Note that this is very preliminary, and I merely got my already freezing
kernel branch not to triple fault, but I would probably apply this patch
to upstream.
What is changed here, is that rather than relying on recursive mapping
for accessing page table frames, it now uses linear translation
(virt=phys+KERNEL_OFFSET). The only problem is that the paging code now
makes assumptions that the entire physical address space remains mapped,
which is not necessarily the case on x86_64 architecturally, even though
systems with RAM more than a PML4 are very rare. We'd probably lazily
(but linearly) map physical address space using huge pages.
Previously context::switch used compare_and_swap for acquiring the
global context switch lock, but given its deprecation in more recent
Rust versions, it has been replaced with compare_exchange_weak (which
can be further optimized on some architectures).
It also replaces panic!() with abort() in switch_finish_hook, because
unwinding from assembly is not that fun.
The reason for these types of rewrites, is that more recent Rust
compilers have started to deprecate naked functions that consist of more
than only a single asm block, as they can trigger all sorts of UB.
Previously there was a triple fault, due to a combination of reasons
(e.g. rsp and rbp being ordered in the struct and in the assembly).
Now, the locks will be held __all the way until the new context__ has
been switched to, which completely eliminates any possibility that the
"pcid fault" originates here.
While I am unsure whether this will work, this could also be an
opportunity to be able to remove CONTEXT_SWITCH_LOCK fully.
This is due to a warning in more recent compilers, which forbid anything
but a single inline assembly block, in naked functions. It does
unfortunately triple fault right now, but I hope I may be able to fix it
soon.
So, when I first introduced io_uring, it was not compiled with the
`multi_core` kernel feature, mainly to make development easier (I
thought). However, since io_uring allows multiple simultaneous system
calls, we cannot longer make the in-kernel contexts block, for example
when receiving a message from a pipe, if there can be multiple such
requests simultaneously.
This has required me to change WaitCondition into allowing multiple
simultaneous tasks; although, it introduces a potential race condition:
since a future can only return Pending and not block directly before
releasing the lock (condvar logic), we need some way to make sure that
nothing happens after the context finds out that it has to wait, and the
actual waiting. If a message is pushed in between, and the waker is
called (Context::unblock), just before it was going to block itself,
then we miss the message, and potentially cause a deadlock.
Fortunately, in order to block and unblock contexts, we need to
exclusively lock the context. So, what we can do to ensure that waking
while running is no longer a no-op, is to introduce a "wake flag", which
is set only if the context is currently running, and Runnable.
But, this still caused all weird kinds of hard-to-debug problems, with
arbitrary CPU exceptions and possibly memory corruption. The reason for
this, is that the context switching logic uses really unsafe operations,
which is why context switching (at the moment) requires an exclusive
lock. Before this commit, it would modify the `running` field after the
lock had been released, which obviously can cause a data race, when the
regular context waker code that is run within a system call, locks the
context but not the global switching lock.
The solution was to make sure that the locks were held, all the way
until the actual switching, which was done in assembly. There can still
be a race condition here, since it modifies memory containing registers
after the lock has been released, even if it may be behind &mut on
another context, which can be UB, but it has not contributed to any
actual bugs... yet.
* I have not yet done that rigorous testing, but it appears to work well
enough, and I have not encountered the bug after like 10 tries.
When mapping one (from) virtual address range to another (to) virtual
address range, be mindful of which mapper type to use for each range.
Before this, the same mapper type was used for both ranges. This meant
that if from and to were different (as in not both kernel virtual
addresses or user virtual addresses) then it would appear that either
from or to was not mapped previously and the kernel would panic.
Jeremy Soller