xref: /linux/io_uring/zcrx.c (revision 1e91c98bc9a8ef8198e73151b2a118cd3748925d)
1 // SPDX-License-Identifier: GPL-2.0
2 #include <linux/kernel.h>
3 #include <linux/errno.h>
4 #include <linux/dma-map-ops.h>
5 #include <linux/mm.h>
6 #include <linux/nospec.h>
7 #include <linux/io_uring.h>
8 #include <linux/netdevice.h>
9 #include <linux/rtnetlink.h>
10 #include <linux/skbuff_ref.h>
11 #include <linux/anon_inodes.h>
12 
13 #include <net/page_pool/helpers.h>
14 #include <net/page_pool/memory_provider.h>
15 #include <net/netlink.h>
16 #include <net/netdev_queues.h>
17 #include <net/netdev_rx_queue.h>
18 #include <net/tcp.h>
19 #include <net/rps.h>
20 
21 #include <trace/events/page_pool.h>
22 
23 #include <uapi/linux/io_uring.h>
24 
25 #include "io_uring.h"
26 #include "kbuf.h"
27 #include "memmap.h"
28 #include "zcrx.h"
29 #include "rsrc.h"
30 
31 #define IO_ZCRX_AREA_SUPPORTED_FLAGS	(IORING_ZCRX_AREA_DMABUF)
32 
33 #define IO_DMA_ATTR (DMA_ATTR_SKIP_CPU_SYNC | DMA_ATTR_WEAK_ORDERING)
34 
35 static inline struct io_zcrx_ifq *io_pp_to_ifq(struct page_pool *pp)
36 {
37 	return pp->mp_priv;
38 }
39 
40 static inline struct io_zcrx_area *io_zcrx_iov_to_area(const struct net_iov *niov)
41 {
42 	struct net_iov_area *owner = net_iov_owner(niov);
43 
44 	return container_of(owner, struct io_zcrx_area, nia);
45 }
46 
47 static inline struct page *io_zcrx_iov_page(const struct net_iov *niov)
48 {
49 	struct io_zcrx_area *area = io_zcrx_iov_to_area(niov);
50 	unsigned niov_pages_shift;
51 
52 	lockdep_assert(!area->mem.is_dmabuf);
53 
54 	niov_pages_shift = area->ifq->niov_shift - PAGE_SHIFT;
55 	return area->mem.pages[net_iov_idx(niov) << niov_pages_shift];
56 }
57 
58 static int io_area_max_shift(struct io_zcrx_mem *mem)
59 {
60 	struct sg_table *sgt = mem->sgt;
61 	struct scatterlist *sg;
62 	unsigned shift = -1U;
63 	unsigned i;
64 
65 	for_each_sgtable_dma_sg(sgt, sg, i)
66 		shift = min(shift, __ffs(sg->length));
67 	return shift;
68 }
69 
70 static int io_populate_area_dma(struct io_zcrx_ifq *ifq,
71 				struct io_zcrx_area *area)
72 {
73 	unsigned niov_size = 1U << ifq->niov_shift;
74 	struct sg_table *sgt = area->mem.sgt;
75 	struct scatterlist *sg;
76 	unsigned i, niov_idx = 0;
77 
78 	for_each_sgtable_dma_sg(sgt, sg, i) {
79 		dma_addr_t dma = sg_dma_address(sg);
80 		unsigned long sg_len = sg_dma_len(sg);
81 
82 		if (WARN_ON_ONCE(sg_len % niov_size))
83 			return -EINVAL;
84 
85 		while (sg_len && niov_idx < area->nia.num_niovs) {
86 			struct net_iov *niov = &area->nia.niovs[niov_idx];
87 
88 			if (net_mp_niov_set_dma_addr(niov, dma))
89 				return -EFAULT;
90 			sg_len -= niov_size;
91 			dma += niov_size;
92 			niov_idx++;
93 		}
94 	}
95 
96 	if (WARN_ON_ONCE(niov_idx != area->nia.num_niovs))
97 		return -EFAULT;
98 	return 0;
99 }
100 
101 static void io_release_dmabuf(struct io_zcrx_mem *mem)
102 {
103 	if (!IS_ENABLED(CONFIG_DMA_SHARED_BUFFER))
104 		return;
105 
106 	if (mem->sgt)
107 		dma_buf_unmap_attachment_unlocked(mem->attach, mem->sgt,
108 						  DMA_FROM_DEVICE);
109 	if (mem->attach)
110 		dma_buf_detach(mem->dmabuf, mem->attach);
111 	if (mem->dmabuf)
112 		dma_buf_put(mem->dmabuf);
113 
114 	mem->sgt = NULL;
115 	mem->attach = NULL;
116 	mem->dmabuf = NULL;
117 }
118 
119 static int io_import_dmabuf(struct io_zcrx_ifq *ifq,
120 			    struct io_zcrx_mem *mem,
121 			    struct io_uring_zcrx_area_reg *area_reg)
122 {
123 	unsigned long off = (unsigned long)area_reg->addr;
124 	unsigned long len = (unsigned long)area_reg->len;
125 	unsigned long total_size = 0;
126 	struct scatterlist *sg;
127 	int dmabuf_fd = area_reg->dmabuf_fd;
128 	int i, ret;
129 
130 	if (off)
