Conversation

Replying to

and

I think your best bet to figure out if this is the issue is disabling the module in your kernel build and disabling the sanity checks for kernel module loading so that the init scripts in vendor can successfully load the stock OS build of the HTC battery module from there.

1

2

Replying to

and

It's entirely possible to build it as a dynamic kernel module and include it in vendor, but it's easier to build it into the kernel which is why most aftermarket OS builds take that approach. Stock OS uses dynamic modules to optimize loading time and make kernels more generic.

1

1

Replying to

and

It's a recurring issue that we deal with across a bunch of drivers. The drivers only work if their firmware and configuration files are usable when they're initialized and they often assume that those are available when they're loaded. If built into the kernel, they're broken.

2

4

Replying to

and

this... requires enough effort that I might just live with partially broken charging. especially given that I still haven't fixed the problem with the USB3 pairs in the connector that makes `adb sideload` very hard to use

1

1

Replying to

and

Based on that error message, I'm fairly certain that the battery driver or battery is the issue though. If you look at drivers/power/htc_battery.c, you can see it's involved in negotiating the charging rate. Since there's thermal throttling, etc. batteries are very involved.

2

2

Replying to

and

honestly I just find the normal part of the AOSP build system bad enough, the thought of doing something with the kernel instills dread

1

6

Replying to

and

(it's the LineageOS kernel build integration, I have absolutely no idea how it works)

1

1

Replying to

and

We use the AOSP way of building it separately and including it in the AOSP build as a prebuilt. Qualcomm has always included a way to run the kernel build in-tree via a wrapper in each kernel source tree as a hack for development. LineageOS turned that into a generic system.

2

2

Replying to

and

I don't like running a totally separate build system within the AOSP build system. It takes a ridiculous amount of time and memory to link the Linux kernel with full LTO and it doesn't use kati so it doesn't handle incremental builds well so I really prefer the AOSP approach.

1

4

Replying to

and

Why LTO? Does it have extended hardening properties or is it just 3% optimizing non-bottlenecks?

2

Replying to

and

Pixels build the kernels with Clang CFI which requires LTO. If you disable dynamic kernel modules, set -fvisibility=hidden as the default and avoid using ThinLTO, you get much better CFI granularity because it can see that nearly everything can't be indirectly called.

6:18 AM · Jan 23, 2021Twitter Web App

Replying to

and

Stock OS uses dynamic kernel modules in order to have a generic base kernel across phones with device-specific modules loaded from vendor. It also speeds up the boot process because loading the dynamic kernel modules is done asynchronously as part of the event-based init process.

1

1

Replying to

and

Qualcomm Atheros Wi-Fi in Snapdragon uses SoftMAC and doesn't really have a Wi-Fi chip but rather it's one of the sandboxed processes on the baseband. It hardly has any functionality in the firmware but rather they implement it all in the kernel driver. It's *really* complex.

1

2

Show replies

Replying to

and

It's infuriating that this needs LTO because conceptually there's no reason it couldn't just be normal linking with some extended symbol markup.

1

Replying to

and

Compiling vs linking is the most critical scalability boundary and compiler folks are determined to tear it down because they love their fancy toy (LTO) and keep building all the actually useful stuff on top of it to push it.

2

Show replies