Still a couple bugs:
* Closing the connection to the server leaves git upload-pack /
receive-pack running, which could be used to DOS.
* Sometimes the data is transferred, but it fails at the end, sometimes
with:
git-remote-tor-annex: <socket: 10>: commitBuffer: resource vanished (Broken pipe)
Must be a race condition around shutdown.
Almost working, but there's a bug in the relaying.
Also, made tor hidden service setup pick a random port, to make it harder
to port scan.
This commit was sponsored by Boyd Stephen Smith Jr. on Patreon.
A bit tricky since Proto doesn't support threads. Rather than adding
threading support to it, ended up using a callback that waits for both
data on a Handle, and incoming messages at the same time.
This commit was sponsored by Denis Dzyubenko on Patreon.
For use with tor hidden services, and perhaps other transports later.
Based on Utility.SimpleProtocol, it's a line-based protocol,
interspersed with transfers of bytestrings of a specified size.
Implementation of the local and remote sides of the protocol is done
using a free monad. This lets monadic code be included here, without
tying it to any particular way to get bytes peer-to-peer.
This adds a dependency on the haskell package "free", although that
was probably pulled in transitively from other dependencies already.
This commit was sponsored by Jeff Goeke-Smith on Patreon.
ghc 8 added backtraces on uncaught errors. This is great, but git-annex was
using error in many places for a error message targeted at the user, in
some known problem case. A backtrace only confuses such a message, so omit it.
Notably, commands like git annex drop that failed due to eg, numcopies,
used to use error, so had a backtrace.
This commit was sponsored by Ethan Aubin.
This avoids needing to bind to the right port before something else
does.
The socket is in /var/run/user/$uid/ which ought to be writable by only
that uid. At least it is on linux systems using systemd.
For Windows, may need to revisit this and use ports or something.
The first version of tor to support sockets for hidden services
was 0.2.6.3. That is not in Debian stable, but is available in
backports.
This commit was sponsored by andrea rota.
Tor unfortunately does not come out of the box configured to let hidden
services register themselves on the fly via the ControlPort.
And, changing the config to enable the ControlPort and a particular type
of auth for it may break something already using the ControlPort, or
lessen the security of the system.
So, this leaves only one option to us: Add a hidden service to the
torrc. git-annex enable-tor does so, and picks an unused high port for
tor to listen on for connections to the hidden service.
It's up to the caller to somehow pick a local port to listen on
that won't be used by something else. That may be difficult to do..
This commit was sponsored by Jochen Bartl on Patreon.
Yesod didn't used to do auth checks for that, but this may have changed.
I don't have a way to reproduce the reported problem yet, but this change
certianly won't hurt anything.
This commit was sponsored by Thom May on Patreon.
Restarting a crashing git process could result in filename encoding issues
when not in a unicode locale, as the restarted processes's handles were not
read in raw mode.
Since rawMode is always used when starting a coprocess, didn't bother
to parameterise it and just always enable it for simplicity.
This commit was sponsored by Jake Vosloo on Patreon.
gpg-agent started deleting its socket file on shutdown, and this tickled an
ugly behavior in removeDirectoryRecursive,
https://github.com/haskell/directory/issues/60
Running removeDirectoryRecursive again on exception avoids the problem.
gpg 2.1.15 (or so) seems to have added some new fields to the --with-colons
--list-secret-keys output. These include "fpr" and "grp", and come before
the "uid" line. So, the parser was giving up before it saw the name. Fix by
continuing to look for the uid line until the next "sec" line.
This commit was sponsored by Ole-Morten,Duesund on Patreon.
This gets rid of quite a lot of ugly hacks around json generation.
I doubt that any real-world json parsers can parse incomplete objects, so
while it's not as nice to need to wait for the complete object, especially
for commands like `git annex info` that take a while, it doesn't seem worth
the added complexity.
This also causes the order of fields within the json objects to be
reordered. Since any real json parser shouldn't care, the only possible
problem would be with ad-hoc parsers of the old json output.
Keeping Text.JSON use for now, because it seems a better fit for most of
the commands, which don't use very structured JSON objects, but just output
whatever fields suites them. But this lets Aeson be used when a more
structured data type is available to serialize to JSON.
