Re: [PATCH rdma-next v2 00/13] DEVX asynchronous events
From: Jason Gunthorpe <jgg@ziepe.ca>
Date: 2019-07-03 20:19:16
On Sun, Jun 30, 2019 at 07:23:21PM +0300, Leon Romanovsky wrote:
From: Leon Romanovsky <redacted> Changelog: v1 -> v2: * Added Saeed's ack to net patches * Patch #2: * Fix to gather user asynchronous events on top of kernel events. * Patch #7: * Fix obj_id to be 32 bits. * Patch #8: * Inline async_event_queue applicable fields into devx_async_event_file. * Move to use bitfields in few places rather than flags. * Shorten name of UAPI attribute. * Patch #10: * Use explicitly 'struct file *' instead of void * * Store struct devx_async_event_file * instead of uobj * on the subscription. * Drop 'is_obj_related' and use list_empty instead. * Drop the temp arrays as part of the subscription API and move to simpler logic. * Revise devx_cleanup_subscription() to be success oriented without the is_close flag. * Leave key level 1 in the tree upon bad flow to prevent a race with IRQ flow. * Fix some styling notes. * Patch #11: * Use rcu read lock also for the un-affiliated event flow. * Improve locking scheme as part of read events. * Return -EIO as soon as destroyed occurred. * Use a better errno as part read event failure when the buffer size was too small. * Upon hot unplug call wake_up_interruptible() unconditionally. * Use eqe->data for affiliated events header. * Fix some styling notes. * Patch #12: * Use rcu read lock also for the first XA layer. * Patch #13: * A new patch to clean up mdev usage from devx code, it can be accessed from ib_dev now. v0 -> v1: * Fix the unbind / hot unplug flows to work properly. * Fix Ref count handling on the eventfd mode in some flow. * Rebased to latest rdma-next Thanksquoted
From Yishai:This series enables RDMA applications that use the DEVX interface to subscribe and read device asynchronous events. The solution is designed to allow extension of events in the future without need to perform any changes in the driver code. To enable that few changes had been done in mlx5_core, it includes: * Reading device event capabilities that are user related (affiliated and un-affiliated) and set the matching mask upon creating the matching EQ. * Enable DEVX/mlx5_ib to register for ANY event instead of the option to get some hard-coded ones. * Enable DEVX/mlx5_ib to get the device raw data for CQ completion events. * Enhance mlx5_core_create/destroy CQ to enable DEVX using them so that CQ events will be reported as well. In mlx5_ib layer the below changes were done: * A new DEVX API was introduced to allocate an event channel by using the uverbs FD object type. * Implement the FD channel operations to enable read/poo/close over it. * A new DEVX API was introduced to subscribe for specific events over an event channel. * Manage an internal data structure over XA(s) to subscribe/dispatch events over the different event channels. * Use from DEVX the mlx5_core APIs to create/destroy a CQ to be able to get its relevant events.
Applied to for-next, thanks Jason