Re: [Bug-tar] --sparse is broken on filesystems where small files may have zero blocks
On Tue, Oct 29, 2013 at 04:01:16PM -0600, Andreas Dilger wrote: On Oct 29, 2013, at 11:37 AM, Jan Kara j...@suse.cz wrote: On Tue 29-10-13 16:27:02, Pavel Raiskup wrote: Well, I now recalled somehow relevant Red Hat bug, sorry I have not mentioned it before: https://bugzilla.redhat.com/show_bug.cgi?id=757557 CC'ing fs-devel: The question is whether that is not a bug in filesystem — whether filesystem should not _always_ return to fstat() block count at least 1 if there are at least some data (even if these data are inlined in inode)? Just for catching the context, this thread starts here: http://lists.gnu.org/archive/html/bug-tar/2013-10/msg00030.html So 'st_blocks' should be the number of blocks allocated to the file, 512-byte units. If we are able to store the whole file within inode, we have no blocks allocated and thus setting st_blocks to 0 looks as a decent thing to do. Looking into filesystems where this is possible (ext4, btrfs, reiserfs) they will all set st_blocks to 0 if the file body is inlined and the size is smaller than 512 bytes. One could consider that any inode with inline data still needs to have the inode allocated to hold the data. ext4 is yet a different matter. It does really report the number of allocated blocks in st_blocks so it will report 0 while data can fit into the inode (whose size is configurable during fs creation, default is 256). In practice that will result in reporting non-zero st_blocks for 512-byte and larger files anyways so there won't be an observable difference between what ext4 and btrfs / reiserfs do. But we might still want to fix up ext4 to be consistent with btrfs and reiserfs so that things are more future-proof. Given that tar and rsync DO exist that silently drop user data for files with st_blocks == 0 (we have seen this in with Lustre tests) it makes sense to fix the ext4_getattr() to set st_blocks = 1 for files with inline data. It is already doing something similar for files with delalloc blocks to work around the same problem. Blocks used to store inode metadata are not considered data blocks, so, these are not reported. In case of delalloc, afaik, it's doing it to reserve space and in case it's not used, well, it's removed from the counter. But, it differs from the inline case since you'll not be using the reserved delalloc blocks (unless the file is extended, of course). I'm not pretty sure about this, but if you account the space used by inline data as a 'block', you might be accounting the same block twice, one as metadata block, and another as data block. Also, the idea of inline data is exactly to avoid block allocation for very small files that fit into inode's literal area, setting st_blocks to 1 would go against this purpose. And also, free-space count might be wrong setting it to 1 too. If that is not a bug in fs, is there possible to detect that particular file is completely sparse? As Joerg wrote, SEEK_DATA / SEEK_HOLE is a proper interface for this at least for systems that support it. Once you have called stat(2) on the file, inode will be in cache anyways so the additional cost of open(2), lseek(2), close(2) won't be that big. For systems that don't support SEEK_DATA / SEEK_HOLE, you can use some heuristic like: if (st.st_size st.st_blksize || st.st_blocks 0) /* Bite the bullet and scan the data for non-zero bytes */ else /* Assume the file is sparse */ I don’t see how this is better than the existing heuristic: if (st.st_blocks 0) /* file has data */ else /* file is sparse */ that tools/applications are already using today. It isn’t possible to retroactively fix tar, rsync, etc. to use SEEK_DATA/SEEK_HOLE, but in the ext4 case the inline data feature is very new and it makes sense to fix this in the kernel. Of course it ALSO makes sense to fix this in userspace to handle the (st.st_size st.st_blksize) case (this is not a lot of overhead) so that the chance of losing data in a variety of kernel/userspace combinations is minimized. Cheers, Andreas -- To unsubscribe from this list: send the line unsubscribe linux-fsdevel in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html -- Carlos
Re: [Bug-tar] --sparse is broken on filesystems where small files may have zero blocks
On Tue, Oct 29, 2013 at 04:27:02PM +0100, Pavel Raiskup wrote: CC'ing fs-devel: The question is whether that is not a bug in filesystem ??? whether filesystem should not _always_ return to fstat() block count at least 1 if there are at least some data (even if these data are inlined in inode)? Just for catching the context, this thread starts here: http://lists.gnu.org/archive/html/bug-tar/2013-10/msg00030.html If that is not a bug in fs, is there possible to detect that particular file is completely sparse? For XFS we only support inline data for non-regular files at the moment, but for example an inode that has data stored just inline will report zero block, and the unfinished patches to support inline regular file data would logically do the same. As already pointed out by Joerg we do have a proper lseek interface to detect if there are holes.
