xref: /openbmc/linux/net/xdp/xdp_umem.c (revision 99b75a4e) 
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 	if (force_copy)
109 		/* For copy-mode, we are done. */
110 		goto out_rtnl_unlock;
111 
112 	if (!dev->netdev_ops->ndo_bpf ||
113 	    !dev->netdev_ops->ndo_xsk_async_xmit) {
114 		err = -EOPNOTSUPP;
115 		goto err_unreg_umem;
116 	}
117 
118 	bpf.command = XDP_SETUP_XSK_UMEM;
119 	bpf.xsk.umem = umem;
120 	bpf.xsk.queue_id = queue_id;
121 
122 	err = dev->netdev_ops->ndo_bpf(dev, &bpf);
123 	if (err)
124 		goto err_unreg_umem;
125 	rtnl_unlock();
126 
127 	dev_hold(dev);
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 out_rtnl_unlock:
137 	rtnl_unlock();
138 	return err;
139 }
140 
141 static void xdp_umem_clear_dev(struct xdp_umem *umem)
142 {
143 	struct netdev_bpf bpf;
144 	int err;
145 
146 	if (umem->zc) {
147 		bpf.command = XDP_SETUP_XSK_UMEM;
148 		bpf.xsk.umem = NULL;
149 		bpf.xsk.queue_id = umem->queue_id;
150 
151 		rtnl_lock();
152 		err = umem->dev->netdev_ops->ndo_bpf(umem->dev, &bpf);
153 		rtnl_unlock();
154 
155 		if (err)
156 			WARN(1, "failed to disable umem!\n");
157 	}
158 
159 	if (umem->dev) {
160 		rtnl_lock();
161 		xdp_clear_umem_at_qid(umem->dev, umem->queue_id);
162 		rtnl_unlock();
163 	}
164 
165 	if (umem->zc) {
166 		dev_put(umem->dev);
167 		umem->zc = false;
168 	}
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 	xdp_umem_clear_dev(umem);
197 
198 	ida_simple_remove(&umem_ida, umem->id);
199 
200 	if (umem->fq) {
201 		xskq_destroy(umem->fq);
202 		umem->fq = NULL;
203 	}
204 
205 	if (umem->cq) {
206 		xskq_destroy(umem->cq);
207 		umem->cq = NULL;
208 	}
209 
210 	xsk_reuseq_destroy(umem);
211 
212 	xdp_umem_unpin_pages(umem);
213 
214 	kfree(umem->pages);
215 	umem->pages = NULL;
216 
217 	xdp_umem_unaccount_pages(umem);
218 	kfree(umem);
219 }
220 
221 static void xdp_umem_release_deferred(struct work_struct *work)
222 {
223 	struct xdp_umem *umem = container_of(work, struct xdp_umem, work);
224 
225 	xdp_umem_release(umem);
226 }
227 
228 void xdp_get_umem(struct xdp_umem *umem)
229 {
230 	refcount_inc(&umem->users);
231 }
232 
233 void xdp_put_umem(struct xdp_umem *umem)
234 {
235 	if (!umem)
236 		return;
237 
238 	if (refcount_dec_and_test(&umem->users)) {
239 		INIT_WORK(&umem->work, xdp_umem_release_deferred);
240 		schedule_work(&umem->work);
241 	}
242 }
243 
244 static int xdp_umem_pin_pages(struct xdp_umem *umem)
245 {
246 	unsigned int gup_flags = FOLL_WRITE;
247 	long npgs;
248 	int err;
249 
250 	umem->pgs = kcalloc(umem->npgs, sizeof(*umem->pgs),
251 			    GFP_KERNEL | __GFP_NOWARN);
252 	if (!umem->pgs)
253 		return -ENOMEM;
254 
255 	down_read(&current->mm->mmap_sem);
256 	npgs = get_user_pages_longterm(umem->address, umem->npgs,
257 				       gup_flags, &umem->pgs[0], NULL);
258 	up_read(&current->mm->mmap_sem);
259 
260 	if (npgs != umem->npgs) {
261 		if (npgs >= 0) {
262 			umem->npgs = npgs;
263 			err = -ENOMEM;
264 			goto out_pin;
265 		}
266 		err = npgs;
267 		goto out_pgs;
268 	}
269 	return 0;
270 
271 out_pin:
272 	xdp_umem_unpin_pages(umem);
273 out_pgs:
274 	kfree(umem->pgs);
275 	umem->pgs = NULL;
276 	return err;
277 }
278 
279 static int xdp_umem_account_pages(struct xdp_umem *umem)
280 {
281 	unsigned long lock_limit, new_npgs, old_npgs;
282 
283 	if (capable(CAP_IPC_LOCK))
284 		return 0;
285 
286 	lock_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
287 	umem->user = get_uid(current_user());
288 
289 	do {
290 		old_npgs = atomic_long_read(&umem->user->locked_vm);
291 		new_npgs = old_npgs + umem->npgs;
292 		if (new_npgs > lock_limit) {
293 			free_uid(umem->user);
294 			umem->user = NULL;
295 			return -ENOBUFS;
296 		}
297 	} while (atomic_long_cmpxchg(&umem->user->locked_vm, old_npgs,
298 				     new_npgs) != old_npgs);
299 	return 0;
300 }
301 
302 static int xdp_umem_reg(struct xdp_umem *umem, struct xdp_umem_reg *mr)
303 {
304 	u32 chunk_size = mr->chunk_size, headroom = mr->headroom;
305 	unsigned int chunks, chunks_per_page;
306 	u64 addr = mr->addr, size = mr->len;
307 	int size_chk, err, i;
308 
309 	if (chunk_size < XDP_UMEM_MIN_CHUNK_SIZE || chunk_size > PAGE_SIZE) {
310 		/* Strictly speaking we could support this, if:
311 		 * - huge pages, or*
312 		 * - using an IOMMU, or
313 		 * - making sure the memory area is consecutive
314 		 * but for now, we simply say "computer says no".