Mostly the username is only used for the git committer or other display
purposes, and we can just fall back to a dummy value in these cases.
The only remaining place where an error is thrown is when starting local
pairing, which needs the username to be known.
Sadly my bug report about this is not going to get fixed it seems, so
I have to drag around a whole added module file just to deal with it.
https://github.com/haskell/directory/issues/52
It started exporting a isSymbolicLink which supports windows. But,
git-annex does no use symlinks on windows yet and this conflicts with the
function by the same name from unix-compat, so hide it.
This fixes behavior in this situation:
l1 <- lockShared Nothing "lck"
l2 <- lockShared Nothing "lck"
dropLock l1
dropLock l2
Before, the lock was dropped upon the second dropLock call, but the fd
remained open, and would never be closed while the program was running.
Fixed by a rather round-about method, but it should work well enough.
It would have been simpler to open open the shared lock once, and not open
it again in the second call to lockShared. But, that's difficult to do
atomically.
This also affects Windows and PID locks, not just posix locks.
In the case of pid locks, multiple calls to waitLock within the same
process are allowed because the side lock is locked using a posix lock,
and so multiple exclusive locks can be taken in the same process. So,
this change fixes a similar problem with pid locks.
l1 <- waitLock (Seconds 1) "lck"
l2 <- waitLock (Seconds 1) "lck"
dropLock l1
dropLock l2
Here the l2 side lock fd remained open but not locked,
although the pid lock file was removed. After this change, the second
dropLock will close both fds to the side lock, and delete the pidlock.
According to https://github.com/redneb/disk-free-space/issues/3 ,
disk-free-space should be at least as portable as my homegrown code was.
One change I noticed is, getDiskSize was not implemented for windows
in the old code, and should work now.
This lets readonly repos be used. If a repo is readonly, we can ignore the
keys database, because nothing that we can do will change the state of the
repo anyway.
* Removed the webapp-secure build flag, rolling it into the webapp build
flag.
* Removed the quvi and tahoe build flags, which only adds aeson to
the core dependencies.
* Removed the feed build flag, which only adds feed to the core
dependencies.
Build flags have cost in both code complexity and also make Setup configure
have to work harder to find a usable set of build flags when some
dependencies are missing.
This reverts commit d14770ca9c.
That changed the type of error from an IOError to something else, so broke
stuff that was catching IOErrors.
So back to a UserError, but be explicit this time that's what it's
throwing.
Using fail here causes a "user error" exception to be thrown, which implies
the user is at fault in its wording, which is incorrect.
Also audited for other uses of fail in git-annex; the others are in monadic
contexts where fail may not throw an exception, and involve user input, so
kept them as-is.
The repo path is typically relative, not absolute, so
providing it to absPathFrom doesn't yield an absolute path.
This is not a bug, just unclear documentation.
Indeed, there seem to be no reason to simplifyPath here, which absPathFrom
does, so instead just combine the repo path and the TopFilePath.
Also, removed an export of the TopFilePath constructor; asTopFilePath
is provided to construct one as-is.
Several tricky parts:
* When the conflict is just between the same key being locked and unlocked,
the unlocked version wins, and the file is not renamed in this case.
* Need to update associated file map when conflict resolution renames
an unlocked file.
* git merge runs the smudge filter on the conflicting file, and actually
overwrites the file with the same content it had before, and so
invalidates its inode cache. This makes it difficult to know when it's
safe to remove such files as conflict cruft, without going so far as to
compare their entire contents.
Dealt with this by preventing the smudge filter from populating the file
when a merge is run. However, that also prevents the smudge filter being
run for non-conflicting files, so eg moving a file won't put its new
content into place.
* Ideally, if a merge or a merge conflict resolution renames an unlocked
file, the file in the work tree can just be moved, rather than copying
the content to a new worktree file.
This is attempted to be done in merge conflict resolution, but
due to git merge's behavior of running smudge filters, what actually
seems to happen is the old worktree file with the content is deleted and
rewritten as a pointer file, so doesn't get reused.
So, this is probably not as efficient as it optimally could be.