Re: [Bug-tar] --sparse is broken on filesystems where small files may have zero blocks
#define ST_IS_SPARSE(st) \ (ST_NBLOCKS (st)\ - ((st).st_size / ST_NBLOCKSIZE + ((st).st_size % ST_NBLOCKSIZE != 0))) +((st).st_size / ST_NBLOCKSIZE \ + + ((st).st_size % ST_NBLOCKSIZE != 0 \ + (st).st_size / ST_NBLOCKSIZE != 0))) May the st.st_size / ST_NBLOCKSIZE be greater than 1 and data still stored in inode directly? Seems like on ext4 filesystem it is not possible [1] but does anybody know about exception? [1] https://ext4.wiki.kernel.org/index.php/Ext4_Disk_Layout#Inline_Data Pavel
Re: [Bug-tar] --sparse is broken on filesystems where small files may have zero blocks
Hi Paul, On Mon, Oct 28, 2013 at 09:19:58PM -0700, Paul Eggert wrote: Thanks for the bug report. I pushed the following patch to paxutils and it should propagate into GNU tar in the next release. From 63493234ec38ad606a5f726bc82e4fe5d8661cab Mon Sep 17 00:00:00 2001 From: Paul Eggert egg...@cs.ucla.edu Date: Mon, 28 Oct 2013 21:16:53 -0700 Subject: [PATCH] paxutils: support --sparse with tiny files on Netapp filers * lib/system.h (ST_IS_SPARSE): Port to NFS + Netapp filers, where a tiny file can have zero blocks but nonzero size. Problem reported by Andrew J. Schorr in http://lists.gnu.org/archive/html/bug-tar/2013-10/msg00030.html. --- lib/system.h | 9 - 1 file changed, 8 insertions(+), 1 deletion(-) diff --git a/lib/system.h b/lib/system.h index ef46267..e7f531c 100644 --- a/lib/system.h +++ b/lib/system.h @@ -389,9 +389,16 @@ extern int errno; # define ST_NBLOCKSIZE 512 #endif +/* Network Appliance file systems store small files directly in the + inode if st_size = 64; in this case the number of blocks can be + zero. Perhaps other file systems have similar problems; so, + somewhat arbitrarily, do not consider a file to be sparse if + it has no blocks but st_size ST_NBLOCKSIZE. */ #define ST_IS_SPARSE(st) \ (ST_NBLOCKS (st)\ - ((st).st_size / ST_NBLOCKSIZE + ((st).st_size % ST_NBLOCKSIZE != 0))) +((st).st_size / ST_NBLOCKSIZE \ + + ((st).st_size % ST_NBLOCKSIZE != 0 \ + (st).st_size / ST_NBLOCKSIZE != 0))) /* Declare standard functions. */ Thanks for the patch, but I don't think that fixes the problem in sparse.c:sparse_scan_file where it says if (ST_NBLOCKS (st-stat) == 0) offset = st-stat.st_size; else ... Regards, Andy
Re: [Bug-tar] --sparse is broken on filesystems where small files may have zero blocks
Hi, On Tue, Oct 29, 2013 at 08:03:47AM -0400, Andrew J. Schorr wrote: Thanks for the patch, but I don't think that fixes the problem in sparse.c:sparse_scan_file where it says if (ST_NBLOCKS (st-stat) == 0) offset = st-stat.st_size; else ... To be clear, I can see that the fix to ST_IS_SPARSE should cause dump_regular_file to be called instead of sparse_dump_file, but I still wonder if it is wise to leave this logic in place. At the very last, I think a comment would be helpful to explain that this test is valid only because ST_IS_SPARSE has already succeeded. Regards, Andy
Re: [Bug-tar] --sparse is broken on filesystems where small files may have zero blocks
To be clear, I can see that the fix to ST_IS_SPARSE should cause dump_regular_file to be called instead of sparse_dump_file, but I still wonder if it is wise to leave this logic in place. At the very last, I think a comment would be helpful to explain that this test is valid only because ST_IS_SPARSE has already succeeded. At least for the check for zero blocks in sparse file: It is intentional because it makes the processing of completely sparse files to be done in constant time (try to archive 'file' from `truncate -s 10G file`). This could be documented possibly. Otherwise, I would not say that there is unclear that sparse_dump_file is supposed to be called only against real sparse files.
