IB/uverbs: Allow all DESTROY commands to succeed after disassociate

The disassociate function was broken by design because it failed all
commands. This prevents userspace from calling destroy on a uobject after
it has detected a device fatal error and thus reclaiming the resources in
userspace is prevented.

This fix is now straightforward, when anything destroys a uobject that is
not the user the object remains on the IDR with a NULL context and object
pointer. All lookup locking modes other than DESTROY will fail. When the
user ultimately calls the destroy function it is simply dropped from the
IDR while any related information is returned.

Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>

drivers/infiniband/core/rdma_core.c

@@ -180,7 +180,7 @@ static int uverbs_destroy_uobject(struct ib_uobject *uobj,
 	assert_uverbs_usecnt(uobj, UVERBS_LOOKUP_WRITE);
 
 	if (uobj->object) {
-		ret = uobj->type->type_class->remove_commit(uobj, reason);
+		ret = uobj->type->type_class->destroy_hw(uobj, reason);
 		if (ret) {
 			if (ib_is_destroy_retryable(ret, reason, uobj))
 				return ret;
@@ -204,10 +204,13 @@ static int uverbs_destroy_uobject(struct ib_uobject *uobj,
 
 	/*
 	 * For DESTROY the usecnt is held write locked, the caller is expected
-	 * to put it unlock and put the object when done with it.
+	 * to put it unlock and put the object when done with it. Only DESTROY
+	 * can remove the IDR handle.
 	 */
 	if (reason != RDMA_REMOVE_DESTROY)
 		atomic_set(&uobj->usecnt, 0);
+	else
+		uobj->type->type_class->remove_handle(uobj);
 
 	if (!list_empty(&uobj->list)) {
 		spin_lock_irqsave(&ufile->uobjects_lock, flags);
@@ -554,8 +557,8 @@ static void alloc_abort_idr_uobject(struct ib_uobject *uobj)
 	spin_unlock(&uobj->ufile->idr_lock);
 }
 
-static int __must_check remove_commit_idr_uobject(struct ib_uobject *uobj,
-						  enum rdma_remove_reason why)
+static int __must_check destroy_hw_idr_uobject(struct ib_uobject *uobj,
+					       enum rdma_remove_reason why)
 {
 	const struct uverbs_obj_idr_type *idr_type =
 		container_of(uobj->type, struct uverbs_obj_idr_type,
@@ -573,20 +576,28 @@ static int __must_check remove_commit_idr_uobject(struct ib_uobject *uobj,
 	if (why == RDMA_REMOVE_ABORT)
 		return 0;
 
-	alloc_abort_idr_uobject(uobj);
-	/* Matches the kref in alloc_commit_idr_uobject */
-	uverbs_uobject_put(uobj);
+	ib_rdmacg_uncharge(&uobj->cg_obj, uobj->context->device,
+			   RDMACG_RESOURCE_HCA_OBJECT);
 
 	return 0;
 }
 
+static void remove_handle_idr_uobject(struct ib_uobject *uobj)
+{
+	spin_lock(&uobj->ufile->idr_lock);
+	idr_remove(&uobj->ufile->idr, uobj->id);
+	spin_unlock(&uobj->ufile->idr_lock);
+	/* Matches the kref in alloc_commit_idr_uobject */
+	uverbs_uobject_put(uobj);
+}
+
 static void alloc_abort_fd_uobject(struct ib_uobject *uobj)
 {
 	put_unused_fd(uobj->id);
 }
 
-static int __must_check remove_commit_fd_uobject(struct ib_uobject *uobj,
-						 enum rdma_remove_reason why)
+static int __must_check destroy_hw_fd_uobject(struct ib_uobject *uobj,
+					      enum rdma_remove_reason why)
 {
 	const struct uverbs_obj_fd_type *fd_type =
 		container_of(uobj->type, struct uverbs_obj_fd_type, type);
@@ -598,6 +609,10 @@ static int __must_check remove_commit_fd_uobject(struct ib_uobject *uobj,
 	return 0;
 }
 
+static void remove_handle_fd_uobject(struct ib_uobject *uobj)
+{
+}
+
 static int alloc_commit_idr_uobject(struct ib_uobject *uobj)
 {
 	struct ib_uverbs_file *ufile = uobj->ufile;
@@ -741,13 +756,41 @@ void rdma_lookup_put_uobject(struct ib_uobject *uobj,
 	uverbs_uobject_put(uobj);
 }
 
