RDMA/core: Clear driver_udata before destroying objects in rdma_core
PATCH v2 1/1 RDMA/core: Clear driver_udata before destroying objects in rdma_core
When uobject creation fails while copying user output data (i.e., after the HW object has been created), it calls rdma_alloc_abort_uobject() with hw_obj_valid=true which then invokes the object's destroy callback, but passes the uverbs_attr_bundle from the original creation command.
The issue is that drivers are expected to validate the udata and return an error if there's unexpected content, and data from the wrong command counts as "unexpected content", so this ends up causing the driver's destroy call to fail.
A similar issue exists in the rereg_mr path when a new MR is created and the old one is destroyed.
Fix this by clearing driver_udata before invoking the driver's destroy method.
---
Changes in v2:
- Drop redundant NULL check in create_qp() now that uverbs_get_cleared_udata() handles NULL (Leon)
drivers/infiniband/core/ib_core_uverbs.c | 12 ------------ drivers/infiniband/core/rdma_core.c | 2 ++ drivers/infiniband/core/rdma_core.h | 17 +++++++++++------ drivers/infiniband/core/verbs.c | 3 +-- 4 files changed, 14 insertions(+), 20 deletions(-)
drivers/infiniband/core/ib_core_uverbs.c +0 -12
@@ -535,18 +535,6 @@ int uverbs_destroy_def_handler(struct uverbs_attr_bundle *attrs) }EXPORT_SYMBOL (uverbs_destroy_def_handler); -/* -* When calling a destroy function during an error unwind we need to pass in -* the udata that is sanitized of all user arguments. Ie from the driver -* perspective it looks like no udata was passed. -*/ -struct ib_udata *uverbs_get_cleared_udata (struct uverbs_attr_bundle *attrs) -{ - attrs->driver_udata = (struct ib_udata ){}; -return &attrs->driver_udata; -} -EXPORT_SYMBOL_NS_GPL (uverbs_get_cleared_udata,"rdma_core" ); -/** * _uverbs_alloc() - Quickly allocate memory for use with a bundle * @bundle: The bundle
drivers/infiniband/core/rdma_core.c +2 -0
@@ -734,6 +734,7 @@ void rdma_assign_uobject(struct ib_uobject *to_uobj, struct ib_uobject *new_uobj* If this fails then the uobject is still completely valid (though with * a new ID) and we leak it until context close. */ +uverbs_get_cleared_udata (attrs);uverbs_destroy_uobject (to_uobj,RDMA_REMOVE_DESTROY , attrs); }EXPORT_SYMBOL_NS_GPL (rdma_assign_uobject,"rdma_core" );
@@ -750,6 +751,7 @@ void rdma_alloc_abort_uobject(struct ib_uobject *uobj,int ret;if (hw_obj_valid) { +uverbs_get_cleared_udata (attrs); ret = uobj->uapi_object->type_class->destroy_hw ( uobj,RDMA_REMOVE_ABORT , attrs);/*
drivers/infiniband/core/rdma_core.h +11 -6
@@ -71,14 +71,19 @@ int uverbs_output_written(const struct uverbs_attr_bundle *bundle, size_t idx);void setup_ufile_idr_uobject (struct ib_uverbs_file *ufile); -#if IS_ENABLED (CONFIG_INFINIBAND_USER_ACCESS ) -struct ib_udata *uverbs_get_cleared_udata (struct uverbs_attr_bundle *attrs); -#else -static inline struct ib_udata *uverbs_get_cleared_udata (struct uverbs_attr_bundle *attrs) +/* +* When calling a destroy function during an error unwind we need to pass in +* the udata that is sanitized of all user arguments. Ie from the driver +* perspective it looks like no udata was passed. +*/ +static inline struct ib_udata * +uverbs_get_cleared_udata (struct uverbs_attr_bundle *attrs) { -return NULL ; +if (!attrs) +return NULL ; + attrs->driver_udata = (struct ib_udata ){}; +return &attrs->driver_udata; } -#endif /* * This is the runtime description of the uverbs API, used by the syscall
drivers/infiniband/core/verbs.c +1 -2
@@ -1331,8 +1331,7 @@ static struct ib_qp *create_qp(struct ib_device *dev, struct ib_pd *pd,return qp; err_security: - qp->device->ops.destroy_qp ( - qp, uattrs ?uverbs_get_cleared_udata (uattrs) :NULL ); + qp->device->ops.destroy_qp (qp,uverbs_get_cleared_udata (uattrs)); err_create:rdma_restrack_put (&qp->res);xa_destroy (&qp->comp_cntrs);
Signature · 1 line
2.56.0.rc1.315.gc6ed9934b7-goog
[email protected]
AI reviewSashiko has reviewed this patch and found no issues. It looks great!
Quoted 14 lines
When uobject creation fails while copying user output data (i.e., after the HW object has been created), it calls
rdma_alloc_abort_uobject()with hw_obj_valid=true which then invokes the object's destroy callback, but passes theuverbs_attr_bundlefrom the original creation command.The issue is that drivers are expected to validate the udata and return an error if there's unexpected content, and data from the wrong command counts as "unexpected content", so this ends up causing the driver's destroy call to fail.
A similar issue exists in the
rereg_mrpath when a new MR is created and the old one is destroyed.
Quoted 6 lines
Fix this by clearing
driver_udatabefore invoking the driver's destroy method.Fixes 6e0954b11c05 ("RDMA/uverbs: Allow drivers to create a new HW object during rereg_mr")Fixes 0ac8903cbbe6 ("RDMA/core: Allow the ioctl layer to abort a fully created uobject")Signed-off-by Jacob Moroni <[email protected]>
Sashiko has reviewed this patch and found no issues. It looks great!
Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=1
Leon Romanovsky
AppliedQuoted 6 lines · Jacob Moroni
When uobject creation fails while copying user output data (i.e., after the HW object has been created), it calls
rdma_alloc_abort_uobject()with hw_obj_valid=true which then invokes the object's destroy callback, but passes theuverbs_attr_bundlefrom the original creation command.
Quoted 6 lines
The issue is that drivers are expected to validate the udata and return an error if there's unexpected content, and data from the wrong command counts as "unexpected content", so this ends up causing the driver's destroy call to fail.
[...]
Applied, thanks!
[1/1] RDMA/core: Clear driver_udata before destroying objects in rdma_core
https://git.kernel.org/rdma/rdma/c/d37e98844afea3
Best regards,
Signature · 1 line
Leon Romanovsky <[email protected]>