Re: [Bug-tar] --sparse is broken on filesystems where small files may have zero blocks
On Oct 29, 2013, at 7:10 AM, Pavel Raiskup prais...@redhat.com wrote: To be clear, I can see that the fix to ST_IS_SPARSE should cause dump_regular_file to be called instead of sparse_dump_file, but I still wonder if it is wise to leave this logic in place. At the very last, I think a comment would be helpful to explain that this test is valid only because ST_IS_SPARSE has already succeeded. At least for the check for zero blocks in sparse file: It is intentional because it makes the processing of completely sparse files to be done in constant time (try to archive 'file' from `truncate -s 10G file`). This could be documented possibly. Otherwise, I would not say that there is unclear that sparse_dump_file is supposed to be called only against real sparse files. What about this sparse file: $ truncate -s 10G file echo hello file Are there filesystems where the 6 bytes here would be stored in the inode? That would give a large sparse file with zero allocated blocks but not zero content. Tim
Re: [Bug-tar] --sparse is broken on filesystems where small files may have zero blocks
Thanks for the bug report. I pushed the following patch to paxutils and it should propagate into GNU tar in the next release. From 63493234ec38ad606a5f726bc82e4fe5d8661cab Mon Sep 17 00:00:00 2001 From: Paul Eggert egg...@cs.ucla.edu Date: Mon, 28 Oct 2013 21:16:53 -0700 Subject: [PATCH] paxutils: support --sparse with tiny files on Netapp filers * lib/system.h (ST_IS_SPARSE): Port to NFS + Netapp filers, where a tiny file can have zero blocks but nonzero size. Problem reported by Andrew J. Schorr in http://lists.gnu.org/archive/html/bug-tar/2013-10/msg00030.html. --- lib/system.h | 9 - 1 file changed, 8 insertions(+), 1 deletion(-) diff --git a/lib/system.h b/lib/system.h index ef46267..e7f531c 100644 --- a/lib/system.h +++ b/lib/system.h @@ -389,9 +389,16 @@ extern int errno; # define ST_NBLOCKSIZE 512 #endif +/* Network Appliance file systems store small files directly in the + inode if st_size = 64; in this case the number of blocks can be + zero. Perhaps other file systems have similar problems; so, + somewhat arbitrarily, do not consider a file to be sparse if + it has no blocks but st_size ST_NBLOCKSIZE. */ #define ST_IS_SPARSE(st) \ (ST_NBLOCKS (st)\ - ((st).st_size / ST_NBLOCKSIZE + ((st).st_size % ST_NBLOCKSIZE != 0))) +((st).st_size / ST_NBLOCKSIZE \ + + ((st).st_size % ST_NBLOCKSIZE != 0 \ + (st).st_size / ST_NBLOCKSIZE != 0))) /* Declare standard functions. */ -- 1.8.3.1