1 // SPDX-License-Identifier: GPL-2.0
2 /* XDP user-space packet buffer
3 * Copyright(c) 2018 Intel Corporation.
4 */
5
6 #include <linux/init.h>
7 #include <linux/sched/mm.h>
8 #include <linux/sched/signal.h>
9 #include <linux/sched/task.h>
10 #include <linux/uaccess.h>
11 #include <linux/slab.h>
12 #include <linux/bpf.h>
13 #include <linux/mm.h>
14 #include <linux/netdevice.h>
15 #include <linux/rtnetlink.h>
16 #include <linux/idr.h>
17 #include <linux/vmalloc.h>
18
19 #include "xdp_umem.h"
20 #include "xsk_queue.h"
21
22 static DEFINE_IDA(umem_ida);
23
24 static void xdp_umem_unpin_pages(struct xdp_umem *umem)
25 {
26 unpin_user_pages_dirty_lock(umem->pgs, umem->npgs, true);
27
28 kvfree(umem->pgs);
29 umem->pgs = NULL;
30 }
31
32 static void xdp_umem_unaccount_pages(struct xdp_umem *umem)
33 {
34 if (umem->user) {
35 atomic_long_sub(umem->npgs, &umem->user->locked_vm);
36 free_uid(umem->user);
37 }
38 }
39
40 static void xdp_umem_addr_unmap(struct xdp_umem *umem)
41 {
42 vunmap(umem->addrs);
43 umem->addrs = NULL;
44 }
45
46 static int xdp_umem_addr_map(struct xdp_umem *umem, struct page **pages,
47 u32 nr_pages)
48 {
49 umem->addrs = vmap(pages, nr_pages, VM_MAP, PAGE_KERNEL);
50 if (!umem->addrs)
51 return -ENOMEM;
52 return 0;
53 }
54
55 static void xdp_umem_release(struct xdp_umem *umem)
56 {
57 umem->zc = false;
58 ida_free(&umem_ida, umem->id);
59
60 xdp_umem_addr_unmap(umem);
61 xdp_umem_unpin_pages(umem);
62
63 xdp_umem_unaccount_pages(umem);
64 kfree(umem);
65 }
66
67 static void xdp_umem_release_deferred(struct work_struct *work)
68 {
69 struct xdp_umem *umem = container_of(work, struct xdp_umem, work);
70
71 xdp_umem_release(umem);
72 }
73
74 void xdp_get_umem(struct xdp_umem *umem)
75 {
76 refcount_inc(&umem->users);
77 }
78
79 void xdp_put_umem(struct xdp_umem *umem, bool defer_cleanup)
80 {
81 if (!umem)
82 return;
83
84 if (refcount_dec_and_test(&umem->users)) {
85 if (defer_cleanup) {
86 INIT_WORK(&umem->work, xdp_umem_release_deferred);
87 schedule_work(&umem->work);
88 } else {
89 xdp_umem_release(umem);
90 }
91 }
92 }
93
94 static int xdp_umem_pin_pages(struct xdp_umem *umem, unsigned long address)
95 {
96 unsigned int gup_flags = FOLL_WRITE;
97 long npgs;
98 int err;
99
100 umem->pgs = kvcalloc(umem->npgs, sizeof(*umem->pgs), GFP_KERNEL | __GFP_NOWARN);
101 if (!umem->pgs)
102 return -ENOMEM;
103
104 mmap_read_lock(current->mm);
105 npgs = pin_user_pages(address, umem->npgs,
106 gup_flags | FOLL_LONGTERM, &umem->pgs[0]);
107 mmap_read_unlock(current->mm);
108
109 if (npgs != umem->npgs) {
110 if (npgs >= 0) {
111 umem->npgs = npgs;
112 err = -ENOMEM;
113 goto out_pin;
114 }
115 err = npgs;
116 goto out_pgs;
117 }
118 return 0;
119
120 out_pin:
121 xdp_umem_unpin_pages(umem);
122 out_pgs:
123 kvfree(umem->pgs);
124 umem->pgs = NULL;
125 return err;
126 }
127
128 static int xdp_umem_account_pages(struct xdp_umem *umem)
129 {
130 unsigned long lock_limit, new_npgs, old_npgs;
131
132 if (capable(CAP_IPC_LOCK))
133 return 0;
134
135 lock_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
136 umem->user = get_uid(current_user());
137
138 do {
139 old_npgs = atomic_long_read(&umem->user->locked_vm);
140 new_npgs = old_npgs + umem->npgs;
