There is no risk of getting an inconsistent result here: if the ID
returned by queryValidPathId() is deleted from the database
concurrently, subsequent queries involving that ID will simply fail
(since IDs are never reused).
In the Hydra build farm we fairly regularly get SQLITE_PROTOCOL errors
(e.g., "querying path in database: locking protocol"). The docs for
this error code say that it "is returned if some other process is
messing with file locks and has violated the file locking protocol
that SQLite uses on its rollback journal files." However, the SQLite
source code reveals that this error can also occur under high load:
if( cnt>5 ){
int nDelay = 1; /* Pause time in microseconds */
if( cnt>100 ){
VVA_ONLY( pWal->lockError = 1; )
return SQLITE_PROTOCOL;
}
if( cnt>=10 ) nDelay = (cnt-9)*238; /* Max delay 21ms. Total delay 996ms */
sqlite3OsSleep(pWal->pVfs, nDelay);
}
i.e. if certain locks cannot be not acquired, SQLite will retry a
number of times before giving up and returing SQLITE_PROTOCOL. The
comments say:
Circumstances that cause a RETRY should only last for the briefest
instances of time. No I/O or other system calls are done while the
locks are held, so the locks should not be held for very long. But
if we are unlucky, another process that is holding a lock might get
paged out or take a page-fault that is time-consuming to resolve,
during the few nanoseconds that it is holding the lock. In that case,
it might take longer than normal for the lock to free.
...
The total delay time before giving up is less than 1 second.
On a heavily loaded machine like lucifer (the main Hydra server),
which often has dozens of processes waiting for I/O, it seems to me
that a page fault could easily take more than a second to resolve.
So, let's treat SQLITE_PROTOCOL as SQLITE_BUSY and retry the
transaction.
Issue NixOS/hydra#14.
On a system with multiple CPUs, running Nix operations through the
daemon is significantly slower than "direct" mode:
$ NIX_REMOTE= nix-instantiate '<nixos>' -A system
real 0m0.974s
user 0m0.875s
sys 0m0.088s
$ NIX_REMOTE=daemon nix-instantiate '<nixos>' -A system
real 0m2.118s
user 0m1.463s
sys 0m0.218s
The main reason seems to be that the client and the worker get moved
to a different CPU after every call to the worker. This patch adds a
hack to lock them to the same CPU. With this, the overhead of going
through the daemon is very small:
$ NIX_REMOTE=daemon nix-instantiate '<nixos>' -A system
real 0m1.074s
user 0m0.809s
sys 0m0.098s
For instance, it's pointless to keep copy-from-other-stores running if
there are no other stores, or download-using-manifests if there are no
manifests. This also speeds things up because we don't send queries
to those substituters.
Otherwise subsequent invocations of "--repair" will keep rebuilding
the path. This only happens if the path content differs between
builds (e.g. due to timestamps).
It is surprisingly impossible to check if a mountpoint is a bind mount
on Linux, and in my previous commit I forgot to check if /nix/store was
even a mountpoint at all. statvfs.f_flag is not populated with MS_BIND
(and even if it were, my check was wrong in the previous commit).
Luckily, the semantics of mount with MS_REMOUNT | MS_BIND make both
checks unnecessary: if /nix/store is not a mountpoint, then mount will
fail with EINVAL, and if /nix/store is not a bind-mount, then it will
not be made writable. Thus, if /nix/store is not a mountpoint, we fail
immediately (since we don't know how to make it writable), and if
/nix/store IS a mountpoint but not a bind-mount, we fail at first write
(see below for why we can't check and fail immediately).
Note that, due to what is IMO buggy behavior in Linux, calling mount
with MS_REMOUNT | MS_BIND on a non-bind readonly mount makes the
mountpoint appear writable in two places: In the sixth (but not the
10th!) column of mountinfo, and in the f_flags member of struct statfs.
All other syscalls behave as if the mount point were still readonly (at
least for Linux 3.9-rc1, but I don't think this has changed recently or
is expected to soon). My preferred semantics would be for MS_REMOUNT |
MS_BIND to fail on a non-bind mount, as it doesn't make sense to remount
a non bind-mount as a bind mount.
/nix/store could be a read-only bind mount even if it is / in its own filesystem, so checking the 4th field in mountinfo is insufficient.
Signed-off-by: Shea Levy <shea@shealevy.com>
It turns out that in multi-user Nix, a builder may be able to do
ln /etc/shadow $out/foo
Afterwards, canonicalisePathMetaData() will be applied to $out/foo,
causing /etc/shadow's mode to be set to 444 (readable by everybody but
writable by nobody). That's obviously Very Bad.
