On macOS the system tar has builtin support for lzma while xz isn't
available as a separate binary. There's no builtin package manager
there available either so having to install lzma (without nix) would be
rather painful.
This should handle installation scenarios we can handle with
anything resembling confidence. Goal is approximating the existing
setup--not enforcing a best-practice...
Approaches (+ installer-handled, - manual) and configs each covers:
+ no change needed; /nix OK on boot volume:
All pre-Catalina (regardless of T2 or FileVault use)
+ create new unencrypted volume:
Catalina, pre-T2, no FileVault
+ create new encrypted-at-rest volume:
Catalina, pre-T2, FileVault
Catalina, T2, no FileVault
- require user to pre-create encrypted volume
Catalina, T2, FileVault
The default login shell for users on macOS 10.15 changed from bash to
zsh. So while generally nonstandard we need to configure it to make nix
function out of the box on macOS.
Starting macOS 10.15 /nix can't be creasted directly anymore due to the
readonly filesystem, but synthetic.conf was introduced to enable
creating mountpoints or symlinks for special usecases like package
managers.
The install-multi-user script uses blue, green, and red colors, as
well as bold and underline, to add helpful formatting that helps
structure its rather voluminous output.
Unfortunately, the terminal escape sequences it uses are not quite
well-formed. The relevant information is all there, just obscured
by some extra noise, a leading parameter `38`. Empirically, the
result is:
* On macOS, in both Terminal.app and iTerm2, the spurious `38` is
ignored, the rest of the escape sequence is applied, and the colors
show up as intended.
* On Linux, in at least gnome-terminal and xterm, the spurious `38`
and the next parameter after it are ignored, and what's left is
applied. So in the sequence `38;4;32`, the 4 (underline) is
ignored but the 32 (green) takes effect; in a more typical sequence
like `38;34`, the 34 (blue) is ignored and nothing happens.
These codes are all unchanged since this script's origins as a
Darwin-only script -- so the fact that they work fine in common macOS
terminals goes some way to explain how the bug arose.
Happily, we can make the colors work as intended by just deleting the
extra `38;`. Tested in all four terminals mentioned above; the new
codes work correctly on all of them, and on the two macOS terminals
they work exactly the same as before.
---
In a bit more technical detail -- perhaps more than anyone, me
included, ever wanted to know, but now that I've gone and learned it
I'll write it down anyway :) -- here's what's happening in these codes:
An ECMA-48 "control sequence" begins with `\033[` aka "CSI", contains
any number of parameters as semicolon-separated decimal numbers (plus
sometimes other wrinkles), and ends with a byte from 0x40..0x7e. In
our case, with `m` aka "SGR", "Select Graphic Rendition".
An SGR control sequence `\033[...m` sets colors, fonts, text styles,
etc. In particular a parameter `31` means red, `32` green, `34` blue,
`4` underline, and `0` means reset to normal. Those are all we use.
There is also a `38`. This is used for setting colors too... but it
needs arguments. `38;5;nn` is color nn from a 256-color palette, and
`38;2;rr;gg;bb` has the given RGB values.
There is no meaning defined for `38;1` or `38;34` etc. On seeing a
parameter `38` followed by an unrecognized argument for it, apparently
some implementations (as seen on macOS) discard only the `38` and
others (as seen on Linux) discard the argument too before resuming.
After installing Nix, I found that all the files and directories
initially copied into the store were writable, with mode 644 or 755:
drwxr-xr-x 9 root root 4096 Dec 31 1969 /nix/store/ddmmzn4ggz1f66lwxjy64n89864yj9w9-nix-2.3.3
The reason is that that's how they were in the unpacked tarball, and
the install-multi-user script used `rsync -p` without doing anything
else to affect the permissions.
The plain `install` script for a single-user install takes care to
do a `chmod -R a-w` on each store path copied. We could do the same
here with one more command; or we can pass `--chmod` to rsync, to
have it write the files with the desired modes in the first place.
Tested the new `rsync` command on both a Linux machine with a
reasonably-modern rsync (3.1.3) and a Mac with its default, ancient,
rsync 2.6.9, and it works as expected on both. Thankfully the latter
is just new enough to have `--chmod`, which dates to rsync 2.6.7.
