On Tue, Apr 09, 2019 at 12:45:25PM -0400, Mathieu Desnoyers wrote: > ----- On Apr 9, 2019, at 12:40 PM, paulmck paulmck@xxxxxxxxxxxxx wrote: > > > On Tue, Apr 09, 2019 at 11:56:03AM -0400, Mathieu Desnoyers wrote: > >> ----- On Apr 9, 2019, at 11:40 AM, Joel Fernandes, Google joel@xxxxxxxxxxxxxxxxx > >> wrote: > >> > >> > On Mon, Apr 08, 2019 at 01:24:47PM -0400, Mathieu Desnoyers wrote: > >> >> ----- On Apr 8, 2019, at 11:46 AM, paulmck paulmck@xxxxxxxxxxxxx wrote: > >> >> > >> >> > On Mon, Apr 08, 2019 at 10:49:32AM -0400, Mathieu Desnoyers wrote: > >> >> >> ----- On Apr 8, 2019, at 10:22 AM, paulmck paulmck@xxxxxxxxxxxxx wrote: > >> >> >> > >> >> >> > On Mon, Apr 08, 2019 at 09:05:34AM -0400, Mathieu Desnoyers wrote: > >> >> >> >> ----- On Apr 7, 2019, at 10:27 PM, paulmck paulmck@xxxxxxxxxxxxx wrote: > >> >> >> >> > >> >> >> >> > On Sun, Apr 07, 2019 at 09:07:18PM +0000, Joel Fernandes wrote: > >> >> >> >> >> On Sun, Apr 07, 2019 at 04:41:36PM -0400, Mathieu Desnoyers wrote: > >> >> >> >> >> > > >> >> >> >> >> > ----- On Apr 7, 2019, at 3:32 PM, Joel Fernandes, Google joel@xxxxxxxxxxxxxxxxx > >> >> >> >> >> > wrote: > >> >> >> >> >> > > >> >> >> >> >> > > On Sun, Apr 07, 2019 at 03:26:16PM -0400, Mathieu Desnoyers wrote: > >> >> >> >> >> > >> ----- On Apr 7, 2019, at 9:59 AM, paulmck paulmck@xxxxxxxxxxxxx wrote: > >> >> >> >> >> > >> > >> >> >> >> >> > >> > On Sun, Apr 07, 2019 at 06:39:41AM -0700, Paul E. McKenney wrote: > >> >> >> >> >> > >> >> On Sat, Apr 06, 2019 at 07:06:13PM -0400, Joel Fernandes wrote: > >> >> >> >> >> > >> > > >> >> >> >> >> > >> > [ . . . ] > >> >> >> >> >> > >> > > >> >> >> >> >> > >> >> > > diff --git a/include/asm-generic/vmlinux.lds.h > >> >> >> >> >> > >> >> > > b/include/asm-generic/vmlinux.lds.h > >> >> >> >> >> > >> >> > > index f8f6f04c4453..c2d919a1566e 100644 > >> >> >> >> >> > >> >> > > --- a/include/asm-generic/vmlinux.lds.h > >> >> >> >> >> > >> >> > > +++ b/include/asm-generic/vmlinux.lds.h > >> >> >> >> >> > >> >> > > @@ -338,6 +338,10 @@ > >> >> >> >> >> > >> >> > > KEEP(*(__tracepoints_ptrs)) /* Tracepoints: pointer array */ \ > >> >> >> >> >> > >> >> > > __stop___tracepoints_ptrs = .; \ > >> >> >> >> >> > >> >> > > *(__tracepoints_strings)/* Tracepoints: strings */ \ > >> >> >> >> >> > >> >> > > + . = ALIGN(8); \ > >> >> >> >> >> > >> >> > > + __start___srcu_struct = .; \ > >> >> >> >> >> > >> >> > > + *(___srcu_struct_ptrs) \ > >> >> >> >> >> > >> >> > > + __end___srcu_struct = .; \ > >> >> >> >> >> > >> >> > > } \ > >> >> >> >> >> > >> >> > > >> >> >> >> >> > >> >> > This vmlinux linker modification is not needed. I tested without it and srcu > >> >> >> >> >> > >> >> > torture works fine with rcutorture built as a module. Putting further prints > >> >> >> >> >> > >> >> > in kernel/module.c verified that the kernel is able to find the srcu structs > >> >> >> >> >> > >> >> > just fine. You could squash the below patch into this one or apply it on top > >> >> >> >> >> > >> >> > of the dev branch. > >> >> >> >> >> > >> >> > >> >> >> >> >> > >> >> Good point, given that otherwise FORTRAN named common blocks would not > >> >> >> >> >> > >> >> work. > >> >> >> >> >> > >> >> > >> >> >> >> >> > >> >> But isn't one advantage of leaving that stuff in the RO_DATA_SECTION() > >> >> >> >> >> > >> >> macro that it can be mapped read-only? Or am I suffering from excessive > >> >> >> >> >> > >> >> optimism? > >> >> >> >> >> > >> > > >> >> >> >> >> > >> > And to answer the other question, in the case where I am suffering from > >> >> >> >> >> > >> > excessive optimism, it should be a separate commit. Please see below > >> >> >> >> >> > >> > for the updated original commit thus far. > >> >> >> >> >> > >> > > >> >> >> >> >> > >> > And may I have your Tested-by? > >> >> >> >> >> > >> > >> >> >> >> >> > >> Just to confirm: does the cleanup performed in the modules going > >> >> >> >> >> > >> notifier end up acting as a barrier first before freeing the memory ? > >> >> >> >> >> > >> If not, is it explicitly stated that a barrier must be issued before > >> >> >> >> >> > >> module unload ? > >> >> >> >> >> > >> > >> >> >> >> >> > > > >> >> >> >> >> > > You mean rcu_barrier? It is mentioned in the documentation that this is the > >> >> >> >> >> > > responsibility of the module writer to prevent delays for all modules. > >> >> >> >> >> > > >> >> >> >> >> > It's a srcu barrier yes. Considering it would be a barrier specific