xref: /openbmc/linux/net/xdp/xdp_umem.c (revision 5e012745)
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 
18 #include "xdp_umem.h"
19 #include "xsk_queue.h"
20 
21 #define XDP_UMEM_MIN_CHUNK_SIZE 2048
22 
23 static DEFINE_IDA(umem_ida);
24 
25 void xdp_add_sk_umem(struct xdp_umem *umem, struct xdp_sock *xs)
26 {
27 	unsigned long flags;
28 
29 	spin_lock_irqsave(&umem->xsk_list_lock, flags);
30 	list_add_rcu(&xs->list, &umem->xsk_list);
31 	spin_unlock_irqrestore(&umem->xsk_list_lock, flags);
32 }
33 
34 void xdp_del_sk_umem(struct xdp_umem *umem, struct xdp_sock *xs)
35 {
36 	unsigned long flags;
37 
38 	spin_lock_irqsave(&umem->xsk_list_lock, flags);
39 	list_del_rcu(&xs->list);
40 	spin_unlock_irqrestore(&umem->xsk_list_lock, flags);
41 }
42 
43 /* The umem is stored both in the _rx struct and the _tx struct as we do
44  * not know if the device has more tx queues than rx, or the opposite.
45  * This might also change during run time.
46  */
47 static int xdp_reg_umem_at_qid(struct net_device *dev, struct xdp_umem *umem,
48 			       u16 queue_id)
49 {
50 	if (queue_id >= max_t(unsigned int,
51 			      dev->real_num_rx_queues,
52 			      dev->real_num_tx_queues))
53 		return -EINVAL;
54 
55 	if (queue_id < dev->real_num_rx_queues)
56 		dev->_rx[queue_id].umem = umem;
57 	if (queue_id < dev->real_num_tx_queues)
58 		dev->_tx[queue_id].umem = umem;
59 
60 	return 0;
61 }
62 
63 struct xdp_umem *xdp_get_umem_from_qid(struct net_device *dev,
64 				       u16 queue_id)
65 {
66 	if (queue_id < dev->real_num_rx_queues)
67 		return dev->_rx[queue_id].umem;
68 	if (queue_id < dev->real_num_tx_queues)
69 		return dev->_tx[queue_id].umem;
70 
71 	return NULL;
72 }
73 EXPORT_SYMBOL(xdp_get_umem_from_qid);
74 
75 static void xdp_clear_umem_at_qid(struct net_device *dev, u16 queue_id)
76 {
77 	if (queue_id < dev->real_num_rx_queues)
78 		dev->_rx[queue_id].umem = NULL;
79 	if (queue_id < dev->real_num_tx_queues)
80 		dev->_tx[queue_id].umem = NULL;
81 }
82 
83 int xdp_umem_assign_dev(struct xdp_umem *umem, struct net_device *dev,
84 			u16 queue_id, u16 flags)
85 {
86 	bool force_zc, force_copy;
87 	struct netdev_bpf bpf;
88 	int err = 0;
89 
90 	ASSERT_RTNL();
91 
92 	force_zc = flags & XDP_ZEROCOPY;
93 	force_copy = flags & XDP_COPY;
94 
95 	if (force_zc && force_copy)
96 		return -EINVAL;
97 
98 	if (xdp_get_umem_from_qid(dev, queue_id))
99 		return -EBUSY;
100 
101 	err = xdp_reg_umem_at_qid(dev, umem, queue_id);
102 	if (err)
103 		return err;
104 
105 	umem->dev = dev;
106 	umem->queue_id = queue_id;
107 
108 	dev_hold(dev);
109 
110 	if (force_copy)
111 		/* For copy-mode, we are done. */
112 		return 0;
113 
114 	if (!dev->netdev_ops->ndo_bpf ||
115 	    !dev->netdev_ops->ndo_xsk_async_xmit) {
116 		err = -EOPNOTSUPP;
117 		goto err_unreg_umem;
118 	}
119 
120 	bpf.command = XDP_SETUP_XSK_UMEM;
121 	bpf.xsk.umem = umem;
122 	bpf.xsk.queue_id = queue_id;
123 
124 	err = dev->netdev_ops->ndo_bpf(dev, &bpf);
125 	if (err)
126 		goto err_unreg_umem;
127 
128 	umem->zc = true;
129 	return 0;
130 
131 err_unreg_umem:
132 	if (!force_zc)
133 		err = 0; /* fallback to copy mode */
134 	if (err)
135 		xdp_clear_umem_at_qid(dev, queue_id);
136 	return err;
137 }
138 
139 void xdp_umem_clear_dev(struct xdp_umem *umem)
140 {
141 	struct netdev_bpf bpf;
142 	int err;
143 
144 	ASSERT_RTNL();
145 
146 	if (!umem->dev)
147 		return;
148 
149 	if (umem->zc) {
150 		bpf.command = XDP_SETUP_XSK_UMEM;
151 		bpf.xsk.umem = NULL;
152 		bpf.xsk.queue_id = umem->queue_id;
153 
154 		err = umem->dev->netdev_ops->ndo_bpf(umem->dev, &bpf);
155 
156 		if (err)
157 			WARN(1, "failed to disable umem!\n");
158 	}
159 
160 	xdp_clear_umem_at_qid(umem->dev, umem->queue_id);
