xref: /openbmc/linux/net/xdp/xdp_umem.c (revision 4e95bc26) 
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 	force_zc = flags & XDP_ZEROCOPY;
91 	force_copy = flags & XDP_COPY;
92 
93 	if (force_zc && force_copy)
94 		return -EINVAL;
95 
96 	rtnl_lock();
97 	if (xdp_get_umem_from_qid(dev, queue_id)) {
98 		err = -EBUSY;
99 		goto out_rtnl_unlock;
100 	}
101 
102 	err = xdp_reg_umem_at_qid(dev, umem, queue_id);
103 	if (err)
104 		goto out_rtnl_unlock;
105 
106 	umem->dev = dev;
107 	umem->queue_id = queue_id;
108 
109 	dev_hold(dev);
110 
111 	if (force_copy)
112 		/* For copy-mode, we are done. */
113 		goto out_rtnl_unlock;
114 
115 	if (!dev->netdev_ops->ndo_bpf ||
116 	    !dev->netdev_ops->ndo_xsk_async_xmit) {
117 		err = -EOPNOTSUPP;
118 		goto err_unreg_umem;
119 	}
120 
121 	bpf.command = XDP_SETUP_XSK_UMEM;
122 	bpf.xsk.umem = umem;
123 	bpf.xsk.queue_id = queue_id;
124 
125 	err = dev->netdev_ops->ndo_bpf(dev, &bpf);
126 	if (err)
127 		goto err_unreg_umem;
128 	rtnl_unlock();
129 
130 	umem->zc = true;
131 	return 0;
132 
133 err_unreg_umem:
134 	if (!force_zc)
135 		err = 0; /* fallback to copy mode */
136 	if (err)
137 		xdp_clear_umem_at_qid(dev, queue_id);
138 out_rtnl_unlock:
139 	rtnl_unlock();
140 	return err;
141 }
142 
143 void xdp_umem_clear_dev(struct xdp_umem *umem)
144 {
145 	struct netdev_bpf bpf;
146 	int err;
147 
148 	ASSERT_RTNL();
149 
150 	if (!umem->dev)
151 		return;
152 
153 	if (umem->zc) {
154 		bpf.command = XDP_SETUP_XSK_UMEM;
155 		bpf.xsk.umem = NULL;
156 		bpf.xsk.queue_id = umem->queue_id;
157 
158 		err = umem->dev->netdev_ops->ndo_bpf(umem->dev, &bpf);
159 
160 		if (err)
161 			WARN(1, "failed to disable umem!\n");
162 	}
163 
164 	xdp_clear_umem_at_qid(umem->dev, umem->queue_id);
165 
166 	dev_put(umem->dev);
167 	umem->dev = NULL;
168 	umem->zc = false;
169 }
170 
171 static void xdp_umem_unpin_pages(struct xdp_umem *umem)
172 {
173 	unsigned int i;
174 
175 	for (i = 0; i < umem->npgs; i++) {
176 		struct page *page = umem->pgs[i];
177 
178 		set_page_dirty_lock(page);
179 		put_page(page);
180 	}
181 
182 	kfree(umem->pgs);
183 	umem->pgs = NULL;
184 }
185 
186 static void xdp_umem_unaccount_pages(struct xdp_umem *umem)
187 {
188 	if (umem->user) {
189 		atomic_long_sub(umem->npgs, &umem->user->locked_vm);
190 		free_uid(umem->user);
191 	}
192 }
193 
194 static void xdp_umem_release(struct xdp_umem *umem)
195 {
196 	rtnl_lock();
197 	xdp_umem_clear_dev(umem);
198 	rtnl_unlock();
199 
200 	ida_simple_remove(&umem_ida, umem->id);
201 
202 	if (umem->fq) {
203 		xskq_destroy(umem->fq);
204 		umem->fq = NULL;
205 	}
206 
207 	if (umem->cq) {
208 		xskq_destroy(umem->cq);
209 		umem->cq = NULL;
210 	}
211 
212 	xsk_reuseq_destroy(umem);
213 
214 	xdp_umem_unpin_pages(umem);
215 
216 	kfree(umem->pages);
217 	umem->pages = NULL;
218 
219 	xdp_umem_unaccount_pages(umem);
220 	kfree(umem);
221 }
222 
223 static void xdp_umem_release_deferred(struct work_struct *work)
224 {
225 	struct xdp_umem *umem = container_of(work, struct xdp_umem, work);
226 
227 	xdp_umem_release(umem);
228 }
229 
230 void xdp_get_umem(struct xdp_umem *umem)
231 {
232 	refcount_inc(&umem->users);
233 }
234 
235 void xdp_put_umem(struct xdp_umem *umem)
236 {
237 	if (!umem)
238 		return;
239 
240 	if (refcount_dec_and_test(&umem->users)) {
241 		INIT_WORK(&umem->work, xdp_umem_release_deferred);
242 		schedule_work(&umem->work);
243 	}
244 }
245 
246 static int xdp_umem_pin_pages(struct xdp_umem *umem)
247 {
248 	unsigned int gup_flags = FOLL_WRITE;
249 	long npgs;
250 	int err;
251 
252 	umem->pgs = kcalloc(umem->npgs, sizeof(*umem->pgs),
253 			    GFP_KERNEL | __GFP_NOWARN);
254 	if (!umem->pgs)
255 		return -ENOMEM;
256 
257 	down_read(&current->mm->mmap_sem);
258 	npgs = get_user_pages(umem->address, umem->npgs,
259 			      gup_flags | FOLL_LONGTERM, &umem->pgs[0], NULL);