131 		return -EINVAL;
132 	if (WARN_ON_ONCE(!ifq->dev))
133 		return -EFAULT;
134 	if (!IS_ENABLED(CONFIG_DMA_SHARED_BUFFER))
135 		return -EINVAL;
136 
137 	mem->is_dmabuf = true;
138 	mem->dmabuf = dma_buf_get(dmabuf_fd);
139 	if (IS_ERR(mem->dmabuf)) {
140 		ret = PTR_ERR(mem->dmabuf);
141 		mem->dmabuf = NULL;
142 		goto err;
143 	}
144 
145 	mem->attach = dma_buf_attach(mem->dmabuf, ifq->dev);
146 	if (IS_ERR(mem->attach)) {
147 		ret = PTR_ERR(mem->attach);
148 		mem->attach = NULL;
149 		goto err;
150 	}
151 
152 	mem->sgt = dma_buf_map_attachment_unlocked(mem->attach, DMA_FROM_DEVICE);
153 	if (IS_ERR(mem->sgt)) {
154 		ret = PTR_ERR(mem->sgt);
155 		mem->sgt = NULL;
156 		goto err;
157 	}
158 
159 	for_each_sgtable_dma_sg(mem->sgt, sg, i)
160 		total_size += sg_dma_len(sg);
161 
162 	if (total_size != len) {
163 		ret = -EINVAL;
164 		goto err;
165 	}
166 
167 	mem->size = len;
168 	return 0;
169 err:
170 	io_release_dmabuf(mem);
171 	return ret;
172 }
173 
174 static unsigned long io_count_account_pages(struct page **pages, unsigned nr_pages)
175 {
176 	struct folio *last_folio = NULL;
177 	unsigned long res = 0;
178 	int i;
179 
180 	for (i = 0; i < nr_pages; i++) {
181 		struct folio *folio = page_folio(pages[i]);
182 
183 		if (folio == last_folio)
184 			continue;
185 		last_folio = folio;
186 		res += folio_nr_pages(folio);
187 	}
188 	return res;
189 }
190 
191 static int io_import_umem(struct io_zcrx_ifq *ifq,
192 			  struct io_zcrx_mem *mem,
193 			  struct io_uring_zcrx_area_reg *area_reg)
194 {
195 	struct page **pages;
196 	int nr_pages, ret;
197 
198 	if (area_reg->dmabuf_fd)
199 		return -EINVAL;
200 	if (!area_reg->addr)
201 		return -EFAULT;
202 	pages = io_pin_pages((unsigned long)area_reg->addr, area_reg->len,
203 				   &nr_pages);
204 	if (IS_ERR(pages))
205 		return PTR_ERR(pages);
206 
207 	ret = sg_alloc_table_from_pages(&mem->page_sg_table, pages, nr_pages,
208 					0, (unsigned long)nr_pages << PAGE_SHIFT,
209 					GFP_KERNEL_ACCOUNT);
210 	if (ret) {
211 		unpin_user_pages(pages, nr_pages);
212 		kvfree(pages);
213 		return ret;
214 	}
215 
216 	mem->account_pages = io_count_account_pages(pages, nr_pages);
217 	ret = io_account_mem(ifq->user, ifq->mm_account, mem->account_pages);
218 	if (ret < 0)
219 		mem->account_pages = 0;
220 
221 	mem->sgt = &mem->page_sg_table;
222 	mem->pages = pages;
223 	mem->nr_folios = nr_pages;
224 	mem->size = area_reg->len;
225 	return ret;
226 }
227 
228 static void io_release_area_mem(struct io_zcrx_mem *mem)
229 {
230 	if (mem->is_dmabuf) {
231 		io_release_dmabuf(mem);
232 		return;
233 	}
234 	if (mem->pages) {
235 		unpin_user_pages(mem->pages, mem->nr_folios);
236 		sg_free_table(mem->sgt);
237 		mem->sgt = NULL;
238 		kvfree(mem->pages);
239 	}
240 }
241 
242 static int io_import_area(struct io_zcrx_ifq *ifq,
243 			  struct io_zcrx_mem *mem,
244 			  struct io_uring_zcrx_area_reg *area_reg)
245 {
246 	int ret;
247 
248 	if (area_reg->flags & ~IO_ZCRX_AREA_SUPPORTED_FLAGS)
249 		return -EINVAL;
250 	if (area_reg->rq_area_token)
251 		return -EINVAL;
252 	if (area_reg->__resv2[0] || area_reg->__resv2[1])
253 		return -EINVAL;
254 
255 	ret = io_validate_user_buf_range(area_reg->addr, area_reg->len);
256 	if (ret)
257 		return ret;
258 	if (area_reg->addr & ~PAGE_MASK || area_reg->len & ~PAGE_MASK)
259 		return -EINVAL;
260 
261 	if (area_reg->flags & IORING_ZCRX_AREA_DMABUF)
262 		return io_import_dmabuf(ifq, mem, area_reg);
263 	return io_import_umem(ifq, mem, area_reg);
264 }
265 
266 static void io_zcrx_unmap_area(struct io_zcrx_ifq *ifq,
267 				struct io_zcrx_area *area)
268 {
269 	int i;
270 