161 
162 	dev_put(umem->dev);
163 	umem->dev = NULL;
164 	umem->zc = false;
165 }
166 
167 static void xdp_umem_unpin_pages(struct xdp_umem *umem)
168 {
169 	unsigned int i;
170 
171 	for (i = 0; i < umem->npgs; i++) {
172 		struct page *page = umem->pgs[i];
173 
174 		set_page_dirty_lock(page);
175 		put_page(page);
176 	}
177 
178 	kfree(umem->pgs);
179 	umem->pgs = NULL;
180 }
181 
182 static void xdp_umem_unaccount_pages(struct xdp_umem *umem)
183 {
184 	if (umem->user) {
185 		atomic_long_sub(umem->npgs, &umem->user->locked_vm);
186 		free_uid(umem->user);
187 	}
188 }
189 
190 static void xdp_umem_release(struct xdp_umem *umem)
191 {
192 	rtnl_lock();
193 	xdp_umem_clear_dev(umem);
194 	rtnl_unlock();
195 
196 	ida_simple_remove(&umem_ida, umem->id);
197 
198 	if (umem->fq) {
199 		xskq_destroy(umem->fq);
200 		umem->fq = NULL;
201 	}
202 
203 	if (umem->cq) {
204 		xskq_destroy(umem->cq);
205 		umem->cq = NULL;
206 	}
207 
208 	xsk_reuseq_destroy(umem);
209 
210 	xdp_umem_unpin_pages(umem);
211 
212 	kfree(umem->pages);
213 	umem->pages = NULL;
214 
215 	xdp_umem_unaccount_pages(umem);
216 	kfree(umem);
217 }
218 
219 static void xdp_umem_release_deferred(struct work_struct *work)
220 {
221 	struct xdp_umem *umem = container_of(work, struct xdp_umem, work);
222 
223 	xdp_umem_release(umem);
224 }
225 
226 void xdp_get_umem(struct xdp_umem *umem)
227 {
228 	refcount_inc(&umem->users);
229 }
230 
231 void xdp_put_umem(struct xdp_umem *umem)
232 {
233 	if (!umem)
234 		return;
235 
236 	if (refcount_dec_and_test(&umem->users)) {
237 		INIT_WORK(&umem->work, xdp_umem_release_deferred);
238 		schedule_work(&umem->work);
239 	}
240 }
241 
242 static int xdp_umem_pin_pages(struct xdp_umem *umem)
243 {
244 	unsigned int gup_flags = FOLL_WRITE;
245 	long npgs;
246 	int err;
247 
248 	umem->pgs = kcalloc(umem->npgs, sizeof(*umem->pgs),
249 			    GFP_KERNEL | __GFP_NOWARN);
250 	if (!umem->pgs)
251 		return -ENOMEM;
252 
253 	down_read(&current->mm->mmap_sem);
254 	npgs = get_user_pages(umem->address, umem->npgs,
255 			      gup_flags | FOLL_LONGTERM, &umem->pgs[0], NULL);
256 	up_read(&current->mm->mmap_sem);
257 
258 	if (npgs != umem->npgs) {
259 		if (npgs >= 0) {
260 			umem->npgs = npgs;
261 			err = -ENOMEM;
262 			goto out_pin;
263 		}
264 		err = npgs;
265 		goto out_pgs;
266 	}
267 	return 0;
268 
269 out_pin:
270 	xdp_umem_unpin_pages(umem);
271 out_pgs:
272 	kfree(umem->pgs);
273 	umem->pgs = NULL;
274 	return err;
275 }
276 
277 static int xdp_umem_account_pages(struct xdp_umem *umem)
278 {
279 	unsigned long lock_limit, new_npgs, old_npgs;
280 
281 	if (capable(CAP_IPC_LOCK))
282 		return 0;
283 
284 	lock_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
285 	umem->user = get_uid(current_user());
286 
287 	do {
288 		old_npgs = atomic_long_read(&umem->user->locked_vm);
289 		new_npgs = old_npgs + umem->npgs;
290 		if (new_npgs > lock_limit) {
291 			free_uid(umem->user);
292 			umem->user = NULL;
293 			return -ENOBUFS;
294 		}
295 	} while (atomic_long_cmpxchg(&umem->user->locked_vm, old_npgs,
296 				     new_npgs) != old_npgs);
297 	return 0;
298 }
299 
300 static int xdp_umem_reg(struct xdp_umem *umem, struct xdp_umem_reg *mr)
301 {
302 	u32 chunk_size = mr->chunk_size, headroom = mr->headroom;
303 	unsigned int chunks, chunks_per_page;
304 	u64 addr = mr->addr, size = mr->len;
305 	int size_chk, err, i;
306 
307 	if (chunk_size < XDP_UMEM_MIN_CHUNK_SIZE || chunk_size > PAGE_SIZE) {
308 		/* Strictly speaking we could support this, if:
309 		 * - huge pages, or*
310 		 * - using an IOMMU, or
311 		 * - making sure the memory area is consecutive
312 		 * but for now, we simply say "computer says no".
