On Wed, Dec 15, 2021 at 2:25 PM Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote: > > On Wed, Dec 15, 2021 at 11:49:51AM -0800, Peter Oskolkov wrote: > > > TL;DR: our models are different here. In your model a single server > > can have a bunch of workers interacting with it; in my model only a > > single RUNNING worker is assigned to a server, which it wakes when it > > blocks. > > So part of the problem is that none of that was evident from the code. > It is also completely different from the scheduler code it lives in, > making it double confusing. > > After having read the code, I still had no clue what so ever how it was > supposed to be used. Which is where my reverse engineering started :/ I posted a doc patch: https://lore.kernel.org/lkml/20211122211327.5931-6-posk@xxxxxxxxxx/ a lib patch with userspace code: https://lore.kernel.org/lkml/20211122211327.5931-5-posk@xxxxxxxxxx/ and a doc patch for the lib/userspace code: https://lore.kernel.org/lkml/20211122211327.5931-7-posk@xxxxxxxxxx/ I spent at least two weeks polishing the lib patch and the docs, much more if previous patchsets are to be taken into account. Yes, they are confusing, and most likely answer all of the wrong questions, but I did try to make my approach as clear as possible... I apologize if that was not very successful... > > > More details: > > > > "Working servers" cannot get wakeups, because a "working server" has a > > single RUNNING worker attached to it. When a worker blocks, it wakes > > its attached server and becomes a detached blocked worker (same is > > true if the worker is "preempted": it blocks and wakes its assigned > > server). > > But who would do the preemption if the server isn't allowed to run? > > > Blocked workers upon wakeup do this, in order: > > > > - always add themselves to the runnable worker list (the list is > > shared among ALL servers, it is NOT per server); > > That seems like a scalability issue. And, as said, it is completely > alien when compared to the way Linux itself does scheduling. > > > - wake a server pointed to by idle_server_ptr, if not NULL; > > - sleep, waiting for a wakeup from a server; > > > > Server S, upon becoming IDLE (no worker to run, or woken on idle > > server list) does this, in order, in userspace (simplified, see > > umcg_get_idle_worker() in > > https://lore.kernel.org/lkml/20211122211327.5931-5-posk@xxxxxxxxxx/): > > - take a userspace (spin) lock (so the steps below are all within a > > single critical section): > > Don't ever suggest userspace spinlocks, they're horrible crap. This can easily be a mutex, not really important (although for very short critical sections with only memory reads/writes, like here, spin locks often perform better, in our experience). > > > - compare_xchg(idle_server_ptr, NULL, S); > > - if failed, there is another server in idle_server_ptr, so S adds > > itself to the userspace idle server list, releases the lock, goes to > > sleep; > > - if succeeded: > > - check the runnable worker list; > > - if empty, release the lock, sleep; > > - if not empty: > > - get the list > > - xchg(idle_server_ptr, NULL) (either S removes itself, or > > a worker in the kernel does it first, does not matter); > > - release the lock; > > - wake server S1 on idle server list. S1 goes through all > > of these steps. > > > > The protocol above serializes the userspace dealing with the idle > > server ptr/list. Wakeups in the kernel will be caught if there are > > idle servers. Yes, the protocol in the userspace is complicated (more > > complicated than outlined above, as the reaped idle/runnable worker > > list from the kernel is added to the userspace idle/runnable worker > > list), but the kernel side is very simple. I've tested this > > interaction extensively, I'm reasonably sure that no worker wakeups > > are lost. > > Sure, but also seems somewhat congestion prone :/ The whole critical section under the loc is just several memory read/write operations, so very short. And workers are removed from the kernel's list of runnable/woken workers all at once; and the server processing the runnable worker list knows how many of them are now available to run, so the appropriate number of idle servers can be woken (not yet implemented in my lib patch). So yes, this can be a bottleneck, but there are ways to make it less and less likely (by making the userspace more complicated; but this is not a concern).
