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.
On Xtensa, backtrace can not recover the two most significant bits of
the address, as the window call size is encoded in these bits.
Because of this, __builtin_return_address modifies these MSBs to
match those of the callee, "fixing" the address. An unfortunate side
effect is that the zero return address, which usually terminates the
backtrace, gets converted to 0x40000000. While there is a valid
instruction at this address, its occurrence in the backtrace is
highly unlikely: this is the first instruction of WindowOverflow4
vector, and IDF apps switch VECBASE to an IRAM location very early at
startup.
Problem:
The new API esp_bt_mem_release() that was added freed BTDM data to heap from esp_bt_controller_mem_release().
Now with the BT memory optimization commit ee787085f9,
the BTDM data is optimized and reduced to only 32 bytes which is not sufficient amount to be added to heap.
So, using the API leads to assert saying that the region is too small.
Solution:
Modify heap_caps_add_region_with_caps to return ESP_ERR_INVALID_SIZE in case the range is too small to create a new heap.
Do not assert if return value is ESP_ERR_INVALID_SIZE
This also fixes using API esp_bt_controller_mem_release() with ESP_BT_MODE_BTDM
Signed-off-by: Hrishikesh Dhayagude <hrishi@espressif.com>
Add back a feature that was available in the old heap implementation
in release/v2.1 and earlier: keep track of which task allocates each
block from the heap. The task handle is conditionally added as
another word in the heap poisoning header under this configuration
option CONFIG_HEAP_TASK_TRACKING.
To allow custom monitoring and debugging code to be added, add helper
functions in multi_heap.c and multi_heap_poisoning.c to provide access
to information in the block headers.
Add esp_heap_debug_dump_totals() to monitor heap usage
esp_heap_debug_dump_totals() dumps into a user-provided data structure
a summary of the amound of heap memory in region type that is used by
each task. Optionally it will also dump into another data structure
the metadata about each allocated block for a given list of tasks or
for all tasks (limited by available space).
Address change requests on PR #1498
This set of changes fixes the files in e3b702c to just add the
CONFIG_HEAP_TASK_TRACKING option without adding the new function
heap_caps_get_per_task_info() in case that is the only portion of the
PR that will be accepted. Part of the change is to remove the new .c
and .h files containing that function and to remove the line to
compile it from components/heap/component.mk since it should not have
been included in e3b702c. One or more additional commits to add the
new function will follow.
The other changes here:
- uint32_t get_all_caps() moves to heap_private.h
- replace "void* foo" with "void *foo"
- add braces around single-line "if" blocks
- replace tab characters with spaces
Address change requests on PR #1498, part 2
This set of changes fixes the files in cdf32aa to add the new function
heap_caps_get_per_task_info() with its new name and to add the line to
compile it in components/heap/component.mk. This does not address all
the suggested changes because there are some needing further
discussion.
This commit does not include the suggested change to move the
declaration of the new function into esp_heap_caps.h because the new
function references TaskHandle_t so esp_heap_caps.h would have to
include freertos/FreeRTOS.h and freertos/task.h, but FreeRTOS.h
includes esp_heap_caps.h through two other header files which results
in compilation errors because not all of FreeRTOS.h has been read yet.
Change heap_caps_get_per_task_info() to take struct of params
In addition to moving the large number of function parameters into a
struct as the single parameter, the following changes were made:
- Clear out the totals for any prepopulated tasks so the app code
doesn't have to do it.
- Rather than partitioning the per-task totals into a predetermined
set of heap capabilities, take a list of <caps,mask> pairs to
compare the caps to the heap capabilities as masked. This lets the
caller configure the desired partitioning, or none.
- Allow the totals array pointer or the blocks array pointer to be
NULL to indicate not to collect that part of the information.
- In addition to returning the total space allocated by each task,
return the number of blocks allocated by each task.
- Don't need to return the heap capabilities as part of the details
for each block since the heap region (and therefore its
capabilities) can be determined from the block address.
- Renamed heap_task_info.h to esp_heap_task_info.h to fit the naming
convention, and renamed the structs for totals and block details to
better fit the revised function name.
- Provide full Doxygen commenting for the function and parameter
structs.
Add copyright header to new files
Merges https://github.com/espressif/esp-idf/pull/1498
Felipe Neves