1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3 * Devmem TCP
4 *
5 * Authors: Mina Almasry <almasrymina@google.com>
6 * Willem de Bruijn <willemdebruijn.kernel@gmail.com>
7 * Kaiyuan Zhang <kaiyuanz@google.com
8 */
9
10 #include <linux/dma-buf.h>
11 #include <linux/genalloc.h>
12 #include <linux/mm.h>
13 #include <linux/netdevice.h>
14 #include <linux/types.h>
15 #include <net/netdev_queues.h>
16 #include <net/netdev_rx_queue.h>
17 #include <net/page_pool/helpers.h>
18 #include <net/page_pool/memory_provider.h>
19 #include <net/sock.h>
20 #include <net/tcp.h>
21 #include <trace/events/page_pool.h>
22
23 #include "devmem.h"
24 #include "mp_dmabuf_devmem.h"
25 #include "page_pool_priv.h"
26
27 /* Device memory support */
28
29 static DEFINE_XARRAY_FLAGS(net_devmem_dmabuf_bindings, XA_FLAGS_ALLOC1);
30
31 static const struct memory_provider_ops dmabuf_devmem_ops;
32
net_devmem_dmabuf_free_chunk_owner(struct gen_pool * genpool,struct gen_pool_chunk * chunk,void * not_used)33 static void net_devmem_dmabuf_free_chunk_owner(struct gen_pool *genpool,
34 struct gen_pool_chunk *chunk,
35 void *not_used)
36 {
37 struct dmabuf_genpool_chunk_owner *owner = chunk->owner;
38
39 kvfree(owner->area.niovs);
40 kfree(owner);
41 }
42
net_devmem_get_dma_addr(const struct net_iov * niov)43 static dma_addr_t net_devmem_get_dma_addr(const struct net_iov *niov)
44 {
45 struct dmabuf_genpool_chunk_owner *owner;
46
47 owner = net_devmem_iov_to_chunk_owner(niov);
48 return owner->base_dma_addr +
49 ((dma_addr_t)net_iov_idx(niov) << PAGE_SHIFT);
50 }
51
net_devmem_dmabuf_binding_release(struct percpu_ref * ref)52 static void net_devmem_dmabuf_binding_release(struct percpu_ref *ref)
53 {
54 struct net_devmem_dmabuf_binding *binding =
55 container_of(ref, struct net_devmem_dmabuf_binding, ref);
56
57 INIT_WORK(&binding->unbind_w, __net_devmem_dmabuf_binding_free);
58 schedule_work(&binding->unbind_w);
59 }
60
__net_devmem_dmabuf_binding_free(struct work_struct * wq)61 void __net_devmem_dmabuf_binding_free(struct work_struct *wq)
62 {
63 struct net_devmem_dmabuf_binding *binding = container_of(wq, typeof(*binding), unbind_w);
64
65 size_t size, avail;
66
67 gen_pool_for_each_chunk(binding->chunk_pool,
68 net_devmem_dmabuf_free_chunk_owner, NULL);
69
70 size = gen_pool_size(binding->chunk_pool);
71 avail = gen_pool_avail(binding->chunk_pool);
72
73 if (!WARN(size != avail, "can't destroy genpool. size=%zu, avail=%zu",
74 size, avail))
75 gen_pool_destroy(binding->chunk_pool);
76
77 dma_buf_unmap_attachment_unlocked(binding->attachment, binding->sgt,
78 binding->direction);
79 dma_buf_detach(binding->dmabuf, binding->attachment);
80 dma_buf_put(binding->dmabuf);
81 xa_destroy(&binding->bound_rxqs);
82 percpu_ref_exit(&binding->ref);
83 kvfree(binding->tx_vec);
84 kfree(binding);
85 }
86
87 struct net_iov *
net_devmem_alloc_dmabuf(struct net_devmem_dmabuf_binding * binding)88 net_devmem_alloc_dmabuf(struct net_devmem_dmabuf_binding *binding)
89 {
90 struct dmabuf_genpool_chunk_owner *owner;
91 unsigned long dma_addr;
92 struct net_iov *niov;
93 ssize_t offset;
94 ssize_t index;
95
96 dma_addr = gen_pool_alloc_owner(binding->chunk_pool, PAGE_SIZE,
97 (void **)&owner);
98 if (!dma_addr)
99 return NULL;
100
