There was a somewhat promiscuous pattern in the kernel where IPC
mechanisms would do something that might effect the current thread
choice, then check _must_switch_threads() (or occasionally
__must_switch_threads -- don't ask, the distinction is being replaced
by real English words), sometimes _is_in_isr() (but not always, even
in contexts where that looks like it would be a mistake), and then
call _Swap() if everything is OK, otherwise releasing the irq_lock().
Sometimes this was done directly, sometimes via the inverted test,
sometimes (poll, heh) by doing the test when the thread state was
modified and then needlessly passing the result up the call stack to
the point of the _Swap().
And some places were just calling _reschedule_threads(), which did all
this already.
Unify all this madness. The old _reschedule_threads() function has
split into two variants: _reschedule_yield() and
_reschedule_noyield(). The latter is the "normal" one that respects
the cooperative priority of the current thread (i.e. it won't switch
out even if there is a higher priority thread ready -- the current
thread has to pend itself first), the former is used in the handful of
places where code was doing a swap unconditionally, just to preserve
precise behavior across the refactor. I'm not at all convinced it
should exist...
Signed-off-by: Andy Ross <andrew.j.ross@intel.com>
Assertions should never be used to test for error conditions, such as
checking for overflows. It should only be used to test for invariants.
Signed-off-by: Leandro Pereira <leandro.pereira@intel.com>
If a large size is requested, the expression `size += sizeof(...)`
might overflow, leading to a small block being requested and returned
by k_malloc().
Use a GCC builtin to trap the overflow and return NULL in this case.
Signed-off-by: Leandro Pereira <leandro.pereira@intel.com>
We would like to offer the capability to have memory pool heap data
structures that are usable from user mode threads. The current
k_mem_pool implementation uses IRQ locking and system-wide membership
lists that make it incompatible with user mode constraints.
However, much of the existing memory pool code can be abstracted to some
common functions that are used by both k_mem_pool and the new
sys_mem_pool implementations.
The sys_mem_pool implementation has the following differences:
* The alloc/free APIs work directly with pointers, no internal memory
block structures are exposed to the end user. A pointer to the source
pool is provided for allocation, but freeing memory just requires the
pointer and nothing else.
* k_mem_pool uses IRQ locks and required very fine-grained locking in
order to not affect system latency. sys_mem_pools just use a semaphore
to protect the pool data structures at the API level, since there aren't
implications for system responsiveness with this kind of concurrency
control.
* sys_mem_pools do not support the notion of timeouts for requesting
memory.
* sys_mem_pools are specified at compile time with macros, just like
kernel memory pools. Alternative forms of specification at runtime
will be a later enhancement.
Signed-off-by: Andrew Boie <andrew.p.boie@intel.com>
The xtensa-asm2 work included a patch that added nano_internal.h
includes in lots of places that needed to have _Swap defined, because
it had to break a cycle and this no longer got pulled in from the arch
headers.
Unfortunately those new includes created new and more amusing cycles
elsewhere which led to breakage on other platforms.
Break out the _Swap definition (only) into a separate header and use
that instead. Cleaner. Seems not to have any more hidden gotchas.
Signed-off-by: Andy Ross <andrew.j.ross@intel.com>
_Swap() is defined in nano_internal.h. Everything calls _Swap().
Pretty much nothing that called _Swap() included nano_internal.h,
expecting it to be picked up automatically through other headers (as
it happened, from the kernel arch-specific include file). A new
_Swap() is going to need some other symbols in the inline definition,
so I needed to break that cycle. Now nothing sees _Swap() defined
anymore. Put nano_internal.h everywhere it's needed.
Our kernel includes remain a big awful yucky mess. This makes things
more correct but no less ugly. Needs cleanup.
Signed-off-by: Andy Ross <andrew.j.ross@intel.com>
We don't need to store the full k_mem_block, rather just the
k_mem_block_id. In effect, this saves 4 bytes of memory per allocated
memory chunk. Also take advantage of the newly introduced
k_mem_pool_free_id API here.
