Is it possible to detect `kAddressBits` at runtime?

[rojkov]

As noted in #33 page tables on Aarch64 can have either 3 or 4 levels. tcmalloc assumes 4 levels for Aarch64 (kAddressBits = 48), but for example RaspberryOS for Aarch64 runs on a kernel with CONFIG_PGTABLE_LEVELS=3. As result all programs using tcmalloc fail without patching it to set kAddressBits = 39.

Is it possible to detect kAddressBits at runtime?

[ckennelly]

It would be possible to make kAddressBits be runtime configurable (albeit at a slight performance cost, since we'd need to load the constant rather than use an immediate)--but I think one challenge is reliably detecting the value at runtime.

Other than reading /proc/config.gz--which would be a bit involved while allocating memory--the only strategy that I'm aware of seems to be trial and error. I seem to recall a proposed kernel patch that would have exposed it (maybe as a proc file?) but I don't think it ever landed.

[vchrisb]

maybe the following does help: openjdk/jdk#40
And the thread on the topic: http://openjdk.5641.n7.nabble.com/ZGC-aarch64-Unable-to-allocate-heap-for-certain-Linux-kernel-configurations-td420728.html

I think this is how the folks ad openjdk are handling it by trial and error.

[fandahao17]

+1 after struggling with a mysterious SIGABRT for a whole day and finally finding this issue. As a low-level library, tcmalloc may be depended by projects whose developers even don't know its existence. If detect the correct value of kAddressBits is hard, could we detect whether kAddressBits is misconfigured at runtime so we can abort in a more user-friendly way?

[PKizzle]

As checking for the kernel build config in /proc/config.gz requires CONFIG_IKCONFIG_PROC to be set, the best option would be for tcmalloc to use the proposed trial and error approach. Also a way to manually force kAddressBits to be set during compilation would be appreciated (without any automatic detection).

[tristanmorgan]

Code that detects the address space size.. https://github.com/ptitSeb/box64/blob/e42001b2bc030f93bdba582bf12d6eac63fae345/src/custommem.c#L1444

[ketanhwr]

V8 does a best-effort runtime detection (with fallback to compile time constants) for determining virtual address space size, can something similar be tried in tcmalloc during initialization?

https://github.com/v8/v8/blob/14.1-lkgr/src/sandbox/sandbox.cc#L39-L99