It was holding on to a Value* (i.e. a std::shared_ptr<ValidPathInfo>*)
outside of the pathInfoCache lock, so the std::shared_ptr could be
destroyed between the release of the lock and the decrement of the
std::shared_ptr refcount. This can happen if more than
'path-info-cache-size' paths are added in the meantime, *or* if
clearPathInfoCache() is called. The hydra-queue-runner queue monitor
thread periodically calls the later, so is likely to trigger a crash.
Fixes https://github.com/NixOS/hydra/issues/542.
Doing so prevents emacs tags from working, as well as makes the code extremely
confusing for a newbie.
In the prior state, if someone wants to find the definition of "ExprApp" for
example, a grep through the code reveals nothing. Since the definition could be
hiding in numerous ".h" files, it's really difficult to find. This personally
took me several hours to figure out.
This can be iterated on and currently leaves out settings we know we
want to forward, but it fixes#1713 and fixes#1935 and isn't
fundamentally broken like the status quo. Future changes are suggested
in a comment.
Flex's regexes have an annoying feature: the dot matches everything
except a newline. This causes problems for expressions like:
"${0}\
"
where the backslash-newline combination matches this rule instead of the
intended one mentioned in the comment:
<STRING>\$|\\|\$\\ {
/* This can only occur when we reach EOF, otherwise the above
(...|\$[^\{\"\\]|\\.|\$\\.)+ would have triggered.
This is technically invalid, but we leave the problem to the
parser who fails with exact location. */
return STR;
}
However, the parser actually accepts the resulting token sequence
('"' DOLLAR_CURLY 0 '}' STR '"'), which is a problem because the lexer
rule didn't assign anything to yylval. Ultimately this leads to a crash
when dereferencing a NULL pointer in ExprConcatStrings::bindVars().
The fix does change the syntax of the language in some corner cases
but I think it's only turning previously invalid (or crashing) syntax
to valid syntax. E.g.
"a\
b"
and
''a''\
b''
were previously syntax errors but now both result in "a\nb".
Found by afl-fuzz.
This allows building armv[67]l-linux derivations on compatible aarch64
machines. Failure to add the architecture may result from missing
hardware support, in which case we can't run 32-bit binaries and don't
need to restrict them with seccomp anyway,
This allows specifying additional systems that a machine is able to
build for. This may apply on some armv7-capable aarch64 processors, or
on systems using qemu-user with binfmt-misc to support transparent
execution of foreign-arch programs.
This removes the previous hard-coded assumptions about which systems are
ABI-compatible with which other systems, and instead relies on the user
to specify any additional platforms that they have ensured compatibility
for and wish to build for locally.
NixOS should probably add i686-linux on x86_64-linux systems for this
setting by default.
Otherwise, running e.g.
nix-instantiate --eval -E --strict 'builtins.replaceStrings [""] ["X"] "abc"'
would just hang in an infinite loop.
Found by afl-fuzz.
First attempt of this was reverted in e2d71bd186 because it caused
another infinite loop, which is fixed now and a test added.
This is important since this is given as an example.
Other patterns containing "empty search string" will still
be handled differently on different platforms ("asdf|")
but that's less of an issue.
The overhead of sandbox builds is a problem on NixOS (since building a
NixOS configuration involves a lot of small derivations) but not for
typical non-NixOS use cases. So outside of NixOS we can enable it.
Issue #179.
The assertion is broken because there is no one-to-one mapping from
length of a base64 string to the length of the output.
E.g.
"1q69lz7Empb06nzfkj651413n9icx0njmyr3xzq1j9q=" results in a 32-byte output.
"1q69lz7Empb06nzfkj651413n9icx0njmyr3xzq1j9qy" results in a 33-byte output.
To reproduce, evaluate:
builtins.derivationStrict {
name = "0";
builder = "0";
system = "0";
outputHashAlgo = "sha256";
outputHash = "1q69lz7Empb06nzfkj651413n9icx0njmyr3xzq1j9qy";
}
Found by afl-fuzz.
Otherwise, running e.g.
nix-instantiate --eval -E --strict 'builtins.replaceStrings [""] ["X"] "abc"'
would just hang in an infinite loop.
Found by afl-fuzz.
Instead of having lexicographicOrder() create a temporary sorted array
of Attr*:s and copying attr names from that, copy the attr names
first and then sort that.