271 	guard(mutex)(&ifq->pp_lock);
272 	if (!area->is_mapped)
273 		return;
274 	area->is_mapped = false;
275 
276 	for (i = 0; i < area->nia.num_niovs; i++)
277 		net_mp_niov_set_dma_addr(&area->nia.niovs[i], 0);
278 
279 	if (area->mem.is_dmabuf) {
280 		io_release_dmabuf(&area->mem);
281 	} else {
282 		dma_unmap_sgtable(ifq->dev, &area->mem.page_sg_table,
283 				  DMA_FROM_DEVICE, IO_DMA_ATTR);
284 	}
285 }
286 
287 static int io_zcrx_map_area(struct io_zcrx_ifq *ifq, struct io_zcrx_area *area)
288 {
289 	int ret;
290 
291 	guard(mutex)(&ifq->pp_lock);
292 	if (area->is_mapped)
293 		return 0;
294 
295 	if (!area->mem.is_dmabuf) {
296 		ret = dma_map_sgtable(ifq->dev, &area->mem.page_sg_table,
297 				      DMA_FROM_DEVICE, IO_DMA_ATTR);
298 		if (ret < 0)
299 			return ret;
300 	}
301 
302 	ret = io_populate_area_dma(ifq, area);
303 	if (ret && !area->mem.is_dmabuf)
304 		dma_unmap_sgtable(ifq->dev, &area->mem.page_sg_table,
305 				  DMA_FROM_DEVICE, IO_DMA_ATTR);
306 	if (ret == 0)
307 		area->is_mapped = true;
308 	return ret;
309 }
310 
311 static void io_zcrx_sync_for_device(struct page_pool *pool,
312 				    struct net_iov *niov)
313 {
314 #if defined(CONFIG_HAS_DMA) && defined(CONFIG_DMA_NEED_SYNC)
315 	dma_addr_t dma_addr;
316 
317 	unsigned niov_size;
318 
319 	if (!dma_dev_need_sync(pool->p.dev))
320 		return;
321 
322 	niov_size = 1U << io_pp_to_ifq(pool)->niov_shift;
323 	dma_addr = page_pool_get_dma_addr_netmem(net_iov_to_netmem(niov));
324 	__dma_sync_single_for_device(pool->p.dev, dma_addr + pool->p.offset,
325 				     niov_size, pool->p.dma_dir);
326 #endif
327 }
328 
329 #define IO_RQ_MAX_ENTRIES		32768
330 
331 #define IO_SKBS_PER_CALL_LIMIT	20
332 
333 struct io_zcrx_args {
334 	struct io_kiocb		*req;
335 	struct io_zcrx_ifq	*ifq;
336 	struct socket		*sock;
337 	unsigned		nr_skbs;
338 };
339 
340 static const struct memory_provider_ops io_uring_pp_zc_ops;
341 
342 static inline atomic_t *io_get_user_counter(struct net_iov *niov)
343 {
344 	struct io_zcrx_area *area = io_zcrx_iov_to_area(niov);
345 
346 	return &area->user_refs[net_iov_idx(niov)];
347 }
348 
349 static bool io_zcrx_put_niov_uref(struct net_iov *niov)
350 {
351 	atomic_t *uref = io_get_user_counter(niov);
352 	int old;
353 
354 	old = atomic_read(uref);
355 	do {
356 		if (unlikely(old == 0))
357 			return false;
358 	} while (!atomic_try_cmpxchg(uref, &old, old - 1));
359 
360 	return true;
361 }
362 
363 static void io_zcrx_get_niov_uref(struct net_iov *niov)
364 {
365 	atomic_inc(io_get_user_counter(niov));
366 }
367 
368 static void io_fill_zcrx_offsets(struct io_uring_zcrx_offsets *offsets)
369 {
370 	offsets->head = offsetof(struct io_uring, head);
371 	offsets->tail = offsetof(struct io_uring, tail);
372 	offsets->rqes = ALIGN(sizeof(struct io_uring), L1_CACHE_BYTES);
373 }
374 
375 static int io_allocate_rbuf_ring(struct io_ring_ctx *ctx,
376 				 struct io_zcrx_ifq *ifq,
377 				 struct io_uring_zcrx_ifq_reg *reg,
378 				 struct io_uring_region_desc *rd,
379 				 u32 id)
380 {
381 	u64 mmap_offset;
382 	size_t off, size;
383 	void *ptr;
384 	int ret;
385 
386 	io_fill_zcrx_offsets(&reg->offsets);
387 	off = reg->offsets.rqes;
388 	size = off + sizeof(struct io_uring_zcrx_rqe) * reg->rq_entries;
389 	if (size > rd->size)
390 		return -EINVAL;
391 
392 	mmap_offset = IORING_MAP_OFF_ZCRX_REGION;
393 	mmap_offset += id << IORING_OFF_PBUF_SHIFT;
394 
395 	ret = io_create_region(ctx, &ifq->region, rd, mmap_offset);
396 	if (ret < 0)
397 		return ret;
398 
399 	ptr = io_region_get_ptr(&ifq->region);
400 	ifq->rq_ring = (struct io_uring *)ptr;
401 	ifq->rqes = (struct io_uring_zcrx_rqe *)(ptr + off);