On a systemd-based Linux distribution: If the user has previously had multi-user Nix installed on the system, removed it and then reinstalled multi-user Nix again the old nix-daemon.service will still be running when `scripts/install-systemd-multi-user.sh` tries to start it which results in nothing being done and the old daemon continuing its run.
When a normal user then tries to use Nix through the daemon the nix binary will fail to connect to the nix-daemon as it does not belong to the currently installed Nix system. See below for steps to reproduce the issue that motivated this change.
$ sh <(curl https://nixos.org/nix/install) --daemon
$ sudo rm -rf /etc/nix /nix /root/.nix-profile /root/.nix-defexpr /root/.nix-channels /home/nix-installer/.nix-profile /home/nix-installer/.nix-defexpr /home/nix-installer/.nix-channels ~/.nix-channels ~/.nix-defexpr/ ~/.nix-profile /etc/profile.d/nix.sh.backup-before-nix /etc/profile.d/nix.sh; sed -i '/added by Nix installer$/d' ~/.bash_profile
$ unset NIX_REMOTE
$ sh <(curl https://nixos.org/nix/install) --daemon
└$ export NIX_REMOTE=daemon
└$ nix-env -iA nixpkgs.hello
installing 'hello-2.10'
error: cannot connect to daemon at '/nix/var/nix/daemon-socket/socket': No such file or directory
(use '--show-trace' to show detailed location information)
└$ sudo systemctl restart nix-daemon.service
└$ nix-env -iA nixpkgs.hello
installing 'hello-2.10'
these paths will be fetched (6.09 MiB download, 27.04 MiB unpacked):
/nix/store/2g75chlbpxlrqn15zlby2dfh8hr9qwbk-hello-2.10
/nix/store/aag9d1y4wcddzzrpfmfp9lcmc7skd7jk-glibc-2.27
copying path '/nix/store/aag9d1y4wcddzzrpfmfp9lcmc7skd7jk-glibc-2.27' from 'https://cache.nixos.org'...
copying path '/nix/store/2g75chlbpxlrqn15zlby2dfh8hr9qwbk-hello-2.10' from 'https://cache.nixos.org'...
building '/nix/store/w9adagg6vlikr799nkkqc9la5hbbpgmi-user-environment.drv'...
created 2 symlinks in user environment
In the multi-user install script, we originally made sure no previous
references to Nix existed. This prevented any previous installs from
contaminating the new install. However, some users need the ability to
repair their existing Nix installation without uninstalling all
references to Nix. This change allows users with existing Nix
installations to use the installer, while still outputing a warning
message on the dangers of this. As a result, the multi-user install
script work much more like the single-user install script has worked
in the past.
This is a requirement for macOS Catalina users now that
/Library/LaunchDaemons/org.nixos.nix-daemon.plisg is not managed by
the Nix store. If there is ever a change to the .plist, all users will
need to rerun this install script to get the new changes. Otherwise,
changes to the launch daemon will require manual interventions.
On Catalina, the /nix filesystem might not be mounted at start time.
To avoid this service not starting, we need to keep the launch agent
outside of the Nix store. A wait4pid will hold for our /nix dir to be
mounted.
Fixes#3125.
If a network proxy configuration is detected, setup an override
systemd unit file for nix-daemon service with the non-empty
proxy variables.
Proxy detection is performed by looking for http/https/ftp proxy and no
proxy variables in user environment
With macOS catalina, we can no longer modify the root system
volume (#2925). macOS provides a system configuration file in
synthetic.conf(5) to create empty root directories. This can be used
to mount /nix to a separate volume. As a result, this directory will
need to already exist prior to installation. Instead, check for
/nix/store and /nix/var for a live Nix installation.
When we are in single user mode, we still want to have access to
profiles. This way things in Nixpkgs that rely on them getting set
accurately are done in both cases. The point where I hit this is with
using aspell which looks in NIX_PROFILES:
https://github.com/NixOS/nixpkgs/blob/master/pkgs/development/libraries/aspell/default.nix
Before this patch, NIX_PROFILES was never set in single user mode!
This corrects that.
Having max-jobs = 32 ($NIX_USER_COUNT is hardcoded to that value) may
severely overload the machine. The nix.conf(5) manual page says max-jobs
defaults to 1, so let's use that value.
