Skip to contentSkip to conversation

Detection of sparse files is broken on btrfs

PATCH Pavel Raiskup 3 participants 6 messages Jan 9, 2018 – Jan 10, 2018 1 message not in the archive Btrfs file system

lore
Share & tools
lore

A message that is not in the archive (sent off-list or not yet received): [email protected]

#1 lore raw
Quoted 2 lines · CET Mark H Weaver

I propose that we revisit this bug and fix it. We clearly cannot assume that st_blocks == 0 implies that the file contains only zeroes.

. only on btrfs, as far as we know, because of some race condition. So what about special casing that filesystem, where we can lseek() for holes anyway? Since I would prefer fixing btrfs, I'm CC'ing devels again.

I'm attaching tar patch (public domain, use as you wish) mostly for discussion about the idea (I can or anybody finalize the ifdef-hell, etc.). Note this fixes the failing sparse03.at for me (Fedora 27 x86_64 + btrfs).

references for btrfs guys:
https://www.mail-archive.com/[email protected]/msg05453.html
https://www.spinics.net/lists/linux-btrfs/msg56768.html

Pavel

src/sparse.c +47 -0
index d41c0ea..d0a7a55 100644
@@ -18,6 +18,7 @@
 #include <system.h>
 #include <inttostr.h>
 #include <quotearg.h>
+#include <sys/statfs.h>
 #include "common.h"
 
 struct tar_sparse_file;
@@ -261,12 +262,58 @@ sparse_scan_file_raw (struct tar_sparse_file *file)
   return tar_sparse_scan (file, scan_end, NULL);
 }
 
+enum sparse_fs_behavior
+  {
+    sparse_fs_behavior_init = 0,
+    sparse_fs_behavior_fine,
+    sparse_fs_behavior_uncertain
+  };
+
+static enum sparse_fs_behavior
+check_sparse_behavior (int fd)
+{
+  struct statfs buf;
+  if (fstatfs (fd, &buf))
+    return sparse_fs_behavior_fine;
+
+  if (buf.f_type == 0x9123683e)
+    return sparse_fs_behavior_uncertain; /* btrfs */
+
+  return sparse_fs_behavior_fine;
+}
+
+static bool
+wholesparse_detection_prohibited (struct tar_stat_info *st)
+{
+  static dev_t cached_device = 0;
+  static enum sparse_fs_behavior behavior;
+
+  if (behavior == sparse_fs_behavior_init
+      || cached_device != st->stat.st_dev)
+    {
+      cached_device = st->stat.st_dev;
+      behavior = check_sparse_behavior (st->fd);
+    }
+
+  return behavior == sparse_fs_behavior_uncertain;
+}
+
+
 static bool
 sparse_scan_file_wholesparse (struct tar_sparse_file *file)
 {
   struct tar_stat_info *st = file->stat_info;
   struct sp_array sp = {0, 0};
 
+  /* Some file-systems report st_blksize=0 for files which have some
+     inode-inlined data.  This is, per bug-tar@, rather unfortunate
+     behavior, but we need to deal with these filesystems somehow.  So,
+     let's prohibit the "wholesparse" detection method for such filesystems,
+     and let's hope that 'SEEK_HOLE/SEEK_DATA' works (if not, we fallback to
+     slow-but-safe 'raw' method anyway).  */
+  if (wholesparse_detection_prohibited (file->stat_info))
+    return false;
+
   /* Note that this function is called only for truly sparse files of size >= 1
      block size (checked via ST_IS_SPARSE before).  See the thread
      http://www.mail-archive.com/[email protected]/msg04209.html for more info */
#2 lore raw
Quoted 3 lines · Pavel Raiskup

So
what about special casing that filesystem, where we can lseek() for holes anyway?

If we can lseek for holes, then why not just do that? We shouldn't need special-case code for btrfs per se. Any filesystem where we can lseek for holes should take advantage of that optimization.

#3 lore raw
Quoted 5 lines · CET Paul Eggert
Quoted 2 lines · Pavel Raiskup

So what about special casing that filesystem, where we can lseek() for holes anyway?

If we can lseek for holes, then why not just do that?

Checking whether lseek() actually works costs some additional syscalls _per sparse_ file; checking for ST_NBLOCKS() is without this penalty.

Quoted 2 lines

We shouldn't need special-case code for btrfs per se. Any filesystem where we can lseek for holes should take advantage of that optimization.

It is done so actually, the 'wholesparse' is another optimization on top of that (but usable also in cases where SEEK_HOLE isn't defined at all).

Pavel

#4 lore raw
Quoted 9 lines · Pavel Raiskup
Quoted 5 lines · CET Paul Eggert
Quoted 2 lines · Pavel Raiskup

So what about special casing that filesystem, where we can lseek() for holes anyway?

If we can lseek for holes, then why not just do that?

Checking whether lseek() actually works costs some additional syscalls _per sparse_ file; checking for ST_NBLOCKS() is without this penalty.

Well, star does this since a long time and the penalty is a few microseconds.

"~A

#5 lore raw
Quoted 7 lines · CET Joerg Schilling
Quoted 5 lines · Pavel Raiskup
Quoted 4 lines · CET Paul Eggert
Quoted 2 lines · Pavel Raiskup

So what about special casing that filesystem, where we can lseek() for holes anyway?

Quoted 6 lines
Quoted 4 lines
Quoted 1 line

If we can lseek for holes, then why not just do that?

Checking whether lseek() actually works costs some additional syscalls _per sparse_ file; checking for ST_NBLOCKS() is without this penalty.

Well, star does this since a long time and the penalty is a few microseconds.

It would be interesting to see how network filesystems are affected.

Pavel

#6 lore raw
In reply to Paul Eggert
Quoted 3 lines · Paul Eggert

If we can lseek for holes, then why not just do that? We shouldn't need special-case code for btrfs per se. Any filesystem where we can lseek for holes should take advantage of that optimization.

This is what star uses since 13 years ;-)

Jörg

Signature · 3 lines
 EMail:[email protected]                    (home) Jörg Schilling D-13353 Berlin
    [email protected] (work) Blog: http://schily.blogspot.com/
 URL: http://cdrecord.org/private/ http://sf.net/projects/schilytools/files/'