Hi Roland, I understand that you were reluctant to review these patches as long as there was an ongoing debate on whether or not the i_mmap_mutex should be changed into a spinlock. It seems that the debate concluded with the decision to change it into a rwsem [1], as apparently this provides the optimal performance with the new optimistic spinning patch [2]. I believe this means that there will be no problem adding paging support to the RDMA stack that depends on sleepable MMU notifiers. Changes from V0: http://marc.info/?l=linux-rdma&m=139375790322547&w=2 - Rebased against latest upstream / for-next branch. - Removed dependency on patches that were accepted upstream. - Removed pre-patches that were accepted upstream [3]. - Add extended uverb call for querying device (patch 1) and use kernel device attributes to report ODP capabilities through the new uverb entry instead of having a special verb. - Allow upgrading page access permissions during page faults. - Minor fixes to issues that came up during regression testing of the patches. The following set of patches implements on-demand paging (ODP) support in the RDMA stack and in the mlx5_ib Infiniband driver. What is on-demand paging? Applications register memory with an RDMA adapter using system calls, and subsequently post IO operations that refer to the corresponding virtual addresses directly to HW. Until now, this was achieved by pinning the memory during the registration calls. The goal of on demand paging is to avoid pinning the pages of registered memory regions (MRs). This will allow users the same flexibility they get when swapping any other part of their processes address spaces. Instead of requiring the entire MR to fit in physical memory, we can allow the MR to be larger, and only fit the current working set in physical memory. This can make programming with RDMA much simpler. Today, developers that are working with more data than their RAM can hold need either to deregister and reregister memory regions throughout their process's life, or keep a single memory region and copy the data to it. On demand paging will allow these developers to register a single MR at the beginning of their process's life, and let the operating system manage which pages needs to be fetched at a given time. In the future, we might be able to provide a single memory access key for each process that would provide the entire process's address as one large memory region, and the developers wouldn't need to register memory regions at all. How does page faults generally work? With pinned memory regions, the driver would map the virtual addresses to bus addresses, and pass these addresses to the HCA to associate them with the new MR. With ODP, the driver is now allowed to mark some of the pages in the MR as not-present. When the HCA attempts to perform memory access for a communication operation, it notices the page is not present, and raises a page fault event to the driver. In addition, the HCA performs whatever operation is required by the transport protocol to suspend communication until the page fault is resolved. Upon receiving the page fault interrupt, the driver first needs to know on which virtual address the page fault occurred, and on what memory key. When handling send/receive operations, this information is inside the work queue. The driver reads the needed work queue elements, and parses them to gather the address and memory key. For other RDMA operations, the event generated by the HCA only contains the virtual address and rkey, as there are no work queue elements involved. Having the rkey, the driver can find the relevant memory region in its data structures, and calculate the actual pages needed to complete the operation. It then uses get_user_pages to retrieve the needed pages back to the memory, obtains dma mapping, and passes the addresses to the HCA. Finally, the driver notifies the HCA it can continue operation on the queue pair that encountered the page fault. The pages that get_user_pages returned are unpinned immediately by releasing their reference. How are invalidations handled? The patches add infrastructure to subscribe the RDMA stack as an mmu notifier client [4]. Each process that uses ODP register a notifier client. When receiving page invalidation notifications, they are passed to the mlx5_ib driver, which updates the HCA with new, not-present mappings. Only after flushing the HCA's page table caches the notifier returns, allowing the kernel to release the pages. What operations are supported? Currently only send, receive and RDMA write operations are supported on the RC protocol, and also send operations on the UD protocol. We hope to implement support for other transports and operations in the future. The structure of the patchset Patches 1-6: The first set of patches adds page fault support to the IB core layer, allowing MRs to be registered