101 offset = dma_addr - owner->base_dma_addr;
102 index = offset / PAGE_SIZE;
103 niov = &owner->area.niovs[index];
104
105 niov->desc.pp_magic = 0;
106 niov->desc.pp = NULL;
107 atomic_long_set(&niov->desc.pp_ref_count, 0);
108
109 return niov;
110 }
111
net_devmem_free_dmabuf(struct net_iov * niov)112 void net_devmem_free_dmabuf(struct net_iov *niov)
113 {
114 struct net_devmem_dmabuf_binding *binding = net_devmem_iov_binding(niov);
115 unsigned long dma_addr = net_devmem_get_dma_addr(niov);
116
117 if (WARN_ON(!gen_pool_has_addr(binding->chunk_pool, dma_addr,
118 PAGE_SIZE)))
119 return;
120
121 gen_pool_free(binding->chunk_pool, dma_addr, PAGE_SIZE);
122 }
123
net_devmem_unbind_dmabuf(struct net_devmem_dmabuf_binding * binding)124 void net_devmem_unbind_dmabuf(struct net_devmem_dmabuf_binding *binding)
125 {
126 struct netdev_rx_queue *rxq;
127 unsigned long xa_idx;
128 unsigned int rxq_idx;
129
130 xa_erase(&net_devmem_dmabuf_bindings, binding->id);
131
132 /* Ensure no tx net_devmem_lookup_dmabuf() are in flight after the
133 * erase.
134 */
135 synchronize_net();
136
137 if (binding->list.next)
138 list_del(&binding->list);
139
140 xa_for_each(&binding->bound_rxqs, xa_idx, rxq) {
141 const struct pp_memory_provider_params mp_params = {
142 .mp_priv = binding,
143 .mp_ops = &dmabuf_devmem_ops,
144 };
145
146 rxq_idx = get_netdev_rx_queue_index(rxq);
147
148 __net_mp_close_rxq(binding->dev, rxq_idx, &mp_params);
149 }
150
151 percpu_ref_kill(&binding->ref);
152 }
153
net_devmem_bind_dmabuf_to_queue(struct net_device * dev,u32 rxq_idx,struct net_devmem_dmabuf_binding * binding,struct netlink_ext_ack * extack)154 int net_devmem_bind_dmabuf_to_queue(struct net_device *dev, u32 rxq_idx,
155 struct net_devmem_dmabuf_binding *binding,
156 struct netlink_ext_ack *extack)
157 {
158 struct pp_memory_provider_params mp_params = {
159 .mp_priv = binding,
160 .mp_ops = &dmabuf_devmem_ops,
161 };
162 struct netdev_rx_queue *rxq;
163 u32 xa_idx;
164 int err;
165
166 err = __net_mp_open_rxq(dev, rxq_idx, &mp_params, extack);
167 if (err)
168 return err;
169
170 rxq = __netif_get_rx_queue(dev, rxq_idx);
171 err = xa_alloc(&binding->bound_rxqs, &xa_idx, rxq, xa_limit_32b,
172 GFP_KERNEL);
173 if (err)
174 goto err_close_rxq;
175
176 return 0;
177
178 err_close_rxq:
179 __net_mp_close_rxq(dev, rxq_idx, &mp_params);
180 return err;
181 }
182
183 struct net_devmem_dmabuf_binding *
net_devmem_bind_dmabuf(struct net_device * dev,struct device * dma_dev,enum dma_data_direction direction,unsigned int dmabuf_fd,struct netdev_nl_sock * priv,struct netlink_ext_ack * extack)184 net_devmem_bind_dmabuf(struct net_device *dev,
185 struct device *dma_dev,
186 enum dma_data_direction direction,
187 unsigned int dmabuf_fd, struct netdev_nl_sock *priv,
188 struct netlink_ext_ack *extack)
189 {
190 struct net_devmem_dmabuf_binding *binding;
191 static u32 id_alloc_next;
192 struct scatterlist *sg;
193 struct dma_buf *dmabuf;
194 unsigned int sg_idx, i;
195 unsigned long virtual;
196 int err;
197
198 if (!dma_dev) {
199 NL_SET_ERR_MSG(extack, "Device doesn't support DMA");
200 return ERR_PTR(-EOPNOTSUPP);
201 }
202
203 dmabuf = dma_buf_get(dmabuf_fd);
204 if (IS_ERR(dmabuf))
205 return ERR_CAST(dmabuf);
206