If that becomes a problem, could look into running the merge in a separate
worktree and updating the real worktree more efficiently, similarly to the
direct mode merge. However, the direct mode merge had a lot of bugs, and
I'd rather not use that more error-prone method unless really needed.
Writes are optimised by queueing up multiple writes when possible.
The queue is flushed after the Annex monad action finishes. That makes it
happen on program termination, and also whenever a nested Annex monad action
finishes.
Reads are optimised by checking once (per AnnexState) if the database
exists. If the database doesn't exist yet, all reads return mempty.
Reads also cause queued writes to be flushed, so reads will always be
consistent with writes (as long as they're made inside the same Annex monad).
A future optimisation path would be to determine when that's not necessary,
which is probably most of the time, and avoid flushing unncessarily.
Design notes for this commit:
- separate reads from writes
- reuse a handle which is left open until program
exit or until the MVar goes out of scope (and autoclosed then)
- writes are queued
- queue is flushed periodically
- immediate queue flush before any read
- auto-flush queue when database handle is garbage collected
- flush queue on exit from Annex monad
(Note that this may happen repeatedly for a single database connection;
or a connection may be reused for multiple Annex monad actions,
possibly even concurrent ones.)
- if database does not exist (or is empty) the handle
is not opened by reads; reads instead return empty results
- writes open the handle if it was not open previously
Caused by AMP.. Since I've finally upgraded my dev laptop to 7.10,
I may start missing imports that are not needed with it but are with older
versions..
http://bugs.debian.org/807341
* Fix insecure temporary permissions when git-annex repair is used in
in a corrupted git repository.
Other calls to withTmpDir didn't leak any potentially private data,
but repair clones the git repository to a temp directory which is made
using the user's umask. Thus, it might expose a git repo that is
otherwise locked down.
* Fix potential denial of service attack when creating temp dirs.
Since withTmpDir used easily predictable temporary directory names,
an attacker could create foo.0, foo.1, etc and as long as it managed to
keep ahead of it, could prevent it from ever returning.
I'd rate this as a low utility DOS attack. Most attackers in a position
to do this could just fill up the disk /tmp is on to prevent anything
from writing temp files. And few parts of git-annex use withTmpDir
anyway, so DOS potential is quite low.
Examined all callers of withTmpDir and satisfied myself that
switching to mkdtmp and so getting a mode 700 temp dir wouldn't break any
of them.
Note that withTmpDirIn continues to not force temp dir to 700.
But it's only used for temp directories inside .git/annex/wherever/
so that is not a problem.
Also re-audited all other uses of temp files and dirs in git-annex.
The Keys database can hold multiple inode caches for a given key. One for
the annex object, and one for each pointer file, which may not be hard
linked to it.
Inode caches for a key are recorded when its content is added to the annex,
but only if it has known pointer files. This is to avoid the overhead of
maintaining the database when not needed.
When the smudge filter outputs a file's content, the inode cache is not
updated, because git's smudge interface doesn't let us write the file. So,
dropping will fall back to doing an expensive verification then. Ideally,
git's interface would be improved, and then the inode cache could be
updated then too.
Had everything available, just didn't combine the progress meter with the
other places progress is sent to update it. (And to a remote repo already
did show progress.)
Most special remotes should already display progress meters with -J,
same as without it. One exception to this is the web, since it relies on
wget/curl progress display without -J. Still todo..
There's a potential race, but it's detected and just results in the other
process failing to take the side lock, so possibly retrying one second
later on. The race window is quite narrow so the extra delay is minor.
Left the side lock files mode 666 because an interruption can leave a side
lock file created by another user for a shared repository. When this
happens, the non-owning user can't delete it (+t) but can still lock it,
and so the code falls back to acting as it did before this commit.
This is less portable, since currently sidelocks rely on /dev/shm.
But, I've seen crazy lustre inconsistencies that make me not trust the
link() method at all, so what can you do.
I have a strace taken on a lustre filesystem on which link() returned 0,
but didn't actually succeed, since the file already existed.
One of the linux man pages recommended using link followed by checking like
this. I was reading it yesterday, but cannot find it now.
Fixes a recent-ish build warning on about 64 bit vs non.
