On Fri, Apr 8, 2022 at 6:12 PM Alexei Starovoitov <alexei.starovoitov@xxxxxxxxx> wrote: > > On Fri, Apr 08, 2022 at 04:37:02PM -0700, Joanne Koong wrote: > > > > > > + * > > > > > > + * void bpf_free(struct bpf_dynptr *ptr) > > > > > > > > > > thinking about the next patch set that will add storing this malloc > > > > > dynptr into the map, bpf_free() will be a lie, right? As it will only > > > > > decrement a refcnt, not necessarily free it, right? So maybe just > > > > > generic bpf_dynptr_put() or bpf_malloc_put() or something like that is > > > > > a bit more "truthful"? > > > > I like the simplicity of bpf_free(), but I can see how that might be > > > > confusing. What are your thoughts on "bpf_dynptr_free()"? Since when > > > > we get into dynptrs that are stored in maps vs. dynptrs stored > > > > locally, calling bpf_dynptr_free() frees (invalidates) your local > > > > dynptr even if it doesn't free the underlying memory if it still has > > > > valid refcounts on it? To me, "malloc" and "_free" go more intuitively > > > > together as a pair. > > > > > > Sounds good to me (though let's use _dynptr() as a suffix > > > consistently). I also just realized that maybe we should call > > > bpf_malloc() a bpf_malloc_dynptr() instead. I can see how we might > > > want to enable plain bpf_malloc() with statically known size (similar > > > to statically known bpf_ringbuf_reserve()) for temporary local malloc > > > with direct memory access? So bpf_malloc_dynptr() would be a > > > dynptr-enabled counterpart to fixed-sized bpf_malloc()? And then > > > bpf_free() will work with direct pointer returned from bpf_malloc(), > > > while bpf_free_dynptr() will work with dynptr returned from > > > bpf_malloc_dynptr(). > > I see! What is the advantage of a plain bpf_malloc()? Is it that it's > > a more ergonomic API (you get back a direct pointer to the data > > instead of getting back a dynptr and then having to call > > bpf_dynptr_data to get direct access) and you don't have to allocate > > extra bytes for refcounting? > > > > I will rename this to bpf_malloc_dynptr() and bpf_free_dynptr(). > > Let's make it consistent with kptr. Those helpers will be: > bpf_kptr_alloc(btf_id, flags, &ptr) > bpf_kptr_get > bpf_kptr_put > > bpf_dynptr_alloc(byte_size, flags, &dynptr); I don't have strong feelings about this naming, but bpf_ringbuf_reserve_dynptr() is a bit of counter-example with a convention of using "_dynptr" suffix for variations of API that *produce* dynptrs as an output. bpf_dynptr_alloc() sounds like we are allocating struct bpf_dynptr itself, not a memory to which bpf_dynptr points. But I'm don't have perfect naming scheme. > bpf_dynptr_put(dynptr); > would fit the best. > > Output arg being first doesn't match anything we had. > let's keep it last. yep, agree > > zero-alloc or plain kmalloc can be indicated by the flag. > kzalloc() in the kernel is just static inline that adds __GFP_ZERO to flags. > We don't need bpf_dynptr_alloc and bpf_dynptr_zalloc as two helpers. > The latter can be a static inline helper in a bpf program. yeah, sure, my point was that zero-initialization is a better default > > Similar to Andrii's concern I feel that bpf_dynptr_free() would be misleading.
