xref: /linux/net/xdp/xdp_umem.c (revision b45e0c30bc58fb6fcaa42f1d1d813cefb8ab4117) 
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 #define XDP_UMEM_MIN_CHUNK_SIZE 2048
23 
24 static DEFINE_IDA(umem_ida);
25 
26 void xdp_add_sk_umem(struct xdp_umem *umem, struct xdp_sock *xs)
27 {
28 	unsigned long flags;
29 
30 	spin_lock_irqsave(&umem->xsk_list_lock, flags);
31 	list_add_rcu(&xs->list, &umem->xsk_list);
32 	spin_unlock_irqrestore(&umem->xsk_list_lock, flags);
33 }
34 
35 void xdp_del_sk_umem(struct xdp_umem *umem, struct xdp_sock *xs)
36 {
37 	unsigned long flags;
38 
39 	spin_lock_irqsave(&umem->xsk_list_lock, flags);
40 	list_del_rcu(&xs->list);
41 	spin_unlock_irqrestore(&umem->xsk_list_lock, flags);
42 }
43 
44 /* The umem is stored both in the _rx struct and the _tx struct as we do
45  * not know if the device has more tx queues than rx, or the opposite.
46  * This might also change during run time.
47  */
48 static int xdp_reg_umem_at_qid(struct net_device *dev, struct xdp_umem *umem,
49 			       u16 queue_id)
50 {
51 	if (queue_id >= max_t(unsigned int,
52 			      dev->real_num_rx_queues,
53 			      dev->real_num_tx_queues))
54 		return -EINVAL;
55 
56 	if (queue_id < dev->real_num_rx_queues)
57 		dev->_rx[queue_id].umem = umem;
58 	if (queue_id < dev->real_num_tx_queues)
59 		dev->_tx[queue_id].umem = umem;
60 
61 	return 0;
62 }
63 
64 struct xdp_umem *xdp_get_umem_from_qid(struct net_device *dev,
65 				       u16 queue_id)
66 {
67 	if (queue_id < dev->real_num_rx_queues)
68 		return dev->_rx[queue_id].umem;
69 	if (queue_id < dev->real_num_tx_queues)
70 		return dev->_tx[queue_id].umem;
71 
72 	return NULL;
73 }
74 EXPORT_SYMBOL(xdp_get_umem_from_qid);
75 
76 static void xdp_clear_umem_at_qid(struct net_device *dev, u16 queue_id)
77 {
78 	if (queue_id < dev->real_num_rx_queues)
79 		dev->_rx[queue_id].umem = NULL;
80 	if (queue_id < dev->real_num_tx_queues)
81 		dev->_tx[queue_id].umem = NULL;
82 }
83 
84 int xdp_umem_assign_dev(struct xdp_umem *umem, struct net_device *dev,
85 			u16 queue_id, u16 flags)
86 {
87 	bool force_zc, force_copy;
88 	struct netdev_bpf bpf;
89 	int err = 0;
90 
91 	ASSERT_RTNL();
92 
93 	force_zc = flags & XDP_ZEROCOPY;
94 	force_copy = flags & XDP_COPY;
95 
96 	if (force_zc && force_copy)
97 		return -EINVAL;
98 
99 	if (xdp_get_umem_from_qid(dev, queue_id))
100 		return -EBUSY;
101 
102 	err = xdp_reg_umem_at_qid(dev, umem, queue_id);
103 	if (err)
104 		return err;
105 
106 	umem->dev = dev;
107 	umem->queue_id = queue_id;
108 
109 	if (flags & XDP_USE_NEED_WAKEUP) {
110 		umem->flags |= XDP_UMEM_USES_NEED_WAKEUP;
111 		/* Tx needs to be explicitly woken up the first time.
112 		 * Also for supporting drivers that do not implement this
113 		 * feature. They will always have to call sendto().
