Re: [Bug-wget] [PATCH] Implement --keep-badhash to keep Metalink's files with a bad hash

[Matthew White]

On Thu, 04 Aug 2016 17:08:20 +0200
Tim Ruehsen <address@hidden> wrote:

> On Thursday, August 4, 2016 12:40:49 PM CEST Matthew White wrote:
> > On Thu, 04 Aug 2016 12:08:50 +0200
> > 
> > Tim Ruehsen <address@hidden> wrote:
> > > On Wednesday, August 3, 2016 2:40:14 PM CEST Matthew White wrote:
> > > > On Wed, 03 Aug 2016 14:05:06 +0200
> > > > 
> > > > Tim Ruehsen <address@hidden> wrote:
> > > > > On Tuesday, August 2, 2016 11:27:28 AM CEST Matthew White wrote:
> > > > > > On Mon, 01 Aug 2016 16:32:30 +0200
> > > > > > 
> > > > > > Tim Ruehsen <address@hidden> wrote:
> > > > > > > On Saturday, July 30, 2016 9:41:48 PM CEST Matthew White wrote:
> > > > > > > > Hello!
> > > > > > > > 
> > > > > > > > I think that sometimes it could help to keep downloaded
> > > > > > > > Metalink's
> > > > > > > > files
> > > > > > > > which have a bad hash.
> > > > > > > > 
> > > > > > > > The default wget behaviour is to delete such files.
> > > > > > > > 
> > > > > > > > This patch provides a way to keep files which have a bad hash
> > > > > > > > through
> > > > > > > > the
> > > > > > > > option --keep-badhash. It appends the suffix .badhash to the
> > > > > > > > file
> > > > > > > > name,
> > > > > > > > except without overwriting existing files. In the latter case,
> > > > > > > > an
> > > > > > > > unique
> > > > > > > > suffix is appended after .badhash.
> > > > > > > > 
> > > > > > > > I made this patch working on the following branch:
> > > > > > > > master (latest 20cac2c5ab3d63aacfba35fb10878a2d490e2377)
> > > > > > > > git://git.savannah.gnu.org/wget.git
> > > > > > > > 
> > > > > > > > What do you think?
> > > > > > > 
> > > > > > > Hi Matthew,
> > > > > > > 
> > > > > > > good work !
> > > > > > > 
> > > > > > > While your FSF assignment is underway (PM), we can continue
> > > > > > > polishing.
> > > > > > > 
> > > > > > > Didn't test your code yet,  but from what I see, there are just
> > > > > > > these
> > > > > > > peanuts:
> > > > > > > 
> > > > > > > 1.
> > > > > > > +  bhash = malloc (strlen (name) + strlen (".badhash") + 1);
> > > > > > > +  strcat (strcpy (bhash, name), ".badhash");
> > > > > > > 
> > > > > > > Please consider using concat_strings() from util.h.
> > > > > > > And please leave an empty line between var declaration and code -
> > > > > > > just
> > > > > > > for
> > > > > > > readability.
> > > > > > > 
> > > > > > > 2.
> > > > > > > Please add 'link' and 'unlink' to bootstrap.conf. Gnulib will
> > > > > > > emulate
> > > > > > > these on platforms where they are not available (or have a
> > > > > > > slightly
> > > > > > > different behavior). I guess we simply forgot 'unlink' when we
> > > > > > > switched
> > > > > > > to gnulib.
> > > > > > > 
> > > > > > > 3.
> > > > > > > The (GNU) commit messages should ideally look like:
> > > > > > > 
> > > > > > > one line of short description
> > > > > > > <empty line>
> > > > > > > file changes
> > > > > > > <empty line>
> > > > > > > long description
> > > > > > > 
> > > > > > > For example:
> > > > > > > Add new option --keep-badhash
> > > > > > > 
> > > > > > > * src/init.c: Add keepbadhash
> > > > > > > * src/main.c: Add keep-badhash
> > > > > > > * src/options.h: Add keep_badhash
> > > > > > > * doc/wget.texi: Add docs for --keep-badhash
> > > > > > > * src/metalink.h: Add prototypes badhash_suffix(),
> > > > > > > badhash_or_remove()
> > > > > > > * src/metalink.c: New functions badhash_suffix(),
> > > > > > > badhash_or_remove().
> > > > > > > 
> > > > > > >   (retrieve_from_metalink): Call append .badhash()
> > > > > > > 
> > > > > > > With --keep-badhash, append .badhash to Metalink's files with
> > > > > > > checksum
> > > > > > > mismatch, except without overwriting existing files.
> > > > > > > 