Fortunately, this fails in NixOS's default configuration because
/nix/store is a bind mount, so "ln" will fail with "Invalid
cross-device link". It also fails if hard-link restrictions are
enabled, so a workaround is:
echo 1 > /proc/sys/fs/protected_hardlinks
The solution is to check that all files in $out are owned by the build
user. This means that innocuous operations like "ln
${pkgs.foo}/some-file $out/" are now rejected, but that already failed
in chroot builds anyway.
With this flag, if any valid derivation output is missing or corrupt,
it will be recreated by using a substitute if available, or by
rebuilding the derivation. The latter may use hash rewriting if
chroots are not available.
Using the immutable bit is problematic, especially in conjunction with
store optimisation. For instance, if the garbage collector deletes a
file, it has to clear its immutable bit, but if the file has
additional hard links, we can't set the bit afterwards because we
don't know the remaining paths.
So now that we support having the entire Nix store as a read-only
mount, we may as well drop the immutable bit. Unfortunately, we have
to keep the code to clear the immutable bit for backwards
compatibility.
It turns out that the immutable bit doesn't work all that well. A
better way is to make the entire Nix store a read-only bind mount,
i.e. by doing
$ mount --bind /nix/store /nix/store
$ mount -o remount,ro,bind /nix/store
(This would typically done in an early boot script, before anything
from /nix/store is used.)
Since Nix needs to be able to write to the Nix store, it now detects
if /nix/store is a read-only bind mount and then makes it writable in
a private mount namespace.
To implement binary caches efficiently, Hydra needs to be able to map
the hash part of a store path (e.g. "gbg...zr7") to the full store
path (e.g. "/nix/store/gbg...kzr7-subversion-1.7.5"). (The binary
cache mechanism uses hash parts as a key for looking up store paths to
ensure privacy.) However, doing a search in the Nix store for
/nix/store/<hash>* is expensive since it requires reading the entire
directory. queryPathFromHashPart() prevents this by doing a cheap
database lookup.
queryValidPaths() combines multiple calls to isValidPath() in one.
This matters when using the Nix daemon because it reduces latency.
For instance, on "nix-env -qas \*" it reduces execution time from 5.7s
to 4.7s (which is indistinguishable from the non-daemon case).
Getting substitute information using the binary cache substituter has
non-trivial latency overhead. A package or NixOS system configuration
can have hundreds of dependencies, and in the worst case (when the
local info cache is empty) we have to do a separate HTTP request for
each of these. If the ping time to the server is t, getting N info
files will take tN seconds; e.g., with a ping time of 0.1s to
nixos.org, sequentially downloading 1000 info files (a typical NixOS
config) will take at least 100 seconds.
To fix this problem, the binary cache substituter can now perform
requests in parallel. This required changing the substituter
interface to support a function querySubstitutablePathInfos() that
queries multiple paths at the same time, and rewriting queryMissing()
to take advantage of parallelism. (Due to local caching,
parallelising queryMissing() is sufficient for most use cases, since
it's almost always called before building a derivation and thus fills
the local info cache.)
For example, parallelism speeds up querying all 1056 paths in a
particular NixOS system configuration from 116s to 2.6s. It works so
well because the eccentricity of the top-level derivation in the
dependency graph is only 9. So we only need 10 round-trips (when
using an unlimited number of parallel connections) to get everything.
Currently we do a maximum of 150 parallel connections to the server.
Thus it's important that the binary cache server (e.g. nixos.org) has
a high connection limit. Alternatively we could use HTTP pipelining,
but WWW::Curl doesn't support it and libcurl has a hard-coded limit of
5 requests per pipeline.
We can't open a SQLite database if the disk is full. Since this
prevents the garbage collector from running when it's most needed, we
reserve some dummy space that we can free just before doing a garbage
collection. This actually revives some old code from the Berkeley DB
days.
Fixes#27.
Make the garbage collector more concurrent by deleting valid paths
outside the region where we're holding the global GC lock. This
should greatly reduce the time during which new builds are blocked,
since the deletion accounts for the vast majority of the time spent in
the GC.
To ensure that this is safe, the valid paths are invalidated and
renamed to some arbitrary path while we're holding the lock. This
ensures that we when we finally delete the path, it's not a (newly)
valid or locked path.
Not all SQLite builds have the function sqlite3_table_column_metadata.
We were only using it in a schema upgrade check for compatibility with
databases that were probably never seen in the wild. So remove it.
