On Wed, Nov 20, 2024 at 04:51:36PM -0800, Alexei Starovoitov wrote: > On Tue, Nov 19, 2024 at 4:22 AM Matt Bobrowski <mattbobrowski@xxxxxxxxxx> wrote: > > > > Hi, > > > > Currently, we have BPF kfuncs which allow BPF programs to add and > > remove elements from a BPF linked list. However, we're currently > > missing other simple capabilities, like being able to iterate over the > > elements within the BPF linked lists. What is our current appetite > > with regards to adding new BPF kfuncs that support this kind of > > capability to BPF linked lists? > > What kind of kfuncs do you have in mind for link lists ? At this point, it'd have to be some kind of iterator based BPF kfunc that allows a BPF program to traverse over the supplied BPF linked list, coupled with a delete based BPF kfunc such that elements from the list can be removed whilst performing the iteration i.e. list_for_each_safe/list_del. Now that I know you're not completely opposed to adding such BPF kfuncs, I can concretely start thinking about what this will actually end up looking like. But essentially, it'd need to be BPF kfuncs that support those 2 previously mentioned capabilities, being traversal and arbitrary removal of an element whilst performing the traversal. > So far the only user of bpf_rbtree is sched-ext. > It was used in one scheduler and the experience was painful. > There is a chance that we will remove rbtree and link list > support from the verifier to reduce complexity. > So new link list kfuncs may be ok, but potentially not for too long. Noted. > > I know that we're now somewhat advocating for using BPF arenas > > whenever and wherever possible, especially when it comes to building > > out and supporting more complicated data structures in BPF. However, > > IMO BPF linked lists still have their place. Specifically, and as of > > now, I'd argue that the BPF linked list implementation could be > > considered more memory efficient when compared to a BPF arena backed > > linked list implementation. This is purely due to the fact that the > > BPF linked list implementation can perform more constrained memory > > allocations for elements via bpf_obj_new_impl() based on the demand, > > whereas for a BPF arena based implementation a BPF program needs to > > allocate memory upfront in terms of the number of pages (modulo the > > fact that not all pages for the BPF arena will necessarily be reserved > > upfront). The fact that allocations are performed in terms of > > multiples of PAGE_SIZE can lead to unnecessary memory wastage. > > I don't follow this logic. > bpf_mem_alloc is relying on slab that relies on page alloc. > So either arena or bpf_ma allocates a page at a time. > So from that pov the cost is the same. Oh, what? So, both are actually performing full page sized allocations whenever there's a need to fetch more memory? My shallow understanding at this point was that the BPF specific memory allocator simply acts as a front-end cache to kmalloc(), and depending on the size of your allocation request i.e. via bpf_obj_new_impl for example, depends on what freelist that allocation is fulfilled from. Any needs for refilling a freelist due to exhaustion pressure are performed at freelist size granularity i.e. 16, 32, 64, 128, 256..., 4096. So, based on my current understanding and assuming that your BPF program at most requests 128 byte sized allocations via bpf_obj_new_impl, the BPF specific memory allocator will via kmalloc() only allocate that much more memory at a time when needing to refill the freelist caches (maybe it'd be a multiple of that because of batching or something, but I don't fully understand the details of that just yet). In comparison to using something like a BPF arena, if I had to build out my own in-BPF program memory allocator that was backed by a BPF arena, in the need of a refill like situation, the allocation would be performed at page size granularity as that is the amount of memory that the BPF arena pulls in at a time. This is what I was basically trying to get at, but my trail of thought may be flawed by my lack of understanding on how something actually works in practice. > But in practice bpf_ma needs extra 8 bytes for every allocation > whereas arena allocs don't have this overhead. Makes sense, freelist objects need to be tracked somehow by the BPF specific memory allocator... > Right now arena allocs need to be sleepable and that is > a severe limitation for tracing use cases. > We're working on lifting that. Once that happens > allocs in arena will be more usable than bpf_ma. Cool, sonuds good. I'm following this closely too BTW, because at some point I think we'll end up using BPF arenas for a bunch of our stuff. > kptr-s in arena is another required feature. > There were few proposals. So it will be done as well. Eventually. > So new link list kfuncs are ok, but might get removed in the future. Let me think about this a little more and send out an RFC patch series when time permits.
