unix_private_data *) channel->private_data;
 	EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
 
-	ret = __unix_discard(data->dev,
-			(channel->flags & CHANNEL_FLAGS_BLOCK_DEVICE),
-			(off_t)(block) * channel->block_size + data->offset,
-			(off_t)(count) * channel->block_size);
+	if (channel->flags & CHANNEL_FLAGS_BLOCK_DEVICE) {
+#ifdef BLKDISCARD
+		__u64 range[2];
+
+		range[0] = (__u64)(block) * channel->block_size + data->offset;
+		range[1] = (__u64)(count) * channel->block_size;
+
+		ret = ioctl(data->dev, BLKDISCARD, &range);
+#else
+		goto unimplemented;
+#endif
+	} else {
+#if defined(HAVE_FALLOCATE) && defined(FALLOC_FL_PUNCH_HOLE)
+		/*
+		 * If we are not on block device, try to use punch hole
+		 * to reclaim free space.
+		 */
+		ret = fallocate(data->dev,
+				FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
+				(off_t)(block) * channel->block_size + data->offset,
+				(off_t)(count) * channel->block_size);
+#else
+		goto unimplemented;
+#endif
+	}
 	if (ret < 0) {
 		if (errno == EOPNOTSUPP)
 			goto unimplemented;
#1 lore raw
Quoted 9 lines · Darrick J. Wong

From: Darrick J. Wong <[email protected]>

Revert bcca9876a3428c10417c660b78933e6e70e8a5f5, because fallocate(PUNCH_HOLE) on block devices was changed to use zeroout instead of discard shortly after block device fallocate was merged. zeroout isn't necessarily a "drop storage" operation like discard is, so we prefer to use that on block devices.

Thanks, applied.

- Ted

#2 lore raw
Quoted 9 lines · Theodore Y. Ts'o
Quoted 8 lines · Darrick J. Wong

From: Darrick J. Wong <[email protected]>

Revert bcca9876a3428c10417c660b78933e6e70e8a5f5, because fallocate(PUNCH_HOLE) on block devices was changed to use zeroout instead of discard shortly after block device fallocate was merged. zeroout isn't necessarily a "drop storage" operation like discard is, so we prefer to use that on block devices.

Quoted 5 lines
Quoted 1 line

Thanks, applied.

- Ted

I just noticed this patch, sorry. I think we can still use fallocate, but we need to set the right flags to make sure it uses discard instead of zeroout. See fs/block_dev.c

	switch (mode) {
	case FALLOC_FL_ZERO_RANGE:
	case FALLOC_FL_ZERO_RANGE | FALLOC_FL_KEEP_SIZE:
		error = blkdev_issue_zeroout(bdev, start >> 9, len >> 9,
					    GFP_KERNEL, BLKDEV_ZERO_NOUNMAP);
		break;
	case FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE:
		error = blkdev_issue_zeroout(bdev, start >> 9, len >> 9,
					     GFP_KERNEL, BLKDEV_ZERO_NOFALLBACK);
		break;
	case FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE | FALLOC_FL_NO_HIDE_STALE:
		error = blkdev_issue_discard(bdev, start >> 9, len >> 9,
					     GFP_KERNEL, 0);
		break;
	default:
		return -EOPNOTSUPP;
	}

So if we want a discard (meaning we want to unallocate the blocks without necessarily making sure we can't read stale data from it) we have to use FALLOC_FL_NO_HIDE_STALE.

So the flags would be FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE | FALLOC_FL_NO_HIDE_STALE

Ted, Darrick what do you think ? Can we keep the bcca9876a3428c10417c660b78933e6e70e8a5f5 commit and just change the flags ?

-Lukas

#3 lore raw
Quoted 38 lines · Lukas Czerner
Quoted 9 lines · Theodore Y. Ts'o
Quoted 8 lines · Darrick J. Wong

From: Darrick J. Wong <[email protected]>

Revert bcca9876a3428c10417c660b78933e6e70e8a5f5, because fallocate(PUNCH_HOLE) on block devices was changed to use zeroout instead of discard shortly after block device fallocate was merged. zeroout isn't necessarily a "drop storage" operation like discard is, so we prefer to use that on block devices.

Quoted 5 lines
Quoted 1 line

Thanks, applied.

- Ted

I just noticed this patch, sorry. I think we can still use fallocate, but we need to set the right flags to make sure it uses discard instead of zeroout. See fs/block_dev.c

	switch (mode) {
	case FALLOC_FL_ZERO_RANGE:
	case FALLOC_FL_ZERO_RANGE | FALLOC_FL_KEEP_SIZE:
		error = blkdev_issue_zeroout(bdev, start >> 9, len >> 9,
					    GFP_KERNEL, BLKDEV_ZERO_NOUNMAP);
		break;
	case FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE:
		error = blkdev_issue_zeroout(bdev, start >> 9, len >> 9,
					     GFP_KERNEL, BLKDEV_ZERO_NOFALLBACK);
		break;
	case FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE | FALLOC_FL_NO_HIDE_STALE:
		error = blkdev_issue_discard(bdev, start >> 9, len >> 9,
					     GFP_KERNEL, 0);
		break;
	default:
		return -EOPNOTSUPP;
	}
Quoted 5 lines

So if we want a discard (meaning we want to unallocate the blocks without necessarily making sure we can't read stale data from it) we have to use FALLOC_FL_NO_HIDE_STALE.

So the flags would be FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE | FALLOC_FL_NO_HIDE_STALE

Userspace isn't allowed to pass in _NO_HIDE_STALE; see FALLOC_FL_SUPPORTED_MASK in include/linux/falloc.h.

The behavior of "no hide stale" isn't defined in the manpages; it's merely a reserved code point.

--D

Quoted 5 lines

Ted, Darrick what do you think ? Can we keep the bcca9876a3428c10417c660b78933e6e70e8a5f5 commit and just change the flags ?

-Lukas