> > > > > > > Without --keep-badhash, remove downloaded files with checksum
> > > > > > > mismatch
> > > > > > > (this conforms to the old behaviour).
> > > > > > > 
> > > > > > > [This also applies to to your other patches]
> > > > > > > 
> > > > > > > 4.
> > > > > > > Please not only write 'keepbadhash', but also what you did (e.g.
> > > > > > > remove,
> > > > > > > rename, add, ...), see my example above.
> > > > > > > 
> > > > > > > Those ChangeLog entries should allow finding changes / faults /
> > > > > > > regressions
> > > > > > > etc. even when a versioning system is not at hand, e.g. when
> > > > > > > unpacking
> > > > > > > the
> > > > > > > sources from a tarball. (Not debatable that consulting commit
> > > > > > > messages
> > > > > > > directly is much more powerful.)
> > > > > > > 
> > > > > > > 5.
> > > > > > > You write "except without overwriting existing files", maybe you
> > > > > > > should
> > > > > > > mention appending a counter instead of overwriting existent files
> > > > > > > !?
> > > > > > > 
> > > > > > > 
> > > > > > > Regards, Tim
> > > > > > 
> > > > > > Thanks Tim!
> > > > > > 
> > > > > > I really needed your guidance. I sent the modified patches to
> > > > > > address@hidden
> > > > > > 
> > > > > > I believe there are more things to fix.
> > > > > > 
> > > > > > Consider the following after applying the attached patch:
> > > > > > * src/metalink.c (retrieve_from_metalink): line 124: If the download
> > > > > > is
> > > > > > interrupted (output_stream isn't NULL), and there are more urls for
> > > > > > the
> > > > > > same file (we are still in the download loop), switching to the next
> > > > > > url
> > > > > > should resume the download (instead than start it over).
> > > > > > 
> > > > > > In this patch I added a fix to rename/remove the fully downloaded
> > > > > > file
> > > > > > (output_stream is NULL), but there was an error (probably checksum)
> > > > > > and
> > > > > > we
> > > > > > are still in the download loop (more urls for the same file). But as
> > > > > > said
> > > > > > before, switching to the next url should continue the download: *
> > > > > > src/metalink.c (retrieve_from_metalink): line 131: Rename/remove
> > > > > > fully
> > > > > > downloaded file on error
> > > > > > 
> > > > > > I still have to investigate the problem...
> > > > > 
> > > > > So, I wait with 0001-... !?
> > > > > 
> > > > > 0002-... has been pushed (extended with 'symlink').
> > > > > 
> > > > > Thanks for your contribution.
> > > > > 
> > > > > Tim
> > > > 
> > > > Ok for 0002. Thanks.
> > > > 
> > > > There's no problem applying the 0001.
> > > 
> > > Applied and pushed !
> > > 
> > > > The "if the stream got interrupted, then restart the download with the
> > > > next
> > > > url" (output_stream isn't NULL) was already there before the patch 0001.
> > > > 
> > > > [debug is required to know when output_stream isn't NULL]
> > > > 
> > > > commit 7fad76db4cdb7a0fe7e5aa0dd88f5faaf8f4cdc8
> > > > * src/metalink.c (retrieve_from_metalink): line 124: 'if
> > > > (output_stream)'
> > > > remove file and restart download with next url
> > > > 
> > > > With the 0001, if --keep-badhash is used the file is renamed instead
> > > > than
> > > > removed, even when "output_stream is NULL".
> > > > 
> > > > I have to look through the "stream interrupted" situation.
> > > > 
> > > > I'm guessing that the download is not resumed.
> > > > 
> > > > What do you suggest?
> > > 
> > > We need a  document where we define wanted behavior, create tests and
> > > amend the code.
> > > 
> > > Regards, Tim
> > 
> > Ok, I just documented a test. See the attached tarball.
> > 
> > Also there is a patch attached (there's a copy in the tarball too),
> > providing the suggested bugfix.
> 
> I made that test without your patch => crash due to badhash_or_remove(NULL).