402 
403 	return 0;
404 }
405 
406 static void io_free_rbuf_ring(struct io_zcrx_ifq *ifq)
407 {
408 	io_free_region(ifq->user, &ifq->region);
409 	ifq->rq_ring = NULL;
410 	ifq->rqes = NULL;
411 }
412 
413 static void io_zcrx_free_area(struct io_zcrx_ifq *ifq,
414 			      struct io_zcrx_area *area)
415 {
416 	io_zcrx_unmap_area(ifq, area);
417 	io_release_area_mem(&area->mem);
418 
419 	if (area->mem.account_pages)
420 		io_unaccount_mem(ifq->user, ifq->mm_account,
421 				 area->mem.account_pages);
422 
423 	kvfree(area->freelist);
424 	kvfree(area->nia.niovs);
425 	kvfree(area->user_refs);
426 	kfree(area);
427 }
428 
429 static int io_zcrx_append_area(struct io_zcrx_ifq *ifq,
430 				struct io_zcrx_area *area)
431 {
432 	if (ifq->area)
433 		return -EINVAL;
434 	ifq->area = area;
435 	return 0;
436 }
437 
438 static int io_zcrx_create_area(struct io_zcrx_ifq *ifq,
439 			       struct io_uring_zcrx_area_reg *area_reg,
440 			       struct io_uring_zcrx_ifq_reg *reg)
441 {
442 	int buf_size_shift = PAGE_SHIFT;
443 	struct io_zcrx_area *area;
444 	unsigned nr_iovs;
445 	int i, ret;
446 
447 	if (reg->rx_buf_len) {
448 		if (!is_power_of_2(reg->rx_buf_len) ||
449 		     reg->rx_buf_len < PAGE_SIZE)
450 			return -EINVAL;
451 		buf_size_shift = ilog2(reg->rx_buf_len);
452 	}
453 
454 	ret = -ENOMEM;
455 	area = kzalloc_obj(*area);
456 	if (!area)
457 		goto err;
458 	area->ifq = ifq;
459 
460 	ret = io_import_area(ifq, &area->mem, area_reg);
461 	if (ret)
462 		goto err;
463 
464 	if (buf_size_shift > io_area_max_shift(&area->mem)) {
465 		ret = -ERANGE;
466 		goto err;
467 	}
468 
469 	ifq->niov_shift = buf_size_shift;
470 	nr_iovs = area->mem.size >> ifq->niov_shift;
471 	area->nia.num_niovs = nr_iovs;
472 
473 	ret = -ENOMEM;
474 	area->nia.niovs = kvmalloc_objs(area->nia.niovs[0], nr_iovs,
475 					GFP_KERNEL_ACCOUNT | __GFP_ZERO);
476 	if (!area->nia.niovs)
477 		goto err;
478 
479 	area->freelist = kvmalloc_array(nr_iovs, sizeof(area->freelist[0]),
480 					GFP_KERNEL_ACCOUNT | __GFP_ZERO);
481 	if (!area->freelist)
482 		goto err;
483 
484 	area->user_refs = kvmalloc_objs(area->user_refs[0], nr_iovs,
485 					GFP_KERNEL_ACCOUNT | __GFP_ZERO);
486 	if (!area->user_refs)
487 		goto err;
488 
489 	for (i = 0; i < nr_iovs; i++) {
490 		struct net_iov *niov = &area->nia.niovs[i];
491 
492 		niov->owner = &area->nia;
493 		area->freelist[i] = i;
494 		atomic_set(&area->user_refs[i], 0);
495 		niov->type = NET_IOV_IOURING;
496 	}
497 
498 	area->free_count = nr_iovs;
499 	/* we're only supporting one area per ifq for now */
500 	area->area_id = 0;
501 	area_reg->rq_area_token = (u64)area->area_id << IORING_ZCRX_AREA_SHIFT;
502 	spin_lock_init(&area->freelist_lock);
503 
504 	ret = io_zcrx_append_area(ifq, area);
505 	if (!ret)
506 		return 0;
507 err:
508 	if (area)
509 		io_zcrx_free_area(ifq, area);
510 	return ret;
511 }
512 
513 static struct io_zcrx_ifq *io_zcrx_ifq_alloc(struct io_ring_ctx *ctx)
514 {
515 	struct io_zcrx_ifq *ifq;
516 
517 	ifq = kzalloc_obj(*ifq);
518 	if (!ifq)
519 		return NULL;
520 
521 	ifq->if_rxq = -1;
522 	spin_lock_init(&ifq->rq_lock);
523 	mutex_init(&ifq->pp_lock);
524 	refcount_set(&ifq->refs, 1);
525 	refcount_set(&ifq->user_refs, 1);
526 	return ifq;
527 }
528 
529 static void io_zcrx_drop_netdev(struct io_zcrx_ifq *ifq)
530 {
531 	guard(mutex)(&ifq->pp_lock);
532 
533 	if (!ifq->netdev)
534 		return;
535 	netdev_put(ifq->netdev, &ifq->netdev_tracker);
536 	ifq->netdev = NULL;
537 }
538 
539 static void io_close_queue(struct io_zcrx_ifq *ifq)
540 {
541 	struct net_device *netdev;
542 	netdevice_tracker netdev_tracker;
543 	struct pp_memory_provider_params p = {