315 		 */
316 		return -EINVAL;
317 	}
318 
319 	if (!is_power_of_2(chunk_size))
320 		return -EINVAL;
321 
322 	if (!PAGE_ALIGNED(addr)) {
323 		/* Memory area has to be page size aligned. For
324 		 * simplicity, this might change.
325 		 */
326 		return -EINVAL;
327 	}
328 
329 	if ((addr + size) < addr)
330 		return -EINVAL;
331 
332 	chunks = (unsigned int)div_u64(size, chunk_size);
333 	if (chunks == 0)
334 		return -EINVAL;
335 
336 	chunks_per_page = PAGE_SIZE / chunk_size;
337 	if (chunks < chunks_per_page || chunks % chunks_per_page)
338 		return -EINVAL;
339 
340 	headroom = ALIGN(headroom, 64);
341 
342 	size_chk = chunk_size - headroom - XDP_PACKET_HEADROOM;
343 	if (size_chk < 0)
344 		return -EINVAL;
345 
346 	umem->address = (unsigned long)addr;
347 	umem->chunk_mask = ~((u64)chunk_size - 1);
348 	umem->size = size;
349 	umem->headroom = headroom;
350 	umem->chunk_size_nohr = chunk_size - headroom;
351 	umem->npgs = size / PAGE_SIZE;
352 	umem->pgs = NULL;
353 	umem->user = NULL;
354 	INIT_LIST_HEAD(&umem->xsk_list);
355 	spin_lock_init(&umem->xsk_list_lock);
356 
357 	refcount_set(&umem->users, 1);
358 
359 	err = xdp_umem_account_pages(umem);
360 	if (err)
361 		return err;
362 
363 	err = xdp_umem_pin_pages(umem);
364 	if (err)
365 		goto out_account;
366 
367 	umem->pages = kcalloc(umem->npgs, sizeof(*umem->pages), GFP_KERNEL);
368 	if (!umem->pages) {
369 		err = -ENOMEM;
370 		goto out_account;
371 	}
372 
373 	for (i = 0; i < umem->npgs; i++)
374 		umem->pages[i].addr = page_address(umem->pgs[i]);
375 
376 	return 0;
377 
378 out_account:
379 	xdp_umem_unaccount_pages(umem);
380 	return err;
381 }
382 
383 struct xdp_umem *xdp_umem_create(struct xdp_umem_reg *mr)
384 {
385 	struct xdp_umem *umem;
386 	int err;
387 
388 	umem = kzalloc(sizeof(*umem), GFP_KERNEL);
389 	if (!umem)
390 		return ERR_PTR(-ENOMEM);
391 
392 	err = ida_simple_get(&umem_ida, 0, 0, GFP_KERNEL);
393 	if (err < 0) {
394 		kfree(umem);
395 		return ERR_PTR(err);
396 	}
397 	umem->id = err;
398 
399 	err = xdp_umem_reg(umem, mr);
400 	if (err) {
401 		ida_simple_remove(&umem_ida, umem->id);
402 		kfree(umem);
403 		return ERR_PTR(err);
404 	}
405 
406 	return umem;
407 }
408 
409 bool xdp_umem_validate_queues(struct xdp_umem *umem)
410 {
411 	return umem->fq && umem->cq;
412 }
413