xref: /openbmc/linux/net/xdp/xdp_umem.c (revision 5a170e9e) 
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 
17 #include "xdp_umem.h"
18 #include "xsk_queue.h"
19 
20 #define XDP_UMEM_MIN_CHUNK_SIZE 2048
21 
22 void xdp_add_sk_umem(struct xdp_umem *umem, struct xdp_sock *xs)
23 {
24 	unsigned long flags;
25 
26 	spin_lock_irqsave(&umem->xsk_list_lock, flags);
27 	list_add_rcu(&xs->list, &umem->xsk_list);
28 	spin_unlock_irqrestore(&umem->xsk_list_lock, flags);
29 }
30 
31 void xdp_del_sk_umem(struct xdp_umem *umem, struct xdp_sock *xs)
32 {
33 	unsigned long flags;
34 
35 	spin_lock_irqsave(&umem->xsk_list_lock, flags);
36 	list_del_rcu(&xs->list);
37 	spin_unlock_irqrestore(&umem->xsk_list_lock, flags);
38 }
39 
40 /* The umem is stored both in the _rx struct and the _tx struct as we do
41  * not know if the device has more tx queues than rx, or the opposite.
42  * This might also change during run time.
43  */
44 static void xdp_reg_umem_at_qid(struct net_device *dev, struct xdp_umem *umem,
45 				u16 queue_id)
46 {
47 	if (queue_id < dev->real_num_rx_queues)
48 		dev->_rx[queue_id].umem = umem;
49 	if (queue_id < dev->real_num_tx_queues)
50 		dev->_tx[queue_id].umem = umem;
51 }
52 
53 struct xdp_umem *xdp_get_umem_from_qid(struct net_device *dev,
54 				       u16 queue_id)
55 {
56 	if (queue_id < dev->real_num_rx_queues)
57 		return dev->_rx[queue_id].umem;
58 	if (queue_id < dev->real_num_tx_queues)
59 		return dev->_tx[queue_id].umem;
60 
61 	return NULL;
62 }
63 
64 static void xdp_clear_umem_at_qid(struct net_device *dev, u16 queue_id)
65 {
66 	if (queue_id < dev->real_num_rx_queues)
67 		dev->_rx[queue_id].umem = NULL;
68 	if (queue_id < dev->real_num_tx_queues)
69 		dev->_tx[queue_id].umem = NULL;
70 }
71 
72 int xdp_umem_assign_dev(struct xdp_umem *umem, struct net_device *dev,
73 			u16 queue_id, u16 flags)
74 {
75 	bool force_zc, force_copy;
76 	struct netdev_bpf bpf;
77 	int err = 0;
78 
79 	force_zc = flags & XDP_ZEROCOPY;
80 	force_copy = flags & XDP_COPY;
81 
82 	if (force_zc && force_copy)
83 		return -EINVAL;
84 
85 	rtnl_lock();
86 	if (xdp_get_umem_from_qid(dev, queue_id)) {
87 		err = -EBUSY;
88 		goto out_rtnl_unlock;
89 	}
90 
91 	xdp_reg_umem_at_qid(dev, umem, queue_id);
92 	umem->dev = dev;
93 	umem->queue_id = queue_id;
94 	if (force_copy)
95 		/* For copy-mode, we are done. */
96 		goto out_rtnl_unlock;
97 
98 	if (!dev->netdev_ops->ndo_bpf ||
99 	    !dev->netdev_ops->ndo_xsk_async_xmit) {
100 		err = -EOPNOTSUPP;
101 		goto err_unreg_umem;
102 	}
103 
104 	bpf.command = XDP_SETUP_XSK_UMEM;
105 	bpf.xsk.umem = umem;
106 	bpf.xsk.queue_id = queue_id;
107 
108 	err = dev->netdev_ops->ndo_bpf(dev, &bpf);
109 	if (err)
110 		goto err_unreg_umem;
111 	rtnl_unlock();
112 
113 	dev_hold(dev);
114 	umem->zc = true;
115 	return 0;
116 
117 err_unreg_umem:
118 	xdp_clear_umem_at_qid(dev, queue_id);