544 		.mp_ops = &io_uring_pp_zc_ops,
545 		.mp_priv = ifq,
546 	};
547 
548 	scoped_guard(mutex, &ifq->pp_lock) {
549 		netdev = ifq->netdev;
550 		netdev_tracker = ifq->netdev_tracker;
551 		ifq->netdev = NULL;
552 	}
553 
554 	if (netdev) {
555 		if (ifq->if_rxq != -1) {
556 			netdev_lock(netdev);
557 			netif_mp_close_rxq(netdev, ifq->if_rxq, &p);
558 			netdev_unlock(netdev);
559 		}
560 		netdev_put(netdev, &netdev_tracker);
561 	}
562 	ifq->if_rxq = -1;
563 }
564 
565 static void io_zcrx_ifq_free(struct io_zcrx_ifq *ifq)
566 {
567 	io_close_queue(ifq);
568 
569 	if (ifq->area)
570 		io_zcrx_free_area(ifq, ifq->area);
571 	free_uid(ifq->user);
572 	if (ifq->mm_account)
573 		mmdrop(ifq->mm_account);
574 	if (ifq->dev)
575 		put_device(ifq->dev);
576 
577 	io_free_rbuf_ring(ifq);
578 	mutex_destroy(&ifq->pp_lock);
579 	kfree(ifq);
580 }
581 
582 static void io_put_zcrx_ifq(struct io_zcrx_ifq *ifq)
583 {
584 	if (refcount_dec_and_test(&ifq->refs))
585 		io_zcrx_ifq_free(ifq);
586 }
587 
588 static void io_zcrx_return_niov_freelist(struct net_iov *niov)
589 {
590 	struct io_zcrx_area *area = io_zcrx_iov_to_area(niov);
591 
592 	spin_lock_bh(&area->freelist_lock);
593 	area->freelist[area->free_count++] = net_iov_idx(niov);
594 	spin_unlock_bh(&area->freelist_lock);
595 }
596 
597 static void io_zcrx_return_niov(struct net_iov *niov)
598 {
599 	netmem_ref netmem = net_iov_to_netmem(niov);
600 
601 	if (!niov->desc.pp) {
602 		/* copy fallback allocated niovs */
603 		io_zcrx_return_niov_freelist(niov);
604 		return;
605 	}
606 	page_pool_put_unrefed_netmem(niov->desc.pp, netmem, -1, false);
607 }
608 
609 static void io_zcrx_scrub(struct io_zcrx_ifq *ifq)
610 {
611 	struct io_zcrx_area *area = ifq->area;
612 	int i;
613 
614 	if (!area)
615 		return;
616 
617 	/* Reclaim back all buffers given to the user space. */
618 	for (i = 0; i < area->nia.num_niovs; i++) {
619 		struct net_iov *niov = &area->nia.niovs[i];
620 		int nr;
621 
622 		if (!atomic_read(io_get_user_counter(niov)))
623 			continue;
624 		nr = atomic_xchg(io_get_user_counter(niov), 0);
625 		if (nr && !page_pool_unref_netmem(net_iov_to_netmem(niov), nr))
626 			io_zcrx_return_niov(niov);
627 	}
628 }
629 
630 static void zcrx_unregister(struct io_zcrx_ifq *ifq)
631 {
632 	if (refcount_dec_and_test(&ifq->user_refs)) {
633 		io_close_queue(ifq);
634 		io_zcrx_scrub(ifq);
635 	}
636 	io_put_zcrx_ifq(ifq);
637 }
638 
639 struct io_mapped_region *io_zcrx_get_region(struct io_ring_ctx *ctx,
640 					    unsigned int id)
641 {
642 	struct io_zcrx_ifq *ifq = xa_load(&ctx->zcrx_ctxs, id);
643 
644 	lockdep_assert_held(&ctx->mmap_lock);
645 
646 	return ifq ? &ifq->region : NULL;
647 }
648 
649 static int zcrx_box_release(struct inode *inode, struct file *file)
650 {
651 	struct io_zcrx_ifq *ifq = file->private_data;
652 
653 	if (WARN_ON_ONCE(!ifq))
654 		return -EFAULT;
655 	zcrx_unregister(ifq);
656 	return 0;
657 }
658 
659 static const struct file_operations zcrx_box_fops = {
660 	.owner		= THIS_MODULE,
661 	.release	= zcrx_box_release,
662 };
663 
664 static int zcrx_export(struct io_ring_ctx *ctx, struct io_zcrx_ifq *ifq,
665 		       struct zcrx_ctrl *ctrl, void __user *arg)
666 {
667 	struct zcrx_ctrl_export *ce = &ctrl->zc_export;
668 	struct file *file;
669 	int fd = -1;
670 
671 	if (!mem_is_zero(ce, sizeof(*ce)))
672 		return -EINVAL;
673 	fd = get_unused_fd_flags(O_CLOEXEC);
674 	if (fd < 0)
675 		return fd;
676 
677 	ce->zcrx_fd = fd;
678 	if (copy_to_user(arg, ctrl, sizeof(*ctrl))) {
679 		put_unused_fd(fd);
680 		return -EFAULT;
681 	}
682 
683 	refcount_inc(&ifq->refs);
684 	refcount_inc(&ifq->user_refs);