141 if (new_npgs > lock_limit) {
142 free_uid(umem->user);
143 umem->user = NULL;
144 return -ENOBUFS;
145 }
146 } while (atomic_long_cmpxchg(&umem->user->locked_vm, old_npgs,
147 new_npgs) != old_npgs);
148 return 0;
149 }
150
151 #define XDP_UMEM_FLAGS_VALID ( \
152 XDP_UMEM_UNALIGNED_CHUNK_FLAG | \
153 XDP_UMEM_TX_SW_CSUM | \
154 XDP_UMEM_TX_METADATA_LEN | \
155 0)
156
157 static int xdp_umem_reg(struct xdp_umem *umem, struct xdp_umem_reg *mr)
158 {
159 bool unaligned_chunks = mr->flags & XDP_UMEM_UNALIGNED_CHUNK_FLAG;
160 u32 chunk_size = mr->chunk_size, headroom = mr->headroom;
161 u64 addr = mr->addr, size = mr->len;
162 u32 chunks_rem, npgs_rem;
163 u64 chunks, npgs;
164 int err;
165
166 if (chunk_size < XDP_UMEM_MIN_CHUNK_SIZE || chunk_size > PAGE_SIZE) {
167 /* Strictly speaking we could support this, if:
168 * - huge pages, or*
169 * - using an IOMMU, or
170 * - making sure the memory area is consecutive
171 * but for now, we simply say "computer says no".
172 */
173 return -EINVAL;
174 }
175
176 if (mr->flags & ~XDP_UMEM_FLAGS_VALID)
177 return -EINVAL;
178
179 if (!unaligned_chunks && !is_power_of_2(chunk_size))
180 return -EINVAL;
181
182 if (!PAGE_ALIGNED(addr)) {
183 /* Memory area has to be page size aligned. For
184 * simplicity, this might change.
185 */
186 return -EINVAL;
187 }
188
189 if ((addr + size) < addr)
190 return -EINVAL;
191
192 npgs = div_u64_rem(size, PAGE_SIZE, &npgs_rem);
193 if (npgs_rem)
194 npgs++;
195 if (npgs > U32_MAX)
196 return -EINVAL;
197
198 chunks = div_u64_rem(size, chunk_size, &chunks_rem);
199 if (!chunks || chunks > U32_MAX)
200 return -EINVAL;
201
202 if (!unaligned_chunks && chunks_rem)
203 return -EINVAL;
204
205 if (headroom >= chunk_size - XDP_PACKET_HEADROOM)
206 return -EINVAL;
207
208 if (mr->flags & XDP_UMEM_TX_METADATA_LEN) {
209 if (mr->tx_metadata_len >= 256 || mr->tx_metadata_len % 8)
210 return -EINVAL;
211 umem->tx_metadata_len = mr->tx_metadata_len;
212 }
213
214 umem->size = size;
215 umem->headroom = headroom;
216 umem->chunk_size = chunk_size;
217 umem->chunks = chunks;
218 umem->npgs = npgs;
219 umem->pgs = NULL;
220 umem->user = NULL;
221 umem->flags = mr->flags;
222
223 INIT_LIST_HEAD(&umem->xsk_dma_list);
224 refcount_set(&umem->users, 1);
225
226 err = xdp_umem_account_pages(umem);
227 if (err)
228 return err;
229
230 err = xdp_umem_pin_pages(umem, (unsigned long)addr);
231 if (err)
232 goto out_account;
233
234 err = xdp_umem_addr_map(umem, umem->pgs, umem->npgs);
235 if (err)
236 goto out_unpin;
237
238 return 0;
239
240 out_unpin:
241 xdp_umem_unpin_pages(umem);
242 out_account:
243 xdp_umem_unaccount_pages(umem);
244 return err;
245 }
246
247 struct xdp_umem *xdp_umem_create(struct xdp_umem_reg *mr)
248 {
249 struct xdp_umem *umem;
250 int err;
251
252 umem = kzalloc(sizeof(*umem), GFP_KERNEL);
253 if (!umem)
254 return ERR_PTR(-ENOMEM);
255
256 err = ida_alloc(&umem_ida, GFP_KERNEL);
257 if (err < 0) {
258 kfree(umem);
259 return ERR_PTR(err);
260 }
261 umem->id = err;
262
263 err = xdp_umem_reg(umem, mr);
264 if (err) {
265 ida_free(&umem_ida, umem->id);
266 kfree(umem);
267 return ERR_PTR(err);
268 }
269
270 return umem;
271 }
272
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.