19 февр. 2021 г., в 16:34, Lukas Czerner [off-list ref] написал(а):

On Fri, Feb 19, 2021 at 02:49:16PM +0300, Alexey Lyashkov wrote:

quoted

Lukas,

because e2fsprogs have an bad assumption about IO size for the O_DIRECT case.
and because library uses a code like

quoted

quoted

set_block_size(1k);
seek(fs, 1);
read_block();

quoted

quoted

quoted

which caused an 1k read inside of 4k disk block size not aligned by block size, which is prohibited and caused an error report.

Reference to patch.
https://patchwork.ozlabs.org/project/linux-ext4/patch/20201023112659.1559-1-artem.blagodarenko@gmail.com/

Alright, I skimmed through your patch proposal and I am not sure I
completely understand the problem because you have not provided the code
adding O_DIRECT support for e2image.

debugfs -D … will hit same problem.

However I think that it is a reasonable assumption to make that there is
not going to be a file system on a block device such that the fs blocksize
is smaller than device sector size. You can't create such fs with mke2fs
and you can't mount such file system either.

This is don’t need to be create this FS, calling an ext2_open2 is enough.

All that said I can now see that there is a problem in case of mke2fs
and debugfs when used with O_DIRECT (-D) on a file system image with 1k
block size stored on a file in the host file system on the block device
with sector size larger than 1k (...I am getting Inception flashbacks now)

if you have open a large (>256T) device with debugfs without -D you will be see a large swap.
once this FS want to consume a ~10G for bitmaps and other parts.
with cached read you memory consumption increased by two.

btw.
you can easy replicate it with losetup which able to specify a block size.

In fact I can confirm that indeed, both mke2fs and debugfs will fail in
such scenario. The question is whether we care enough to support
O_DIRECT in such situations. Personally I don't care enough about this.
However it would be nice to at least have a check (probably in
ext2fs_open2, unix_open_channel or such) and notify user about the
problem.

it’s not a tools problem. It’s problem of e2fsprogs library as ext2_open2 affected by this bug.
But this is not a single function where bug lives.

Note that this conversation does not affect my patch since
ext2fs_mmp_read() does not use the unix_io infrastructure.

It’s good to convert to use it.

Alex

-Lukas

quoted

Alex

quoted

19 февр. 2021 г., в 13:57, Lukas Czerner [off-list ref] написал(а):

On Fri, Feb 19, 2021 at 01:08:17PM +0300, Alexey Lyashkov wrote:

quoted

Andreas,

What about to disable a O_DIRECT global on any block devices in the e2fsprogs library as this don’t work on 4k disk drives at all ?
Instead of fixing an O_DIRECT access with patches sends early.

Why would it not work at all ? This is a fix for a specific problem and
I am not currently aware of ony other problems e2fsprogs should have
with 4k sector size drives. Do you have a specific problem in mind ?

Thanks!
-Lukas

quoted

Alex

quoted

19 февр. 2021 г., в 1:20, Andreas Dilger [off-list ref] написал(а):

On Feb 18, 2021, at 2:51 AM, Lukas Czerner [off-list ref] wrote:

quoted

Currently the mmp block is read using O_DIRECT to avoid any caching that
may be done by the VM. However when working with regular files this
creates alignment issues when the device of the host file system has
sector size larger than the blocksize of the file system in the file
we're working with.

This can be reproduced with t_mmp_fail test when run on the device with
4k sector size because the mke2fs fails when trying to read the mmp
block.

Fix it by disabling O_DIRECT when working with regular files. I don't
think there is any risk of doing so since the file system layer, unlike
shared block device, should guarantee cache consistency.

Signed-off-by: Lukas Czerner <redacted>
Reviewed-by: Eric Sandeen <redacted>

Reviewed-by: Andreas Dilger <redacted>

quoted

---
v2: Fix comment - it avoids problems when the sector size is larger not
smaller than blocksize

lib/ext2fs/mmp.c | 22 +++++++++++-----------
1 file changed, 11 insertions(+), 11 deletions(-)

diff --git a/lib/ext2fs/mmp.c b/lib/ext2fs/mmp.c
index c21ae272..cca2873b 100644
--- a/lib/ext2fs/mmp.c
+++ b/lib/ext2fs/mmp.c

@@ -57,21 +57,21 @@ errcode_t ext2fs_mmp_read(ext2_filsys fs, blk64_t mmp_blk, void *buf)

	 * regardless of how the io_manager is doing reads, to avoid caching of
	 * the MMP block by the io_manager or the VM.  It needs to be fresh. */
	if (fs->mmp_fd <= 0) {
+		struct stat st;
		int flags = O_RDWR | O_DIRECT;

-retry:
+		/*
+		 * There is no reason for using O_DIRECT if we're working with
+		 * regular file. Disabling it also avoids problems with
+		 * alignment when the device of the host file system has sector
+		 * size larger than blocksize of the fs we're working with.
+		 */
+		if (stat(fs->device_name, &st) == 0 &&
+		    S_ISREG(st.st_mode))
+			flags &= ~O_DIRECT;
+
		fs->mmp_fd = open(fs->device_name, flags);
		if (fs->mmp_fd < 0) {
-			struct stat st;
-
-			/* Avoid O_DIRECT for filesystem image files if open
-			 * fails, since it breaks when running on tmpfs. */
-			if (errno == EINVAL && (flags & O_DIRECT) &&
-			    stat(fs->device_name, &st) == 0 &&
-			    S_ISREG(st.st_mode)) {
-				flags &= ~O_DIRECT;
-				goto retry;
-			}
			retval = EXT2_ET_MMP_OPEN_DIRECT;
			goto out;
		}
--
2.26.2

Cheers, Andreas