Previously, this would fail at startup for non-NixOS installs:
nix-env --help
The fix for this is to just use "nixManDir" as the value for MANPATH
when spawning "man".
To test this, I’m using the following:
$ nix-build release.nix -A build
$ MANPATH= ./result/bin/nix-env --help
Fixes#1627
nix-store --export, nix-store --dump, and nix dump-path would previously
fail silently if writing the data out failed, because
a) FdSink::write ignored exceptions, and
b) the commands relied on FdSink's destructor, which ignores
exceptions, to flush the data out.
This could cause rather opaque issues with installing nixos, because
nix-store --export would happily proceed even if it couldn't write its
data out (e.g. if nix-store --import on the other side of the pipe
failed).
This commit adds tests that expose these issues in the nix-store
commands, and fixes them for all three.
This was caused by derivations with 'allowSubstitutes = false'. Such
derivations will be built locally. However, if there is another
SubstitionGoal that has the output of the first derivation in its
closure, then the path will be simultaneously built and substituted.
There was a check to catch this situation (via pathIsLockedByMe()),
but it no longer worked reliably because substitutions are now done in
another thread. (Thus the comment 'It can't happen between here and
the lockPaths() call below because we're not allowing multi-threading'
was no longer valid.)
The fix is to handle the path already being locked in both
SubstitutionGoal and DerivationGoal.
All ANSI sequences except color setting are now filtered out. In
particular, terminal resets (such as from NixOS VM tests) are filtered
out.
Also, fix the completely broken tab character handling.
builtins.path allows specifying the name of a path (which makes paths
with store-illegal names now addable), allows adding paths with flat
instead of recursive hashes, allows specifying a filter (so is a
generalization of filterSource), and allows specifying an expected
hash (enabling safe path adding in pure mode).
the case of hydra where the overhead of single threaded encoding is more
noticeable e.g most of the time spent in "Sending inputs"/"Receiving outputs"
is due to compression while the actual upload to the binary cache seems
to be negligible.
This is needed by nixos-install, which uses the Nix store on the
installation CD as a substituter. We don't want to disable signature
checking entirely because substitutes from cache.nixos.org should
still be checked. So now we can pas "local?trusted=1" to mark only the
Nix store in /nix as not requiring signatures.
Fixes#1819.
Instead, if a fixed-output derivation produces has an incorrect output
hash, we now unconditionally move the outputs to the path
corresponding with the actual hash and register it as valid. Thus,
after correcting the hash in the Nix expression (e.g. in a fetchurl
call), the fixed-output derivation doesn't have to be built again.
It would still be good to have a command for reporting the actual hash
of a fixed-output derivation (instead of throwing an error), but
"nix-build --hash" didn't do that.
Following discussion with Shea and Graham. It's a big enough change
from the last release. Also, from a semver perspective, 2.0 makes more
sense because we did remove some interfaces (like nix-pull/nix-push).
Some servers, such as Artifactory, allow uploading with PUT and BASIC
auth. This allows nix copy to work to upload binaries to those
servers.
Worked on together with @adelbertc
In this mode, the following restrictions apply:
* The builtins currentTime, currentSystem and storePath throw an
error.
* $NIX_PATH and -I are ignored.
* fetchGit and fetchMercurial require a revision hash.
* fetchurl and fetchTarball require a sha256 attribute.
* No file system access is allowed outside of the paths returned by
fetch{Git,Mercurial,url,Tarball}. Thus 'nix build -f ./foo.nix' is
not allowed.
Thus, the evaluation result is completely reproducible from the
command line arguments. E.g.
nix build --pure-eval '(
let
nix = fetchGit { url = https://github.com/NixOS/nixpkgs.git; rev = "9c927de4b179a6dd210dd88d34bda8af4b575680"; };
nixpkgs = fetchGit { url = https://github.com/NixOS/nixpkgs.git; ref = "release-17.09"; rev = "66b4de79e3841530e6d9c6baf98702aa1f7124e4"; };
in (import (nix + "/release.nix") { inherit nix nixpkgs; }).build.x86_64-linux
)'
The goal is to enable completely reproducible and traceable
evaluation. For example, a NixOS configuration could be fully
described by a single Git commit hash. 'nixos-rebuild' would do
something like
nix build --pure-eval '(
(import (fetchGit { url = file:///my-nixos-config; rev = "..."; })).system
')
where the Git repository /my-nixos-config would use further fetchGit
calls or Git externals to fetch Nixpkgs and whatever other
dependencies it has. Either way, the commit hash would uniquely
identify the NixOS configuration and allow it to reproduced.