114 		 */
115 		xsk_set_tx_need_wakeup(umem);
116 	}
117 
118 	dev_hold(dev);
119 
120 	if (force_copy)
121 		/* For copy-mode, we are done. */
122 		return 0;
123 
124 	if (!dev->netdev_ops->ndo_bpf || !dev->netdev_ops->ndo_xsk_wakeup) {
125 		err = -EOPNOTSUPP;
126 		goto err_unreg_umem;
127 	}
128 
129 	bpf.command = XDP_SETUP_XSK_UMEM;
130 	bpf.xsk.umem = umem;
131 	bpf.xsk.queue_id = queue_id;
132 
133 	err = dev->netdev_ops->ndo_bpf(dev, &bpf);
134 	if (err)
135 		goto err_unreg_umem;
136 
137 	umem->zc = true;
138 	return 0;
139 
140 err_unreg_umem:
141 	if (!force_zc)
142 		err = 0; /* fallback to copy mode */
143 	if (err)
144 		xdp_clear_umem_at_qid(dev, queue_id);
145 	return err;
146 }
147 
148 void xdp_umem_clear_dev(struct xdp_umem *umem)
149 {
150 	struct netdev_bpf bpf;
151 	int err;
152 
153 	ASSERT_RTNL();
154 
155 	if (!umem->dev)
156 		return;
157 
158 	if (umem->zc) {
159 		bpf.command = XDP_SETUP_XSK_UMEM;
160 		bpf.xsk.umem = NULL;
161 		bpf.xsk.queue_id = umem->queue_id;
162 
163 		err = umem->dev->netdev_ops->ndo_bpf(umem->dev, &bpf);
164 
165 		if (err)
166 			WARN(1, "failed to disable umem!\n");
167 	}
168 
169 	xdp_clear_umem_at_qid(umem->dev, umem->queue_id);
170 
171 	dev_put(umem->dev);
172 	umem->dev = NULL;
173 	umem->zc = false;
174 }
175 
176 static void xdp_umem_unmap_pages(struct xdp_umem *umem)
177 {
178 	unsigned int i;
179 
180 	for (i = 0; i < umem->npgs; i++)
181 		if (PageHighMem(umem->pgs[i]))
182 			vunmap(umem->pages[i].addr);
183 }
184 
185 static int xdp_umem_map_pages(struct xdp_umem *umem)
186 {
187 	unsigned int i;
188 	void *addr;
189 
190 	for (i = 0; i < umem->npgs; i++) {
191 		if (PageHighMem(umem->pgs[i]))
192 			addr = vmap(&umem->pgs[i], 1, VM_MAP, PAGE_KERNEL);
193 		else
194 			addr = page_address(umem->pgs[i]);
195 
196 		if (!addr) {
197 			xdp_umem_unmap_pages(umem);
198 			return -ENOMEM;
199 		}
200 
201 		umem->pages[i].addr = addr;
202 	}
203 
204 	return 0;
205 }
206 
207 static void xdp_umem_unpin_pages(struct xdp_umem *umem)
208 {
209 	unsigned int i;
210 
211 	for (i = 0; i < umem->npgs; i++) {
212 		struct page *page = umem->pgs[i];
213 
214 		set_page_dirty_lock(page);
215 		put_page(page);
216 	}
217 
218 	kfree(umem->pgs);
219 	umem->pgs = NULL;
220 }
221 
222 static void xdp_umem_unaccount_pages(struct xdp_umem *umem)
223 {
224 	if (umem->user) {
225 		atomic_long_sub(umem->npgs, &umem->user->locked_vm);
226 		free_uid(umem->user);
227 	}
228 }
229 
230 static void xdp_umem_release(struct xdp_umem *umem)
231 {
232 	rtnl_lock();
233 	xdp_umem_clear_dev(umem);
234 	rtnl_unlock();
235 
236 	ida_simple_remove(&umem_ida, umem->id);
237 
238 	if (umem->fq) {
239 		xskq_destroy(umem->fq);
240 		umem->fq = NULL;
241 	}
242 
243 	if (umem->cq) {
244 		xskq_destroy(umem->cq);
245 		umem->cq = NULL;
246 	}
247 
