xref: /openbmc/linux/include/rdma/ib_umem_odp.h (revision ae213c44)
1 /*
2  * Copyright (c) 2014 Mellanox Technologies. All rights reserved.
3  *
4  * This software is available to you under a choice of one of two
5  * licenses.  You may choose to be licensed under the terms of the GNU
6  * General Public License (GPL) Version 2, available from the file
7  * COPYING in the main directory of this source tree, or the
8  * OpenIB.org BSD license below:
9  *
10  *     Redistribution and use in source and binary forms, with or
11  *     without modification, are permitted provided that the following
12  *     conditions are met:
13  *
14  *      - Redistributions of source code must retain the above
15  *        copyright notice, this list of conditions and the following
16  *        disclaimer.
17  *
18  *      - Redistributions in binary form must reproduce the above
19  *        copyright notice, this list of conditions and the following
20  *        disclaimer in the documentation and/or other materials
21  *        provided with the distribution.
22  *
23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
30  * SOFTWARE.
31  */
32 
33 #ifndef IB_UMEM_ODP_H
34 #define IB_UMEM_ODP_H
35 
36 #include <rdma/ib_umem.h>
37 #include <rdma/ib_verbs.h>
38 #include <linux/interval_tree.h>
39 
40 struct umem_odp_node {
41 	u64 __subtree_last;
42 	struct rb_node rb;
43 };
44 
45 struct ib_umem_odp {
46 	struct ib_umem umem;
47 	struct ib_ucontext_per_mm *per_mm;
48 
49 	/*
50 	 * An array of the pages included in the on-demand paging umem.
51 	 * Indices of pages that are currently not mapped into the device will
52 	 * contain NULL.
53 	 */
54 	struct page		**page_list;
55 	/*
56 	 * An array of the same size as page_list, with DMA addresses mapped
57 	 * for pages the pages in page_list. The lower two bits designate
58 	 * access permissions. See ODP_READ_ALLOWED_BIT and
59 	 * ODP_WRITE_ALLOWED_BIT.
60 	 */
61 	dma_addr_t		*dma_list;
62 	/*
63 	 * The umem_mutex protects the page_list and dma_list fields of an ODP
64 	 * umem, allowing only a single thread to map/unmap pages. The mutex
65 	 * also protects access to the mmu notifier counters.
66 	 */
67 	struct mutex		umem_mutex;
68 	void			*private; /* for the HW driver to use. */
69 
70 	int notifiers_seq;
71 	int notifiers_count;
72 	int npages;
73 
74 	/* Tree tracking */
75 	struct umem_odp_node	interval_tree;
76 
77 	struct completion	notifier_completion;
78 	int			dying;
79 	struct work_struct	work;
80 };
81 
82 static inline struct ib_umem_odp *to_ib_umem_odp(struct ib_umem *umem)
83 {
84 	return container_of(umem, struct ib_umem_odp, umem);
85 }
86 
87 /*
88  * The lower 2 bits of the DMA address signal the R/W permissions for
89  * the entry. To upgrade the permissions, provide the appropriate
90  * bitmask to the map_dma_pages function.
91  *
92  * Be aware that upgrading a mapped address might result in change of
93  * the DMA address for the page.
94  */
95 #define ODP_READ_ALLOWED_BIT  (1<<0ULL)
96 #define ODP_WRITE_ALLOWED_BIT (1<<1ULL)
97 
98 #define ODP_DMA_ADDR_MASK (~(ODP_READ_ALLOWED_BIT | ODP_WRITE_ALLOWED_BIT))
99 
100 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
101 
102 struct ib_ucontext_per_mm {
103 	struct ib_ucontext *context;
104 	struct mm_struct *mm;
105 	struct pid *tgid;
106 	bool active;
107 
108 	struct rb_root_cached umem_tree;
109 	/* Protects umem_tree */
110 	struct rw_semaphore umem_rwsem;
111 
112 	struct mmu_notifier mn;
113 	unsigned int odp_mrs_count;
114 
115 	struct list_head ucontext_list;
116 	struct rcu_head rcu;
117 };
118 
119 int ib_umem_odp_get(struct ib_umem_odp *umem_odp, int access);
120 struct ib_umem_odp *ib_alloc_odp_umem(struct ib_umem_odp *root_umem,
121 				      unsigned long addr, size_t size);
122 void ib_umem_odp_release(struct ib_umem_odp *umem_odp);
123 
124 int ib_umem_odp_map_dma_pages(struct ib_umem_odp *umem_odp, u64 start_offset,
125 			      u64 bcnt, u64 access_mask,
126 			      unsigned long current_seq);
127 
128 void ib_umem_odp_unmap_dma_pages(struct ib_umem_odp *umem_odp, u64 start_offset,
129 				 u64 bound);
130 
131 typedef int (*umem_call_back)(struct ib_umem_odp *item, u64 start, u64 end,
132 			      void *cookie);
133 /*
134  * Call the callback on each ib_umem in the range. Returns the logical or of
135  * the return values of the functions called.
136  */
137 int rbt_ib_umem_for_each_in_range(struct rb_root_cached *root,
138 				  u64 start, u64 end,
139 				  umem_call_back cb,
140 				  bool blockable, void *cookie);
141 
142 /*
143  * Find first region intersecting with address range.
144  * Return NULL if not found
145  */
146 struct ib_umem_odp *rbt_ib_umem_lookup(struct rb_root_cached *root,
147 				       u64 addr, u64 length);
148 
149 static inline int ib_umem_mmu_notifier_retry(struct ib_umem_odp *umem_odp,
150 					     unsigned long mmu_seq)
151 {
152 	/*
153 	 * This code is strongly based on the KVM code from
154 	 * mmu_notifier_retry. Should be called with
155 	 * the relevant locks taken (umem_odp->umem_mutex
156 	 * and the ucontext umem_mutex semaphore locked for read).
157 	 */
158 
159 	if (unlikely(umem_odp->notifiers_count))
160 		return 1;
161 	if (umem_odp->notifiers_seq != mmu_seq)
162 		return 1;
163 	return 0;
164 }
165 
166 #else /* CONFIG_INFINIBAND_ON_DEMAND_PAGING */
167 
168 static inline int ib_umem_odp_get(struct ib_umem_odp *umem_odp, int access)
169 {
170 	return -EINVAL;
171 }
172 
173 static inline void ib_umem_odp_release(struct ib_umem_odp *umem_odp) {}
174 
175 #endif /* CONFIG_INFINIBAND_ON_DEMAND_PAGING */
176 
177 #endif /* IB_UMEM_ODP_H */
178