without their pages to be pinned. Patch 1 adds an extended verb to query device attributes, and patch 2 adds capability bits, configuration options, and a method for querying whether the paging capabilities from user-space. The next two patches (3-4) make some necessary changes to the ib_umem type. Patches 5 and 6 add paging support and invalidation support respectively. Patches 7-12: This set of patches add small size new functionality to the mlx5 driver and builds toward paging support. Patch 7 make changes to UMR mechanism (an internal mechanism used by mlx5 to update device page mappings). Patch 8 adds infrastructure support for page fault handling to the mlx5_core module. Patch 9 queries the device for paging capabilities, and patch 11 adds a function to do partial device page table updates. Finally, patch 12 adds a helper function to read information from user-space work queues in the driver's context. Patches 13-16: The final part of this patch set finally adds paging support to the mlx5 driver. Patch 13 adds in mlx5_ib the infrastructure to handle page faults coming from mlx5_core. Patch 14 adds the code to handle UD send page faults and RC send and receive page faults. Patch 15 adds support for page faults caused by RDMA write operations, and patch 16 adds invalidation support to the mlx5 driver, allowing pages to be unmapped dynamically. [1] [PATCH 0/5] mm: i_mmap_mutex to rwsem https://lkml.org/lkml/2013/6/24/683 [2] Re: Performance regression from switching lock to rw-sem for anon-vma tree https://lkml.org/lkml/2013/6/17/452 [3] pre-patches that were accepted upstream: a74d241 IB/mlx5: Refactor UMR to have its own context struct 48fea83 IB/mlx5: Set QP offsets and parameters for user QPs and not just for kernel QPs b475598 mlx5_core: Store MR attributes in mlx5_mr_core during creation and after UMR 8605933 IB/mlx5: Add MR to radix tree in reg_mr_callback [4] Integrating KVM with the Linux Memory Management (presentation), Andrea Archangeli http://www.linux-kvm.org/wiki/images/3/33/KvmForum2008%24kdf2008_15.pdf Haggai Eran (11): IB/core: Add an extended user verb to query device attributes IB/core: Replace ib_umem's offset field with a full address IB/core: Add umem function to read data from user-space IB/mlx5: Enhance UMR support to allow partial page table update net/mlx5_core: Add support for page faults events and low level handling IB/mlx5: Implement the ODP capability query verb IB/mlx5: Changes in memory region creation to support on-demand paging IB/mlx5: Add mlx5_ib_update_mtt to update page tables after creation IB/mlx5: Add function to read WQE from user-space IB/mlx5: Page faults handling infrastructure IB/mlx5: Handle page faults Sagi Grimberg (1): IB/core: Add flags for on demand paging support Shachar Raindel (4): IB/core: Add support for on demand paging regions IB/core: Implement support for MMU notifiers regarding on demand paging regions IB/mlx5: Add support for RDMA write responder page faults IB/mlx5: Implement on demand paging by adding support for MMU notifiers drivers/infiniband/Kconfig | 11 + drivers/infiniband/core/Makefile | 1 + drivers/infiniband/core/umem.c | 63 +- drivers/infiniband/core/umem_odp.c | 620 ++++++++++++++++++++ drivers/infiniband/core/umem_rbtree.c | 94 +++ drivers/infiniband/core/uverbs.h | 1 + drivers/infiniband/core/uverbs_cmd.c | 170 ++++-- drivers/infiniband/core/uverbs_main.c | 5 +- drivers/infiniband/hw/amso1100/c2_provider.c | 2 +- drivers/infiniband/hw/ehca/ehca_mrmw.c | 2 +- drivers/infiniband/hw/ipath/ipath_mr.c | 2 +- drivers/infiniband/hw/mlx5/Makefile | 1 + drivers/infiniband/hw/mlx5/main.c | 39 +- drivers/infiniband/hw/mlx5/mem.c | 67 ++- drivers/infiniband/hw/mlx5/mlx5_ib.h | 114 +++- drivers/infiniband/hw/mlx5/mr.c | 303 ++++++++-- drivers/infiniband/hw/mlx5/odp.c | 770 +++++++++++++++++++++++++ drivers/infiniband/hw/mlx5/qp.c | 198 +++++-- drivers/infiniband/hw/nes/nes_verbs.c | 4 +- drivers/infiniband/hw/ocrdma/ocrdma_verbs.c | 2 +- drivers/infiniband/hw/qib/qib_mr.c | 2 +- drivers/net/ethernet/mellanox/mlx5/core/eq.c | 11 +- drivers/net/ethernet/mellanox/mlx5/core/fw.c | 35 +- drivers/net/ethernet/mellanox/mlx5/core/main.c | 8 +- drivers/net/ethernet/mellanox/mlx5/core/qp.c | 134 ++++- include/linux/mlx5/device.h | 73 ++- include/linux/mlx5/driver.h | 20 +- include/linux/mlx5/qp.h | 63 ++ include/rdma/ib_umem.h | 29 +- include/rdma/ib_umem_odp.h | 156 +++++ include/rdma/ib_verbs.h | 47 +- include/uapi/rdma/ib_user_verbs.h | 25 + 32 files changed, 2907 insertions(+), 165 deletions(-) create mode 100644 drivers/infiniband/core/umem_odp.c create mode 100644 drivers/infiniband/core/umem_rbtree.c create mode 100644 drivers/infiniband/hw/mlx5/odp.c create mode 100644 include/rdma/ib_umem_odp.h -- 1.7.11.2 -- To unsubscribe from this list: send the line "unsubscribe linux-rdma" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html