NOTE: Both max-jobs and cores are now being set to their default value,
so they can be removed alltogether.
Sandboxing is now enabled by default on Linux, but is still disabled on
macOS. However, the installer always turned it off to ensure consistent
behavior.
Remove this default configuration, so we fall back to the default
platform-specific value.
Signed-off-by: Austin Seipp <aseipp@pobox.com>
Instead, print a message about what happened and tell the user what can
be done (run "[sudo -i] nix-channel --update nixpkgs" again at a later
time). This change allows installing Nix when you're offline.
Since the multi-user installer is so verbose, the message isn't printed
until the end.
Fixes issue #2650 ("installation without internet connection").
This removes part of the PATH that were being added automatically in multi-user installs:
- $HOME/.nix-profile/lib/kde4/libexec - shouldn't be needed anymore, we are now using kde5
- @localstatedir@/nix/profiles/default/lib/kde4/libexec - same as above
- @localstatedir@/nix/profiles/default - shouldn't ever contain binaries
If the profile is sourced inside a script with `set -u`, the check for
__ETC_PROFILE_NIX_SOURCED and NIX_SSL_CERT_FILE would raise an error.
A simple guard around this check allows the script to operate under
standard environments (where it is fairly reasonable to assume USER
and HOME are set.)
By passing --daemon or --no-daemon, the installer can be forced to
select one or the other installation options, despite what the
automatic detection can provide.
This commit can be backported to 2.0-maintenance because it explicitly
turns off the daemon installation for Linux under systemd.
We use grep instead of an equality check because it is difficult
to extract _just_ the user's note, instead it is prefixed with
some plist junk. This was causing the user note to always be set,
even if there was no reason for it.
Grep would ignore files that didn't exist, but would complain
about files in a directory if the directory didn't exist. Simply check
for the directory first, prior to grepping it.
- darwin installer: delete hardware report, not necessary
- moves os-specific code from the darwin installer to to `poly_*`
functions
- adds profile.d support to the profile targets, which automatically
handles many distros which don't have a /etc/bashrc but do have an
/etc/profile.d
- /bin/bash -> /usr/bin/env bash
- document why each excluded shellcheck check is excluded
- rename the multi-user to Daemon-based
Checking for MANPATH without quotes always returns true, so that it breaks bash-completion for man pages on modern systems without MANPATH environment variable.
The installer will error out if a user's shell configuration includes any mention of ~nix-profile~, even if this is in a comment. This change is designed to do the bare minimum to ignore lines beginning with a `#`.
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
The current behaviour modifies the first writeable file from amongst
.bash_profile, .bash_login and .profile. So .bash_profile (if it is
writable) would be modified even if a user has already sourced nix.sh
in, say, .profile.
This commit introduces a new environment variable,
NIX_INSTALLER_NO_MODIFY_PROFILE. If this is set during installation,
then the modifications are unconditionally skipped.
This is useful for users who have a manually curated set of dotfiles
that they are porting to a new machine. In such scenarios, nix.sh is
already sourced at a place where the user prefers. Without this
change, the nix installer would insist on modifying .bash_profile if
it exists.
This commit also add documentations for both the current behaviour and
the new override.
Currently, man has issues finding man pages for Nix-installed
application (also, `nix-env --help` doesn't work). The issue is caused
by custom `$MANPATH` set by my system. That makes man use it instead of
searching in default location.
Either of next lines workaround the issue:
```sh
unset MANPATH
export MANPATH=$HOME/.nix-profile/share/man:$MANPATH
```
This patch adds the later line to the `nix-profile.sh` if user has
`MANPATH` set. (Not clearing `MANPATH` as that would be disrespect of
user's preferences.)
As a side-effect, host's man might find man pages installed by Nix.
This is primarily to subsume the functionality of the
copy-from-other-stores substituter. For example, in the NixOS
installer, we can now do (assuming we're in the target chroot, and the
Nix store of the installation CD is bind-mounted on /tmp/nix):
$ nix-build ... --option substituters 'local?state=/tmp/nix/var&real=/tmp/nix/store'
However, unlike copy-from-other-stores, this also allows write access
to such a store. One application might be fetching substitutes for
/nix/store in a situation where the user doesn't have sufficient
privileges to create /nix, e.g.:
$ NIX_REMOTE="local?state=/home/alice/nix/var&real=/home/alice/nix/store" nix-build ...
