| Age | Commit message (Collapse) | Author |
|---|
|
+ Fairly considerable speedup, as we don't have to look up the rpc pte
every time separately, instead cacheing them. Adds an architecture
specific limitation to total rpc stack size, though.
|
|
+ Found it in newlib. This allows the compiler to more effectively
optimize away unnecessary register operations. Applied to both kernel
and init.c for a slight speed increase.
I really should move init.c to some other repository though, the
binary files are starting to get annoying
|
|
+ Tests will have to come later
|
|
|
|
|
|
|
|
|
|
|
|
+ Try to write arguments to memory as fast as possible to free up some
registers that the compiler can then play with.
Qemu runs ~1 700 000 rpc's per second, visionfive2 @1GHz (I think) ~800 000.
|
|
|
|
|
|
+ Now 10M alloc/frees succeed, which totals more memory than the virtual
machine has, so no too obvious leaks are occuring. For future
debugging speed, changed 10M to 1M.
+ Also quick fix to rpcs, stacks are now assigned. Not entirely sure why
they worked before this, but good that I found it.
|
|
|
|
|
|
|
|
|
|
|
|
+ The secret is using gravestones. I'll have to write up full
documentation for the feature but essentially riscv lets us encode
whatever we want into page table entries, as long as they're not
active. We use this to encode highest user address that is not a zero,
in that all entries in the top level are either active or gravestones.
When an active entry is removed, it is either a gravestone (if there
are other active entries above it) or it starts a cascade of removing
entries that have been previously removed
Slight runtime overhead to page mapping, pretty major advantage in rpc
calls. Feature will need to be tested more thorougly, and the init
program is sort of a best scenario with just one top level userspace
page table entry active at a time, leading to incredibly fast context
switches.
Current implementation limits a process' max virtual memory to 248 GiB
(in Sv39), but I don't think the missing 8 GiB is that big of a deal.
|
|
+ Both easier to look at and now the userspace doesn't play as big of a
role in the 'benchmarking' with optimizations turned on.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
+ Multiple notify signals at the same time, or queued ones waiting for
ipc_req/fwd to finish would require a notify stack of some kind, which
is maybe too complex for my liking. Not impossible by any means, but I
want to keep things simple.
Let's say a client wants the server to do some async operation. First,
the client calls ipc_req to inform the server that it would like the
operation to be done. The server sets this task in some work queue or
whatever, and returns to the client as quickly as possible. Later on,
when the task is finished, the server does a ipc_req to the client,
with data about what was just finished. The client is responsible for
noting this data down somewhere, either reacting to the operation
directly in the callback, or calling ipc_notify to the thread that
requested the operation. Other possible notifications at the same time
should be handled at the same time, whichever way the client wants to
handle them.
Note that ipc_notify triggers only when the thread it targets is in
its base state, i.e. not doing an ipc_req. This makes things a bit
easier.
|
|
+ Now to actually start implementing shit. :)
|
|
+ In total, a syscall is built up of 6 values, with the first being the
syscall number. Symmetrically, the first value is now a status and the
following five values return "values". This allows us to cram in more
info into the ipc_* functions. The performance difference is
absolutely minimal, at least from my testing in qemu.
|
|
+ Does not mean documentation is done, but it's a nice little
achievement nonetheless.
|
|
|
|
|
|
|
|
|
|
|
|
+ Next step, start documenting contents of each file.
|
|
|
|
+ next up would probably be to start working on process/thread
relationships, largely as documented. Still not entirely sure about
how I want to handle fork, but I suppose I might figure it out at some
point.
|
|
|
|
|
|
|
|
|
|
|
