On Wed, Jul 21, 2021 at 12:56 PM Thierry Delisle <tdelisle@xxxxxxxxxxxx> wrote: > > > Yes, this is naturally supported in the current patchset on the kernel > > side, and is supported in libumcg (to be posted, later when the kernel > > side is settled); internally at Google, some applications use > > different "groups" of workers/servers per NUMA node. > > Good to know. Cforall has the same feature, where we refer to these groups > as "clusters". https://doi.org/10.1002/spe.2925 (Section 7) > > > Please see the attached atomic_stack.h file - I use it in my tests, > > things seem to be working. Specifically, atomic_stack_gc does the > > cleanup. For the kernel side of things, see the third patch in this > > patchset. > > I don't believe the atomic_stack_gc function is robust enough to be > offering > any guarantee. I believe that once a node is unlinked, its next pointer > should > be reset immediately, e.g., by writing 0xDEADDEADDEADDEAD. Do your tests > work > if the next pointer is reset immediately on reclaimed nodes? In my tests reclaimed nodes have their next pointers immediately set to point to the list head. If the kernel gets a node with its @next pointing to something else, then yes, things break down (the kernel kills the process); this has happened occasionally when I had a bug in the userspace code. > > As far as I can tell, the reclaimed nodes in atomic_stack_gc still contain > valid next fields. I believe there is a race which can lead to the kernel > reading reclaimed nodes. If atomic_stack_gc does not reset the fields, > this bug > could be hidden in the testing. Could you, please, provide a bit more details on when/how the race can happen? Servers add themselves to the list, so there can be no races there (servers going idle: add-to-the-list; wait; gc (under a lock); restore @next; do stuff). Workers are trickier, as they can be woken by signals and then block again, but stray signals are so bad here that I'm thinking of actually not letting sleeping workers wake on signals. Other than signals waking queued/unqueued idle workers, are there any other potential races here? > > An more aggressive test is to put each node in a different page and > remove read > permissions when the node is reclaimed. I'm not sure this applies when the > kernel is the one reading. > > > > To keep the kernel side light and simple. To also protect the kernel > > from spinning if userspace misbehaves. Basically, the overall approach > > is to delegate most of the work to the userspace, and keep the bare > > minimum in the kernel. > > I'll try to keep this in mind then. > > > After some thought, I'll suggest a scheme to significantly reduce > complexity. > As I understand, the idle_workers_ptr are linked to form one or more > Multi-Producer Single-Consumer queues. If each head is augmented with a > single > volatile tid-sized word, servers that want to go idle can simply write > their id > in the word. When the kernel adds something to the idle_workers_ptr > list, it > simply does an XCHG with 0 or any INVALID_TID. This scheme only supports > one > server blocking per idle_workers_ptr list. To keep the "kernel side > light and > simple", you can simply ask that any extra servers must synchronize > among each > other to pick which server is responsible for wait on behalf of everyone. > I'm not yet convinced that the single-linked-list approach is infeasible. And if it is, a simple fix would be to have two pointers per list in struct umcg_task: head and next. Thanks, Peter
