Skip to contentSkip to conversation

Why does tail call only unwind the current stack frame instead of resetting the current stack?

DISCUSSION Hsuan-Chi Kuo 3 participants 4 messages Sep 12, 2021 – Sep 13, 2021 BPF

lore
Share & tools
lore
#1 lore raw
Quoted 7 lines · Hsuan-Chi Kuo

Hi,

The function check_max_stack_depth (https://elixir.bootlin.com/linux/latest/source/kernel/bpf/verifier.c#L3574 ) is used to ensure the stack size is no greater than 8KB.

The stack can only grow over 8KB because of tail calls as they only unwind the current stack frame. I wonder why not make tail call reset the stack since what was on the stack

will never be used again?

I think this is just an artifact of current implementation. To do a correct reset of the stack, additional instructions are needed to keep track and accumulate stack size properly, so when a tail call is reached, it knows how much stack size to reset.

Did you hit this limitation? Could you describe your use case in detail?

Quoted 4 lines

Thanks

#2 lore raw

No, I didn't hit this limit.

The current implementation already keeps tracks of the number of tail calls which to me is the same effort of tracking the stack size. I was just wondering if there's any fundamental reason that you can't reset the stack directly. But it seems that there is not.

Thanks

Quoted 24 lines · Yonghong Song
Quoted 7 lines · Hsuan-Chi Kuo

Hi,

The function check_max_stack_depth (https://elixir.bootlin.com/linux/latest/source/kernel/bpf/verifier.c#L3574 ) is used to ensure the stack size is no greater than 8KB.

The stack can only grow over 8KB because of tail calls as they only unwind the current stack frame. I wonder why not make tail call reset the stack since what was on the stack

will never be used again?

I think this is just an artifact of current implementation. To do a correct reset of the stack, additional instructions are needed to keep track and accumulate stack size properly, so when a tail call is reached, it knows how much stack size to reset.

Did you hit this limitation? Could you describe your use case in detail?

Quoted 4 lines

Thanks

#3 lore raw
Quoted 7 lines · Hsuan-Chi Kuo

No, I didn't hit this limit.

The current implementation already keeps tracks of the number of tail calls which to me is the same effort of tracking the stack size. I was just wondering if there's any fundamental reason that you can't reset the stack directly. But it seems that there is not.

The tail_call unwinds the current stack frame.
See comment at line 3585 for details.

And please don't top post.