refactor(ADB): improve internal API

[Rudxain]

• Follow up of refactor(sync): remove set_var in favor of adb_id #653
• See refactor(sync): remove set_var in favor of adb_id #653 (comment)

The goal is to make ADB commands:

• ❤️ more ergonomic: contextual auto-complete, RA suggestions, and pre-parsed stdout
• ✨ less ugly (cleaner code)
• 🧠 more maintainable (easier to debug)
• ⌨️ less typo-prone: not "stringly typed"
• ✅ correct & safer: no arbitrary cmd execution, and no env-vars

I am sorry to whoever is going to review this 😅. I didn't want the diff to be so big 💀

[Rudxain]

After adding an unconditional info logger to the new API, I've found something interesting:

INFO  [src/gui/views/list.rs:157] -----------------------------------------------------------------
INFO  [src/gui/views/list.rs:158] ANDROID_SDK: 0 | DEVICE: fetching devices...
INFO  [src/gui/views/list.rs:162] -----------------------------------------------------------------
ERROR [src/gui/views/list.rs:877] AppsView ready but no phone found
INFO  [src/core/adb.rs:120] Ran command: adb 'shell' 'pm' 'list' 'packages' '-s' '-u'
INFO  [src/core/adb.rs:120] Ran command: adb 'shell' 'pm' 'list' 'packages' '-s' '-e'
INFO  [src/core/adb.rs:120] Ran command: adb 'shell' 'pm' 'list' 'packages' '-s' '-d'

This implies fetch_packages is being called even if there's no device (sub-optimal):

universal-android-debloater-next-generation/src/core/utils.rs

Lines 39 to 59 in 0238982

	pub fn fetch_packages(
	uad_lists: &PackageHashMap,
	device_serial: &str,
	user_id: Option<u16>,
	) -> Vec<PackageRow> {
	let all_sys_packs = AdbCommand::new()
	.shell(device_serial)
	.pm()
	.list_packages_sys(Some(PmListPacksFlag::IncludeUninstalled), user_id)
	.unwrap_or_default();
	let enabled_sys_packs: HashSet<String> = AdbCommand::new()
	.shell(device_serial)
	.pm()
	.list_packages_sys(Some(PmListPacksFlag::OnlyEnabled), user_id)
	.unwrap_or_default()
	.into_iter()
	.collect();
	let disabled_sys_packs: HashSet<String> = AdbCommand::new()
	.shell(device_serial)
	.pm()
	.list_packages_sys(Some(PmListPacksFlag::OnlyDisabled), user_id)

This also seems to happen on the main branch:

universal-android-debloater-next-generation/src/gui/views/list.rs

Lines 854 to 856 in bfda7d0

	async fn load_packages(uad_list: PackageHashMap, user_list: Vec<User>) -> Vec<Vec<PackageRow>> {
	if user_list.len() <= 1 {
	vec![fetch_packages(&uad_list, None)]

universal-android-debloater-next-generation/src/gui/views/list.rs

Lines 159 to 177 in bfda7d0

	Message::LoadUadList(remote) => {
	info!("{:-^65}", "-");
	info!(
	"ANDROID_SDK: {} | DEVICE: {}",
	selected_device.android_sdk, selected_device.model
	);
	info!("{:-^65}", "-");
	self.loading_state = LoadingState::DownloadingList;
	Command::perform(
	Self::init_apps_view(remote, selected_device.clone()),
	Message::LoadPhonePackages,
	)
	}
	Message::LoadPhonePackages((uad_list, list_state)) => {
	self.loading_state = LoadingState::LoadingPackages;
	self.uad_lists.clone_from(&uad_list);
	*list_update_state = list_state;
	Command::perform(
	Self::load_packages(uad_list, selected_device.user_list.clone()),

[Rudxain]

I've found info on adb help:

scripting:
 wait-for[-TRANSPORT]-STATE...
     wait for device to be in a given state
     STATE: device, recovery, rescue, sideload, bootloader, or disconnect
     TRANSPORT: usb, local, or any [default=any]

So now we know the full list of states, which we can turn into a (non_exhaustive) enum!

As a bonus, wait-for-device allows UAD-NG to stop polling adb devices (no need for retry!)

[Rudxain]

I've just realized a minor performance oversight: It's possible to "batch" multiple adb shell commands in one run, but this new API makes batching impossible. An easy fix would be to not implicitly run commands, and allow "finished" builders to become ShellCommand objects for every built argument-list, with an implicit ";" in between each finished command.

So instead of:

// Imagine this is an unrolled loop,
// not actual code written manually

// adb shell pm uninstall --user 0 a.a
ACommand::new()
    .shell("")
    .pm()
    .uninstall(0, PackageId::new("a.a").unwrap());
// adb shell pm uninstall --user 0 b.b
ACommand::new()
    .shell("")
    .pm()
    .uninstall(0, PackageId::new("b.b").unwrap());

It would be:

// adb shell 'pm uninstall --user 0 a.a; pm uninstall --user 0 b.b'
ACommand::new()
    .shell("")
    .pm()
    .uninstall(0, PackageId::new("a.a").unwrap())
    .pm()
    .uninstall(0, PackageId::new("b.b").unwrap())
    .run();

This guarantees adb and mksh are executed only once, even when changing the states of thousands of packs.

The biggest obstacle to this, are commands whose output must be parsed: Batching pm list commands would actually be slower, as we would need to iterate over every pack, trying to find the boundaries between the outputs of distinct flags:

# ambiguity!
adb shell 'pm list packages -e; pm list packages -d'

A hacky solution is to insert a delimiter between outputs:

adb shell '\
pm list packages -e;\
echo;\
pm list packages -d\
'

Now we simply find "\n\n", and we've efficiently resolved the ambiguity! This would work for the "raw" list_packages_sys (not in debug mode) but not for list_packages_sys_parsed. So I'll have to add a dedicated batching API for those purposes.

IMO, we shouldn't worry too much about this. It's needlessly complicated just to "improve" performance. Also, too many args (or long args) are more likely to cause crashes

Suggestions have been applied

[Rudxain]

I forgot to emphasize this, but we have to test the multi-user handling before release. I'm not confident that adb::list_users or sync::list_users_parsed work 100% properly. I don't have devices with multiple users, so I would have to install an emulator or something

[AnonymousWP]

Hey @Rudxain, just wanted to thank you for your tremendous work on this PR, it was probably a lot of work. Feel free to convert one of your comments into an issue that are deemed important enough.

Could someone test this? I don't have multiple users either. Perhaps create an issue and add it to the 1.1.1 milestone.