Signed-off-by: Johan Hedberg <johan.hedberg@intel.com>
The k_mem_pool_free API has no use for the full k_mem_block struct. In
particular, it only needs the k_mem_block_id. Introduce a new API
which takes only this essential struct. This paves the way to
simplify & improve the k_malloc/k_free implementation a bit.
Signed-off-by: Johan Hedberg <johan.hedberg@intel.com>
This test was just wrong. If the current thread did not race with any
others during the allocation process, then the result will be false
because it was detected so earlier in the function. If we did race,
then sure: it might be true now if someone snuck in and freed a block.
But so what? We already have the block we want to break. The
behavior in the code as written was to early-exit from the break loop,
returning a buffer that was larger than the one requested (though
otherwise benign -- we wouldn't leak, just waste memory). No idea
what I was thinking.
Thanks to Du Quanwen for the diagnosis.
Signed-off-by: Andy Ross <andrew.j.ross@intel.com>
This patch amounts to a mostly complete rewrite of the k_mem_pool
allocator, which had been the source of historical complaints vs. the
one easily available in newlib. The basic design of the allocator is
unchanged (it's still a 4-way buddy allocator), but the implementation
has made different choices throughout. Major changes:
Space efficiency: The old implementation required ~2.66 bytes per
"smallest block" in overhead, plus 16 bytes per log4 "level" of the
allocation tree, plus a global tracking struct of 32 bytes and a very
surprising 12 byte overhead (in struct k_mem_block) per active
allocation on top of the returned data pointer. This new allocator
uses a simple bit array as the only per-block storage and places the
free list into the freed blocks themselves, requiring only ~1.33 bits
per smallest block, 12 bytes per level, 32 byte globally and only 4
bytes of per-allocation bookeeping. And it puts more of the generated
tree into BSS, slightly reducing binary sizes for non-trivial pool
sizes (even as the code size itself has increased a tiny bit).
IRQ safe: atomic operations on the store have been cut down to be at
most "4 bit sets and dlist operations" (i.e. a few dozen
instructions), reducing latency significantly and allowing us to lock
against interrupts cleanly from all APIs. Allocations and frees can
be done from ISRs now without limitation (well, obviously you can't
sleep, so "timeout" must be K_NO_WAIT).
Deterministic performance: there is no more "defragmentation" step
that must be manually managed. Block coalescing is done synchronously
at free time and takes constant time (strictly log4(num_levels)), as
the detection of four free "partner bits" is just a simple shift and
mask operation.
Cleaner behavior with odd sizes. The old code assumed that the
specified maximum size would be a power of four multiple of the
minimum size, making use of non-standard buffer sizes problematic.
This implementation re-aligns the sub-blocks at each level and can
handle situations wehre alignment restrictions mean fewer than 4x will
be available. If you want precise layout control, you can still
specify the sizes rigorously. It just doesn't break if you don't.
More portable: the original implementation made use of GNU assembler
macros embedded inline within C __asm__ statements. Not all
toolchains are actually backed by a GNU assembler even when the
support the GNU assembly syntax. This is pure C, albeit with some
hairy macros to expand the compile-time-computed values.
Related changes that had to be rolled into this patch for bisectability:
* The new allocator has a firm minimum block size of 8 bytes (to store
the dlist_node_t). It will "work" with smaller requested min_size
values, but obviously makes no firm promises about layout or how
many will be available. Unfortunately many of the tests were
written with very small 4-byte minimum sizes and to assume exactly
how many they could allocate. Bump the sizes to match the allocator
minimum.
* The mbox and pipes API made use of the internals of k_mem_block and
had to be ported to the new scheme. Blocks no longer store a
backpointer to the pool that allocated them (it's an integer ID in a
bitfield) , so if you want to "nullify" them you have to use the
data pointer.
* test_mbox_api had a bug were it was prematurely freeing k_mem_blocks
that it sent through the mailbox. This worked in the old allocator
because the memory wouldn't be touched when freed, but now we stuff
list pointers in there and the bug was exposed.
* Remove test_mpool_options: the options (related to defragmentation
behavior) tested no longer exist.
Signed-off-by: Andy Ross <andrew.j.ross@intel.com>