This is the method used by the disk-free-space library, and I tested it to
yield the same results on even 10 tb drives on OSX -- so it's getting 64
bit values.
Also, rename lockContent to lockContentExclusive
inAnnexSafe should perhaps be eliminated, and instead use
`lockContentShared inAnnex`. However, I'm waiting on that, as there are
only 2 call sites for inAnnexSafe and it's fiddly.
On Solaris, using f_bsize provided a value that is apparently much larger
than the real block size. The solaris docs for statvfs say
f_bsize is the "preferred" file system block size, and I guess the
filesystem prefers larger blocks, but uses smaller ones or something.
The docs also say that f_frsize is the "fundamental" block size.
Switched to using f_frsize on Linux and kFreeBSD too, since I guess
f_bsize could in theory vary the same way there too. Assuming that Solaris
is not violating the posix spec, I guess the linux man page for statvfs
is not as well written and I misunderstood it.
This is the kind of annoying thing that makes me not want to use a library.
conduitManagerSettings was a perfectly fine name and could have been kept
forever.
Since I want git-annex to keep building on debian stable, I need to still
support the old http-client, which required explicit calls to
closeManager, or use of withManager to get Managers to close at appropriate
times. This is not needed in the new version, and so they added a
deprecation warning. IMHO much too early, because look at the mess I had to
go through to avoid that deprecation warning while supporting both
versions..
When gpg.program is configured, it's used to get the command to run for
gpg. Useful on systems that have only a gpg2 command or want to use it
instead of the gpg command.
It was failing at link time, some problem with terminatePID.
Re-implemented that to not use a C wrapper function, which cleared up the
problem. Removed old EvilLinker hack with must have been related to the
same problem.
Note that I have not tested this with older ghc's. In
f11f7520b5 I mention having tried this
approach before, and getting segfaults.. So, who knows. It seems to work
fine with ghc 7.10 at least.
This is mostly to be able to see how long a command took to run. Also exit
code may be useful.
Unofrtunately, I can't put the command name in there, because it's not
available at this point, and it would be a much larger change to wrap the
ProcessHandle data type to add that. However, it's generally pretty obvious
which process exited from context.
Since _encodeFilePath generates a String that doesn't use the filesystem
encoding, when this exception is caught, we know we already have such a
String, and can just return it as-is.
This was caused by 23e9d3bb77
an Arbitrary String is not necessarily encoded using the filesystem
encoding, and in a non-utf8 locale, encodeBS throws an exception on such a
string. All I could think to do is limit test data to ascii.
This shouldn't be a problem in practice, because the all Strings in
git-annex that are not generated by Arbitrary should be loaded in a way
that does apply the filesystem encoding.
Oh boy, not again. So, another place that the filesystem encoding needs to
be applied. Yay.
In passing, I changed decodeBS so if a NUL is embedded in the input, the
resulting FilePath doesn't get truncated at that NUL. This was needed to
make prop_b64_roundtrips pass, and on reviewing the callers of decodeBS, I
didn't see any where this wouldn't make sense. When a FilePath is used to
operate on the filesystem, it'll get truncated at a NUL anyway, whereas if
a String is being used for something else, it might conceivably have a NUL
in it, and we wouldn't want it to get truncated when going through
decodeBS.
(NB: There may be a speed impact from this change.)
While cryptohash has SHA3 support, it has not been updated for the final
version of the spec. Note that cryptonite has not been ported to all arches
that cryptohash builds on yet.
This fixes a reversion introduced by relative path changes back last winter.
The root cause is simplifyPath "../foo" was incorrectly yielding "foo".
absPathFrom seems quite horrible. Probably most things that use it should
use </> instead.
This is needed because when preferred content matches on files,
the second pass would otherwise want to drop all keys. Using a bloom filter
avoids this, and in the case of a false positive, a key will be left
undropped that preferred content would allow dropping. Chances of that
happening are a mere 1 in 1 million.
I want this as fast as possible, so it can be added to code paths without
slowing them down.
