Enables builds with ASAN to catch memory corruption
bugs faster and in CI. This is an incredibly valuable
instrument that must be used as much as possible.
Somewhat based on jade's work from Lix, though there's a lot that
we have to do differently:
19ae87e5ce
Co-authored-by: Jade Lovelace <lix@jade.fyi>
See #13570 for details --- the idea is that included the store dir in
store paths makes systematic JSON parting with e.g. Serde, Aeson,
nlohmann, or similiar harder.
After talking to Eelco, we are changing the `Derivation` format right
away because not only is `nix derivation` technically experimental, we think it is
also less widely used in practice than, say, `nix path-info`.
Progress on #13570
Add `read-only` setting to `dummy://` store for back compat.
Test by changing an existing test to use this instead, fixing a TODO.
Co-Authored-By: HaeNoe <git@haenoe.party>
Co-authored-by: Eelco Dolstra <edolstra@gmail.com>
This allows the weird network or DNS server fallback mechanism inside
glibc to work, and prevents a "Resolving timed out after 5000
milliseconds" error. Read on for details.
The DNS request stuff (dns-hosts) in glibc uses this fallback procedure
to minimize network RTT in the ideal case while dealing with
ill-behaving networks and DNS servers gracefully (see resolv.conf(5)):
- Use sendmmsg() to send UDP DNS requests for IPv4 and IPv6 in parallel
- If that times out (meaning that none or only one of the responses have
been received), send the requests one by one, waiting for the response
before sending the next request ("single-request")
- If that still times out, try to use a different socket (hence
different address) for each request ("single-request-reopen")
The default timeout inside glibc is 5 seconds. Therefore, setting
connect-timeout, and therefore CURLOPT_CONNECTTIMEOUT to 5 seconds
prevents the single-request fallback, and setting it to even 10 seconds
prevents the single-request-reopen fallback as well.
The fallback decision is saved by glibc, but only thread-locally, and
libcurl starts a new thread for getaddrinfo() for each connection.
Therefore for every connection the fallback starts from sendmmsg() all
over again. And since these are considered to have timed out by libcurl,
even though getaddrinfo() might return a successful result, it is not
cached in libcurl.
While a user could tweak these with resolv.conf(5) options (e.g. using
networking.resolvconf.extraOptions in NixOS), and indeed that is
probably needed to avoid annoying delays, it still means that the
default connect-timeout of 5 is too low. Raise it to give fallback a
chance.
This has been dropped on unstable an nix no longer
compiled with overridden nixpkgs input. On 25.05 these
overrides already do nothing.
Tested with:
nix build .#packages.x86_64-darwin.nix-cli -L --override-input nixpkgs https://releases.nixos.org/nixos/unstable/nixos-25.11pre859555.ab0f3607a6c7/nixexprs.tar.xz
Default deployment target on 25.05 is 11.3, so 10.13
sdk override doesn't have to be updated at all as evident
from the fact that we didn't observe any issues with it.
This reverts commit bdbc739d6e.
Such a change needs more thought put into it. By versioning
shared libraries we'd make a false impression that libraries
themselves are actually versioned and have some sort of stable
ABI, which is not the case.
This will be useful when C bindings become stable, but as long
as they are experimental it does not make sense to set SONAME.
Also this change should not have been backported, since it's
severely breaking.
This implements a special back-compat shim to specifically allow
unbracketed IPv6 addresses in store references. This is something
that is relied upon in the wild and the old parsing logic accepted
both ways (brackets were optional). This patch restores this behavior.
As always, we didn't have any tests for this.
Addresses #13937.
`perf c2c` shows a lot of cacheline conflicts between purely read-only
Store methods (like `parseStorePath()`) and the Sync classes. So
allocate pathInfoCache separately to avoid that.
Calling `drainFD()` will hang if another process has the write side
open, since then the child won't get an EOF. This can happen if we
have multiple threads doing a build, since in that case another thread
may fork a child process that inherits the write side of the first
thread.
We could set O_CLOEXEC on the write side (using pipe2()) but it won't
help here since we don't always do an exec() in the child, e.g. in the
case of builtin builders. (We need a "close-on-fork", not a
"close-on-exec".)
This is relied upon (specifically the `local` store) by existing
tooling [1] and we broke this in 3e7879e6df (which
was first released in 2.31).
To lessen the scope of the breakage we should not normalize "auto" references
and explicitly specified references like "local" or "daemon". It also makes
sense to canonicalize local://,daemon:// to be more compatible with prior
behavior.
[1]: 05e1b3cba2/lib/NOM/Builds.hs (L60-L64)
Exactly why is is correct is a little subtle, because sometimes the
worker is owned by the worker. But the commit message in
e437b08250 explained the situation well
enough: I made that commit message part of the ABI docs, and now it
should be understandable to the next person.
