This can be nice if user has memory for a the temporary buffer and
doesnt want the sorting to allocate or be able to fail.
Also counts are now stack allocated, there isn't really any reason to
allocate them on the heap as 256x 64 bit values only adds up to 2 KiB
This moves locking to the inodes themselves which allows reducing lock
times significantly. Main inodes (ext2 and tmpfs) still do contain a
single big mutex that gets locked during operations but now we have the
architecture to optimize these.
I was modifying the `vaddr` argument while invalidating which lead to
the smp message containing wrong virtual address. This showed up in
really weird bugs from invalid TLB
Initial step of paging now just prepares fast page for heap, actual page
table initialization happens after heap is initialized which allows
x86_64 to never depend on kmalloc for pages.
Processor's stacks are now also spawned with PMM/VMM allocated stacks
instead of kmalloc identity mapped.
ca-certificates:
- update to 2026.03.19
- install to /etc/cacert
- extract individual ceritificates from the bundle
openssl:
- depend on ca-certificates
- install hashed symlinks to individual certs
curl:
- don't depend on ca-certificates; openssl handles this
- set both ca-bundle and ca-path
There is no need for them to be uncached. Having them as uncached killed
the networking performance, over 90% time was spent in kernel out of
which 80% was in checksum calculation and memcpy, half each (measured in
qemu with e1000e)
Add a concept for HashMapFindable instead of manually specifying the
requries expression everywhere.
Allow HashMapFindable also for `remove`, `contains`, `operator[]`, `at`
Make Entry have a const Key. This allows iterator's operator* and
operator-> return values have const keys.
Add a concept for HashSetFindable instead of manually specifying the
requries expression everywhere.
Allow HashSetFindable also for `remove` and `contains`.
Remove non-const return values from iterator; you should never modify
the hashed value in place.
Don't require key to be move assignable, move construction is enough.
There is really no need to have two implementation of the same thing.
Only difference now is that HashMap's value type has to be movable but
this wasn't an issue
Instead of representing the map as vector or linked lists which required
an allocation for every insertion and deallocation for removal, we now
store a single big contiguous block of memory and use hash chains to
handle collisions. This intuitively feels much better although I did not
run any benchmarks.
I don't know why I though the block chain had to be stored fully in the
ThreadBlocker, that did not even fix the problem I was trying to fix
when I last rewrote it. Roll back to doubly linked list of block chain
and now just check that the node is contained within the ThreadBlocker
before removing and after acquiring the ThreadBlocker's lock. Also there
is no need to have a separate lock the node's blocker field. We can just
perform an atomic reads and writes to it. We can still get a blocker
that the node is no longer part of, but this can be resolved with a
simple check. This patch reduces ThreadBlocker's size from over 200
bytes to just 12 bytes +4 bytes padding
This generates much nicer assembly as it does not have to read thread
pointer for every access to TCB (errno, cancel_state, cancelled) and
instead it can read it once and use the same value for all accesses
Instead of calculating bit-by-bit crc32, we now calculate a lookup table
during compile time. The old crc32 calculation was taking almost 50% of
the decompression time.
Also handle multiple symbols at once without outputting to user. It is
much more efficient to output many bytes instead of the up to 258 that a
single symbol can decode to :^)
Also add custom load addresses for x86_64 target. This allows qemu to
load the kernel with -kernel argument. Without these addresses qemu
would refuse to load as it only supports 32 bit ELFs, but as our kernel
starts in 32 bit mode anyway, we can just load it!
We no longer require the user to pass full compressed data in one go,
instead the decompressor reports to the user if it needs more input or
output space.
This was just RefPtr<OpenFileDescription> and descriptor flags.
Descriptor flags only define O_CLOEXEC, so we can just store fd's
cloexec status in a bitmap rather than separate fields. This cuts down
the size of OpenFileDescriptorSet to basically half!
Instead of immediately doing rerender of client data and syncing 60 Hz,
we now only keep track of the damaged regions and also do the rerender
step 60 Hz.
Instead of sending while serializing (what even was that), we serialize
the whole packet into a buffer which can be sent in one go. First of all
this reduces the number of sends by a lot. This also fixes WindowServer
ending up sending partial packets when client is not responsive.
Previously we would just try sending once, if any send failed the send
was aborted while partial packet was already transmitted. This lead to
packet stream being out of sync leading to the client killing itself.
Now we allow 64 KiB outgoing buffer per client. If this buffer ever fills
up, we will not send partial packets.
Kernel syscall API no longer zeros all unused argument registers and
libc now uses inlined syscall macro internally. This significantly
cleans up generated code for basic syscall wrapper functions.
Instead of keeping track of the current time and rescheduling when
interval has passed, keep track of the next expected reschedule time.
This prevents theoretically missing every second pre-emption when
scheduler's timer is interrupting at same rate as the interval.
Bananymous