On Wed, Apr 28, 2021 at 08:13:53AM -0700, Andy Lutomirski wrote: > On Wed, Apr 28, 2021 at 8:05 AM Michel Lespinasse <michel@xxxxxxxxxxxxxx> wrote: > > > > On Wed, Apr 07, 2021 at 08:36:01AM -0700, Andy Lutomirski wrote: > > > On 4/6/21 6:44 PM, Michel Lespinasse wrote: > > > > The page table tree is walked with local irqs disabled, which prevents > > > > page table reclamation (similarly to what fast GUP does). The logic is > > > > otherwise similar to the non-speculative path, but with additional > > > > restrictions: in the speculative path, we do not handle huge pages or > > > > wiring new pages tables. > > > > > > Not on most architectures. Quoting the actual comment in mm/gup.c: > > > > > > > * Before activating this code, please be aware that the following assumptions > > > > * are currently made: > > > > * > > > > * *) Either MMU_GATHER_RCU_TABLE_FREE is enabled, and tlb_remove_table() is used to > > > > * free pages containing page tables or TLB flushing requires IPI broadcast. > > > > > > On MMU_GATHER_RCU_TABLE_FREE architectures, you cannot make the > > > assumption that it is safe to dereference a pointer in a page table just > > > because irqs are off. You need RCU protection, too. > > > > > > You have the same error in the cover letter. > > > > Hi Andy, > > > > Thanks for your comment. At first I thought did not matter, because we > > only enable ARCH_SUPPORTS_SPECULATIVE_PAGE_FAULT on selected > > architectures, and I thought MMU_GATHER_RCU_TABLE_FREE is not set on > > these. But I was wrong - MMU_GATHER_RCU_TABLE_FREE is enabled on X86 > > with paravirt. So I took another look at fast GUP to make sure I > > actually understand it. > > > > This brings a question about lockless_pages_from_mm() - I see it > > disabling interrupts, which it explains is necessary for disabling THP > > splitting IPIs, but I do not see it taking an RCU read lock as would > > be necessary for preventing paga table freeing on > > MMU_GATHER_RCU_TABLE_FREE configs. I figure local_irq_save() > > indirectly takes an rcu read lock somehow ? I think this is something > > I should also mention in my explanation, and I have not seen a good > > description of this on the fast GUP side... > > Sounds like a bug! That being said, based on my extremely limited > understanding of how the common RCU modes work, local_irq_save() > probably implies an RCU lock in at least some cases. Hi Paul! In modern kernels, local_irq_save() does have RCU reader semantics, meaning that synchronize_rcu() will wait for pre-exiting irq-disabled regions. It will also wait for pre-existing bh-disable, preempt-disable, and of course rcu_read_lock() sections of code. But don't try this in older kernels, that is not in kernel in which synchronize_sched() is defined! Thanx, Paul
