- 03 Dec, 2018 16 commits
-
-
Leon Romanovsky authored
[leonro@server ~]$ rdma res show 1: mlx5_0: pd 3 cq 5 qp 4 cm_id 0 mr 0 ctx 0 Signed-off-by:
Leon Romanovsky <leonro@mellanox.com> Signed-off-by:
Doug Ledford <dledford@redhat.com>
-
Leon Romanovsky authored
Add restrack annotations to track allocations of ucontexts. Signed-off-by:
-
Leon Romanovsky authored
Add ability to track allocated ib_ucontext, which are limited resource and worth to be visible by users. Signed-off-by:
-
Doug Ledford authored
Make all of the write() handlers use a consistent flow From Jason, This series unifies all the write handlers to use a flow that is very similar to the ioctl handler flow, including having the same basic assumptions about extensible buffer handling and the same handler function call signature. Along the way this consolidates all the copy_to/from_user into a small set of safe buffer accessor functions tailored to the usage here. These accessors use the new dispatcher-controlled calling convention for ucore data, and support a placement of the response that does not rely on the cmd.response value. Overall this brings in in strong bounds checking to all the write() handlers and consistent enforcement of the zero-fill/zero-check methodology for buffer extension. The end result is a significant complexity reduction for all of the handlers and creates a high degree of uniformity between the write, write_ex, and ioctl handlers and dispatch flow. Thanks Jason Gunthorpe (12): RDMA/uverbs: Remove out_len checks that are now done by the core RDMA/uverbs: Use uverbs_attr_bundle to pass ucore for write/write_ex RDMA/uverbs: Get rid of the 'callback' scheme in the compat path RDMA/uverbs: Use uverbs_response() for remaining response copying RDMA/uverbs: Use uverbs_request() for request copying RDMA/uverbs: Use uverbs_request() and core for write_ex handlers RDMA/uverbs: Fill in the response for IB_USER_VERBS_EX_CMD_MODIFY_QP RDMA/uverbs: Simplify ib_uverbs_ex_query_device RDMA/uverbs: Add a simple iterator interface for reading the command RDMA/uverbs: Use the iterator for ib_uverbs_unmarshall_recv() RDMA/uverbs: Do not check the input length on create_cq/qp paths RDMA/uverbs: Use only attrs for the write() handler signature drivers/infiniband/core/rdma_core.h | 5 +- drivers/infiniband/core/uverbs_cmd.c | 1165 ++++++++++--------------- drivers/infiniband/core/uverbs_main.c | 23 +- drivers/infiniband/core/uverbs_uapi.c | 23 +- include/rdma/uverbs_ioctl.h | 9 +- 5 files changed, 479 insertions(+), 746 deletions(-) Signed-off-by: -
Jason Gunthorpe authored
All of the old arguments can be derived from the uverbs_attr_bundle structure, so get rid of the redundant arguments. Most of the prior work has been removing users of the arguments to allow this to be a simple patch. Signed-off-by:
Jason Gunthorpe <jgg@mellanox.com> Signed-off-by:
-
Jason Gunthorpe authored
If the user did not provide a long enough command buffer then the missing bytes are forced to zero. There is no reason to check the length if a zero value is OK. Signed-off-by:
-
Jason Gunthorpe authored
This has a very complicated memory layout, with two flex arrays. Use the iterator API to make reading it clearer. Signed-off-by:
-
Jason Gunthorpe authored
Several methods have a command with a trailing flex array, and they all open code some extraction scheme. Centralize this into a simple iterator API. Signed-off-by:
-
Jason Gunthorpe authored
We truncate the response structure if there is not enough room in the user buffer so there is no reason to have all the mess with finely managing response_length. Just fully fill the attrs and truncate on copy. Signed-off-by:
-
Jason Gunthorpe authored
A response struct was defined, and userspace is providing it (but not checking it). Fill it in and write it out. Signed-off-by:
-
Jason Gunthorpe authored
The write_ex handlers have this horrible boilerplate in every function to do the zero extend/zero check and min size checks. This is now handled in the core code via the meta-data, and the zero checks are handled by uverbs_request(). Replace all the occurrences. Signed-off-by:
-
Jason Gunthorpe authored
This function properly zero-extends, and zero-checks if the user buffer is not the same size as the kernel command struct. Signed-off-by:
-
Jason Gunthorpe authored
This function properly truncates and zero-fills the response which is the standard used by the ioctl uAPI when working with user data. Signed-off-by:
-
Jason Gunthorpe authored
There is no reason for this. For response processing we simply need to copy, truncate, and zero fill the response into whatever output buffer was provided. Add a function uverbs_response() that does this consistently. Signed-off-by:
-
Jason Gunthorpe authored
This creates a consistent way to access the two core buffers across write and write_ex handlers. Remove the open coded ucore conversion in the write/ex compatibility handlers. Signed-off-by:
-
Jason Gunthorpe authored
write() methods must work with fixed sized structures as that is the only way to know where the udata segment starts. The common udata code now rejects any write() that has a response buffer shorter than the core's response. Thus all the checks of out_len for write methods are redundant and can be removed. Signed-off-by:
-
- 30 Nov, 2018 3 commits
-
-
Guy Levi authored
The current implementation of create QP requires contiguous memory, such a requirement is problematic once the memory is fragmented or the system is low in memory, it causes failures in dma_zalloc_coherent(). This patch takes advantage of the new mlx5_core API which allocates a fragmented buffer. This makes the QP creation much more resilient to memory fragmentation. Data-path code was adapted to the fact that WQEs can cross buffers. We also use the opportunity to fix some cosmetic legacy coding convention errors which were in the feature scope. Signed-off-by:
Guy Levi <guyle@mellanox.com> Reviewed-by:
Majd Dibbiny <majd@mellanox.com> Signed-off-by:
-
Guy Levi authored
The current implementation of create SRQ requires contiguous memory, such a requirement is problematic once the memory is fragmented or the system is low in memory, it causes failures in dma_zalloc_coherent(). This patch takes the advantage of the new mlx5_core API which allocates a fragmented buffer, and makes the SRQ creation much more resilient to memory fragmentation. Data-path code was adapted to the fact that WQEs can cross buffers. Signed-off-by:
Yishai Hadas <yishaih@mellanox.com> Signed-off-by:
-
Chuck Lever authored
FRWR memory registration is done with a series of calls and WRs. 1. ULP invokes ib_dma_map_sg() 2. ULP invokes ib_map_mr_sg() 3. ULP posts an IB_WR_REG_MR on the Send queue Step 2 generates an iova. It is permissible for ULPs to change this iova (with certain restrictions) between steps 2 and 3. rxe_map_mr_sg captures the MR's iova but later when rxe processes the REG_MR WR, it ignores the MR's iova field. If a ULP alters the MR's iova after step 2 but before step 3, rxe never captures that change. When the remote sends an RDMA Read targeting that MR, rxe looks up the R_key, but the altered iova does not match the iova stored in the MR, causing the RDMA Read request to fail. Reported-by:
Anna Schumaker <schumaker.anna@gmail.com> Signed-off-by:
Chuck Lever <chuck.lever@oracle.com> Reviewed-by:
Sagi Grimberg <sagi@grimberg.me> Signed-off-by:
-
- 29 Nov, 2018 5 commits
-
-
Mark Bloch authored
Allow a user to attach a DEVX counter via mlx5 raw flow creation. In order to attach a counter we introduce a new attribute: MLX5_IB_ATTR_CREATE_FLOW_ARR_COUNTERS_DEVX A counter can be attached to multiple flow steering rules. Signed-off-by:
Mark Bloch <markb@mellanox.com> Reviewed-by:
-
Leon Romanovsky authored
QIB driver was added in 2010 with many BUG_ON(), most of them were cleaned out after years of development and usages. It looks like that it is safe now to remove rest of BUG_ONs. Signed-off-by:
Dennis Dalessandro <dennis.dalessandro@intel.com> Signed-off-by:
-
Colin Ian King authored
There is a spelling mistake in a usnic_err error message, fix it. Signed-off-by:
Colin Ian King <colin.king@canonical.com> Signed-off-by:
-
kbuild test robot authored
drivers/infiniband/core/uverbs_cmd.c:1095:1-3: WARNING: PTR_ERR_OR_ZERO can be used Use PTR_ERR_OR_ZERO rather than if(IS_ERR(...)) + PTR_ERR Generated by: scripts/coccinelle/api/ptr_ret.cocci Fixes: 7106a976 ("RDMA/uverbs: Make write() handlers return 0 on success") Signed-off-by:
kbuild test robot <fengguang.wu@intel.com> Signed-off-by:
-
Colin Ian King authored
Fix spelling mistake in usnic_err error message Signed-off-by:
-
- 26 Nov, 2018 10 commits
-
-
Jason Gunthorpe authored
Have the core code initialize the driver_udata if the method has a udata description. This is done using the same create_udata the handler was supposed to call. This makes ioctl consistent with the write and write_ex paths. Signed-off-by:
-
Jason Gunthorpe authored
Now that we have metadata describing the command format the core code can directly compute the udata pointers and all the really ugly ib_uverbs_init_udata() calls can be removed from the handlers. This means all the write() handlers are no longer sensitive to the layout of the command buffer. Signed-off-by:
-
Jason Gunthorpe authored
The core code needs to compute the udata so we may as well pass it in the uverbs_attr_bundle instead of on the stack. This converts the simple case of write_ex() which already has a core calculation. Also change the write() path to use the attrs for ib_uverbs_init_udata() instead of on the stack. This lets the write to write_ex compatibility path continue to follow the lead of the _ex path. Signed-off-by:
-
Jason Gunthorpe authored
The size meta-data in the prior patch describes the smallest acceptable buffer for the write() interface. Globally check this in the core code. This is necessary in the case of write() methods that have a driver udata to prevent computing a negative udata buffer length. The return code of -ENOSPC is chosen here as some of the handlers already use this code, however many other handler use EINVAL. Signed-off-by:
-
Jason Gunthorpe authored
We need the structure sizes to compute the location of the udata in the core code. Annotate the sizes into the new macro language. This is generated largely by script and checked by comparing against the similar list in rdma-core. Signed-off-by:
-
Jason Gunthorpe authored
The uverbs_attr_bundle already contains this pointer, and most methods don't actually need it. Get rid of the redundant function argument. Signed-off-by:
-
Jason Gunthorpe authored
Currently they return the command length, while all other handlers return 0. This makes the write path closer to the write_ex and ioctl path. Signed-off-by:
-
Jason Gunthorpe authored
Now that we can add meta-data to the description of write() methods we need to pass the uverbs_attr_bundle into all write based handlers so future patches can use it as a container for any new data transferred out of the core. This is the first step to bringing the write() and ioctl() methods to a common interface signature. This is a simple search/replace, and we push the attr down into the uobj and other APIs to keep changes minimal. Signed-off-by:
-
Jason Gunthorpe authored
If the struct is used with a driver_udata it should have a trailing driver_data flex array to mark it as having udata. In most cases this forces the end of the struct to be aligned to u64 which is needed to make the trailing driver_data naturally aligned. Unfortunately We have a few cases where the base struct is not aligned to 8 bytes, these are marked with a u32 driver_data and userspace will check for alignment issues when it compiles the driver. Also remove the empty ib_uverbs_modify_qp_resp as nothing uses this. pahole says there is no change to any struct sizes by this change. Signed-off-by:
-
Colin Ian King authored
There is a spelling mistake in the module description text, fix it. Signed-off-by:
-
- 22 Nov, 2018 6 commits
-
-
Parav Pandit authored
When the rdma device is getting removed, get resource info can race with device removal, as below: CPU-0 CPU-1 -------- -------- rdma_nl_rcv_msg() nldev_res_get_cq_dumpit() mutex_lock(device_lock); get device reference mutex_unlock(device_lock); [..] ib_unregister_device() /* Valid reference to * device->dev exists. */ ib_dealloc_device() [..] provider->fill_res_entry(); Even though device object is not freed, fill_res_entry() can get called on device which doesn't have a driver anymore. Kernel core device reference count is not sufficient, as this only keeps the structure valid, and doesn't guarantee the driver is still loaded. Similar race can occur with device renaming and device removal, where device_rename() tries to rename a unregistered device. While this is fine for devices of a class which are not net namespace aware, but it is incorrect for net namespace aware class coming in subsequent series. If a class is net namespace aware, then the below [1] call trace is observed in above situation. Therefore, to avoid the race, keep a reference count and let device unregistration wait until all netlink users drop the reference. [1] Call trace: kernfs: ns required in 'infiniband' for 'mlx5_0' WARNING: CPU: 18 PID: 44270 at fs/kernfs/dir.c:842 kernfs_find_ns+0x104/0x120 libahci i2c_core mlxfw libata dca [last unloaded: devlink] RIP: 0010:kernfs_find_ns+0x104/0x120 Call Trace: kernfs_find_and_get_ns+0x2e/0x50 sysfs_rename_link_ns+0x40/0xb0 device_rename+0xb2/0xf0 ib_device_rename+0xb3/0x100 [ib_core] nldev_set_doit+0x165/0x190 [ib_core] rdma_nl_rcv_msg+0x249/0x250 [ib_core] ? netlink_deliver_tap+0x8f/0x3e0 rdma_nl_rcv+0xd6/0x120 [ib_core] netlink_unicast+0x17c/0x230 netlink_sendmsg+0x2f0/0x3e0 sock_sendmsg+0x30/0x40 __sys_sendto+0xdc/0x160 Fixes: da5c8507 ("RDMA/nldev: add driver-specific resource tracking") Signed-off-by:Parav Pandit <parav@mellanox.com> Signed-off-by:
-
Parav Pandit authored
Currently several rdma_cm module specific functions are declared in core_priv.h file. Now that we have cma_priv.h file specific to rdma_cm kernel module, move them from core_priv.h to cma_priv.h Signed-off-by:
-
Jason Gunthorpe authored
Add annotations to the uverbs_api structure indicating which driver methods are called by the implementation. If the required method is NULL the write API will be not be callable. This effectively duplicates the cmd_mask system, however it does it by expressing invariants required by the core code, not by delegating decision making to the driver. This is another step toward eliminating cmd_mask. Signed-off-by:
-
Jason Gunthorpe authored
Now that we use struct uverbs_uapi to link the method functions to the dispatcher there is no reason to have them be extern symbols. Signed-off-by:
-
Jason Gunthorpe authored
This organizes the write commands into objects and links them to the uverbs_api data structure. The command path is reworked to use uapi instead of its internal structures. The command mask is moved from a runtime check to a registration time check in the uapi. Since the write interface does not have the object ID as part of the command, the radix bins are converted into linear lists to support the lookup. Signed-off-by:
-
Jason Gunthorpe authored
Bringing all uapi entry points into one place lets us deal with them consistently. For instance the write, write_ex and ioctl paths can be disabled when an API is not supported by the driver. This will replace the uverbs_cmd_table static arrays. Signed-off-by:
-
