When two CPUs read the area of the DPORT and the area of the APB, the result is corrupted for the CPU that read the APB area.
And another CPU has valid data.
The method of eliminating this error.
Before reading the registers of the DPORT, make a preliminary reading of the APB register.
In this case, the joint access of the two CPUs to the registers of the APB and the DPORT is successful.
Makes app image booting more reliable (256-bit rather than 8-bit verification.)
Some measurements, time to boot a 655KB app.bin file and run to app_main() execution.
(All for rev 1 silicon, ie no 340ms spurious WDT delay.)
80MHz QIO mode:
before = 300ms
after = 140ms
40MHz DIO mode:
before = 712ms
after = 577ms
40MHz DIO mode, secure boot enabled
before = 1380ms
after = 934ms
(Secure boot involves two ECC signature verifications (partition table, app) that take approx 300ms each with 80MHz CPU.)
The issue that cache entries are not invalidated correctly sometimes
can also be reproduced for non-encrypted flash as well.
This change updates the workaround to do Cache_Flush, enabling it for
non-encrypted flash, and adds a unit test.
Partition/SPI/OTA docs & OTA new functionality
* Update partition, SPI flash & OTA docs to reflect functionality changes
* Refactor OTA implementation to perform checks mentioned in API doc
* Add new functions to OTA API: esp_ota_get_running_partition() & esp_ota_get_next_update_partition() functions
* Add spi_flash_cache2phys() & spi_flash_phys2cache() functions to support esp_ota_get_running_partition()
See merge request !513
Seems doing certain kinds of short reads while flash encryption is
enabled will return stale data. This fixes it, but is probably a
little heavy-handed performance wise.
Without this, it's possible for stale information to be read from
cache via mmap, even if the MMU table entry had been invalidated
prior to writing flash (if the same MMU table entry was re-used after
writing flash.)