313 		 */
314 		return -EINVAL;
315 	}
316 
317 	if (!is_power_of_2(chunk_size))
318 		return -EINVAL;
319 
320 	if (!PAGE_ALIGNED(addr)) {
321 		/* Memory area has to be page size aligned. For
322 		 * simplicity, this might change.
323 		 */
324 		return -EINVAL;
325 	}
326 
327 	if ((addr + size) < addr)
328 		return -EINVAL;
329 
330 	chunks = (unsigned int)div_u64(size, chunk_size);
331 	if (chunks == 0)
332 		return -EINVAL;
333 
334 	chunks_per_page = PAGE_SIZE / chunk_size;
335 	if (chunks < chunks_per_page || chunks % chunks_per_page)
336 		return -EINVAL;
337 
338 	headroom = ALIGN(headroom, 64);
339 
340 	size_chk = chunk_size - headroom - XDP_PACKET_HEADROOM;
341 	if (size_chk < 0)
342 		return -EINVAL;
343 
344 	umem->address = (unsigned long)addr;
345 	umem->chunk_mask = ~((u64)chunk_size - 1);
346 	umem->size = size;
347 	umem->headroom = headroom;
348 	umem->chunk_size_nohr = chunk_size - headroom;
349 	umem->npgs = size / PAGE_SIZE;
350 	umem->pgs = NULL;
351 	umem->user = NULL;
352 	INIT_LIST_HEAD(&umem->xsk_list);
353 	spin_lock_init(&umem->xsk_list_lock);
354 
355 	refcount_set(&umem->users, 1);
356 
357 	err = xdp_umem_account_pages(umem);
358 	if (err)
359 		return err;
360 
361 	err = xdp_umem_pin_pages(umem);
362 	if (err)
363 		goto out_account;
364 
365 	umem->pages = kcalloc(umem->npgs, sizeof(*umem->pages), GFP_KERNEL);
366 	if (!umem->pages) {
367 		err = -ENOMEM;
368 		goto out_account;
369 	}
370 
371 	for (i = 0; i < umem->npgs; i++)
372 		umem->pages[i].addr = page_address(umem->pgs[i]);
373 
374 	return 0;
375 
376 out_account:
377 	xdp_umem_unaccount_pages(umem);
378 	return err;
379 }
380 
381 struct xdp_umem *xdp_umem_create(struct xdp_umem_reg *mr)
382 {
383 	struct xdp_umem *umem;
384 	int err;
385 
386 	umem = kzalloc(sizeof(*umem), GFP_KERNEL);
387 	if (!umem)
388 		return ERR_PTR(-ENOMEM);
389 
390 	err = ida_simple_get(&umem_ida, 0, 0, GFP_KERNEL);
391 	if (err < 0) {
392 		kfree(umem);
393 		return ERR_PTR(err);
394 	}
395 	umem->id = err;
396 
397 	err = xdp_umem_reg(umem, mr);
398 	if (err) {
399 		ida_simple_remove(&umem_ida, umem->id);
400 		kfree(umem);
401 		return ERR_PTR(err);
402 	}
403 
404 	return umem;
405 }
406 
407 bool xdp_umem_validate_queues(struct xdp_umem *umem)
408 {
409 	return umem->fq && umem->cq;
410 }
411