119 	if (!force_zc)
120 		err = 0; /* fallback to copy mode */
121 out_rtnl_unlock:
122 	rtnl_unlock();
123 	return err;
124 }
125 
126 static void xdp_umem_clear_dev(struct xdp_umem *umem)
127 {
128 	struct netdev_bpf bpf;
129 	int err;
130 
131 	if (umem->zc) {
132 		bpf.command = XDP_SETUP_XSK_UMEM;
133 		bpf.xsk.umem = NULL;
134 		bpf.xsk.queue_id = umem->queue_id;
135 
136 		rtnl_lock();
137 		err = umem->dev->netdev_ops->ndo_bpf(umem->dev, &bpf);
138 		rtnl_unlock();
139 
140 		if (err)
141 			WARN(1, "failed to disable umem!\n");
142 	}
143 
144 	if (umem->dev) {
145 		rtnl_lock();
146 		xdp_clear_umem_at_qid(umem->dev, umem->queue_id);
147 		rtnl_unlock();
148 	}
149 
150 	if (umem->zc) {
151 		dev_put(umem->dev);
152 		umem->zc = false;
153 	}
154 }
155 
156 static void xdp_umem_unpin_pages(struct xdp_umem *umem)
157 {
158 	unsigned int i;
159 
160 	for (i = 0; i < umem->npgs; i++) {
161 		struct page *page = umem->pgs[i];
162 
163 		set_page_dirty_lock(page);
164 		put_page(page);
165 	}
166 
167 	kfree(umem->pgs);
168 	umem->pgs = NULL;
169 }
170 
171 static void xdp_umem_unaccount_pages(struct xdp_umem *umem)
172 {
173 	if (umem->user) {
174 		atomic_long_sub(umem->npgs, &umem->user->locked_vm);
175 		free_uid(umem->user);
176 	}
177 }
178 
179 static void xdp_umem_release(struct xdp_umem *umem)
180 {
181 	struct task_struct *task;
182 	struct mm_struct *mm;
183 
184 	xdp_umem_clear_dev(umem);
185 
186 	if (umem->fq) {
187 		xskq_destroy(umem->fq);
188 		umem->fq = NULL;
189 	}
190 
191 	if (umem->cq) {
192 		xskq_destroy(umem->cq);
193 		umem->cq = NULL;
194 	}
195 
196 	xsk_reuseq_destroy(umem);
197 
198 	xdp_umem_unpin_pages(umem);
199 
200 	task = get_pid_task(umem->pid, PIDTYPE_PID);
201 	put_pid(umem->pid);
202 	if (!task)
203 		goto out;
204 	mm = get_task_mm(task);
205 	put_task_struct(task);
206 	if (!mm)
207 		goto out;
208 
209 	mmput(mm);
210 	kfree(umem->pages);
211 	umem->pages = NULL;
212 
213 	xdp_umem_unaccount_pages(umem);
214 out:
215 	kfree(umem);
216 }
217 
218 static void xdp_umem_release_deferred(struct work_struct *work)
219 {
220 	struct xdp_umem *umem = container_of(work, struct xdp_umem, work);
221 
222 	xdp_umem_release(umem);
223 }
224 
225 void xdp_get_umem(struct xdp_umem *umem)
226 {
227 	refcount_inc(&umem->users);
228 }
229 
230 void xdp_put_umem(struct xdp_umem *umem)
231 {
232 	if (!umem)
233 		return;
234 
235 	if (refcount_dec_and_test(&umem->users)) {
236 		INIT_WORK(&umem->work, xdp_umem_release_deferred);
237 		schedule_work(&umem->work);
238 	}
239 }
240 
241 static int xdp_umem_pin_pages(struct xdp_umem *umem)
242 {
243 	unsigned int gup_flags = FOLL_WRITE;
244 	long npgs;
245 	int err;
246 
247 	umem->pgs = kcalloc(umem->npgs, sizeof(*umem->pgs),
248 			    GFP_KERNEL | __GFP_NOWARN);
249 	if (!umem->pgs)
250 		return -ENOMEM;
251 
252 	down_write(&current->mm->mmap_sem);
