The process abstraction has changed to:
setup(task, process)
start(task, process_data)
prefork(task, process, mp)
The prefork() occurs in the main process right before fork.
The file src/nxt_main_process.c is completely free of process
specific logic.
The creation of a process now supports a PROCESS_CREATED state. The
The setup() function of each process can set its state to either
created or ready. If created, a MSG_PROCESS_CREATED is sent to main
process, where external setup can be done (required for rootfs under
container).
The core processes (discovery, controller and router) doesn't need
external setup, then they all proceeds to their start() function
straight away.
In the case of applications, the load of the module happens at the
process setup() time and The module's init() function has changed
to be the start() of the process.
The module API has changed to:
setup(task, process, conf)
start(task, data)
As a direct benefit of the PROCESS_CREATED message, the clone(2) of
processes using pid namespaces now doesn't need to create a pipe
to make the child block until parent setup uid/gid mappings nor it
needs to receive the child pid.
For each request, the worker calls the php_execute_script function
from libphp that changes to the script directory before doing its
work and then restores the process directory before returning. The
chdir(2) calls it performs are unnecessary in Unit design. In simple
benchmarks, profiling shows that the chdir syscall code path (syscall,
FS walk, etc.) is where the CPU spends most of its time.
PHP SAPI semantics requires the script to be run from the script
directory. In Unit's PHP implementation, we have two use cases:
- script
- arbitrary path
The "script" configuration doesn't have much need for a working
directory change: it can be changed once at module initialization.
The module needs to chdir again only if the user's PHP script also
calls chdir to switch to another directory during execution.
If "script" is not used in Unit configuration, we must ensure the
script is run from its directory (thus calling chdir before exec),
but there's no need to restore the working directory later.
Our implementation disables mandatory chdir calls with the SAPI
option SAPI_OPTION_NO_CHDIR, instead calling chdir only when needed.
To detect the user's calls to chdir, a simple "unit" extension is
added that hooks the built-in chdir() PHP call.
For backward compatibility, the Linux capabilities macros exposes v1 semantics
(32-bit) by default. We probe the version at runtime (because of pre-compiled
binaries) but the kernel syscall API is conservative and it doesn't return a
64-bit capability version if the input version is v1.
This patch suppress the kernel > 5.0 dmesg log below:
capability: warning: 'unitd' uses 32-bit capabilities (legacy support in use)
The setuid/setgid syscalls requires root capabilities but if the kernel
supports unprivileged user namespace then the child process has the full
set of capabilities in the new namespace, then we can allow setting "user"
and "group" in such cases (this is a common security use case).
Tests were added to ensure user gets meaningful error messages for
uid/gid mapping misconfigurations.
This is required to avoid include cycles, as some nxt_clone_* functions
depend on the credential structures, but nxt_process depends on clone
structures.
Now the nxt_user_groups_get() function uses getgrouplist(3) when available
(except MacOS, see below). For some platforms, getgrouplist() supports
a method of probing how much groups the user has but the behavior is not
consistent. The method used here consists of optimistically trying to get up
to min(256, NGROUPS_MAX) groups; only if ngroups returned exceeds the original
value, we do a second call. This method can block main's process if LDAP/NDIS+
is in use.
MacOS has getgrouplist(3) but it's buggy. It doesn't update ngroups if the
value passed is smaller than the number of groups the user has. Some
projects (like Go stdlib) call getgrouplist() in a loop, increasing ngroups
until it exceeds the number of groups user belongs to or fail when a limit
is reached. For performance reasons, this is to be avoided and MacOS is
handled in the fallback implementation.
The fallback implementation is the old Unit approach. It saves main's
user groups (getgroups(2)) and then calls initgroups(3) to load application's
groups in main, then does a second getgroups(2) to store the gids and restore
main's groups in the end. Because of initgroups(3)' call to setgroups(2),
this method requires root capabilities. In the case of OSX, which has
small NGROUPS_MAX by default (16), it's not possible to restore main's groups
if it's large; if so, this method fallbacks again: user_cred gids aren't
stored, and the worker process calls initgroups() itself and may block for
some time if LDAP/NDIS+ is in use.
- The mode of testdir was changed to allow reading from other users/groups.
- The java multipart test now uploads the file into an app writable dir.
- The build directory was made readable for other users.
- The python environment test now uses the HOME env var instead of PWD
because the latter is not set by the root shell (/bin/sh) by default.
- The node `node_modules` directory now is copied into the `testdir` instead
of using symlinks.
When using "credential: true", the new namespace starts with a completely
empty uid and gid ranges. Then, any setuid/setgid/setgroups calls using ids
not properly mapped with uidmap and gidmap fields return EINVAL, meaning
the id is not valid inside the new namespace.
Tiago Natel de Moura