+void setup_ufile_idr_uobject(struct ib_uverbs_file *ufile)
+{
+	spin_lock_init(&ufile->idr_lock);
+	idr_init(&ufile->idr);
+}
+
+void release_ufile_idr_uobject(struct ib_uverbs_file *ufile)
+{
+	struct ib_uobject *entry;
+	int id;
+
+	/*
+	 * At this point uverbs_cleanup_ufile() is guaranteed to have run, and
+	 * there are no HW objects left, however the IDR is still populated
+	 * with anything that has not been cleaned up by userspace. Since the
+	 * kref on ufile is 0, nothing is allowed to call lookup_get.
+	 *
+	 * This is an optimized equivalent to remove_handle_idr_uobject
+	 */
+	idr_for_each_entry(&ufile->idr, entry, id) {
+		WARN_ON(entry->object);
+		uverbs_uobject_put(entry);
+	}
+
+	idr_destroy(&ufile->idr);
+}
+
 const struct uverbs_obj_type_class uverbs_idr_class = {
 	.alloc_begin = alloc_begin_idr_uobject,
 	.lookup_get = lookup_get_idr_uobject,
 	.alloc_commit = alloc_commit_idr_uobject,
 	.alloc_abort = alloc_abort_idr_uobject,
 	.lookup_put = lookup_put_idr_uobject,
-	.remove_commit = remove_commit_idr_uobject,
+	.destroy_hw = destroy_hw_idr_uobject,
+	.remove_handle = remove_handle_idr_uobject,
 	/*
 	 * When we destroy an object, we first just lock it for WRITE and
 	 * actually DESTROY it in the finalize stage. So, the problematic
@@ -945,7 +988,8 @@ const struct uverbs_obj_type_class uverbs_fd_class = {
 	.alloc_commit = alloc_commit_fd_uobject,
 	.alloc_abort = alloc_abort_fd_uobject,
 	.lookup_put = lookup_put_fd_uobject,
-	.remove_commit = remove_commit_fd_uobject,
+	.destroy_hw = destroy_hw_fd_uobject,
+	.remove_handle = remove_handle_fd_uobject,
 	.needs_kfree_rcu = false,
 };
 EXPORT_SYMBOL(uverbs_fd_class);

drivers/infiniband/core/rdma_core.h

@@ -110,4 +110,7 @@ int uverbs_finalize_object(struct ib_uobject *uobj,
 			   enum uverbs_obj_access access,
 			   bool commit);
 
+void setup_ufile_idr_uobject(struct ib_uverbs_file *ufile);
+void release_ufile_idr_uobject(struct ib_uverbs_file *ufile);
+
 #endif /* RDMA_CORE_H */

drivers/infiniband/core/uverbs_main.c

@@ -253,6 +253,8 @@ void ib_uverbs_release_file(struct kref *ref)
 	struct ib_device *ib_dev;
 	int srcu_key;
 
+	release_ufile_idr_uobject(file);
+
 	srcu_key = srcu_read_lock(&file->device->disassociate_srcu);
 	ib_dev = srcu_dereference(file->device->ib_dev,
 				  &file->device->disassociate_srcu);
@@ -867,8 +869,6 @@ static int ib_uverbs_open(struct inode *inode, struct file *filp)
 	}
 
 	file->device	 = dev;
-	spin_lock_init(&file->idr_lock);
-	idr_init(&file->idr);
 	kref_init(&file->ref);
 	mutex_init(&file->ucontext_lock);
 
@@ -885,6 +885,8 @@ static int ib_uverbs_open(struct inode *inode, struct file *filp)
 	file->uverbs_cmd_mask = ib_dev->uverbs_cmd_mask;
 	file->uverbs_ex_cmd_mask = ib_dev->uverbs_ex_cmd_mask;
 
+	setup_ufile_idr_uobject(file);
+
 	return nonseekable_open(inode, filp);
 
 err_module:
@@ -904,7 +906,6 @@ static int ib_uverbs_close(struct inode *inode, struct file *filp)
 	struct ib_uverbs_file *file = filp->private_data;
 
 	uverbs_destroy_ufile_hw(file, RDMA_REMOVE_CLOSE);
-	idr_destroy(&file->idr);
 
 	mutex_lock(&file->device->lists_mutex);
 	if (!file->is_closed) {

include/rdma/uverbs_types.h

@@ -61,6 +61,7 @@ enum rdma_lookup_mode {
  * Destruction flow:
  *   lookup_get(exclusive=true) & uverbs_try_lock_object
  *   remove_commit
+ *   remove_handle (optional)
  *   lookup_put(exclusive=true) via rdma_lookup_put_uobject
  *
  * Allocate Error flow #1
@@ -92,8 +93,9 @@ struct uverbs_obj_type_class {
 					 enum rdma_lookup_mode mode);
 	void (*lookup_put)(struct ib_uobject *uobj, enum rdma_lookup_mode mode);
 	/* This does not consume the kref on uobj */
-	int __must_check (*remove_commit)(struct ib_uobject *uobj,
-					  enum rdma_remove_reason why);
+	int __must_check (*destroy_hw)(struct ib_uobject *uobj,
+				       enum rdma_remove_reason why);
+	void (*remove_handle)(struct ib_uobject *uobj);
 	u8    needs_kfree_rcu;
 };