248 	xsk_reuseq_destroy(umem);
249 
250 	xdp_umem_unmap_pages(umem);
251 	xdp_umem_unpin_pages(umem);
252 
253 	kfree(umem->pages);
254 	umem->pages = NULL;
255 
256 	xdp_umem_unaccount_pages(umem);
257 	kfree(umem);
258 }
259 
260 static void xdp_umem_release_deferred(struct work_struct *work)
261 {
262 	struct xdp_umem *umem = container_of(work, struct xdp_umem, work);
263 
264 	xdp_umem_release(umem);
265 }
266 
267 void xdp_get_umem(struct xdp_umem *umem)
268 {
269 	refcount_inc(&umem->users);
270 }
271 
272 void xdp_put_umem(struct xdp_umem *umem)
273 {
274 	if (!umem)
275 		return;
276 
277 	if (refcount_dec_and_test(&umem->users)) {
278 		INIT_WORK(&umem->work, xdp_umem_release_deferred);
279 		schedule_work(&umem->work);
280 	}
281 }
282 
283 static int xdp_umem_pin_pages(struct xdp_umem *umem)
284 {
285 	unsigned int gup_flags = FOLL_WRITE;
286 	long npgs;
287 	int err;
288 
289 	umem->pgs = kcalloc(umem->npgs, sizeof(*umem->pgs),
290 			    GFP_KERNEL | __GFP_NOWARN);
291 	if (!umem->pgs)
292 		return -ENOMEM;
293 
294 	down_read(&current->mm->mmap_sem);
295 	npgs = get_user_pages(umem->address, umem->npgs,
296 			      gup_flags | FOLL_LONGTERM, &umem->pgs[0], NULL);
297 	up_read(&current->mm->mmap_sem);
298 
299 	if (npgs != umem->npgs) {
300 		if (npgs >= 0) {
301 			umem->npgs = npgs;
302 			err = -ENOMEM;
303 			goto out_pin;
304 		}
305 		err = npgs;
306 		goto out_pgs;
307 	}
308 	return 0;
309 
310 out_pin:
311 	xdp_umem_unpin_pages(umem);
312 out_pgs:
313 	kfree(umem->pgs);
314 	umem->pgs = NULL;
315 	return err;
316 }
317 
318 static int xdp_umem_account_pages(struct xdp_umem *umem)
319 {
320 	unsigned long lock_limit, new_npgs, old_npgs;
321 
322 	if (capable(CAP_IPC_LOCK))
323 		return 0;
324 
325 	lock_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
326 	umem->user = get_uid(current_user());
327 
328 	do {
329 		old_npgs = atomic_long_read(&umem->user->locked_vm);
330 		new_npgs = old_npgs + umem->npgs;
331 		if (new_npgs > lock_limit) {
332 			free_uid(umem->user);
333 			umem->user = NULL;
334 			return -ENOBUFS;
335 		}
336 	} while (atomic_long_cmpxchg(&umem->user->locked_vm, old_npgs,
337 				     new_npgs) != old_npgs);
338 	return 0;
339 }
340 
341 static int xdp_umem_reg(struct xdp_umem *umem, struct xdp_umem_reg *mr)
342 {
343 	bool unaligned_chunks = mr->flags & XDP_UMEM_UNALIGNED_CHUNK_FLAG;
344 	u32 chunk_size = mr->chunk_size, headroom = mr->headroom;
345 	unsigned int chunks, chunks_per_page;
346 	u64 addr = mr->addr, size = mr->len;
347 	int size_chk, err;
348 
349 	if (chunk_size < XDP_UMEM_MIN_CHUNK_SIZE || chunk_size > PAGE_SIZE) {
350 		/* Strictly speaking we could support this, if:
351 		 * - huge pages, or*
352 		 * - using an IOMMU, or
353 		 * - making sure the memory area is consecutive
354 		 * but for now, we simply say "computer says no".
