This patch ensures that the shutdown procedure can complete due to the
fact we don't kill kernel processes anymore, and only stop the scheduler
from running after the filesystems unmount procedure.
We also need kernel processes during the shutdown procedure, because we
rely on the WorkQueue threads to run WorkQueue items to complete async
IO requests initiated by filesystem sync & unmounting, etc.
This is also simplifying the code around the killing processes, because
we don't need to worry about edge cases such as the FinalizerTask
anymore.
The process could be long gone by the point the async IO request has
completed so hold a weak reference pointer to the requesting Process and
try get a strong reference only when needed.
This patch is necessary because otherwise async IO requests can hold
Process objects long after they were terminated, which would make it
impossible to perform certain tasks in the system, like killing all user
processes during the shutdown procedure.
We first must flush the superblock through the BlockBasedFileSystem
methods properly and only then clear the DiskCache pointer, to prevent a
possible kernel panic due to nullptr dereference.
Previously we would set the KeyCode correctly to the appropriate
extended keys values, like Home and End, but keep the code point of the
original keys, like 1, 2, 3, etc. Because of this, the keys would just
print the original keys, instead of behaving like the extended ones.
This is a minimal set of changes to allow `serenity.sh build riscv64` to
successfully generate the build environment and start building. This
includes some, but not all, assembly stubs that will be needed later on;
they are currently empty.
Shadow doorbell feature was added in the NVMe spec to improve
the performance of virtual devices.
Typically, ringing a doorbell involves writing to an MMIO register in
QEMU, which can be expensive as there will be a trap for the VM.
Shadow doorbell mechanism was added for the VM to communicate with the
OS when it needs to do an MMIO write, thereby avoiding it when it is
not necessary.
There is no performance improvement with this support in Serenity
at the moment because of the block layer constraint of not batching
multiple IOs. Once the command batching support is added to the block
layer, shadow doorbell support can improve performance by avoiding many
MMIO writes.
Default to old MMIO mechanism if shadow doorbell is not supported.
Introduce a new Struct Doorbell that encapsulates the mmio doorbell
register.
This commit does not introduce any functional changes and it is added
in preparation to adding shadow doorbell support.
This was the root cause of zombie processes showing up randomly and
disappearing after some disk activity, such as running shell commands -
The NVMeIO AsyncBlockDeviceRequest member simply held a pointer to a
Process object, therefore it could keep it alive a for a long time after
it ceased to actually function at all.
While LLD and mold support RELR "packed" relocations on all
architectures, the BFD linker currently only implements them on x86-64
and POWER.
This fixes two issues:
- The Kernel had it enabled even for AArch64 + GCC, which led to the
following being printed: `warning: -z pack-relative-relocs ignored`.
- The userland always had it disabled, even in the supported AArch64 +
Clang/mold scenarios.
Two non-functional changes:
- Remove pointless `-latomic` flag. It was specified via
`add_compile_options`, which only affects compilation and not linking,
so the library was never actually linked into the kernel. In fact, we
do not even build `libatomic` for our toolchain.
- Do not disable `-Wnonnull`. The warning-causing code was fixed at some
point.
This commit also removes `-mstrict-align` from the userland. Our target
AArch64 hardware natively supports unaligned accesses without a
significant performance penalty. Allowing the compiler to insert
unaligned accesses into aligned-as-written code allows for some
performance optimizations in fact. We keep this option turned on in the
kernel to preserve correctness for MMIO, as that might be sensitive to
alignment.
Add the device ID for PCI serial port cards that use the WCH CH351
chip. This device has been tested with real hardware where the serial
debug output could succesfully be received.
Now that support for 32-bit x86 has been removed, we don't have to worry
about the top half of `off_t`/`u64` values being chopped off when we try
to pass them in registers. Therefore, we no longer need the workaround
of pointers to stack-allocated values to syscalls.
Note that this changes the system call ABI, so statically linked
programs will have to be re-linked.
Using the kernel stack is preferable, especially when the examined
strings should be limited to a reasonable length.
This is a small improvement, because if we don't actually move these
strings then we don't need to own heap allocations for them during the
syscall handler function scope.
In addition to that, some kernel strings are known to be limited, like
the hostname string, for these strings we also can use FixedStringBuffer
to store and copy to and from these buffers, without using any heap
allocations at all.
Instead, use the FixedCharBuffer class to ensure we always use a static
buffer storage for these names. This ensures that if a Process or a
Thread were created, there's a guarantee that setting a new name will
never fail, as only copying of strings should be done to that static
storage.