to the > >> >> >> >> >> > srcu domain within that module, I don't see how it would cause delays for > >> >> >> >> >> > "all" modules if we implicitly issue the barrier on module unload. What > >> >> >> >> >> > am I missing ? > >> >> >> >> >> > >> >> >> >> >> Yes you are right. I thought of this after I just sent my email. I think it > >> >> >> >> >> makes sense for srcu case to do and could avoid a class of bugs. > >> >> >> >> > > >> >> >> >> > If there are call_srcu() callbacks outstanding, the module writer still > >> >> >> >> > needs the srcu_barrier() because otherwise callbacks arrive after > >> >> >> >> > the module text has gone, which will be disappoint the CPU when it > >> >> >> >> > tries fetching instructions that are no longer mapped. If there are > >> >> >> >> > no call_srcu() callbacks from that module, then there is no need for > >> >> >> >> > srcu_barrier() either way. > >> >> >> >> > > >> >> >> >> > So if an srcu_barrier() is needed, the module developer needs to > >> >> >> >> > supply it. > >> >> >> >> > >> >> >> >> When you say "callbacks arrive after the module text has gone", > >> >> >> >> I think you assume that free_module() is invoked before the > >> >> >> >> MODULE_STATE_GOING notifiers are called. But it's done in the > >> >> >> >> opposite order: going notifiers are called first, and then > >> >> >> >> free_module() is invoked. > >> >> >> >> > >> >> >> >> So AFAIU it would be safe to issue the srcu_barrier() from the module > >> >> >> >> going notifier. > >> >> >> >> > >> >> >> >> Or am I missing something ? > >> >> >> > > >> >> >> > We do seem to be talking past each other. ;-) > >> >> >> > > >> >> >> > This has nothing to do with the order of events at module-unload time. > >> >> >> > > >> >> >> > So please let me try again. > >> >> >> > > >> >> >> > If a given srcu_struct in a module never has call_srcu() invoked, there > >> >> >> > is no need to invoke rcu_barrier() at any time, whether at module-unload > >> >> >> > time or not. Adding rcu_barrier() in this case adds overhead and latency > >> >> >> > for no good reason. > >> >> >> > >> >> >> Not if we invoke srcu_barrier() for that specific domain. If > >> >> >> call_srcu was never invoked for a srcu domain, I don't see why > >> >> >> srcu_barrier() should be more expensive than a simple check that > >> >> >> the domain does not have any srcu work queued. > >> >> > > >> >> > But that simple check does involve a cache miss for each possible CPU (not > >> >> > just each online CPU), so it is non-trivial, especially on large systems. > >> >> > > >> >> >> > If a given srcu_struct in a module does have at least one call_srcu() > >> >> >> > invoked, it is already that module's responsibility to make sure that > >> >> >> > the code sticks around long enough for the callback to be invoked. > >> >> >> > >> >> >> I understand that when users do explicit dynamic allocation/cleanup of > >> >> >> srcu domains, they indeed need to take care of doing explicit srcu_barrier(). > >> >> >> However, if they do static definition of srcu domains, it would be nice > >> >> >> if we can handle the barriers under the hood. > >> >> > > >> >> > All else being equal, of course. But... > >> >> > > >> >> >> > This means that correct SRCU users that invoke call_srcu() already > >> >> >> > have srcu_barrier() at module-unload time. Incorrect SRCU users, with > >> >> >> > reasonable probability, now get a WARN_ON() at module-unload time, with > >> >> >> > the per-CPU state getting leaked. Before this change, they would (also > >> >> >> > with reasonable probability) instead get an instruction-fetch fault when > >> >> >> > the SRCU callback was invoked after the completion of the module unload. > >> >> >> > Furthermore, in all cases where they would previously have gotten the > >> >> >> > instruction-fetch fault, they now get the WARN_ON(), like this: > >> >> >> > > >> >> >> > if (WARN_ON(rcu_segcblist_n_cbs(&sdp->srcu_cblist))) > >> >> >> > return; /* Forgot srcu_barrier(), so just leak it! */ > >> >> >> > > >> >> >> > So this change already represents an improvement in usability. > >> >> >> > >> >> >> Considering that we can do a srcu_barrier() for the specific domain, > >> >> >> and that it should add no noticeable overhead if there is no queued > >> >> >> callbacks, I don't see a good reason for leaving the srcu_barrier > >> >> >> invocation to the user rather than implicitly doing it from