* Look for both 'brotli' and 'bro' as external command,
since upstream has renamed it in newer versions.
If neither are found, current runtime behavior
is preserved: try to find 'bro' on PATH.
* Limit amount handed to BrotliEncoderCompressStream
to ensure interrupts are processed in a timely manner.
Testing shows negligible performance impact.
(Other compression sinks don't seem to require this)
E.g.
$ time nix cat-store --store https://cache.nixos.org?local-nar-cache=/tmp/nars \
/nix/store/b0w2hafndl09h64fhb86kw6bmhbmnpm1-blender-2.79/share/icons/hicolor/scalable/apps/blender.svg > /dev/null
real 0m4.139s
$ time nix cat-store --store https://cache.nixos.org?local-nar-cache=/tmp/nars \
/nix/store/b0w2hafndl09h64fhb86kw6bmhbmnpm1-blender-2.79/share/icons/hicolor/scalable/apps/blender.svg > /dev/null
real 0m0.024s
(Before, the second call took ~0.220s.)
This will use a NAR listing in
/tmp/nars/b0w2hafndl09h64fhb86kw6bmhbmnpm1.ls containing all metadata,
including the offsets of regular files inside the NAR. Thus, we don't
need to read the entire NAR. (We do read the entire listing, but
that's generally pretty small. We could use a SQLite DB by borrowing
some more code from nixos-channel-scripts/file-cache.hh.)
This is primarily useful when Hydra is serving files from an S3 binary
cache, in particular when you have giant NARs. E.g. we had some 12 GiB
NARs, so accessing individuals files was pretty slow.
propagated-user-env-packages files in nixpkgs aren't all terminated by
newlines, as buildenv expected. Now it does not require a terminating
newline; note that this introduces a behaviour change: propagated user
env packages may now be spread across multiple lines. However, nix
1.11.x still expects them to be on a single line so this shouldn't be
used in nixpkgs for now.
The storeUri variable in the build-remote hook is declared very much to
the start of the main function and a bunch of lines later, the same
variable gets checked via hasPrefix() but it gets assigned *after* that
check when the most suitable machine for the build was choosen.
So I guess this was just a typo in d16fd24973
and what we really want is to either checkd the prefix *after* assigning
storeUri or use bestMachine->storeUri directly.
I choose the latter, because the former could introduce even more
regressions if the try block where the variable gets assigned terminates
early.
Nevertheless, the reason why the log output didn't work is because
hasPrefix() checked for "ssh://" in front of storeUri, but if the
storeUri isn't set correctly (or at all), we don't get the log file
descriptor set up properly, leading to no log output.
I've adjusted the remote-builds test to include a regression test for
this, so that we can make sure we get a build output when using remote
builds.
In addition to that I've tested this with two of my build farms and the
build logs are emitted correctly again.
Signed-off-by: aszlig <aszlig@nix.build>
The name had become a misnomer since it's not only for substitution
from binary caches, but when adding/copying any
(non-content-addressed) path to a store.
This allows specifying the AWS configuration profile to use. E.g.
nix copy --from s3://my-cache?profile=aws-dev-account /nix/store/cf3isrlqavvd5w7rpky1fa8j9lcnlggm-...
As far as we're concerned, not being able to access a file just means
the file is missing. Plus, AWS explicitly goes out of its way to
return a 403 if the file is missing and the requester doesn't have
permission to list the bucket.
Also getting rid of an old hack that Eelco said was only relevant
to an older AWS SDK.
For example, you can write
src = fetchgit ./.;
and if ./. refers to an unclean working tree, that tree will be copied
to the Nix store. This removes the need for "cleanSource".
This will allow bind and connect to 127.0.0.1, which can reduce purity/
security (if you're running a vulnerable service on localhost) but is
also needed for a ton of test suites, so I'm leaving it turned off by
default but allowing certain derivations to turn it on as needed.
It also allows DNS resolution of arbitrary hostnames but I haven't found
a way to avoid that. In principle I'd just want to allow resolving
localhost but that doesn't seem to be possible.
I don't think this belongs under `build-use-sandbox = relaxed` because we
want it on Hydra and I don't think it's the end of the world.
Used to determine symlink size with stat and value with readlink.
This could technically result in garbage if symlink changed between
calls. Also gets around the broken stat implementation in our
network filesystem (returns size + 1 giving a byte of garbage).
