Messages in this thread | | | Subject | Re: [PATCH 07/13] powerpc: Preemptible mmu_gather | From | Benjamin Herrenschmidt <> | Date | Tue, 13 Apr 2010 12:06:40 +1000 |
| |
On Fri, 2010-04-09 at 10:14 +0200, Peter Zijlstra wrote: > > and doing that vaddr collection right along with it in the same batch. > > I think that that would work, Ben, Dave?
Well, ours aren't struct pages.
IE. There's fundamentally 3 things that we are trying to batch here :-)
1- The original mmu_gather: batching the freeing of the actual user pages, so that the TLB flush can be delayed/gathered, plus there might be some micro-improvement in passing the page list to the allocator for freeing all at omce. This is thus purely a batch of struct pages.
2- The batching of the TLB flushes (or hash invalidates in the ppc case) proper, which needs the addition of the vaddr for things like sparc and powerpc since we don't just invalidate the whole bloody thing unlike x86 :-) On powerpc, we actually need more, we need the actual PTE content since it also contains tracking information relative to where things have been put in the hash table.
3- The batching of the freeing of the page table structure, which we want to delay more than batch, ie, the goal here is to delay that freeing using RCU until everybody has stopped walking them. This does rely on RCU grace period being "interrupt safe", ie, there's no rcu_read_lock() in the low level TLB or hash miss code, but that code runs with interrupts off.
Now, 2. has a problem I described earlier, which is that we must not have the possibility of introducing a duplicate in the hash table, thus it must not be possible to put a new PTE in until the previous one has been flushed or bad things would happen. This is why powerpc doesn't use the mmu_gather the way it was originally intended to do both 1. and 2. but really only for 1., while for 2. we use a small batch that only exist between lazy_mmu_enter/exit, since those are always fully enclosed by a pte lock section.
3. As you have noticed, relies on the irq stuff. Plus there seem to be a dubious optimization here with mm_users. Might be worth sorting that out. However, it's a very different goal than 1. and 2. in the sense that batching proper is a minor issue, what we want is synchronization with walkers, and that batching is a way to lower the cost of that synchronization (allocating of the RCU struct etc...).
Cheers, Ben.
| |