| Age | Commit message (Collapse) | Author |
|---|
|
Rather than having pulp_tick() on win32 always fiberize/unfiberize the
thread at the start/end of the tick, just fiberize once at library
initialization time.
The old approach was problematic with nested pulp instances. As in,
when a fiber called pulp_tick() on another pulp_t, in scenarios where
the programmer has arranged for a fiber heirarchy of sorts, the nested
pulp_tick() would once again fiberize the already fiberized thread,
then unfiberize, and poof the magic smoke comes out.
Rather than maintaining some kind of reference count and transparently
fiberize vs. getcurrentfiber in pulp_tick(), I'm just introducing the
explicit initialization to address this limitation. This way
pulp_tick() can just always get the current fiber for the calling
context on win32.
The initializer currently does nothing on *nix/ucontext.h systems,
this is just a win32 issue.
|
|
|
|
|
|
In cases where the run queue was empty, the trampoline context would
immediately enter the caller context. Then the trampoline would be
reused again without being re-setup.
This seemed to work fine on Linux but was causing bus errors on OSX.
So instead now enter_context never used and contexts are always
swapped to preserve the state of the context being left, which
seems to have fixed the problem on OSX.
|
|
I don't have an OSX system to test with, but after some digging
this looks like it might suffice.
|
|
|
|
This adds a rudimentary win32 implementation based on the
fibers system API. It seems to work reasonably under wine when
built w/mingw32.
|
|
The sleep functions, which are few, and the only current means for
fibers to block/enter the scheduler, now take an optional
pulp_mailbox_t * parameter.
When supplied, other fibers may communicate with the sleeping fiber
via pulp_fiber_get_mailslot(). This function checks the destination
fiber for a mailbox and verifies there's space available. On success,
it returns a pointer to the next available mailslot while advancing
the mailboxes mail count.
The caller may then use this mailslot pointer to write directly to the
void * it references in the mailbox. If the receiving fiber took care
to populate its mailbox with slots referencing external memory, the
sender could dereference the mailslot's value to find the external
memory and write larger messages without needing to allocate new space
itself only for the reciever to have to free it shortly.
It's also possible to not do anything at all with the mailslot. Simply
successfully getting a mailslot communicates _something_ to the
recipient by virtue of the count advancing.
The returned mailslot can only be considered valid by the calling
fiber until it sleeps, after that the slot must be treated as
reclaimed from the perspective of the fiber that called
pulp_fiber_get_mailslot().
When the sleeping fiber wakes and returns from its mailbox-enabled
sleep call, it simply looks at the count member of the mailbox to see
if any mail was received. It's up to the implementation what is done
with the contents of the mailbox.
The mailbox count is always reset to 0 automatically at the start of a
mailbox-enabled sleep.
|
|
Analogous to pthread_self()
|
|
Expiring alarms of sleeping fibers in pulp_schedule() has the potential
to prevent pulp_tick() from ever returning if there's always another
fiber expiring @ schedule.
In order to allow simple integration with external event loops, pulp_tick()
is supposed to operate on what is essentially a slice of runnable fibers
then immediately return control to the caller.
|
|
libpulp is a very basic cooperatively-scheduled userspace threading
library. It refers to the userspace threads as fibers, and uses my
other little library libthunk to bind the functions and their calling
environment into tidy little structures having a uniform calling
convention.
At this time there's no build system or anything of the sort, it's
early days and I'll probably be submoduling this into another project
for direct inclusion and compilation.
|
