Andrii Nakryiko <andrii.nakryiko@xxxxxxxxx> writes: > On Sun, Oct 27, 2019 at 5:04 AM Toke Høiland-Jørgensen <toke@xxxxxxxxxx> wrote: >> >> Andrii Nakryiko <andrii.nakryiko@xxxxxxxxx> writes: >> >> > On Thu, Oct 24, 2019 at 6:11 AM Toke Høiland-Jørgensen <toke@xxxxxxxxxx> wrote: >> >> >> >> From: Toke Høiland-Jørgensen <toke@xxxxxxxxxx> >> >> >> >> With the functions added in previous commits that can automatically pin >> >> maps based on their 'pinning' setting, we can support auto-pinning of maps >> >> by the simple setting of an option to bpf_object__open. >> >> >> >> Since auto-pinning only does something if any maps actually have a >> >> 'pinning' BTF attribute set, we default the new option to enabled, on the >> >> assumption that seamless pinning is what most callers want. >> >> >> >> When a map has a pin_path set at load time, libbpf will compare the map >> >> pinned at that location (if any), and if the attributes match, will re-use >> >> that map instead of creating a new one. If no existing map is found, the >> >> newly created map will instead be pinned at the location. >> >> >> >> Programs wanting to customise the pinning can override the pinning paths >> >> using bpf_map__set_pin_path() before calling bpf_object__load(). >> >> >> >> Signed-off-by: Toke Høiland-Jørgensen <toke@xxxxxxxxxx> >> >> --- >> > >> > How have you tested this? From reading the code, all the maps will be >> > pinned irregardless of their .pinning setting? >> >> No, build_pin_path() checks map->pinning :) > > subtle... build_pin_path() definitely doesn't imply that it's a "maybe > build pin path?", but see below for pin_path setting. > >> >> > Please add proper tests to test_progs, testing various modes and >> > overrides. >> >> Can do. >> >> > You keep trying to add more and more knobs :) Please stop doing that, >> > even if we have a good mechanism for extensibility, it doesn't mean we >> > need to increase a proliferation of options. >> >> But I like options! ;) >> >> > Each option has to be tested. In current version of your patches, you >> > have something like 4 or 5 different knobs, do you really want to >> > write tests testing each of them? ;) >> >> Heh, I guess I can cut down the number of options to the number of tests :P >> >> > Another high-level feedback. I think having separate passes over all >> > maps (build_map_pin_paths, reuse, then we already have create_maps) is >> > actually making everything more verbose and harder to extend. I'm >> > thinking about all these as sub-steps of map creation. Can you please >> > try refactoring so all these steps are happening per each map in one >> > place: if map needs to be pinned, check if it can be reused, if not - >> > create it. This actually will allow to handle races better, because >> > you will be able to retry easily, while if it's all spread in >> > independent passes, it becomes much harder. Please consider that. >> >> We'll need at least two passes: set pin_path on open, and check reuse / >> create / pin on load. Don't have any objections to consolidating the >> other passes into create_maps; will fix, along with your comments below. > > for BTF-defined maps, can't we just set a pin_path right when we are > reading map definition? With that, we won't even need to store > .pinning field. Would that work? We could, and I did actually try that. However, I think it is more readable to have it be a separate step: init_user_btf_maps() parses the map def, and after that is done we loop over things to build the pin paths. I'll send a v3 in a bit, you can see for yourself :) -Toke
