When small blocks are recombined to create a single block at a shallower
level, it is sufficient to remove those blocks from the free list. There
is no need to mark those small blocks as allocated in the bitmap.
This, in turn, removes the need to mark small blocks back as unallocated
when splitting up a big blocks as they'll already be so marked.
Only the first small block needs to be marked allocated and the
remaining blocks only need to be added to the free list.
This makes the code smaller and more efficient, especially since those
removed bit manipulations were located within loops.
Signed-off-by: Nicolas Pitre <npitre@baylibre.com>
This turns the free-bit flag into an alloc-bit flag effectively
reversing its semantic. This is to make further changes more natural
and easier to understand.
No need to clear the alloc bits at init time as they're located in .bss
and all clear already.
The code remains functionally equivalent after this change.
Signed-off-by: Nicolas Pitre <npitre@baylibre.com>
The realloc function was a bit too intimate with the mempool accounting.
Abstract that knowledge away and move it where it belongs.
Signed-off-by: Nicolas Pitre <npitre@baylibre.com>
The mempool allocator implementation recursively breaks a memory block
into 4 sub-blocks until it minimally fits the requested memory size.
The size of each sub-blocks is rounded up to the next word boundary to
preserve word alignment on the returned memory, and this is a problem.
Let's consider max_sz = 2072 and n_max = 1. That's our level 0.
At level 1, we get one level-0 block split in 4 sub-blocks whose size
is WB_UP(2072 / 4) = 520. However 4 * 520 = 2080 so we must discard the
4th sub-block since it doesn't fit inside our 2072-byte parent block.
We're down to 3 * 520 = 1560 bytes of usable memory.
Our memory usage efficiency is now 1560 / 2072 = 75%.
At level 2, we get 3 level-1 blocks, and each of them may be split
in 4 sub-blocks whose size is WB_UP(520 / 4) = 132. But 4 * 132 = 528
so the 4th sub-block has to be discarded again.
We're down to 9 * 132 = 1188 bytes of usable memory.
Our memory usage efficiency is now 1188 / 2072 = 57%.
At level 3, we get 9 level-2 blocks, each split into WB_UP(132 / 4)
= 36 bytes. Again 4 * 36 = 144 so the 4th sub-block is discarded.
We're down to 27 * 36 = 972 bytes of usable memory.
Our memory usage efficiency is now 972 / 2072 = 47%.
What should be done instead, is to round _down_ sub-block sizes
not _up_. This way, sub-blocks still align to word boundaries, and
they always fit within their parent block as the total size may
no longer exceed the initial size.
Using the same max_sz = 2072 would yield a memory usage efficiency of
99% at level 3, so let's demo a worst case 2044 instead.
Level 1: 4 sub-blocks of WB_DN(2044 / 4) = 508 bytes.
We're down to 4 * 508 = 2032 bytes of usable memory.
Our memory usage efficiency is now 2032 / 2044 = 99%.
Level 2: 4 * 4 sub-blocks of WB_DN(508 / 4) = 124 bytes.
We're down to 16 * 124 = 1984 bytes of usable memory.
Our memory usage efficiency is now 1984 / 2044 = 97%.
Level 3: 16 * 4 sub-blocks of WB_DN(124 / 4) = 28 bytes.
We're down to 64 * 28 = 1792 bytes of usable memory.
Our memory usage efficiency is now 1792 / 2044 = 88%.
Conclusion: if max_sz is a power of 2 then we get 100% efficiency at
all levens in both cases. But if not, then the rounding-up method has
a far worse degradation curve than the rounding-down method, wasting
more than 50% of memory in some cases.
So let's round sub-block sizes down rather than up, and remove
block_fits() which purpose was to identify sub-blocks that didn't
fit within their parent block and is now useless.
Signed-off-by: Nicolas Pitre <npitre@baylibre.com>
The accounting data stored at the beginning of a memory block used by
malloc must push the returned memory address to a word boundary. This
is already the case on 32-bit systems, but not on 64-bit systems where
e.g. struct k_mem_block_id still has a size of 4.
Signed-off-by: Nicolas Pitre <npitre@baylibre.com>
The "bits" field in struct sys_mem_pool_lvl is unioned with a pointer.
That leaves more space for inline free bits on 64-bit targets.
Let's declare it as an array and adjust its size based on the pointer
size. On 32-bit targets the generated code remains identical.