The computation of urlHash didn't take the name into account, so
subsequent fetchurl calls with the same URL but a different name would
resolve to the same cached store path.
The "name" attribute defaults to "source", which we should use for all
similar functions (e.g. fetchTarball and in Hydra) to ensure that we
get a consistent store path regardless of how the tree is fetched.
"source" is not necessarily a correct label, but using an empty name
is problematic: you get an ugly store path ending in a dash, and it's
impossible to have a fixed-output derivation that produces that path
because ".drv" is not a valid store name.
Fixes#904.
You can now include files via the "builders" option, using the syntax
"@<filename>". Having only one option makes it easier to override
builders completely.
For backward compatibility, the default is "@/etc/nix/machines", or
"@<filename>" for each file name in NIX_REMOTE_SYSTEMS.
This makes it slightly more manageable to see at a glance what in a
build's sandbox profile is unique to the build and what is standard. Also
a first step to factoring more of our Darwin logic into scheme functions
that will allow us a bit more flexibility. And of course less of that
nasty codegen in C++! 😀
This speeds up commands like "nix cat-store". For example:
$ time nix cat-store --store https://cache.nixos.org?local-nar-cache=/tmp/nar-cache /nix/store/i60yncmq6w9dyv37zd2k454g0fkl3arl-systemd-234/etc/udev/udev.conf
real 0m4.336s
$ time nix cat-store --store https://cache.nixos.org?local-nar-cache=/tmp/nar-cache /nix/store/i60yncmq6w9dyv37zd2k454g0fkl3arl-systemd-234/etc/udev/udev.conf
real 0m0.045s
The primary motivation is to allow hydra-server to serve files from S3
binary caches. Previously Hydra had a hack to do "nix-store -r
<path>", but that fetches the entire closure so is prohibitively
expensive.
There is no garbage collection of the NAR cache yet. Also, the entire
NAR is read when accessing a single member file. We could generate the
NAR listing to provide random access.
Note: the NAR cache is indexed by the store path hash, not the content
hash, so NAR caches should not be shared between binary caches, unless
you're sure that all your builds are binary-reproducible.
Probably as a result of a bad merge in
4b8f1b0ec0, we had both a
BinaryCacheStoreAccessor and a
RemoteFSAccessor. BinaryCacheStore::getFSAccessor() returned the
latter, but BinaryCacheStore::addToStore() checked for the
former. This probably caused hydra-queue-runner to download paths that
it just uploaded.
This check spuriously fails for e.g. git@github.com:NixOS/nixpkgs.git,
and even for ssh://git@github.com/NixOS/nixpkgs.git, and is made
redundant by the checks git itself will do when fetching the repo. We
instead pass a -- before passing the URI to git to avoid injection.
I needed this to test ACL/xattr removal in
canonicalisePathMetaData(). Might also be useful if you need to build
old Nixpkgs that doesn't have the required patches to remove
setuid/setgid creation.
The worker threads could exit prematurely if they finished processing
all items while the main thread was still adding items. In particular,
this caused hanging nix-store --serve processes in the build farm.
Also, process items from the main thread.
It was getting too much like whac-a-mole listing all the retriable error
conditions, so we now retry by default and list the cases where retrying
is almost certainly hopeless.
I find the error message 'nix-env --set-flag priority NUMBER PKGNAME'
not as helpful as it could be :
- doesn't share the current priorities
- doesn't say that the command must be run on the already installed
PKGNAME (which is confusing the first time)
- the doc needs careful reading:
"If there are multiple derivations matching a name in args that have the same name (e.g., gcc-3.3.6 and gcc-4.1.1), then the derivation with the highest priority is used."
if one stops reading there, he is screwed. Salvation comes with reading "A derivation can define a priority by declaring the meta.priority attribute. This attribute should be a number, with a higher value denoting a lower priority. The default priority is 0."
To sum it up, lower number wins. I tried to convey this idea in the
message too.
This is a hack to make hydra-queue-runner free its temproots
periodically, thereby ensuring that garbage collection of the
corresponding paths is not blocked until the queue runner is
restarted.
It would be better if temproots could be released earlier than at
process exit. I started working on a RAII object returned by functions
like addToStore() that releases temproots. However, this would be a
pretty massive change so I gave up on it for now.