Avoid the set lookup, and rely on laziness,
drops runtime from 14.37 ns to 11.03 ns according to this criterion benchmark:
import Criterion.Main
import qualified Types.Difference as New
import qualified Types.DifferenceOld as Old
main :: IO ()
main = defaultMain
[ bgroup "hasDifference"
[ bench "new" $ whnf (New.hasDifference New.OneLevelObjectHash) new
, bench "old" $ whnf (Old.hasDifference Old.OneLevelObjectHash) old
]
]
where
s = "fromList [ObjectHashLower, OneLevelObjectHash, OneLevelBranchHash]"
new = New.readDifferences s
old = Old.readDifferences s
A little bit of added boilerplate, but I suppose it's worth it to not
need to worry about set lookup overhead. Note that adding more differences
would slow down the old implementation; the new implementation will run
the same speed.
This reverts commit ef0e3ac22e.
Sebastian thinks best to revert this:
It seems to me the reason I needed to look at activatable sockets
might actually be a networkd bug, and I was in error about patch 0001.
On my machines (without DHCP), networkd quits after configuring the
links. I thought this had to do with network activation, but that was
probably mistaken. This was obscured by my testing the change by doing
systemctl stop/start on networkd; now that I actually unplugged the
network cable, I noticed no DBus messages are triggered by this on
this machine. Your test case might have had a similar problem
(networkd quitting on idle). Might be related to [1].
On another machine (with DHCP) networkd remains active all the time,
and patch 0002 works there. You might want to revert 0001, though:
Suppose someone’s running no manager at all, so that polling would be
required. Because networkd is still listed as activable, we would
refrain from polling – by mistake, because networkd doesn’t seem to
actually go active if we listen on its bus, and it’s listed as
activable even when it’s not configured. Connectivity-related messages
will come in when stopping/starting the service, but not when
unplugging the cable.
This removes a bit of complexity, and should make things faster
(avoids tokenizing Params string), and probably involve less garbage
collection.
In a few places, it was useful to use Params to avoid needing a list,
but that is easily avoided.
Problems noticed while doing this conversion:
* Some uses of Params "oneword" which was entirely unnecessary
overhead.
* A few places that built up a list of parameters with ++
and then used Params to split it!
Test suite passes.
The content file may not be owned by the user running git-annex, in which
case, setting the owner write bit was not enough to let lockContent
act on the file. However, with some core.sharedRepository configs, the file
should be writable by the user's group. So, the thing to do is to call
thawContent on it.
Also cleaned up the code, avoiding creating a lock file if we're going to
open it for create later anyway.
And, if there's an exception while preparing to lock the file, but not at
the point of actually taking the lock, throw an exception, instead of
silently not locking and pretending to succeed.
And, on Windows, always use lock file, even if the repo somehow got into
indirect mode (maybe with cygwin git..)
The one exception is in Utility.Daemon. As long as a process only
daemonizes once, which seems reasonable, and as long as it avoids calling
checkDaemon once it's already running as a daemon, the fcntl locking
gotchas won't be a problem there.
Annex.LockFile has it's own separate lock pool layer, which has been
renamed to LockCache. This is a persistent cache of locks that persist
until closed.
This is not quite done; lockContent stil needs to be converted.
This is certianly a cabal bug for not passing the build options in the
cabal file when building Setup.hs.
And, why oh why did ghc enable this warning by default? So unhappy with
this choice.
That failed on OSX. The temp dir was
/var/folders/fb/pnwjj52n7fg0r9mnvpsfll180000gr/T/downloadurl
and the relative path
../../../../../../Volumes/Visitors/joeyh/git-annex/r/.git/...
Didn't work. I have no clue why, how did OSX manage to break this?
But, the relative path is longer most of the time anyway, so let's
just use the absolute path.
Ambiguous occurrence `bracket'
It could refer to either `Control.Exception.bracket',
imported from `Control.Exception' at Utility/FileMode.hs:14:27-33
(and originally defined in `Control.Exception.Base')
or `Utility.Exception.bracket',
imported from `Utility.Exception' at Utility/FileMode.hs:22:1-24
(and originally defined in `Control.Monad.Catch')
Since we started using this for git repos, when a remote was on another
drive, it resulted in a bogus relative path to it being used by git-annex,
which didn't work.