685 
686 	file = anon_inode_create_getfile("[zcrx]", &zcrx_box_fops,
687 					 ifq, O_CLOEXEC, NULL);
688 	if (IS_ERR(file)) {
689 		put_unused_fd(fd);
690 		zcrx_unregister(ifq);
691 		return PTR_ERR(file);
692 	}
693 
694 	fd_install(fd, file);
695 	return 0;
696 }
697 
698 static int import_zcrx(struct io_ring_ctx *ctx,
699 		       struct io_uring_zcrx_ifq_reg __user *arg,
700 		       struct io_uring_zcrx_ifq_reg *reg)
701 {
702 	struct io_zcrx_ifq *ifq;
703 	struct file *file;
704 	int fd, ret;
705 	u32 id;
706 
707 	if (!(ctx->flags & IORING_SETUP_DEFER_TASKRUN))
708 		return -EINVAL;
709 	if (!(ctx->flags & (IORING_SETUP_CQE32|IORING_SETUP_CQE_MIXED)))
710 		return -EINVAL;
711 	if (reg->if_rxq || reg->rq_entries || reg->area_ptr || reg->region_ptr)
712 		return -EINVAL;
713 	if (reg->flags & ~ZCRX_REG_IMPORT)
714 		return -EINVAL;
715 
716 	fd = reg->if_idx;
717 	CLASS(fd, f)(fd);
718 	if (fd_empty(f))
719 		return -EBADF;
720 
721 	file = fd_file(f);
722 	if (file->f_op != &zcrx_box_fops || !file->private_data)
723 		return -EBADF;
724 
725 	ifq = file->private_data;
726 	refcount_inc(&ifq->refs);
727 	refcount_inc(&ifq->user_refs);
728 
729 	scoped_guard(mutex, &ctx->mmap_lock) {
730 		ret = xa_alloc(&ctx->zcrx_ctxs, &id, NULL, xa_limit_31b, GFP_KERNEL);
731 		if (ret)
732 			goto err;
733 	}
734 
735 	reg->zcrx_id = id;
736 	io_fill_zcrx_offsets(&reg->offsets);
737 	if (copy_to_user(arg, reg, sizeof(*reg))) {
738 		ret = -EFAULT;
739 		goto err_xa_erase;
740 	}
741 
742 	scoped_guard(mutex, &ctx->mmap_lock) {
743 		ret = -ENOMEM;
744 		if (xa_store(&ctx->zcrx_ctxs, id, ifq, GFP_KERNEL))
745 			goto err_xa_erase;
746 	}
747 
748 	return 0;
749 err_xa_erase:
750 	scoped_guard(mutex, &ctx->mmap_lock)
751 		xa_erase(&ctx->zcrx_ctxs, id);
752 err:
753 	zcrx_unregister(ifq);
754 	return ret;
755 }
756 
757 int io_register_zcrx_ifq(struct io_ring_ctx *ctx,
758 			  struct io_uring_zcrx_ifq_reg __user *arg)
759 {
760 	struct pp_memory_provider_params mp_param = {};
761 	struct io_uring_zcrx_area_reg area;
762 	struct io_uring_zcrx_ifq_reg reg;
763 	struct io_uring_region_desc rd;
764 	struct io_zcrx_ifq *ifq;
765 	int ret;
766 	u32 id;
767 
768 	/*
769 	 * 1. Interface queue allocation.
770 	 * 2. It can observe data destined for sockets of other tasks.
771 	 */
772 	if (!capable(CAP_NET_ADMIN))
773 		return -EPERM;
774 
775 	/* mandatory io_uring features for zc rx */
776 	if (!(ctx->flags & IORING_SETUP_DEFER_TASKRUN))
777 		return -EINVAL;
778 	if (!(ctx->flags & (IORING_SETUP_CQE32|IORING_SETUP_CQE_MIXED)))
779 		return -EINVAL;
780 	if (copy_from_user(&reg, arg, sizeof(reg)))
781 		return -EFAULT;
782 	if (!mem_is_zero(&reg.__resv, sizeof(reg.__resv)) || reg.zcrx_id)
783 		return -EINVAL;
784 	if (reg.flags & ZCRX_REG_IMPORT)
785 		return import_zcrx(ctx, arg, &reg);
786 	if (copy_from_user(&rd, u64_to_user_ptr(reg.region_ptr), sizeof(rd)))
787 		return -EFAULT;
788 	if (reg.if_rxq == -1 || !reg.rq_entries || reg.flags)
789 		return -EINVAL;
790 	if (reg.rq_entries > IO_RQ_MAX_ENTRIES) {
791 		if (!(ctx->flags & IORING_SETUP_CLAMP))
792 			return -EINVAL;
793 		reg.rq_entries = IO_RQ_MAX_ENTRIES;
794 	}
795 	reg.rq_entries = roundup_pow_of_two(reg.rq_entries);
796 
797 	if (copy_from_user(&area, u64_to_user_ptr(reg.area_ptr), sizeof(area)))
798 		return -EFAULT;
799 
800 	ifq = io_zcrx_ifq_alloc(ctx);
801 	if (!ifq)
802 		return -ENOMEM;
803 
804 	if (ctx->user) {
805 		get_uid(ctx->user);
806 		ifq->user = ctx->user;
807 	}
808 	if (ctx->mm_account) {
809 		mmgrab(ctx->mm_account);
810 		ifq->mm_account = ctx->mm_account;
811 	}
812 	ifq->rq_entries = reg.rq_entries;