The limits which are set are 32 characters for processes' names and 64
characters for thread names - this is because threads' names could be
more verbose than processes' names.
This class encapsulates a fixed Array with compile-time size definition
for storing ASCII characters.
There are also new Kernel StdLib functions to copy user data into such
objects so this class will be useful later on.
Previously we could get a raw pointer to a Mount object which might be
invalid when actually dereferencing it.
To ensure this could not happen, we should just use a callback that will
be used immediately after finding the appropriate Mount entry, while
holding the mount table lock.
We don't really need this method anymore, because we could just try to
find the mount entry based on the given mount point host custody.
This also allows us to remove the is_vfs_root and root_inode_id methods
from the VirtualFileSystem class.
We could easily encounter a case where we do the following:
```
mkdir -p /tmp2
mount /dev/hda /tmp2
```
would produce a bug that doing `ls /tmp2/tmp2` will give the contents
on `/dev/hda` ext2 root directory and also on `/tmp2/tmp2/tmp2` and so
on.
To prevent this, we must compare the current custody against each mount
entry's custody to ensure their paths match.
This is not useful, as we have literally zero knowledge about where this
inode is actually located at with respect to the entire global path tree
so we could easily encounter a case where we do the following:
```
mkdir -p /tmp2
mount /dev/hda /tmp2
```
and when traversing the /tmp2 directory entries, we will see the root
inode of /dev/hda on "/tmp2/tmp2", even if it was not mounted.
Therefore, we should just plainly give the raw directory entries as they
are written "on the disk". Anything else that needs to exactly know if
there's an underlying mounted filesystem, can just use the stat syscall
instead.
This ensures that the host mount point custody path is not the same like
the new to-be-mounted custody.
A scenario that could happen before adding this check is:
```
mkdir -p /tmp2
mount /dev/hda /tmp2/
mount /dev/hda /tmp2/
mount /dev/hda /tmp2/ # this will fail here
```
and after adding this check, the following scenario is now this:
```
mkdir -p /tmp2
mount /dev/hda /tmp2/
mount /dev/hda /tmp2/ # this will fail here
mount /dev/hda /tmp2/ # this will fail here too
```
Currently, ephemeral port allocation is handled by the
allocate_local_port_if_needed() and protocol_allocate_local_port()
methods. Actually binding the socket to an address (which means
inserting the socket/address pair into a global map) is performed either
in protocol_allocate_local_port() (for ephemeral ports) or in
protocol_listen() (for non-ephemeral ports); the latter will fail with
EADDRINUSE if the address is already used by an existing pair present in
the map.
There used to be a bug where for listen() without an explicit bind(),
the port allocation would conflict with itself: first an ephemeral port
would get allocated and inserted into the map, and then
protocol_listen() would check again for the port being free, find the
just-created map entry, and error out. This was fixed in commit
01e5af487f by passing an additional flag
did_allocate_port into protocol_listen() which specifies whether the
port was just allocated, and skipping the check in protocol_listen() if
the flag is set.
However, this only helps if the socket is bound to an ephemeral port
inside of this very listen() call. But calling bind(sin_port = 0) from
userspace should succeed and bind to an allocated ephemeral port, in the
same was as using an unbound socket for connect() does. The port number
can then be retrieved from userspace by calling getsockname (), and it
should be possible to either connect() or listen() on this socket,
keeping the allocated port number. Also, calling bind() when already
bound (either explicitly or implicitly) should always result in EINVAL.
To untangle this, introduce an explicit m_bound state in IPv4Socket,
just like LocalSocket has already. Once a socket is bound, further
attempt to bind it fail. Some operations cause the socket to implicitly
get bound to an (ephemeral) address; this is implemented by the new
ensure_bound() method. The protocol_allocate_local_port() method is
gone; it is now up to a protocol to assign a port to the socket inside
protocol_bind() if it finds that the socket has local_port() == 0.
protocol_bind() is now called in more cases, such as inside listen() if
the socket wasn't bound before that.
Since this is the block size that file system drivers *should* set,
let's name it the logical block size, just like most file systems such
as ext2 already do anyways.
This never was a logical block size, it always was a device specific
block size. Ideally the block size would change in accordance to
whatever the driver wants to use, but that is a change for the future.
For now, let's get rid of this confusing naming.
This also makes it easier to understand and reference where these
(sometimes rather arbitrary) calculations come from.
This also fixes a bug where group_index_from_block_index assumed 1KiB
blocks.
For a long time, our shutdown procedure has basically been:
- Acquire big process lock.