253 	npgs = get_user_pages(umem->address, umem->npgs,
254 			      gup_flags, &umem->pgs[0], NULL);
255 	up_write(&current->mm->mmap_sem);
256 
257 	if (npgs != umem->npgs) {
258 		if (npgs >= 0) {
259 			umem->npgs = npgs;
260 			err = -ENOMEM;
261 			goto out_pin;
262 		}
263 		err = npgs;
264 		goto out_pgs;
265 	}
266 	return 0;
267 
268 out_pin:
269 	xdp_umem_unpin_pages(umem);
270 out_pgs:
271 	kfree(umem->pgs);
272 	umem->pgs = NULL;
273 	return err;
274 }
275 
276 static int xdp_umem_account_pages(struct xdp_umem *umem)
277 {
278 	unsigned long lock_limit, new_npgs, old_npgs;
279 
280 	if (capable(CAP_IPC_LOCK))
281 		return 0;
282 
283 	lock_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
284 	umem->user = get_uid(current_user());
285 
286 	do {
287 		old_npgs = atomic_long_read(&umem->user->locked_vm);
288 		new_npgs = old_npgs + umem->npgs;
289 		if (new_npgs > lock_limit) {
290 			free_uid(umem->user);
291 			umem->user = NULL;
292 			return -ENOBUFS;
293 		}
294 	} while (atomic_long_cmpxchg(&umem->user->locked_vm, old_npgs,
295 				     new_npgs) != old_npgs);
296 	return 0;
297 }
298 
299 static int xdp_umem_reg(struct xdp_umem *umem, struct xdp_umem_reg *mr)
300 {
301 	u32 chunk_size = mr->chunk_size, headroom = mr->headroom;
302 	unsigned int chunks, chunks_per_page;
303 	u64 addr = mr->addr, size = mr->len;
304 	int size_chk, err, i;
305 
306 	if (chunk_size < XDP_UMEM_MIN_CHUNK_SIZE || chunk_size > PAGE_SIZE) {
307 		/* Strictly speaking we could support this, if:
308 		 * - huge pages, or*
309 		 * - using an IOMMU, or
310 		 * - making sure the memory area is consecutive
311 		 * but for now, we simply say "computer says no".
312 		 */
313 		return -EINVAL;
314 	}
315 
316 	if (!is_power_of_2(chunk_size))
317 		return -EINVAL;
318 
319 	if (!PAGE_ALIGNED(addr)) {
320 		/* Memory area has to be page size aligned. For
321 		 * simplicity, this might change.
322 		 */
323 		return -EINVAL;
324 	}
325 
326 	if ((addr + size) < addr)
327 		return -EINVAL;
328 
329 	chunks = (unsigned int)div_u64(size, chunk_size);
330 	if (chunks == 0)
331 		return -EINVAL;
332 
333 	chunks_per_page = PAGE_SIZE / chunk_size;
334 	if (chunks < chunks_per_page || chunks % chunks_per_page)
335 		return -EINVAL;
336 
337 	headroom = ALIGN(headroom, 64);
338 
339 	size_chk = chunk_size - headroom - XDP_PACKET_HEADROOM;
340 	if (size_chk < 0)
341 		return -EINVAL;
342 
343 	umem->pid = get_task_pid(current, PIDTYPE_PID);
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 		goto out;
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 out:
379 	put_pid(umem->pid);
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 = xdp_umem_reg(umem, mr);
393 	if (err) {
394 		kfree(umem);
395 		return ERR_PTR(err);
396 	}
397 
398 	return umem;
399 }
400 
401 bool xdp_umem_validate_queues(struct xdp_umem *umem)
402 {
403 	return umem->fq && umem->cq;
404 }
405