Re: [PATCH bpf-next 2/7] set_memory: introduce set_memory_[ro|x]_noalias

On Nov 17, 2021, at 3:57 PM, Song Liu [off-list ref] wrote:

quoted

On Nov 17, 2021, at 2:01 PM, Peter Zijlstra [off-list ref] wrote:

On Wed, Nov 17, 2021 at 09:36:27PM +0000, Song Liu wrote:

quoted

quoted

On Nov 16, 2021, at 12:00 AM, Peter Zijlstra [off-list ref] wrote:

On Mon, Nov 15, 2021 at 11:13:42PM -0800, Song Liu wrote:

quoted

These allow setting ro/x for module_alloc() mapping, while leave the
linear mapping rw/nx.

This needs a very strong rationale for *why*. How does this not
trivially circumvent W^X ?

In this case, we want to have multiple BPF programs sharing the 2MB page. 
When the JIT engine is working on one program, we would rather existing
BPF programs on the same page stay on RO+X mapping (the module_alloc() 
address). The solution in this version is to let the JIT engine write to 
the page via linear address. 

An alternative is to only use the module_alloc() address, and flip the 
read-only bit (of the whole 2MB page) back and forth. However, this 
requires some serialization among different JIT jobs. 

Neither options seem acceptible to me as they both violate W^X.

Please have a close look at arch/x86/kernel/alternative.c:__text_poke()
for how we modify active text. I think that or something very similar is
the only option. By having an alias in a special (user) address space
that is not accessible by any other CPU, only the poking CPU can expoit
this (temporary) hole, which is a much larger ask than any of the
proposed options.

I would agree that __text_poke() is a safer option. But in this case, we 
will need the temporary hole to be 2MB in size. Also, we will probably 
hold the temporary mapping for longer time (the whole JITing process). 
Does this sound reasonable?

Actually, the hole is probably not always 2MB in size. But it could be up 
to 2MB in size. 

Song