For example,
$ nix-store -q --roots /nix/store/7phd2sav7068nivgvmj2vpm3v47fd27l-patchelf-0.8pre845_0315148
{temp:1}
denotes that the path is only being kept alive by a temporary root
(i.e. /nix/var/nix/temproots/). Similarly,
$ nix-store --gc --print-roots
...
{memory:9} -> /nix/store/094gpjn9f15ip17wzxhma4r51nvsj17p-curl-7.53.1
shows that curl is being used by some process.
This command shows why a package has another package in its runtime
closure. For example, to see why VLC has libdrm.dev in its closure:
$ nix why-depends nixpkgs.vlc nixpkgs.libdrm.dev
/nix/store/g901z9pcj0n5yy5n6ykxk3qm4ina1d6z-vlc-2.2.5.1:
lib/libvlccore.so.8.0.0: …nfig:/nix/store/405lmx6jl8lp0ad1vrr6j498chrqhz8g-libdrm-2.4.75-d…
/nix/store/s3nm7kd8hlcg0facn2q1ff2n7wrwdi2l-mesa-noglu-17.0.7-dev:
nix-support/propagated-native-build-inputs: …-dev /nix/store/405lmx6jl8lp0ad1vrr6j498chrqhz8g-libdrm-2.4.75-d…
Thus, VLC's lib/libvlccore.so.8.0.0 as well as mesa-noglu's
nix-support/propagated-native-build-inputs cause the dependency.
In particular, process() won't return as long as there are active
items. This prevents work item lambdas from referring to stack frames
that no longer exist.
Since we may use a dedicated file descriptor in the future, this
allows us to change it. So builders can do
if [[ -n $NIX_LOG_FD ]]; then
echo "@nix { message... }" >&$NIX_LOG_FD
fi
Nix can now automatically run the garbage collector during builds or
while adding paths to the store. The option "min-free = <bytes>"
specifies that Nix should run the garbage collector whenever free
space in the Nix store drops below <bytes>. It will then delete
garbage until "max-free" bytes are available.
Garbage collection during builds is asynchronous; running builds are
not paused and new builds are not blocked. However, there also is a
synchronous GC run prior to the first build/substitution.
Currently, no old GC roots are deleted (as in "nix-collect-garbage
-d").
Since file locks are per-process rather than per-file-descriptor, the
garbage collector would always acquire a lock on its own temproots
file and conclude that it's stale.
Without this, substitute info is fetched sequentially, which is
superslow. In the old UI (e.g. nix-build), we call printMissing(),
which calls queryMissing(), thereby preheating the binary cache
cache. But the new UI doesn't do that.
In particular, drop the "build-" and "gc-" prefixes which are
pointless. So now you can say
nix build --no-sandbox
instead of
nix build --no-build-use-sandbox
This is useful for testing commands in isolation.
For example,
$ nix run nixpkgs.geeqie -i -k DISPLAY -k XAUTHORITY -c geeqie
runs geeqie in an empty environment, except for $DISPLAY and
$XAUTHORITY.
E.g.
nix run nixpkgs.hello -c hello --greeting Hallo
Note that unlike "nix-shell --command", no quoting of arguments is
necessary.
"-c" (short for "--command") cannot be combined with "--" because they
both consume all remaining arguments. But since installables shouldn't
start with a dash, this is unlikely to cause problems.
Running "nix run" with a diverted store, e.g.
$ nix run --store local?root=/tmp/nix nixpkgs.hello
stopped working when Nix became multithreaded, because
unshare(CLONE_NEWUSER) doesn't work in multithreaded processes. The
obvious solution is to terminate all other threads first, but 1) there
is no way to terminate Boehm GC marker threads; and 2) it appears that
the kernel has a race where unshare(CLONE_NEWUSER) will still fail for
some indeterminate amount of time after joining other threads.
So instead, "nix run" will now exec() a single-threaded helper ("nix
__run_in_chroot") that performs the actual unshare()/chroot()/exec().
Now that we use threads in lots of places, it's possible for
TunnelLogger::log() to be called asynchronously from other threads
than the main loop. So we need to ensure that STDERR_NEXT messages
don't clobber other messages.
Besides being unused, this function has a bug that it will incorrectly
decode the path component Ubuntu\04016.04.2\040LTS\040amd64 as
"Ubuntu.04.2 LTS amd64" instead of "Ubuntu 16.04.2 LTS amd64".
This adds an argument "rev" specififying the Git commit hash. The
existing argument "rev" is renamed to "ref". The default value for
"ref" is "master". When specifying a hash, it's necessary to specify a
ref since we're not cloning the entire repository but only fetching a
specific ref.