355 		 */
356 		return -EINVAL;
357 	}
358 
359 	if (mr->flags & ~(XDP_UMEM_UNALIGNED_CHUNK_FLAG |
360 			XDP_UMEM_USES_NEED_WAKEUP))
361 		return -EINVAL;
362 
363 	if (!unaligned_chunks && !is_power_of_2(chunk_size))
364 		return -EINVAL;
365 
366 	if (!PAGE_ALIGNED(addr)) {
367 		/* Memory area has to be page size aligned. For
368 		 * simplicity, this might change.
369 		 */
370 		return -EINVAL;
371 	}
372 
373 	if ((addr + size) < addr)
374 		return -EINVAL;
375 
376 	chunks = (unsigned int)div_u64(size, chunk_size);
377 	if (chunks == 0)
378 		return -EINVAL;
379 
380 	if (!unaligned_chunks) {
381 		chunks_per_page = PAGE_SIZE / chunk_size;
382 		if (chunks < chunks_per_page || chunks % chunks_per_page)
383 			return -EINVAL;
384 	}
385 
386 	headroom = ALIGN(headroom, 64);
387 
388 	size_chk = chunk_size - headroom - XDP_PACKET_HEADROOM;
389 	if (size_chk < 0)
390 		return -EINVAL;
391 
392 	umem->address = (unsigned long)addr;
393 	umem->chunk_mask = unaligned_chunks ? XSK_UNALIGNED_BUF_ADDR_MASK
394 					    : ~((u64)chunk_size - 1);
395 	umem->size = size;
396 	umem->headroom = headroom;
397 	umem->chunk_size_nohr = chunk_size - headroom;
398 	umem->npgs = size / PAGE_SIZE;
399 	umem->pgs = NULL;
400 	umem->user = NULL;
401 	umem->flags = mr->flags;
402 	INIT_LIST_HEAD(&umem->xsk_list);
403 	spin_lock_init(&umem->xsk_list_lock);
404 
405 	refcount_set(&umem->users, 1);
406 
407 	err = xdp_umem_account_pages(umem);
408 	if (err)
409 		return err;
410 
411 	err = xdp_umem_pin_pages(umem);
412 	if (err)
413 		goto out_account;
414 
415 	umem->pages = kcalloc(umem->npgs, sizeof(*umem->pages), GFP_KERNEL);
416 	if (!umem->pages) {
417 		err = -ENOMEM;
418 		goto out_pin;
419 	}
420 
421 	err = xdp_umem_map_pages(umem);
422 	if (!err)
423 		return 0;
424 
425 	kfree(umem->pages);
426 
427 out_pin:
428 	xdp_umem_unpin_pages(umem);
429 out_account:
430 	xdp_umem_unaccount_pages(umem);
431 	return err;
432 }
433 
434 struct xdp_umem *xdp_umem_create(struct xdp_umem_reg *mr)
435 {
436 	struct xdp_umem *umem;
437 	int err;
438 
439 	umem = kzalloc(sizeof(*umem), GFP_KERNEL);
440 	if (!umem)
441 		return ERR_PTR(-ENOMEM);
442 
443 	err = ida_simple_get(&umem_ida, 0, 0, GFP_KERNEL);
444 	if (err < 0) {
445 		kfree(umem);
446 		return ERR_PTR(err);
447 	}
448 	umem->id = err;
449 
450 	err = xdp_umem_reg(umem, mr);
451 	if (err) {
452 		ida_simple_remove(&umem_ida, umem->id);
453 		kfree(umem);
454 		return ERR_PTR(err);
455 	}
456 
457 	return umem;
458 }
459 
460 bool xdp_umem_validate_queues(struct xdp_umem *umem)
461 {
462 	return umem->fq && umem->cq;
463 }
464