Nix sometimes outputs a warning message like this:
```
directory /nix does not exist; creating it by running ‘?? using sudo
```
... when it really meant to output something that looked like this:
```
directory /nix does not exist; creating it by running 'mkdir -m 0755 /nix && chown gabriel /nix' using sudo
```
The reason why is due to some bizarre behavior in Bash where it will translate anything of the form `$x’` to `??`, leading to the incorrect warning message. I don't know what is the origin of this Bash behavior, but the easiest fix is to just use ASCII quotes instead of unicode quotes.
Just wasted a couple hours chasing shadows because the nix store got
corrupted and there was no indication of that anywhere.
Since an install is one-time only, might as well verify. Optimization
showed that the copied files aren't read-only; fixed that as well.
Also, use /bin/sh since there's a good chance that this script will be
run on systems without /bin/bash
This is currently only used by the Hydra queue runner rework, but like
eff5021eaa it presumably will be useful
for the C++ rewrite of nix-push and
download-from-binary-cache. (@shlevy)
The `set -e` at the top of the script causes the installation to fail to
complete if the shell profile is not writeable. Checking file existence
only is not enough.
The --insecure flag to curl tells curl not to bother checking if the TLS
certificate presented by the server actually matches the hostname
requested, and actually is issued by a trusted CA chain. This almost
entirely negates any benefit from using TLS in the first place.
This removes the --insecure flag to ensure we actually have a secure
connection to the intended hostname before downloading binaries.
Manually tested locally within a dev-shell; was able to download
binaries from https://cache.nixos.org without issue.
[Note: --insecure was only used for fetching NARs, whose integrity is
verified by Nix anyway using the hash from the .narinfo. But if we can
fetch the .narinfo without --insecure, we can also fetch the .nar, so
there is not much point to using --insecure. --Eelco]
Some benchmarking suggested this as a good value. Running
$ benchmark -f ... -t 25 -- sh -c 'rm -f /nix/var/nix/binary-cache*; nix-store -r /nix/store/x5z8a2yvz8h6ccmhwrwrp9igg03575jg-nixos-15.09.git.5fd87e1M.drv --dry-run --option binary-caches-parallel-connections <N>'
gave the following mean elapsed times for these values of N:
N=10: 3.3541
N=20: 2.9320
N=25: 2.6690
N=30: 2.9417
N=50: 3.2021
N=100: 3.5718
N=150: 4.2079
Memory usage is also reduced (N=150 used 186 MB, N=25 only 68 MB).
Closes#708.
This makes that option even more insecure, by also not checking the SSL host.
But without this parameter, one can still get SSL errors even when
"verify-https-binary-caches" is false, which is unexpected IMO.
This reverts commit 76f985b92d. We
shouldn't mess with $MANPATH, because on some "man" implementations
(like NixOS'), the default value on $MANPATH is derived from $PATH. So
if you set $MANPATH, you lose the default locations.
This is not strictly needed for integrity (since we already include
the NAR hash in the fingerprint) but it helps against endless data
attacks [1]. (However, this will also require
download-from-binary-cache.pl to bail out if it receives more than the
specified number of bytes.)
[1] https://isis.poly.edu/~jcappos/papers/cappos_mirror_ccs_08.pdf
In some cases the bash builtin command "cd" can print the variable $CWD
to stdout. This caused the install script to fail while copying files
because the source path was wrong.
Fixes#476.
We only need to sign the store path, NAR hash and references (the
"fingerprint"). Everything else is irrelevant to security. For
instance, the compression algorithm or the hash of the compressed NAR
don't matter as long as the contents of the uncompressed NAR are
correct.
(Maybe we should include derivers in the fingerprint, but they're
broken and nobody cares about them. Also, it might be nice in the
future if .narinfos contained signatures from multiple independent
signers. But that's impossible if the deriver is included in the
fingerprint, since everybody will tend to have a different deriver for
the same store path.)
Also renamed the "Signature" field to "Sig" since the format changed
in an incompatible way.
