1
0
mirror of https://github.com/trezor/trezor-firmware.git synced 2024-12-29 17:48:10 +00:00
trezor-firmware/docs/core/misc/fragmentation.md
2021-05-06 13:14:21 +02:00

4.4 KiB

Memory fragmentation management

Trezor-core memory is managed by a mark-and-sweep garbage collector. Throughout the run-time of the firmware, the memory space gets increasingly fragmented as the GC sweep is initiated at arbitrary points.

To combat fragmentation, we attempt to thoroughly clear the memory space after finishing every workflow, and keep only a limited set of modules alive at all times. These must take care to not hold external references.

Always active modules

The following modules are kept loaded at all times:

  • trezor
  • trezor.utils
  • storage
  • storage.common
  • storage.cache
  • storage.device
  • storage.fido2
  • trezor.pin - held alive because the function show_pin_timeout is registered as a callback for trezorconfig and storage unlock operations
  • usb

The above modules are only allowed to import C modules (trezorconfig, trezorutils, trezorcrypto, etc.) or each other. We currently do not have any automation to enforce this, so please be careful when editing them.

Presizing

To save storage, Micropython only preallocates 1 slot in a module dict. Most of our modules use more slots than that. This means that the dict is reallocated, possibly several times. This is inconvenient at most times, but especially undesirable when it would happen to an always-active module at some point at run-time. The allocator would put the newly reallocated dict somewhere in the middle of the GC arena, and it would stay there.

This does happen in practice: e.g., when you import trezor.strings, a new reference strings is inserted into the trezor module.

For this reason, we call utils.presize_module on trezor and storage at first import time. The sizes are determined empirically and it might be necessary to raise them in the future.

The backing storage for sys.modules can also be reallocated at run-time. We configure Micropython to preallocate 160 slots in mpconfigport.h variable MICROPY_LOADED_MODULES_DICT_SIZE. This is asserted at the end of unimport in trezor.utils, so if we ever need more modules than that, the test suite should catch it.

Top-level and function-local imports

In order to keep the imported image size in check, in certain places we avoid importing something at top-level, and instead import it in a function which actually needs the functionality. That way the module can be imported without immediately pulling in all of its possible dependencies.

The following imports trezor.ui at import time - when importing module, trezor.ui is always imported, regardless of whether anyone calls the function draw_foo:

# module.py
import trezor.ui

def draw_foo():
    trezor.ui.display.draw_text("Foo")

The following defers the import until the function is called:

# module.py

def draw_foo():
    import trezor.ui

    trezor.ui.display.draw_text("Foo")

The general rules of thumb are as follows:

C modules can always be imported.

These do not take any space in RAM.

Always-active modules can always be imported.

They are always active, so we do not need to worry about allocating.

In apps.*, we prefer clarity over optimization.

It might still be useful to, e.g., avoid importing trezor.ui.layouts for operations that are sometimes silent, but it is not too important. All of the application code is scrubbed from memory when the workflow exits.

In system modules, we are extra careful.

This means apps.base, apps.common, and everything outside the apps namespace.

A module should only import on top-level if the import is either:

  • C module or an always active module,
  • a module that is expected to already be imported when this module is loaded (this is often the case in apps.common -- e.g., trezor.workflow is not always active, but is presumed active as soon as session is up),
  • small module without further dependencies,
  • something without which the whole module doesn't make sense (this is usually the case with layout code: apps.common.confirm doesn't make sense without importing trezor.ui)

Avoid importing trezor.ui.

The trezor.ui namespace is one of the largest in the codebase, not counting application code. Importing the trezor.ui module alone is not a big problem, but pulling in anything from trezor.ui.layouts or trezor.ui.components usually means loading the full UI machinery. We only want to do that if we are sure that whoever is importing us is going to be drawing things.