Example usage:
builtins.fetchgit {
url = https://github.com/NixOS/nixpkgs.git;
ref = "release-16.03";
rev = "c1c0484041ab6f9c6858c8ade80a8477c9ae4442";
};
The package list is now cached in
~/.cache/nix/package-search.json. This gives a substantial speedup to
"nix search" queries. For example (on an SSD):
First run: (no package search cache, cold page cache)
$ time nix search blender
Attribute name: nixpkgs.blender
Package name: blender
Version: 2.78c
Description: 3D Creation/Animation/Publishing System
real 0m6.516s
Second run: (package search cache populated)
$ time nix search blender
Attribute name: nixpkgs.blender
Package name: blender
Version: 2.78c
Description: 3D Creation/Animation/Publishing System
real 0m0.143s
In particular, don't use base-64, which we don't support. (We do have
base-32 redirects for hysterical reasons.)
Also, add a test for the hashed mirror feature.
This doesn't work in read-only mode, ensuring that operations like
nix path-info --store https://cache.nixos.org -S nixpkgs.hello
(asking for the closure size of nixpkgs.hello in cache.nixos.org) work
when nixpkgs.hello doesn't exist in the local store.
On second though this was annoying. E.g. "nix log nixpkgs.hello" would
build/download Hello first, even though the log can be fetched
directly from the binary cache.
May need to revisit this.
This allows builds to call setuid binaries. This was previously
possible until we started using seccomp. Turns out that seccomp by
default disallows processes from acquiring new privileges. Generally,
any use of setuid binaries (except those created by the builder
itself) is by definition impure, but some people were relying on this
ability for certain tests.
Example:
$ nix build '(with import <nixpkgs> {}; runCommand "foo" {} "/run/wrappers/bin/ping -c 1 8.8.8.8; exit 1")' --no-allow-new-privileges
builder for ‘/nix/store/j0nd8kv85hd6r4kxgnwzvr0k65ykf6fv-foo.drv’ failed with exit code 1; last 2 log lines:
cannot raise the capability into the Ambient set
: Operation not permitted
$ nix build '(with import <nixpkgs> {}; runCommand "foo" {} "/run/wrappers/bin/ping -c 1 8.8.8.8; exit 1")' --allow-new-privileges
builder for ‘/nix/store/j0nd8kv85hd6r4kxgnwzvr0k65ykf6fv-foo.drv’ failed with exit code 1; last 6 log lines:
PING 8.8.8.8 (8.8.8.8) 56(84) bytes of data.
64 bytes from 8.8.8.8: icmp_seq=1 ttl=46 time=15.2 ms
Fixes#1429.
Functions like copyClosure() had 3 bool arguments, which creates a
severe risk of mixing up arguments.
Also, implement copyClosure() using copyPaths().
Cygwin sqlite3 is patched to call SetDllDirectory("/usr/bin") on init, which
affects the current process and is inherited by child processes. It causes
DLLs to be loaded from /usr/bin/ before $PATH, which breaks all sorts of
things. A typical failures would be header/lib version mismatches (e.g.
openssl when running checkPhase on openssh). We'll just set it back to the
default value.
Note that this is a problem with the cygwin version of sqlite3 (currently
3.18.0). nixpkgs doesn't have the problematic patch.
There's no reason to restrict this to Error exceptions. This shouldn't
matter to #1407 since the repl doesn't catch non-Error exceptions
anyway, but you never know...
Recently aws-sdk-cpp quietly switched to using S3 virtual host URIs
(https://github.com/aws/aws-sdk-cpp/commit/69d9c53882), i.e. it sends
requests to http://<bucket>.<region>.s3.amazonaws.com rather than
http://<region>.s3.amazonaws.com/<bucket>. However this interacts
badly with curl connection reuse. For example, if we do the following:
1) Check whether a bucket exists using GetBucketLocation.
2) If it doesn't, create it using CreateBucket.
3) Do operations on the bucket.
then 3) will fail for a minute or so with a NoSuchBucket exception,
presumably because the server being hit is a fallback for cases when
buckets don't exist.
Disabling the use of virtual hosts ensures that 3) succeeds
immediately. (I don't know what S3's consistency guarantees are for
bucket creation, but in practice buckets appear to be available
immediately.)
Linus Heckemann
Jörg Thalheim
Maximilian Bosch