Make print() builtin synchronous

[ajor]

print(@map) is designed to print all entries in a map. It currently does this by passing a MapID reference to userspace, where bpftrace iterates over the map entries to print them out.

The problem is that this makes print(@map) an asynchronous operation, which can produce unexpected results:

# bpftrace -e 'BEGIN { @x[0] = 1; print(@x); @x[0] = 2; exit() }'
Attaching 1 probe...
@x[0]: 2

Now that BPF has support for iterating over map elements in the kernel, we will be able to reimplement print(@map) as a synchronous operation by writing all the map keys+values to the ring buffer instead of a single MapID reference.

I think this could be implemented with minimal new codegen once #3003 is merged - by rewriting the AST of any print(@map) call into for ($kv : @map) { print($kv) }. Some work will be needed to ensure the entries are still formatted and sorted correctly.

The previous script should produce a more expected result in the future:

# bpftrace -e 'BEGIN { @x[0] = 1; print(@x); @x[0] = 2; exit() }'
Attaching 1 probe...
@x[0]: 1

This also applies to some other builtin functions which operate on maps asynchronously: clear(), zero()

[danobi]

While it's true changing from async to sync is backwards compatible, I'm a little worried about large maps overflowing the perg/ring buffer. In theory the user could be relying on the ability to print a huge map.

We could always add some way for the user to opt in, but I also dislike these kinds of knobs in general.

Seems like a tradeoff and I don't know the right answer atm.

[danobi]

Tangentially related: rust and golang have a edition/version mechanisms for these kinds of subtle changes. Basically you tell the compiler which edition/version you want to build against, and compiler will maintain semantics of that version. Newer code, when you cargo new or go mod init will target the latest version.

A mechanism like that could help us here. But it's probably overkill. Something to think about though.

[danobi]

(Sorry for stream of consciousness)

Maybe we could add an aprint() alias for the old behavior? And print() gets sync by default. I imagine this is beneficial for nearly everyone while also giving pathological users a way out.

[ajor]

Maybe we could add an aprint() alias for the old behavior? And print() gets sync by default. I imagine this is beneficial for nearly everyone while also giving pathological users a way out.

In general, I'm more in favour of adding new syntax instead of new bulitin functions as I think it tends to be more flexible. What about introducing a way to take references to maps and allowing them to be printed, e.g.:

$map_ref = &@mymap;
print($map_ref);

It could also be useful in other scenarios, e.g. for double buffering:

$map_ref = $cond ? &@map1 : &@map2;
$map_ref[blah] = foo;