Signed-off-by: Nicolas Pitre <npitre@baylibre.com>
Minimum alignment and rounding must be done on a word boundary. Let's
replace _ALIGN4() with WB_UP() which is equivalent on 32-bit targets,
and 64-bit aware.
Also enforce a minimal alignment on the memory pool. This is making
a difference mostly on64-bit targets where the widely used 4-byte
alignment is not sufficient.
The _ALIGN4() macro has no users left so it is removed.
Signed-off-by: Nicolas Pitre <npitre@baylibre.com>
move misc/mempool.h to sys/mempool.h and
create a shim for backward-compatibility.
No functional changes to the headers.
A warning in the shim can be controlled with CONFIG_COMPAT_INCLUDES.
Related to #16539
Signed-off-by: Anas Nashif <anas.nashif@intel.com>
move misc/mempool_base.h to sys/mempool_base.h and
create a shim for backward-compatibility.
No functional changes to the headers.
A warning in the shim can be controlled with CONFIG_COMPAT_INCLUDES.
Related to #16539
Signed-off-by: Anas Nashif <anas.nashif@intel.com>
move misc/__assert.h to sys/__assert.h and
create a shim for backward-compatibility.
No functional changes to the headers.
A warning in the shim can be controlled with CONFIG_COMPAT_INCLUDES.
Related to #16539
Signed-off-by: Anas Nashif <anas.nashif@intel.com>
The free block bitmap uses either extra memory specified by a pointer
in struct sys_mem_pool_lvl or the space occupied by that pointer
directly if the bitmap length is small enough to fit it.
But the test is wrong. the inline bitmap should be used if the number
of required bits is smaller or _equal_ to the pointer size. Not doing so
would wrongly bounce the free block bitmap to extra memory when the
number of blocks is exactly 32, which is in disagreement with
Z_MPOOL_LBIT_WORDS() that correctly returns 0 in that case.
In theory that mean that this bug would causes an overflow of the free
block bitmap whenever one level has exactly 32 blocks. But right now
there is a separate bug fixed separately that over-sizes the extra block
bitmap mitigating this bug.
Signed-off-by: Nicolas Pitre <npitre@baylibre.com>
The block_fits() predicate was borked. It would check that a block
fits within the bounds of the whole heap. But that's not enough:
because of alignment changes between levels the sub-blocks may be
adjusted forward. It needs to fit inside the PARENT block that it was
split from.
What could happen at runtime is that the last subblocks of a
misaligned parent block would overlap memory from subsequent blocks,
or even run off the end of the heap. That's bad.
Change the API of block_fits() a little so it can extract the parent
region and do this properly.
Fixes#15279. Passes test introduced in #16728 to demonstrate what
seems like the same issue.
Signed-off-by: Andy Ross <andrew.j.ross@intel.com>
In z_sys_mem_pool_block_alloc() the size of the first level block
allocation is rounded up to the next 4-bite boundary. This means one
or more of the trailing blocks could overlap the free block bitmap.
Let's consider this code from kernel.h:
#define K_MEM_POOL_DEFINE(name, minsz, maxsz, nmax, align) \
char __aligned(align) _mpool_buf_##name[_ALIGN4(maxsz * nmax) \
+ _MPOOL_BITS_SIZE(maxsz, minsz, nmax)]; \
The static pool allocation rounds up the product of maxsz and nmax not
size of individual blocks. If we have, say maxsz = 10 and nmax = 20,
the result of _ALIGN4(10 * 20) is 200. That's the offset at which the
free block bitmap will be located.
However, because z_sys_mem_pool_block_alloc() does this:
lsizes[0] = _ALIGN4(p->max_sz);
Individual level 0 blocks will have a size of 12 not 10. That means
the 17th block will extend up to offset 204, 18th block up to 216, 19th
block to 228, and 20th block to 240. So 4 out of the 20 blocks are
overflowing the static pool area and 3 of them are even located
completely outside of it.
In this example, we have only 20 blocks that can't be split so there is
no extra free block bitmap allocation beyond the bitmap embedded in the
sys_mem_pool_lvl structure. This means that memory corruption will
happen in whatever data is located alongside the _mpool_buf_##name
array. But even with, say, 40 blocks, or larger blocks, the extra bitmap
size would be small compared to the extent of the overflow, and it would
get corrupted too of course.
And the data corruption will happen even without allocating any memory
since z_sys_mem_pool_base_init() stores free_list pointer nodes into
those blocks, which in turn may get corrupted if that other data is
later modified instead.
