| Age | Commit message (Collapse) | Author |
|---|
|
|
|
|
|
This is meant to facilitate a syscall for returning from interrupts, which
will actually work in the general case as opposed to the current hack, which
only works if the interrupt occured during a syscall (which is correct... for
now).
|
|
This is just for simplicity's sake.
I think I could even omit the `movw $TSS, (GdtTss + 2)` and have the linker
fill that out as a relocation, but that would probably be more complex overall.
|
|
|
|
Once again, needed for the interrupt return syscall.
Seems to have made the kernel slower in timebench?
|
|
I want to implement a new syscall for returning from "interrupts" (my
crappy take on Plan 9's notes / UNIX signals), and I've realized that
I can't use SYSRET for that, as it clobbers RCX and R11.
This hasn't been an issue so far, as I've only been switching into usermode
to return from syscalls, but now I might be switching into it at arbitrary
moments (right after an interrupt handler recovers from a page fault, etc).
I could make it so I switch between IRET and SYSRET depending on the context,
but I don't think it's worth it.
|
|
On the assembly side, ensure the stack frame looks always the same, by pushing
a fake "error code" for the interrupts that don't generate one.
On the C side, use a struct instead of magic indices into an "array", and
make it consistent with the current style.
|
|
|
|
I'm storing more information in less space. Win-win.
|
|
This will both let me save space in the allocation header, and make the
debugprint more readable.
|
|
For a while during development it managed to be faster than the old allocator
despite having more overhead, probably because it makes better use of the
cache. It no longer is - not sure why. The code definitely could be optimized
to hell, though, and it'd probably recover its original edge.
Non-power-of-2 sizeclasses would be really useful, as I could then e.g.
fit four processes into a page instead of three, but I think implementing them
"well" would be too complicated for now.
In hindsight arbitrary allocation descriptions were probably a bad idea too.
I've kept them in for now, but I'll switch to an enum of allocation types soon.
I think I can fit all the per-allocation state (size+type) in 2 bytes.
There are so few allocations in the kernel that I no longer care about the
backtraces, so I'm leaving them out.
The "large" allocations could probably be tested with execbuf in userland, so
I might actually implement them soon too.
|
|
Yay for guaranteed O(1) insertions, even if iostress seems a bit slower.
|
|
|
|
* firstfreepage now updates properly, preventing a crash (oops)
* kfree only wipes the length of the allocation, not the entire page - which
should make it easier to see the performance impact of the pagealloc changes
|
|
The postqueue functions remain as-is, as that's a more "specialized" interface.
They're mostly wrappers around queue.h, though.
|
|
|
|
I'm still not sure if I should use sys/queue.h for this.
But yeah, this is more consistent, and it will also let me switch over to O(1)
insertions later on.
|
|
Keeping its old name for now to make things easier for myself.
This might just be replaced by sys/queue.h soon.
|
|
|
|
The bitmap functions now accept page addresses so I don't have to handle
raw bitmap indexes, which was kinda complex.
kmalloc_sanity is now not visible to other code as it wasn't really that useful
in the first place.
|
|
This will likely be changed back, but for the time being it will let me
implement a better allocator without too much effort.
|
|
To use the same testing methodology as when I've introduced request slots:
before:
/ $ iostress 1 1000000 0 > /dev/vtty
run 0: 2585203
1000000 calls, 0 bytes. avg 2585203
after:
/ $ iostress 1 1000000 0 > /dev/vtty
run 0: 2783171
1000000 calls, 0 bytes. avg 2783171
This is around a 7.7% slowdown - that I hope to fix with a better malloc.
While this doesn't really make the code that much simpler, it doesn't feel
like the right approach in the first place
|
|
* I'm being more strict about the linked list state to hopefully ensure
I'm not leaking any references.
* vfsreq_create was renamed to vfsreq_dispatchcopy as that name feels more
clear. It copies its argument, and dispatches it.
* Requests for user backends are now handled more like requests for kernel
backends - there's an accept() function that accepts a request.
|
|
|
|
|
|
|
|
|
|
|
|
I probably should've tested DUP_WRONLY too, now that I think about it. TODO?
|
|
It was a dumb hack that wasn't even necessary - an error when mounting should
shadow over the mountpoint anyways.
|
|
|
|
|
|
|
|
what the fuck was I thinking when I originally implemented that?
|
|
|
|
|
|
time() will probably end up doing io. That sounded bad at first, but Plan 9
does that too (see /sys/src/libc/9sys/nsec.c), so it's probably fine.
I might need better service management soon. Also, dunno what it should return
before it makes contact with NTP.
I could implement RTC support, but eh. Doesn't feel that necessary.
I'll also need to remember how the hell threading works, so it can talk with
the ntp daemon on another thread.
|
|
|
|
|
|
the vm isn't getting an NMI for any real reason anyways, so I might as well
abuse it
|
|
|
|
|
|
|
|
|
|
|
|
pretty useless, but the list of all the TODOs is getting unwieldy, so I might
as well cross some crap off that list
|
|
I can now actually curl an entire page :^)
|
|
/usr/share/hosts because i don't have /etc/ yet and i don't feel like creating it.
|
|
|
