| Message ID | 20180810021444.14014-1-jgg@ziepe.ca (mailing list archive) |
|---|---|
| Headers | show |
| Series | Allow dissociation after destroy for the ioctl methods | expand |
On Thu, Aug 09, 2018 at 08:14:34PM -0600, Jason Gunthorpe wrote: > From: Jason Gunthorpe <jgg@mellanox.com> > > The previous series fixed this for the write methods, but the ioctl > methods had an additional problem where the ioctl parser relies on > datastructures stored in the driver's rodata. > > This series solves this by revising how the parsing tree is stored in > memory. A new datastructure called 'uverbs_api' is introduced that builds > the rodata description into a proper kernel radix tree with an optimized > small value 'bkey' for attributes. > > The new datastructure is designed to allow a driver to disassociate from > the runtime api description so that non-driver methods like destroy can > continue to work. > > The major rework in datastructures and parsing of ioctl provides a nice > code simplification, source code size reduction and I measure > approximately a 6% reduction in ioctl syscall time overhead. > > The downside is that this uses a bit more memory in the radix tree, but > that is likely negligible for RDMA. > > This data structure design is intended to support future series that will > rework how the rodata description is stored to add more features, reduce > its size, and reduce the complexity required to build the rodata. > > v1: > - For loop bug fig from Leon > - Missing check in lookup_get_fd_uobject from Leon > - Use rcu_assign_pointer when setting handler for sparse, from Leon > - Dead code from Michael > - Various comment and commit message edits > > More testing has shown that the prior series introduced a bug related to > module unload of drivers when working with the mlx5 driver specific > objects. This series is needed to fix that bug, so I would like to apply it on > Friday. > > Jason Gunthorpe (10): > IB/uverbs: Have the core code create the uverbs_root_spec > IB/uverbs: Build the specs into a radix tree at runtime > IB/uverbs: Use uverbs_api to manage the object type inside the uobject > IB/uverbs: Provide implementation private memory for the > uverbs_attr_bundle > IB/uverbs: Remove the ib_uverbs_attr pointer from each attr > IB/uverbs: Add a simple allocator to uverbs_attr_bundle > IB/uverbs: Use uverbs_alloc for allocations > IB/uverbs: Use uverbs_api to unmarshal ioctl commands > IB/uverbs: Remove struct uverbs_root_spec and all supporting code > IB/uverbs: Do not check for device disassociation during ioctl applied to for-next Jason
From: Jason Gunthorpe <jgg@mellanox.com> The previous series fixed this for the write methods, but the ioctl methods had an additional problem where the ioctl parser relies on datastructures stored in the driver's rodata. This series solves this by revising how the parsing tree is stored in memory. A new datastructure called 'uverbs_api' is introduced that builds the rodata description into a proper kernel radix tree with an optimized small value 'bkey' for attributes. The new datastructure is designed to allow a driver to disassociate from the runtime api description so that non-driver methods like destroy can continue to work. The major rework in datastructures and parsing of ioctl provides a nice code simplification, source code size reduction and I measure approximately a 6% reduction in ioctl syscall time overhead. The downside is that this uses a bit more memory in the radix tree, but that is likely negligible for RDMA. This data structure design is intended to support future series that will rework how the rodata description is stored to add more features, reduce its size, and reduce the complexity required to build the rodata. v1: - For loop bug fig from Leon - Missing check in lookup_get_fd_uobject from Leon - Use rcu_assign_pointer when setting handler for sparse, from Leon - Dead code from Michael - Various comment and commit message edits More testing has shown that the prior series introduced a bug related to module unload of drivers when working with the mlx5 driver specific objects. This series is needed to fix that bug, so I would like to apply it on Friday. Jason Gunthorpe (10): IB/uverbs: Have the core code create the uverbs_root_spec IB/uverbs: Build the specs into a radix tree at runtime IB/uverbs: Use uverbs_api to manage the object type inside the uobject IB/uverbs: Provide implementation private memory for the uverbs_attr_bundle IB/uverbs: Remove the ib_uverbs_attr pointer from each attr IB/uverbs: Add a simple allocator to uverbs_attr_bundle IB/uverbs: Use uverbs_alloc for allocations IB/uverbs: Use uverbs_api to unmarshal ioctl commands IB/uverbs: Remove struct uverbs_root_spec and all supporting code IB/uverbs: Do not check for device disassociation during ioctl drivers/infiniband/core/Makefile | 5 +- drivers/infiniband/core/rdma_core.c | 145 ++-- drivers/infiniband/core/rdma_core.h | 60 +- drivers/infiniband/core/uverbs.h | 2 +- drivers/infiniband/core/uverbs_ioctl.c | 675 ++++++++++-------- drivers/infiniband/core/uverbs_ioctl_merge.c | 664 ----------------- drivers/infiniband/core/uverbs_main.c | 40 +- drivers/infiniband/core/uverbs_std_types.c | 33 - .../core/uverbs_std_types_counters.c | 20 +- drivers/infiniband/core/uverbs_uapi.c | 346 +++++++++ drivers/infiniband/hw/mlx5/devx.c | 83 +-- drivers/infiniband/hw/mlx5/main.c | 45 +- drivers/infiniband/hw/mlx5/mlx5_ib.h | 1 + include/rdma/ib_verbs.h | 4 +- include/rdma/uverbs_ioctl.h | 301 ++++---- include/rdma/uverbs_std_types.h | 30 +- include/rdma/uverbs_types.h | 9 +- 17 files changed, 1083 insertions(+), 1380 deletions(-) delete mode 100644 drivers/infiniband/core/uverbs_ioctl_merge.c create mode 100644 drivers/infiniband/core/uverbs_uapi.c