207 binding = kzalloc_node(sizeof(*binding), GFP_KERNEL,
208 dev_to_node(&dev->dev));
209 if (!binding) {
210 err = -ENOMEM;
211 goto err_put_dmabuf;
212 }
213
214 binding->dev = dev;
215 xa_init_flags(&binding->bound_rxqs, XA_FLAGS_ALLOC);
216
217 err = percpu_ref_init(&binding->ref,
218 net_devmem_dmabuf_binding_release,
219 0, GFP_KERNEL);
220 if (err < 0)
221 goto err_free_binding;
222
223 mutex_init(&binding->lock);
224
225 binding->dmabuf = dmabuf;
226 binding->direction = direction;
227
228 binding->attachment = dma_buf_attach(binding->dmabuf, dma_dev);
229 if (IS_ERR(binding->attachment)) {
230 err = PTR_ERR(binding->attachment);
231 NL_SET_ERR_MSG(extack, "Failed to bind dmabuf to device");
232 goto err_exit_ref;
233 }
234
235 binding->sgt = dma_buf_map_attachment_unlocked(binding->attachment,
236 direction);
237 if (IS_ERR(binding->sgt)) {
238 err = PTR_ERR(binding->sgt);
239 NL_SET_ERR_MSG(extack, "Failed to map dmabuf attachment");
240 goto err_detach;
241 }
242
243 if (direction == DMA_TO_DEVICE) {
244 binding->tx_vec = kvmalloc_objs(struct net_iov *,
245 dmabuf->size / PAGE_SIZE);
246 if (!binding->tx_vec) {
247 err = -ENOMEM;
248 goto err_unmap;
249 }
250 }
251
252 /* For simplicity we expect to make PAGE_SIZE allocations, but the
253 * binding can be much more flexible than that. We may be able to
254 * allocate MTU sized chunks here. Leave that for future work...
255 */
256 binding->chunk_pool = gen_pool_create(PAGE_SHIFT,
257 dev_to_node(&dev->dev));
258 if (!binding->chunk_pool) {
259 err = -ENOMEM;
260 goto err_tx_vec;
261 }
262
263 virtual = 0;
264 for_each_sgtable_dma_sg(binding->sgt, sg, sg_idx) {
265 dma_addr_t dma_addr = sg_dma_address(sg);
266 struct dmabuf_genpool_chunk_owner *owner;
267 size_t len = sg_dma_len(sg);
268 struct net_iov *niov;
269
270 owner = kzalloc_node(sizeof(*owner), GFP_KERNEL,
271 dev_to_node(&dev->dev));
272 if (!owner) {
273 err = -ENOMEM;
274 goto err_free_chunks;
275 }
276
277 owner->area.base_virtual = virtual;
278 owner->base_dma_addr = dma_addr;
279 owner->area.num_niovs = len / PAGE_SIZE;
280 owner->binding = binding;
281
282 err = gen_pool_add_owner(binding->chunk_pool, dma_addr,
283 dma_addr, len, dev_to_node(&dev->dev),
284 owner);
285 if (err) {
286 kfree(owner);
287 err = -EINVAL;
288 goto err_free_chunks;
289 }
290
291 owner->area.niovs = kvmalloc_objs(*owner->area.niovs,
292 owner->area.num_niovs);
293 if (!owner->area.niovs) {
294 err = -ENOMEM;
295 goto err_free_chunks;
296 }
297
298 for (i = 0; i < owner->area.num_niovs; i++) {
299 niov = &owner->area.niovs[i];
300 niov->type = NET_IOV_DMABUF;
301 niov->owner = &owner->area;
302 page_pool_set_dma_addr_netmem(net_iov_to_netmem(niov),
303 net_devmem_get_dma_addr(niov));
304 if (direction == DMA_TO_DEVICE)
305 binding->tx_vec[owner->area.base_virtual / PAGE_SIZE + i] = niov;
306 }
307
308 virtual += len;
309 }
310
311 err = xa_alloc_cyclic(&net_devmem_dmabuf_bindings, &binding->id,
312 binding, xa_limit_32b, &id_alloc_next,
313 GFP_KERNEL);
314 if (err < 0)
315 goto err_free_chunks;
316
317 list_add(&binding->list, &priv->bindings);
318
319 return binding;
320
321 err_free_chunks:
322 gen_pool_for_each_chunk(binding->chunk_pool,
323 net_devmem_dmabuf_free_chunk_owner, NULL);
324 gen_pool_destroy(binding->chunk_pool);