I was bitten one time too many by Python modifying the Nix store by
creating *.pyc files when run as root. On Linux, we can prevent this
by setting the immutable bit on files and directories (as in ‘chattr
+i’). This isn't supported by all filesystems, so it's not an error
if setting the bit fails. The immutable bit is cleared by the garbage
collector before deleting a path. The only tricky aspect is in
optimiseStore(), since it's forbidden to create hard links to an
immutable file. Thus optimiseStore() temporarily clears the immutable
bit before creating the link.
* Buffer the HashSink. This speeds up hashing a bit because it
prevents lots of calls to the hash update functions (e.g. nix-hash
went from 9.3s to 8.7s of user time on the closure of my
/var/run/current-system).
‘nix-store --export’.
* Add a Perl module that provides the functionality of
‘nix-copy-closure --to’. This is used by build-remote.pl so it no
longer needs to start a separate nix-copy-closure process. Also, it
uses the Perl API to do the export, so it doesn't need to start a
separate nix-store process either. As a result, nix-copy-closure
and build-remote.pl should no longer fail on very large closures due
to an "Argument list too long" error. (Note that having very many
dependencies in a single derivation can still fail because the
environment can become too large. Can't be helped though.)
This should also fix:
nix-instantiate: ./../boost/shared_ptr.hpp:254: T* boost::shared_ptr<T>::operator->() const [with T = nix::StoreAPI]: Assertion `px != 0' failed.
which was caused by hashDerivationModulo() calling the ‘store’
object (during store upgrades) before openStore() assigned it.
derivations added to the store by clients have "correct" output
paths (meaning that the output paths are computed by hashing the
derivation according to a certain algorithm). This means that a
malicious user could craft a special .drv file to build *any*
desired path in the store with any desired contents (so long as the
path doesn't already exist). Then the attacker just needs to wait
for a victim to come along and install the compromised path.
For instance, if Alice (the attacker) knows that the latest Firefox
derivation in Nixpkgs produces the path
/nix/store/1a5nyfd4ajxbyy97r1fslhgrv70gj8a7-firefox-5.0.1
then (provided this path doesn't already exist) she can craft a .drv
file that creates that path (i.e., has it as one of its outputs),
add it to the store using "nix-store --add", and build it with
"nix-store -r". So the fake .drv could write a Trojan to the
Firefox path. Then, if user Bob (the victim) comes along and does
$ nix-env -i firefox
$ firefox
he executes the Trojan injected by Alice.
The fix is to have the Nix daemon verify that derivation outputs are
correct (in addValidPath()). This required some refactoring to move
the hash computation code to libstore.
while checking the contents, since this operation can take a very
long time to finish. Also, fill in missing narSize fields in the DB
while doing this.
even with a very long busy timeout, because SQLITE_BUSY is also
returned to resolve deadlocks. This should get rid of random
"database is locked" errors. This is kind of hard to test though.
* Fix a horrible bug in deleteFromStore(): deletePathWrapped() should
be called after committing the transaction, not before, because the
commit might not succeed.
race with other processes that add new referrers to a path,
resulting in the garbage collector crashing with "foreign key
constraint failed". (Nix/4)
* Make --gc --print-dead etc. interruptible.
* If a path has disappeared, check its referrers first, and don't try
to invalidate paths that have valid referrers. Otherwise we get a
foreign key constraint violation.
* Read the whole Nix store directory instead of statting each valid
path, which is slower.
* Acquire the global GC lock.
faster than the old mode when fsyncs are enabled, because it only
performs an fsync() when doing a checkpoint, rather than at every
commit. Some timings for doing a "nix-instantiate /etc/nixos/nixos
-A system" after modifying the stdenv setup script:
42.5s - SQLite 3.6.23 with truncate mode and fsync
3.4s - SQLite 3.6.23 with truncate mode and no fsync
32.1s - SQLite 3.7.0 with truncate mode and fsync
16.8s - SQLite 3.7.0 with WAL mode and fsync, auto-checkpoint
every 1000 pages
8.3s - SQLite 3.7.0 with WAL mode and fsync, auto-checkpoint
every 8192 pages
1.7s - SQLite 3.7.0 with WAL mode and no fsync
The default is now to use WAL mode with fsyncs. Because WAL doesn't
work on remote filesystems such as NFS (as it uses shared memory),
truncate mode can be re-enabled by setting the "use-sqlite-wal"
option to false.
doesn't work because the garbage collector doesn't actually look at
locks. So r22253 was stupid. Use addTempRoot() instead. Also,
locking the temporary directory in exportPath() was silly because it
isn't even in the store.