the > >> >> >> module going notifier. > >> >> > > >> >> > Now, I could automatically add an indicator of whether or not a > >> >> > call_srcu() had happened, but then again, that would either add a > >> >> > call_srcu() scalability bottleneck or again require a scan of all possible > >> >> > CPUs... to figure out if it was necessary to scan all possible CPUs. > >> >> > > >> >> > Or is scanning all possible CPUs down in the noise in this case? Or > >> >> > am I missing a trick that would reduce the overhead? > >> >> > >> >> Module unloading implicitly does a synchronize_rcu (for RCU-sched), and > >> >> a stop_machine. So I would be tempted to say that overhead of iteration > >> >> over all CPUs might not matter that much considering the rest. > >> >> > >> >> About notifying that a call_srcu has happened for the srcu domain in a > >> >> scalable fashion, let's see... We could have a flag "call_srcu_used" > >> >> for each call_srcu domain. Whenever call_srcu is invoked, it would > >> >> load that flag. It sets it on first use. > >> >> > >> >> The idea here is to only use that flag when srcu_barrier is performed > >> >> right before the srcu domain cleanup (it could become part of that > >> >> cleanup). Else, using it in all srcu_barrier() might be tricky, because > >> >> we may then need to add memory barriers or locking to the call_srcu > >> >> fast-path, which is an overhead we try to avoid. > >> >> > >> >> However, if we only use that flag as part of the srcu domain cleanup, > >> >> it's already prohibited to invoke call_srcu concurrently with the > >> >> cleanup of the same domain, so I don't think we would need any > >> >> memory barriers in call_srcu. > >> > > >> > About the last part of your email, it seems to that if after call_srcu has > >> > returned, if the module could be unloaded on some other CPU - then it would > >> > need to see the flag stored by the preceding call_srcu, so I believe there > >> > would be a memory barrier between the two opreations (call_srcu and module > >> > unload). > >> > >> In order for the module unload not to race against module execution, it needs > >> to happen after the call_srcu in a way that is already ordered by other means, > >> else module unload races against the module code. > >> > >> > > >> > Also about doing the unconditional srcu_barrier, since a module could be > >> > unloaded at any time - don't all SRCU using modules need to invoke > >> > srcu_barrier() during their clean up anyway so we are incurring the barrier > >> > overhead anyway? Or, am I missing a design pattern here? It seems to me > >> > rcutorture module definitely calls srcu_barrier() before it is unloaded. > >> > >> I think a valid approach which is even simpler might be: if a module statically > >> defines a SRCU domain, it should be expected to use it. So adding a > >> srcu_barrier() > >> to its module going notifier should not hurt. The rare case where a module > >> defines > >> a static SRCU domain *and* does not actually use it with call_srcu() does not > >> seem that usual, and not worth optimizing for. > >> > >> Thoughts ? > > > > Most SRCU users use only synchronize_srcu(), and don't ever use > > call_srcu(). Which is not too surprising given that call_srcu() showed > > up late in the game. > > > > But something still bothers me about this, and I am not yet sure > > what. One thing that seems to reduce anxiety somewhat is doing the > > srcu_barrier() on all calls to cleanup_srcu_struct() rather than just > > those invoked from the modules infrastructure, but I don't see why at > > the moment. > > Indeed, providing similar guarantees for the dynamic allocation case > would be nice. > > The one thing that is making me anxious here is use-cases where > users would decide to chain their call_srcu(). Then they would > need as many srcu_barrier() as chain hops. This would be a valid > reason for leaving invocation of srcu_barrier() to the user and > not hide it under the hood. > > Thoughts ? The current state is not horrible, so my thought would be to give it some time to see if better thoughts arise. Either way, cleanup_srcu_struct() keeps its current checks for callbacks still being in flight, which is why I believe that the current state is not horrible. ;-) Thanx, Paul _______________________________________________ amd-gfx mailing list amd-gfx@xxxxxxxxxxxxxxxxxxxxx https://lists.freedesktop.org/mailman/listinfo/amd-gfx