325 err_tx_vec:
326 kvfree(binding->tx_vec);
327 err_unmap:
328 dma_buf_unmap_attachment_unlocked(binding->attachment, binding->sgt,
329 direction);
330 err_detach:
331 dma_buf_detach(dmabuf, binding->attachment);
332 err_exit_ref:
333 percpu_ref_exit(&binding->ref);
334 err_free_binding:
335 kfree(binding);
336 err_put_dmabuf:
337 dma_buf_put(dmabuf);
338 return ERR_PTR(err);
339 }
340
net_devmem_lookup_dmabuf(u32 id)341 struct net_devmem_dmabuf_binding *net_devmem_lookup_dmabuf(u32 id)
342 {
343 struct net_devmem_dmabuf_binding *binding;
344
345 rcu_read_lock();
346 binding = xa_load(&net_devmem_dmabuf_bindings, id);
347 if (binding) {
348 if (!net_devmem_dmabuf_binding_get(binding))
349 binding = NULL;
350 }
351 rcu_read_unlock();
352
353 return binding;
354 }
355
net_devmem_get_net_iov(struct net_iov * niov)356 void net_devmem_get_net_iov(struct net_iov *niov)
357 {
358 net_devmem_dmabuf_binding_get(net_devmem_iov_binding(niov));
359 }
360
net_devmem_put_net_iov(struct net_iov * niov)361 void net_devmem_put_net_iov(struct net_iov *niov)
362 {
363 net_devmem_dmabuf_binding_put(net_devmem_iov_binding(niov));
364 }
365
net_devmem_get_binding(struct sock * sk,unsigned int dmabuf_id)366 struct net_devmem_dmabuf_binding *net_devmem_get_binding(struct sock *sk,
367 unsigned int dmabuf_id)
368 {
369 struct net_devmem_dmabuf_binding *binding;
370 struct net_device *dst_dev;
371 struct dst_entry *dst;
372 int err = 0;
373
374 binding = net_devmem_lookup_dmabuf(dmabuf_id);
375 if (!binding || !binding->tx_vec) {
376 err = -EINVAL;
377 goto out_err;
378 }
379
380 rcu_read_lock();
381 dst = __sk_dst_get(sk);
382 /* If dst is NULL (route expired), attempt to rebuild it. */
383 if (unlikely(!dst)) {
384 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) {
385 err = -EHOSTUNREACH;
386 goto out_unlock;
387 }
388 dst = __sk_dst_get(sk);
389 if (unlikely(!dst)) {
390 err = -ENODEV;
391 goto out_unlock;
392 }
393 }
394
395 /* The dma-addrs in this binding are only reachable to the corresponding
396 * net_device.
397 */
398 dst_dev = dst_dev_rcu(dst);
399 if (unlikely(!dst_dev) || unlikely(dst_dev != binding->dev)) {
400 err = -ENODEV;
401 goto out_unlock;
402 }
403
404 rcu_read_unlock();
405 return binding;
406
407 out_unlock:
408 rcu_read_unlock();
409 out_err:
410 if (binding)
411 net_devmem_dmabuf_binding_put(binding);
412
413 return ERR_PTR(err);
414 }
415
416 struct net_iov *
net_devmem_get_niov_at(struct net_devmem_dmabuf_binding * binding,size_t virt_addr,size_t * off,size_t * size)417 net_devmem_get_niov_at(struct net_devmem_dmabuf_binding *binding,
418 size_t virt_addr, size_t *off, size_t *size)
419 {
420 if (virt_addr >= binding->dmabuf->size)
421 return NULL;
422
423 *off = virt_addr % PAGE_SIZE;
424 *size = PAGE_SIZE - *off;
425
426 return binding->tx_vec[virt_addr / PAGE_SIZE];
427 }
428
429 /*** "Dmabuf devmem memory provider" ***/
430
mp_dmabuf_devmem_init(struct page_pool * pool)431 int mp_dmabuf_devmem_init(struct page_pool *pool)
432 {
433 struct net_devmem_dmabuf_binding *binding = pool->mp_priv;
434
435 if (!binding)
436 return -EINVAL;
437
438 /* dma-buf dma addresses do not need and should not be used with
439 * dma_sync_for_cpu/device. Force disable dma_sync.