- Switch framebuffer to Kernel debug console.
- Sync and lock all file systems so that disk caches are flushed and
files are in a good state.
- Use firmware and architecture-specific functionality to perform
hardware shutdown.
This naive and simple shutdown procedure has multiple issues:
- No processes are terminated properly, meaning they cannot perform more
complex cleanup work. If they were in the middle of I/O, for instance,
only the data that already reached the Kernel is written to disk, and
data corruption due to unfinished writes can therefore still occur.
- No file systems are unmounted, meaning that any important unmount work
will never happen. This is important for e.g. Ext2, which has
facilites for detecting improper unmounts (see superblock's s_state
variable) and therefore requires a proper unmount to be performed.
This was also the starting point for this PR, since I wanted to
introduce basic Ext2 file system checking and unmounting.
- No hardware is properly shut down beyond what the system firmware does
on its own.
- Shutdown is performed within the write() call that asked the Kernel to
change its power state. If the shutdown procedure takes longer (i.e.
when it's done properly), this blocks the process causing the shutdown
and prevents any potentially-useful interactions between Kernel and
userland during shutdown.
In essence, current shutdown is a glorified system crash with minimal
file system cleanliness guarantees.
Therefore, this commit is the first step in improving our shutdown
procedure. The new shutdown flow is now as follows:
- From the write() call to the power state SysFS node, a new task is
started, the Power State Switch Task. Its only purpose is to change
the operating system's power state. This task takes over shutdown and
reboot duties, although reboot is not modified in this commit.
- The Power State Switch Task assumes that userland has performed all
shutdown duties it can perform on its own. In particular, it assumes
that all kinds of clean process shutdown have been done, and remaining
processes can be hard-killed without consequence. This is an important
separation of concerns: While this commit does not modify userland, in
the future SystemServer will be responsible for performing proper
shutdown of user processes, including timeouts for stubborn processes
etc.
- As mentioned above, the task hard-kills remaining user processes.
- The task hard-kills all Kernel processes except itself and the
Finalizer Task. Since Kernel processes can delay their own shutdown
indefinitely if they want to, they have plenty opportunity to perform
proper shutdown if necessary. This may become a problem with
non-cooperative Kernel tasks, but as seen two commits earlier, for now
all tasks will cooperate within a few seconds.
- The task waits for the Finalizer Task to clean up all processes.
- The task hard-kills and finalizes the Finalizer Task itself, meaning
that it now is the only remaining process in the system.
- The task syncs and locks all file systems, and then unmounts them. Due
to an unknown refcount bug we currently cannot unmount the root file
system; therefore the task is able to abort the clean unmount if
necessary.
- The task performs platform-dependent hardware shutdown as before.
This commit has multiple remaining issues (or exposed existing ones)
which will need to be addressed in the future but are out of scope for
now:
- Unmounting the root filesystem is impossible due to remaining
references to the inodes /home and /home/anon. I investigated this
very heavily and could not find whoever is holding the last two
references.
- Userland cannot perform proper cleanup, since the Kernel's power state
variable is accessed directly by tools instead of a proper userland
shutdown procedure directed by SystemServer.
The recently introduced Firmware/PowerState procedures are removed
again, since all of the architecture-independent code can live in the
power state switch task. The architecture-specific code is kept,
however.
Once we move to a more proper shutdown procedure, processes other than
the finalizer task must be able to perform cleanup and finalization
duties, not only because the finalizer task itself needs to be cleaned
up by someone. This global variable, mirroring the early boot flags,
allows a future shutdown process to perform cleanup on its own.
Note that while this *could* be considered a weakening in security, the
attack surface is minimal and the results are not dramatic. To exploit
this, an attacker would have to gain a Kernel write primitive to this
global variable (bypassing KASLR among other things) and then gain some
way of calling the relevant functions, all of this only to destroy some
other running process. The same effect can be achieved with LPE which
can often be gained with significantly simpler userspace exploits (e.g.
of setuid binaries).
Since we never check a kernel process's state like a userland process,
it's possible for a kernel process to ignore the fact that someone is
trying to kill it, and continue running. This is not desireable if we
want to properly shutdown all processes, including Kernel ones.
This is correct since unmount doesn't treat bind mounts specially. If we
don't do this, unmounting bind mounts will call
prepare_for_last_unmount() on the guest FS much too early, which will
most likely fail due to a busy file system.
Previously, we started parsing the ELF file again in a completely
different place, and without the partial mapping that we do while
validating.
Instead of doing manual parsing in two places, just capture the
requested stack size right after we validated it.