Drgn reads zeros for memory allocated by vm_map_ram() about drgn HOT 4 CLOSED

As an FYI, looks like there will be an upstream fix for this!
https://lore.kernel.org/linux-mm/[email protected]/T/#u

from drgn.

This is on the live kernel, right? My wild guess is that this is a bug in /proc/kcore, because drgn isn't super fancy about reading memory. Do you get back the correct results if you read with access_remote_vm(prog["init_mm"].address_of_(), address, size) instead of prog.read()? That will translate the virtual address to a physical address from the kernel page table and then read based on the physical address. If that works, it's very likely a bug in /proc/kcore.

from drgn.

Yeah - live kernel. I tracked it through the /proc/kcore reads with some printf debugging and was coming to that conclusion myself.

The access_remote_vm() worked! So I guess it's time to look into the kernel for the bug.

I'm rather impressed that crash is doing the virt to phys translation manually given that /proc/kcore seems to have an ELF segment for this region of memory. Maybe the original /dev/mem did not, and crash was architected around that?

from drgn.

The issue must be in vread() not supporting memory which comes from vm_map_ram(), if the allocation is small it comes from vb_alloc().

from drgn.