On Tue, Sep 15, 2020 at 10:50:40AM -0400, Peter Xu wrote: > On Mon, Sep 14, 2020 at 08:28:51PM -0300, Jason Gunthorpe wrote: > > Yes, this stuff does pin_user_pages_fast() and MADV_DONTFORK > > together. It sets FOLL_FORCE and FOLL_WRITE to get an exclusive copy > > of the page and MADV_DONTFORK was needed to ensure that a future fork > > doesn't establish a COW that would break the DMA by moving the > > physical page over to the fork. DMA should stay with the process that > > called pin_user_pages_fast() (Is MADV_DONTFORK still needed with > > recent years work to GUP/etc? It is a pretty terrible ancient thing) > > ... Now I'm more confused on what has happened.
I'm going to try to confirm that the MADV_DONTFORK is actually being done by userspace properly, more later. > It means, as long as the rdma region has VM_WRITE set (which I think of no > reason on why it shouldn't...), then it should have the write bit in the COWed > page entry. If so, the page should be stable and I don't undersdand why > another COW could even trigger and how the code path in the "trial cow" patch > is triggered. All the regions the test are doing DMA to will be simple process writable anonymous VMA's from malloc() > Or, the VMA is without VM_WRITE due to some reason? Sorry I probably know > nothing about RDMA, more information on that side might help too. E.g., is the > hardware going to walk the software process page table too when doing RDMA (or > is IOMMU page table used, or none)? It does pin_user_pages_fast(), gets a list of DMA addresses for the pages and then programs the hardware. The pin remains for a very long time and the HW does DMA to those pages independently. Userspace will write to the memory and trigger DMA reads and HW will do DMA writes and trigger something close to an eventfd to let userspace know to check the DMA'd data. Very similar to how an in-kernel driver works. It is similar to VFIO in how it uses pin_user_pages_fast(). Symptoms look to be like the DMA's are not arriving. As before, the requirement is that once a process as done pin_user_pages() the physical page stays with the process. If there is a fork() and a COW then the current memory stays with the original process and the fork'd child gets the copy. MADV_DONTFORK is expected to ensure this.. We haven't been able to narrow to a reproduction that doesn't require alot of hardware unfortunately. It seems oddly sensitive, maybe due to memory layout triggering a COW.. Jason