813 
814 	scoped_guard(mutex, &ctx->mmap_lock) {
815 		/* preallocate id */
816 		ret = xa_alloc(&ctx->zcrx_ctxs, &id, NULL, xa_limit_31b, GFP_KERNEL);
817 		if (ret)
818 			goto ifq_free;
819 	}
820 
821 	ret = io_allocate_rbuf_ring(ctx, ifq, &reg, &rd, id);
822 	if (ret)
823 		goto err;
824 
825 	ifq->netdev = netdev_get_by_index_lock(current->nsproxy->net_ns, reg.if_idx);
826 	if (!ifq->netdev) {
827 		ret = -ENODEV;
828 		goto err;
829 	}
830 	netdev_hold(ifq->netdev, &ifq->netdev_tracker, GFP_KERNEL);
831 
832 	ifq->dev = netdev_queue_get_dma_dev(ifq->netdev, reg.if_rxq);
833 	if (!ifq->dev) {
834 		ret = -EOPNOTSUPP;
835 		goto netdev_put_unlock;
836 	}
837 	get_device(ifq->dev);
838 
839 	ret = io_zcrx_create_area(ifq, &area, &reg);
840 	if (ret)
841 		goto netdev_put_unlock;
842 
843 	if (reg.rx_buf_len)
844 		mp_param.rx_page_size = 1U << ifq->niov_shift;
845 	mp_param.mp_ops = &io_uring_pp_zc_ops;
846 	mp_param.mp_priv = ifq;
847 	ret = netif_mp_open_rxq(ifq->netdev, reg.if_rxq, &mp_param, NULL);
848 	if (ret)
849 		goto netdev_put_unlock;
850 	netdev_unlock(ifq->netdev);
851 	ifq->if_rxq = reg.if_rxq;
852 
853 	reg.zcrx_id = id;
854 
855 	scoped_guard(mutex, &ctx->mmap_lock) {
856 		/* publish ifq */
857 		ret = -ENOMEM;
858 		if (xa_store(&ctx->zcrx_ctxs, id, ifq, GFP_KERNEL))
859 			goto err;
860 	}
861 
862 	reg.rx_buf_len = 1U << ifq->niov_shift;
863 
864 	if (copy_to_user(arg, &reg, sizeof(reg)) ||
865 	    copy_to_user(u64_to_user_ptr(reg.region_ptr), &rd, sizeof(rd)) ||
866 	    copy_to_user(u64_to_user_ptr(reg.area_ptr), &area, sizeof(area))) {
867 		ret = -EFAULT;
868 		goto err;
869 	}
870 	return 0;
871 netdev_put_unlock:
872 	netdev_unlock(ifq->netdev);
873 err:
874 	scoped_guard(mutex, &ctx->mmap_lock)
875 		xa_erase(&ctx->zcrx_ctxs, id);
876 ifq_free:
877 	zcrx_unregister(ifq);
878 	return ret;
879 }
880 
881 static struct net_iov *__io_zcrx_get_free_niov(struct io_zcrx_area *area)
882 {
883 	unsigned niov_idx;
884 
885 	lockdep_assert_held(&area->freelist_lock);
886 
887 	niov_idx = area->freelist[--area->free_count];
888 	return &area->nia.niovs[niov_idx];
889 }
890 
891 void io_unregister_zcrx_ifqs(struct io_ring_ctx *ctx)
892 {
893 	struct io_zcrx_ifq *ifq;
894 
895 	lockdep_assert_held(&ctx->uring_lock);
896 
897 	while (1) {
898 		scoped_guard(mutex, &ctx->mmap_lock) {
899 			unsigned long id = 0;
900 
901 			ifq = xa_find(&ctx->zcrx_ctxs, &id, ULONG_MAX, XA_PRESENT);
902 			if (ifq)
903 				xa_erase(&ctx->zcrx_ctxs, id);
904 		}
905 		if (!ifq)
906 			break;
907 		zcrx_unregister(ifq);
908 	}
909 
910 	xa_destroy(&ctx->zcrx_ctxs);
911 }
912 
913 static inline u32 io_zcrx_rqring_entries(struct io_zcrx_ifq *ifq)
914 {
915 	u32 entries;
916 
917 	entries = smp_load_acquire(&ifq->rq_ring->tail) - ifq->cached_rq_head;
918 	return min(entries, ifq->rq_entries);
919 }
920 
921 static struct io_uring_zcrx_rqe *io_zcrx_get_rqe(struct io_zcrx_ifq *ifq,
922 						 unsigned mask)
923 {
924 	unsigned int idx = ifq->cached_rq_head++ & mask;
925 
926 	return &ifq->rqes[idx];
927 }
928 
929 static inline bool io_parse_rqe(struct io_uring_zcrx_rqe *rqe,
930 				struct io_zcrx_ifq *ifq,
931 				struct net_iov **ret_niov)
932 {
933 	__u64 off = READ_ONCE(rqe->off);
934 	unsigned niov_idx, area_idx;
935 	struct io_zcrx_area *area;
936 
937 	area_idx = off >> IORING_ZCRX_AREA_SHIFT;
938 	niov_idx = (off & ~IORING_ZCRX_AREA_MASK) >> ifq->niov_shift;
939 
940 	if (unlikely(rqe->__pad || area_idx))
941 		return false;
942 	area = ifq->area;
943 
944 	if (unlikely(niov_idx >= area->nia.num_niovs))
945 		return false;
946 	niov_idx = array_index_nospec(niov_idx, area->nia.num_niovs);