The SIGSEGV is nasty:
* src/metalink.c (retrieve_from_metalink: line 166): 'output_stream = fopen 
(mfile->name, "ab");':
  Introduced with commit e0b60fd0736c7f25642b0cb7cdfc0dd74f1d6d6f: Triggers 
when --continue is used;
  SIGSEGV since commit 943a6d585ff4a0afd974ab2e4838c9d2a63cf21a: Appending to 
file leads to badhash_suffix(NULL) or badhash_or_remove(NULL) segmentation fault

Bugfix A:
I attached a patch to avoid the badhash_suffix(NULL) and 
badhash_or_remove(NULL) segmentation faults. But I'm still investigating the 
problem caused by appending to file 'output_stream = fopen (mfile->name, 
"ab");'.

Bugfix B:
'[Bug-wget] [PATCH] Support metalink:file elements with a "path/file" format'
http://lists.gnu.org/archive/html/bug-wget/2016-08/msg00021.html
This supersedes Bugfix A, it enables metalink:file "path/file" name format, 
also fixes the bug you saw in valgrind log (read below).

You discussed the dangers of Bugfix B, I will answer soon:
http://lists.gnu.org/archive/html/bug-wget/2016-08/msg00046.html

> With your patch, valgrind says:
> 
> ==15068== Syscall param open(filename) points to unaddressable byte(s)
> ==15068==    at 0x6848F40: __open_nocancel (syscall-template.S:84)
> ==15068==    by 0x67DFF7D: _IO_file_open (fileops.c:221)
> ==15068==    by 0x67E009F: _IO_file_fopen@@GLIBC_2.2.5 (fileops.c:328)
> ==15068==    by 0x67D4F03: __fopen_internal (iofopen.c:86)
> ==15068==    by 0x434FAE: retrieve_from_metalink (metalink.c:184)
> ==15068==    by 0x406C63: main (main.c:2025)
> ==15068==  Address 0x0 is not stack'd, malloc'd or (recently) free'd
> ==15068== 
> 
> Can you reproduce this ?

The bug you describe in the valgrind log (not the segmentation fault) it's 
there since commit 20cac2c5ab3d63aacfba35fb10878a2d490e2377. It's caused by 
leaving filename unset:
* src/metalink.c (retrieve_from_metalink): src/utils.c (unique_create): Cannot 
handle "path/file" format, so filename is left unset
* src/metalink.c (retrieve_from_metalink): Since commit 
e0b60fd0736c7f25642b0cb7cdfc0dd74f1d6d6f: With --continue, filename is left 
unset

Bugfix B (discussed above):
'[Bug-wget] [PATCH] Support metalink:file elements with a "path/file" format'
http://lists.gnu.org/archive/html/bug-wget/2016-08/msg00021.html

This describes more a user's side of view of the problem:
http://lists.gnu.org/archive/html/bug-wget/2016-08/msg00052.html

You discussed the dangers of enabling "path/file" (see Bugfix B above):
http://lists.gnu.org/archive/html/bug-wget/2016-08/msg00046.html


Vanilla commit 20cac2c5ab3d63aacfba35fb10878a2d490e2377 (before my patches)

$ valgrind --leak-check=full wget [-c] 
--input-metalink=test-pathfile-sha256.meta4

Bug:
* src/metalink.c (retrieve_from_metalink): src/utils.c (unique_create): Cannot 
handle "path/file" format, so filename is left unset

retrieve_from_metalink (metalink.c:171):
local_file = fopen (filename, "rb");

main (main.c:2022):
retr_err = retrieve_from_metalink (metalink);

==15784== Syscall param open(filename) points to unaddressable byte(s)
==15784==    at 0x69D4330: __open_nocancel (syscall-template.S:84)
==15784==    by 0x696C3AD: _IO_file_open (fileops.c:221)
==15784==    by 0x696C4F2: _IO_file_fopen@@GLIBC_2.2.5 (fileops.c:328)
==15784==    by 0x6960DA3: __fopen_internal (iofopen.c:86)
==15784==    by 0x44952E: retrieve_from_metalink (metalink.c:171)
==15784==    by 0x430BA4: main (main.c:2022)
==15784==  Address 0x0 is not stack'd, malloc'd or (recently) free'd
==15784==
...
==15784== ERROR SUMMARY: 2 errors from 1 contexts (suppressed: 0 from 0)


$ valgrind --leak-check=full wget [-c] --input-metalink=test-file-sha256.meta4

==15792== ERROR SUMMARY: 0 errors from 0 contexts (suppressed: 0 from 0)

> 
> Tim

I also attached the tests I have used.

PS: gosh, the testing took some serious time...

Soon,
Matthew

-- 
Matthew White <address@hidden>

0001-Bugfix-Avoid-segmentation-fault-in-Metalink-module.patch
Description: Text Data

wget-tests.tar.xz
Description: application/xz

pgpgtWKh_n_Fh.pgp
Description: PGP signature