I don't quite understand the cause of the deadlock. It only occurred
when git-annex-shell transferinfo was being spawned over ssh to feed
download transfer progress back. And if I removed this line from
feedprogressback, the deadlock didn't occur:
bytes <- readSV v
The problem was not a leaked FD, as far as I could see. So what was it?
I don't know.
Anyway, this is a nice clean implementation, that avoids the deadlock.
Just fork off the async threads to handle filtering the stdout and stderr,
and let them clean up their handles whenever they decide to exit.
I've verified that the handles do get promptly closed, although a little
later than I would expect. Presumably that "little later" is what
was making waiting on the threads deadlock.
Despite the late exit, the last line of stdout and stderr appears where
I'd want it to, so I guess this is ok..
Stderr reader blocks waiting for all stderr, and so blocks the process ever
exiting.
I tried several ways to get around this, but no success yet. For now,
disable the stderr reader entirely.
It sounds worse than it is. ;)
Some external special remotes may run commands that display progress on
stderr. If git-annex is run with --quiet, this should filter out such
displays while letting the errors through.
Came up with a generic way to filter out progress messages while keeping
errors, for commands that use stderr for both.
--json mode will disable command outputs too.
New approach is to do it the expensive way for the first 100 paths
on the command line, but then assume the user doesn't care about order too
much and fall back to the cheap way that does not preserve order.
In this situation, curl -o exits successfully without creating the output
file.
There was already a workaround for curl file:/// but I did not realize this
also affected regular url downloads.
To fix it, pre-create the destination file before starting curl.
Since we cannot always know the size of an url before trying to download
it, let's always do this.
Note that since curl is told -C -, we have to consider if this
makes curl try to do a ranged download, which might fail on some servers
where a regular download would have succeeded. My testing indicates
this isn't a problem; since the file is empty, curl seems to not try to
do a ranged download.
Original report: https://github.com/datalad/datalad/issues/79
Curl bug report: https://github.com/bagder/curl/issues/183
The fix is to stop using w82s, which does not properly reconstitute unicode
strings. Instrad, use utf8 bytestring to get the [Word8] to base64. This
passes unicode through perfectly, including any invalid filesystem encoded
characters.
Note that toB64 / fromB64 are also used for creds and cipher
embedding. It would be unfortunate if this change broke those uses.
For cipher embedding, note that ciphers can contain arbitrary bytes (should
really be using ByteString.Char8 there). Testing indicated it's not safe to
use the new fromB64 there; I think that characters were incorrectly
combined.
For credpair embedding, the username or password could contain unicode.
Before, that unicode would fail to round-trip through the b64.
So, I guess this is not going to break any embedded creds that worked
before.
This bug may have affected some creds before, and if so,
this change will not fix old ones, but should fix new ones at least.
hGetSomeString reads one byte at a time, so unicode bytes are not composed.
The problem comes when outputting that to the console with hPut; that
tried to apply the handle's encoding, and so we get mojibake.
Instead, use ByteStrings, and only convert it to a string for parsing, not
for display.
Note that there are a couple of other things that use hGetSomeString,
which I've left as-is for now.
This reverts commit a7f05c007b.
Consider: relPathDirToFile (absPathFrom "/tmp/repo/xxx" "y/bar") "/tmp/repo/.git/annex/objects/xxx"
This needs to always yield "../../../.git/annex/objects/xxx" but on
Windows, it is "..\\..\\/tmp/repo/.git/annex/objects/xxx"
This is necessary for interop between inode caches created on unix and
windows. Which is more important than supporting inodecaches for large keys
with the wrong size, which are broken anyway.
There should be no slowdown from this change, except on Windows.
Avoid using fileSize which maxes out at just 2 gb on Windows.
Instead, use hFileSize, which doesn't have a bounded size.
Fixes support for files > 2 gb on Windows.
Note that the InodeCache code only needs to compare a file size,
so it doesn't matter it the file size wraps. So it has been
left as-is. This was necessary both to avoid invalidating existing inode
caches, and because the code passed FileStatus around and would have become
more expensive if it called getFileSize.
This commit was sponsored by Christian Dietrich.
