The implicit promise of heap_alloc_caps() and friends is that the memory it
returns is fit for the purpose as requested in the caps field. Before
this commit, that did not happen; e.g. DMA-capable memory wass returned
from a correct region, but not aligned/sized to something the DMA subsystem
can handle.
This commit adds an API to the esp_mm component that is then used by the
heap component to adjust allocation alignment, caps and size dependent on
the hardware requirement of the requested allocation caps.
This return value is used by the tlsf_walk_pool
function to be notified of the need to interrupt
the ongoing traversal of the currently traversed
heap.
Introduce new APIs in essp_heap_caps.h:
- heap_caps_walk()
- heap_caps_walk_all()
Those functions are triggering a callback for all blocks
(allocated or free) of memory present in heaps meeting
the set of capabilities passed as parameter (or all heaps
for heap_caps_walk_all() function)
test_walker.c added to test the new functionality in
test_apps/heap_test/
Implement a function to start tracking the minimum free size from the
moment the function is called.
Implement a function to stop tracking the minimum free size and restore
the minimum free size value calculated since startup.
Implement the tests related to this new feature.
Previously, the hash map was a doubly linked list but was never
using the characteristics of it.
This commit switches the hash map to a singly linked list instead
This commit also fixes memory leak in heap trace tests by setting
hashmap size to 10 for the tests (instead of the default value of 250)
See https://github.com/espressif/esp-idf/issues/11173
1. move startup_stack attr from soc_memory_type_desc_t to soc_memory_region_t and
remove unused aliased_iram field
2. all of the last level of RAM is retention dma accessible on esp32c3
3. remove esp32c2 and later chips retention dma accessible memory caps
4. allow allocate memory from RTC_RAM with MALLOC_CAP_EXEC cap
Add test configuration to run all tests with heap component in the flash.
Add reference to this new configuration in performance section of the documentation.
- and place all added functions and vairables related to the hashmap in RAM
when the config is enabled only.
- add number of hash map entry as a Kconfig value and remove the hash map init function.
This prevents the user from allocating the hash map in flash and pass the pointer to the
init function (as the heap trace manipulate the hash map from functions placed in IRAM).
- add max linear value to the KConfig to make it configurable by the users.
- protect access to static variable "tracing"
- remove unecessary field in heap_trace_hashmap_entry_t
- Fix "test get allocated size"
- Add tests for the free / alloc hooks
- Call alloc function hook on malloc/realloc/calloc base functions
- Add caps parameter to the allocation hook function
- Call TAILQ_INSERT_TAIL in linked_list_setup to add unused records from the tail of the list
- Fix test "heap trace leak check" to expect that after a free, the record is zeroed instead of checking that
the whole list of records is moved by one index in the array.
- Use esp_rom_printf() under lock instead of printf() since it does not rely on interrupts.
- When PSRAM is enabled, the heap integrity check APIs
(`heap_caps_check_integrity` and `heap_caps_check_integrity_all`)
take more time to execute and keep the interrupts disabled for a
longer time.
Add the definition of tlsf_check_hook() in multi_heap if MULTI_HEAP_POISONING
is set. This definition calls the multi_heap_internal_check_block_poisoning()
to check the memory of a free block for corruption. If the light poisoinng is
set this function returns true. If the comprehensive poisoning is set, this
function will check that all byte of memory in the memory chunk passed as parameter
are set to the right FILL pattern.
This commit marks all functions in interrupt_controller_hal.h, cpu_ll.h and cpu_hal.h as deprecated.
Users should use functions from esp_cpu.h instead.
Regression was introduced in 32408b718f, which disallowed
addition of heap region with following condition:
`new_start < start && new_end == start`
This caused issues in Bluetooth APIs `esp_bt_mem_release` or `esp_bt_controller_mem_release`.
This commit fixes the problem and also adds API documentation for supported memory address
ranges in heap add region APIs.
This commit updates the chip independent system chapters of the
programming guide for esp32s3.
Signed-off-by: Sudeep Mohanty <sudeep.mohanty@espressif.com>
heap: ported tlsf allocator into multi heap
heap_host_tests: added tlsf allocator into host test
heap_host_test: update freebytes after using free
heap_tests: tlsf now passing on host tests without poisoning
multi_heap: added support for memalign using tlsf implementation
heap_caps: removed heap_caps_aligned_free
heap/test: fixed broken aligned alloc test build
heap: added poisoning pattern when blocks are being merged
heap/tests: added timing tests for memory allocation
heap: reduced tlsf structure overhead
heap/tlsf: made all short functions inside of tlsf module as inline to improve timings
heap: moved tlsf heap routines outside of flash memory
newlib: linked multiheap memalign with newlib memalign function
heap: moved block member functions to a separate file so multi_heap can use the functions
heap/test: improved the tlsf timing test
heap/test: added memalign on aligned alloc tests
heap: moved tlsf configuration constants to a separated file
heap: added random allocations test with timings
heap: modified the calculation of heap free bytes
heap: make aligned free true deprecated functions and update their documentation
heap: add extra assert after successive mallocs on small allocation host test
heap: remove legacy aligned alloc implementation.
performance: added malloc and free time performance default values
Configurable option to use IRAM as byte accessible memory (in single core mode) using
load-store (non-word aligned and non-word size IRAM access specific) exception handlers.
This allows to use IRAM for use-cases where certain performance penalty
(upto 170 cpu cycles per load or store operation) is acceptable. Additional configuration
option has been provided to redirect mbedTLS specific in-out content length buffers to
IRAM (in single core mode), allows to save 20KB per TLS connection.
