Alexei and Daniel, I just wanted to let you guys know that I'm working
on an alignment tracker in the BPF verifier.
After trying several approaches I think what is going to work is to
maintain state like this:
1) For non-pointer registers, we record what we can prove is the
minimum alignment of the value held in the register.
So for example:
r5 <<= 2
would result in a min_align value of '4'.
These alignment values assist us when check_packet_ptr_add() has to
transition a pointer register and allocate an ID to it.
2) Packet pointer registers have a base alignment (which is something
relative to NET_IP_ALIGN).
Then there is something called an auxiliary offset alignment.
Any time we add some non-constant value to a pointer, we apply the
value's min alignment to the pointer register's auxiliary offset
alignment.
Then check_pkt_ptr_alignment has it's logic adjusted such that it
takes all of this new information into account.
First, it makes the existing test:
if ((NET_IP_ALIGN + reg->off + off) % size != 0) {
except that NET_IP_ALIGN is replaced with the packet pointer base
alignment (which we'll set in the context load helpers, thus putting
the NET_IP_ALIGN detail back into the networking code).
So that turns into something like:
if ((reg->ptr_base_align + reg->off + off) % size != 0) {
Next, if an ID has been assigned, we have to also check the auxiliary
alignment:
if (reg->id && (reg->aux_off_align % size) != 0) {
Otherwise, we can prove that the size access will work.
I think in order for this to work properly, we also have to stop
"forgetting" the reg->off value when we assign an ID to a pointer
register. However, the reg->range we still have to always kill in
this situation.
Anyways, I'll play with this design and see what happens... Feedback
is of course welcome.