440 */
441 pool->dma_sync = false;
442 pool->dma_sync_for_cpu = false;
443
444 if (pool->p.order != 0)
445 return -E2BIG;
446
447 net_devmem_dmabuf_binding_get(binding);
448 return 0;
449 }
450
mp_dmabuf_devmem_alloc_netmems(struct page_pool * pool,gfp_t gfp)451 netmem_ref mp_dmabuf_devmem_alloc_netmems(struct page_pool *pool, gfp_t gfp)
452 {
453 struct net_devmem_dmabuf_binding *binding = pool->mp_priv;
454 struct net_iov *niov;
455 netmem_ref netmem;
456
457 niov = net_devmem_alloc_dmabuf(binding);
458 if (!niov)
459 return 0;
460
461 netmem = net_iov_to_netmem(niov);
462
463 page_pool_set_pp_info(pool, netmem);
464
465 pool->pages_state_hold_cnt++;
466 trace_page_pool_state_hold(pool, netmem, pool->pages_state_hold_cnt);
467 return netmem;
468 }
469
mp_dmabuf_devmem_destroy(struct page_pool * pool)470 void mp_dmabuf_devmem_destroy(struct page_pool *pool)
471 {
472 struct net_devmem_dmabuf_binding *binding = pool->mp_priv;
473
474 net_devmem_dmabuf_binding_put(binding);
475 }
476
mp_dmabuf_devmem_release_page(struct page_pool * pool,netmem_ref netmem)477 bool mp_dmabuf_devmem_release_page(struct page_pool *pool, netmem_ref netmem)
478 {
479 long refcount = atomic_long_read(netmem_get_pp_ref_count_ref(netmem));
480
481 if (WARN_ON_ONCE(!netmem_is_net_iov(netmem)))
482 return false;
483
484 if (WARN_ON_ONCE(refcount != 1))
485 return false;
486
487 page_pool_clear_pp_info(netmem);
488
489 net_devmem_free_dmabuf(netmem_to_net_iov(netmem));
490
491 /* We don't want the page pool put_page()ing our net_iovs. */
492 return false;
493 }
494
mp_dmabuf_devmem_nl_fill(void * mp_priv,struct sk_buff * rsp,struct netdev_rx_queue * rxq)495 static int mp_dmabuf_devmem_nl_fill(void *mp_priv, struct sk_buff *rsp,
496 struct netdev_rx_queue *rxq)
497 {
498 const struct net_devmem_dmabuf_binding *binding = mp_priv;
499 int type = rxq ? NETDEV_A_QUEUE_DMABUF : NETDEV_A_PAGE_POOL_DMABUF;
500
501 return nla_put_u32(rsp, type, binding->id);
502 }
503
mp_dmabuf_devmem_uninstall(void * mp_priv,struct netdev_rx_queue * rxq)504 static void mp_dmabuf_devmem_uninstall(void *mp_priv,
505 struct netdev_rx_queue *rxq)
506 {
507 struct net_devmem_dmabuf_binding *binding = mp_priv;
508 struct netdev_rx_queue *bound_rxq;
509 unsigned long xa_idx;
510
511 xa_for_each(&binding->bound_rxqs, xa_idx, bound_rxq) {
512 if (bound_rxq == rxq) {
513 xa_erase(&binding->bound_rxqs, xa_idx);
514 if (xa_empty(&binding->bound_rxqs)) {
515 mutex_lock(&binding->lock);
516 binding->dev = NULL;
517 mutex_unlock(&binding->lock);
518 }
519 break;
520 }
521 }
522 }
523
524 static const struct memory_provider_ops dmabuf_devmem_ops = {
525 .init = mp_dmabuf_devmem_init,
526 .destroy = mp_dmabuf_devmem_destroy,
527 .alloc_netmems = mp_dmabuf_devmem_alloc_netmems,
528 .release_netmem = mp_dmabuf_devmem_release_page,
529 .nl_fill = mp_dmabuf_devmem_nl_fill,
530 .uninstall = mp_dmabuf_devmem_uninstall,
531 };
532