260 	up_read(&current->mm->mmap_sem);
261 
262 	if (npgs != umem->npgs) {
263 		if (npgs >= 0) {
264 			umem->npgs = npgs;
265 			err = -ENOMEM;
266 			goto out_pin;
267 		}
268 		err = npgs;
269 		goto out_pgs;
270 	}
271 	return 0;
272 
273 out_pin:
274 	xdp_umem_unpin_pages(umem);
275 out_pgs:
276 	kfree(umem->pgs);
277 	umem->pgs = NULL;
278 	return err;
279 }
280 
281 static int xdp_umem_account_pages(struct xdp_umem *umem)
282 {
283 	unsigned long lock_limit, new_npgs, old_npgs;
284 
285 	if (capable(CAP_IPC_LOCK))
286 		return 0;
287 
288 	lock_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
289 	umem->user = get_uid(current_user());
290 
291 	do {
292 		old_npgs = atomic_long_read(&umem->user->locked_vm);
293 		new_npgs = old_npgs + umem->npgs;
294 		if (new_npgs > lock_limit) {
295 			free_uid(umem->user);
296 			umem->user = NULL;
297 			return -ENOBUFS;
298 		}
299 	} while (atomic_long_cmpxchg(&umem->user->locked_vm, old_npgs,
300 				     new_npgs) != old_npgs);
301 	return 0;
302 }
303 
304 static int xdp_umem_reg(struct xdp_umem *umem, struct xdp_umem_reg *mr)
305 {
306 	u32 chunk_size = mr->chunk_size, headroom = mr->headroom;
307 	unsigned int chunks, chunks_per_page;
308 	u64 addr = mr->addr, size = mr->len;
309 	int size_chk, err, i;
310 
311 	if (chunk_size < XDP_UMEM_MIN_CHUNK_SIZE || chunk_size > PAGE_SIZE) {
312 		/* Strictly speaking we could support this, if:
313 		 * - huge pages, or*
314 		 * - using an IOMMU, or
315 		 * - making sure the memory area is consecutive
316 		 * but for now, we simply say "computer says no".
317 		 */
318 		return -EINVAL;
319 	}
320 
321 	if (!is_power_of_2(chunk_size))
322 		return -EINVAL;
323 
324 	if (!PAGE_ALIGNED(addr)) {
325 		/* Memory area has to be page size aligned. For
326 		 * simplicity, this might change.
327 		 */
328 		return -EINVAL;
329 	}
330 
331 	if ((addr + size) < addr)
332 		return -EINVAL;
333 
334 	chunks = (unsigned int)div_u64(size, chunk_size);
335 	if (chunks == 0)
336 		return -EINVAL;
337 
338 	chunks_per_page = PAGE_SIZE / chunk_size;
339 	if (chunks < chunks_per_page || chunks % chunks_per_page)
340 		return -EINVAL;
341 
342 	headroom = ALIGN(headroom, 64);
343 
344 	size_chk = chunk_size - headroom - XDP_PACKET_HEADROOM;
345 	if (size_chk < 0)
346 		return -EINVAL;
347 
348 	umem->address = (unsigned long)addr;
349 	umem->chunk_mask = ~((u64)chunk_size - 1);
350 	umem->size = size;
351 	umem->headroom = headroom;
352 	umem->chunk_size_nohr = chunk_size - headroom;
353 	umem->npgs = size / PAGE_SIZE;
354 	umem->pgs = NULL;
355 	umem->user = NULL;
356 	INIT_LIST_HEAD(&umem->xsk_list);
357 	spin_lock_init(&umem->xsk_list_lock);
358 
359 	refcount_set(&umem->users, 1);
360 
361 	err = xdp_umem_account_pages(umem);
362 	if (err)
363 		return err;
364 
365 	err = xdp_umem_pin_pages(umem);
366 	if (err)
367 		goto out_account;
368 
369 	umem->pages = kcalloc(umem->npgs, sizeof(*umem->pages), GFP_KERNEL);
370 	if (!umem->pages) {
371 		err = -ENOMEM;
372 		goto out_account;
373 	}
374 
375 	for (i = 0; i < umem->npgs; i++)
376 		umem->pages[i].addr = page_address(umem->pgs[i]);
377 
378 	return 0;
379 
380 out_account:
381 	xdp_umem_unaccount_pages(umem);
382 	return err;
383 }
384 
385 struct xdp_umem *xdp_umem_create(struct xdp_umem_reg *mr)
386 {
387 	struct xdp_umem *umem;
388 	int err;
389 
390 	umem = kzalloc(sizeof(*umem), GFP_KERNEL);
391 	if (!umem)
392 		return ERR_PTR(-ENOMEM);
393 
394 	err = ida_simple_get(&umem_ida, 0, 0, GFP_KERNEL);
395 	if (err < 0) {
396 		kfree(umem);
397 		return ERR_PTR(err);
398 	}
399 	umem->id = err;
400 
401 	err = xdp_umem_reg(umem, mr);
402 	if (err) {
403 		ida_simple_remove(&umem_ida, umem->id);
404 		kfree(umem);
405 		return ERR_PTR(err);
406 	}
407 
408 	return umem;
409 }
410 
411 bool xdp_umem_validate_queues(struct xdp_umem *umem)
412 {
413 	return umem->fq && umem->cq;
414 }
415