cpu/esp32: support of multiple heaps for newlib malloc

[gschorcht]

Contribution description

This PR adds support of multiple heaps if module esp_idf_heap is not used.

For that purpose the following unused DRAM sections are added as additional heap segments:

0x3ffae6e0 ... 0x3ffaffff   heap1   6 kiB
0x3ffe0440 ... 0x3ffe3fff   heap2   14 kiB
0x3ffe4350 ... 0x3fffffff   heap3   111 kiB

With these additional heap segments, the overall heap size increases from 180 kiB to 315 kiB for the tests/heap_cmd application.

The standard heap is located in:

0x3ffb0000 ... 0x3ffdffff

It's size depends on other data since .data and .bss sections are also placed there. Therefore, the heap starts at _bss_end.

Testing procedure

Use the following command to flash an ESP32 board:

CFLAGS='-DNUMBER_OF_TESTS=1 -DCHUNK_SIZE=1000' BOARD=esp32-wroom-32 \
make -C tests/malloc flash

Check that all heap segments are used:

CHUNK_SIZE: 1000
NUMBER_OF_TESTS: 1
Allocated 1000 Bytes at 0x0x3ffb2900, total 1000
...
Allocated 1000 Bytes at 0x0x3ffdf900, total 182440
Allocated 1000 Bytes at 0x0x3ffae6e8, total 183448
...
Allocated 1000 Bytes at 0x0x3ffafa98, total 188488
Allocated 1000 Bytes at 0x0x3ffe0448, total 189496
...
Allocated 1000 Bytes at 0x0x3ffe3b68, total 203608
Allocated 1000 Bytes at 0x0x3ffe4358, total 204616
...
Allocated 1000 Bytes at 0x0x3ffffb78, total 316504
Allocations count: 314

Issues/PRs references

Depends on PR #13516
Requires PR RIOT-OS/riotdocker#102 to work correctly.

[benpicco]

I guess https://doc.riot-os.org/group__cpu__esp32.html#esp32_manual_toolchain_installation will need an update too.

Looks good, please rebase (and remove the trailing whitespace while you're at it, so Travis is happy).

[gschorcht]

I guess https://doc.riot-os.org/group__cpu__esp32.html#esp32_manual_toolchain_installation will need an update too.

No. I just forgot to push the master. The documentation doesn't refer a certain commit as riotdocker does. For simplicity's sake, the master is always used. I have pushed the master so that it corresponds to the version used for riotdocker.

[benpicco]

I thought you would have squashed the whitespace fix directly 😉
Anyway, @kaspar030 needs to update Murdock first so CI will not fail.

[gschorcht]

Anyway, @kaspar030 needs to update Murdock first so CI will not fail.

Indeed, tests/malloc failed 😟 I didn't think that tests/malloc would recognize that the chunk allocation with the old fat malloc is wrong.

[benpicco]

My suspicion is correct - tests/malloc works now, so the toolchain was upgraded - pleas squash btw.

[benpicco]

The update has finally reached every CI worker I think - this needs another rebase though :)

[benpicco]

ping @gschorcht 😉

[benpicco]

Please squash

[gschorcht]

Murdock failed because of the following tests that are not related to this PR (output).

    tests/progress_bar/nrf52dk:gnu
    examples/micropython/nrf52dk:gnu
    tests/float/nrf52dk:gnu

Even stranger is that Murdock only did some tests on the nrf52dk board, but none on esp32-wroom-32. I don't understand how tests to run are selected.

[gschorcht]

@benpicco Do you think we should trigger Murdock again with the label CI disable test cache to force all tests to run on ESP32? Or should we simply remove label CI run tests to pass Murdock compilation?

[benpicco]

The tests on nrf52dk are flaky, I don't know why - but looks like now everything worked.

[gschorcht]

The tests on nrf52dk are flaky, I don't know why - but looks like now everything worked.

Yeah, but only because no test was run because of the test cache. Should we run the test again with the label CI clear test cache so we can be sure that the tests on the ESP32 will still run after the rebase?

[benpicco]

I ran the test on a local esp32 and it showed no errors. But if you want, you can run Murdock again.

[gschorcht]

🤔 It probably makes no sense to run all tests with static data again. The only test that probably uses dynamic memory is tests/malloc. So lets merge before the hard feature freeze.

[gschorcht]

@benpicco Thanks for reviewing, testing and merging.