Fixing this issue is simple: rounding on the static pool allocation is
"misparenthesized". Let's turn
_ALIGN4(maxsz * nmax)
into
_ALIGN4(maxsz) * nmax
But that's not sufficient.
In z_sys_mem_pool_base_init() we have:
size_t buflen = p->n_max * p->max_sz, sz = p->max_sz;
u32_t *bits = (u32_t *)((u8_t *)p->buf + buflen);
Considering the same parameters as above, here we're locating the extra
free block bitmap at offset `buflen` which is 20 * 10 = 200, again below
the reach of the last 4 memory blocks. If the number of blocks gets past
the size of the embedded bitmap, it will overlap memory blocks.
Also, the block_ptr() call used here to initialize the free block linked
list uses unrounded p->max_sz, meaning that it is initially not locating
dlist nodes within the same block boundaries as what is expected from
z_sys_mem_pool_block_alloc(). This opens the possibility for allocated
adjacent blocks to overwrite dlist nodes, leading to random crashes in
the future.
So a complete fix must round up p->max_sz here too.
Given that runtime usage of max_sz should always be rounded up, it is
then preferable to round it up once at compile time instead and avoid
further mistakes of that sort. The existing _ALIGN4() usage on p->max_sz
at run time are then redundant.
Signed-off-by: Nicolas Pitre <npitre@baylibre.com>
Permission management no longer necessary, the former
parameter for the mutex is now simply ignored.
Signed-off-by: Andrew Boie <andrew.p.boie@intel.com>
Do not perform early level usage check. This can lead to situation
where block is seen as available on level when it was taken from
the other context.
Fixes: #14504
Signed-off-by: Pawel Dunaj <pawel.dunaj@nordicsemi.no>
Update reserved function names starting with one underscore, replacing
them as follows:
'_k_' with 'z_'
'_K_' with 'Z_'
'_handler_' with 'z_handl_'
'_Cstart' with 'z_cstart'
'_Swap' with 'z_swap'
This renaming is done on both global and those static function names
in kernel/include and include/. Other static function names in kernel/
are renamed by removing the leading underscore. Other function names
not starting with any prefix listed above are renamed starting with
a 'z_' or 'Z_' prefix.
Function names starting with two or three leading underscores are not
automatcally renamed since these names will collide with the variants
with two or three leading underscores.
Various generator scripts have also been updated as well as perf,
linker and usb files. These are
drivers/serial/uart_handlers.c
include/linker/kobject-text.ld
kernel/include/syscall_handler.h
scripts/gen_kobject_list.py
scripts/gen_syscall_header.py
Signed-off-by: Patrik Flykt <patrik.flykt@intel.com>
MISRA rules (see #11425) forbid recursive algorithms. In the case of
rb_walk(), it's not actually used anywhere but a test right now, so we
can simply disable the API when CONFIG_MISRA_SANE is defined. Mempool
had a (IMHO, fairly clever) tail recursive loop in bfree_recombine()
which can be trivially transformed into an only slightly uglier
iterative version.
Signed-off-by: Andy Ross <andrew.j.ross@intel.com>
MISRA rules (see #9892) forbid alloca() and family, even though those
features can be valuable performance and memory size optimizations
useful to Zephyr.
Introduce a MISRA_SANE kconfig, which when true enables a gcc error
condition whenever a variable length array is used.
When enabled, the mempool code will use a theoretical-maximum array
size on the stack instead of one tailored to the current pool
configuration.
The rbtree code will do similarly, but because the theoretical maximum
is quite a bit larger (236 bytes on 32 bit platforms) the array is
placed into struct rbtree instead so it can live in static data (and
also so I don't have to go and retune all the test stack sizes!).
Current code only uses at most two of these (one in the scheduler when
SCHED_SCALABLE is selected, and one for dynamic kernel objects when
USERSPACE and DYNAMIC_OBJECTS are set).
This tunable is false by default, but is selected in a single test (a
subcase of tests/kernel/common) for coverage. Note that the I2C and
SPI subsystems contain uncorrected VLAs, so a few platforms need to be
blacklisted with a filter.
Signed-off-by: Andy Ross <andrew.j.ross@intel.com>
lib/ was starting to get messy and inconsitent. Files being either
dumped in the root or in sub-directories without a clear plan.
Move all library components into one single folder and call it 'os'.
Signed-off-by: Anas Nashif <anas.nashif@intel.com>
Nicolas Pitre