947 
948 	*ret_niov = &area->nia.niovs[niov_idx];
949 	return true;
950 }
951 
952 static void io_zcrx_ring_refill(struct page_pool *pp,
953 				struct io_zcrx_ifq *ifq)
954 {
955 	unsigned int mask = ifq->rq_entries - 1;
956 	unsigned int entries;
957 
958 	guard(spinlock_bh)(&ifq->rq_lock);
959 
960 	entries = io_zcrx_rqring_entries(ifq);
961 	entries = min_t(unsigned, entries, PP_ALLOC_CACHE_REFILL);
962 	if (unlikely(!entries))
963 		return;
964 
965 	do {
966 		struct io_uring_zcrx_rqe *rqe = io_zcrx_get_rqe(ifq, mask);
967 		struct net_iov *niov;
968 		netmem_ref netmem;
969 
970 		if (!io_parse_rqe(rqe, ifq, &niov))
971 			continue;
972 		if (!io_zcrx_put_niov_uref(niov))
973 			continue;
974 
975 		netmem = net_iov_to_netmem(niov);
976 		if (!page_pool_unref_and_test(netmem))
977 			continue;
978 
979 		if (unlikely(niov->desc.pp != pp)) {
980 			io_zcrx_return_niov(niov);
981 			continue;
982 		}
983 
984 		io_zcrx_sync_for_device(pp, niov);
985 		net_mp_netmem_place_in_cache(pp, netmem);
986 	} while (--entries);
987 
988 	smp_store_release(&ifq->rq_ring->head, ifq->cached_rq_head);
989 }
990 
991 static void io_zcrx_refill_slow(struct page_pool *pp, struct io_zcrx_ifq *ifq)
992 {
993 	struct io_zcrx_area *area = ifq->area;
994 
995 	spin_lock_bh(&area->freelist_lock);
996 	while (area->free_count && pp->alloc.count < PP_ALLOC_CACHE_REFILL) {
997 		struct net_iov *niov = __io_zcrx_get_free_niov(area);
998 		netmem_ref netmem = net_iov_to_netmem(niov);
999 
1000 		net_mp_niov_set_page_pool(pp, niov);
1001 		io_zcrx_sync_for_device(pp, niov);
1002 		net_mp_netmem_place_in_cache(pp, netmem);
1003 	}
1004 	spin_unlock_bh(&area->freelist_lock);
1005 }
1006 
1007 static netmem_ref io_pp_zc_alloc_netmems(struct page_pool *pp, gfp_t gfp)
1008 {
1009 	struct io_zcrx_ifq *ifq = io_pp_to_ifq(pp);
1010 
1011 	/* pp should already be ensuring that */
1012 	if (unlikely(pp->alloc.count))
1013 		goto out_return;
1014 
1015 	io_zcrx_ring_refill(pp, ifq);
1016 	if (likely(pp->alloc.count))
1017 		goto out_return;
1018 
1019 	io_zcrx_refill_slow(pp, ifq);
1020 	if (!pp->alloc.count)
1021 		return 0;
1022 out_return:
1023 	return pp->alloc.cache[--pp->alloc.count];
1024 }
1025 
1026 static bool io_pp_zc_release_netmem(struct page_pool *pp, netmem_ref netmem)
1027 {
1028 	struct net_iov *niov;
1029 
1030 	if (WARN_ON_ONCE(!netmem_is_net_iov(netmem)))
1031 		return false;
1032 
1033 	niov = netmem_to_net_iov(netmem);
1034 	net_mp_niov_clear_page_pool(niov);
1035 	io_zcrx_return_niov_freelist(niov);
1036 	return false;
1037 }
1038 
1039 static int io_pp_zc_init(struct page_pool *pp)
1040 {
1041 	struct io_zcrx_ifq *ifq = io_pp_to_ifq(pp);
1042 	int ret;
1043 
1044 	if (WARN_ON_ONCE(!ifq))
1045 		return -EINVAL;
1046 	if (WARN_ON_ONCE(ifq->dev != pp->p.dev))
1047 		return -EINVAL;
1048 	if (WARN_ON_ONCE(!pp->dma_map))
1049 		return -EOPNOTSUPP;
1050 	if (pp->p.order + PAGE_SHIFT != ifq->niov_shift)
1051 		return -EINVAL;
1052 	if (pp->p.dma_dir != DMA_FROM_DEVICE)
1053 		return -EOPNOTSUPP;
1054 
1055 	ret = io_zcrx_map_area(ifq, ifq->area);
1056 	if (ret)
1057 		return ret;
1058 
1059 	refcount_inc(&ifq->refs);
1060 	return 0;
1061 }
1062 
1063 static void io_pp_zc_destroy(struct page_pool *pp)
1064 {
1065 	io_put_zcrx_ifq(io_pp_to_ifq(pp));
1066 }
1067 
1068 static int io_pp_nl_fill(void *mp_priv, struct sk_buff *rsp,
1069 			 struct netdev_rx_queue *rxq)
1070 {
1071 	struct nlattr *nest;
1072 	int type;
1073 
1074 	type = rxq ? NETDEV_A_QUEUE_IO_URING : NETDEV_A_PAGE_POOL_IO_URING;
1075 	nest = nla_nest_start(rsp, type);
1076 	if (!nest)
1077 		return -EMSGSIZE;
1078 	nla_nest_end(rsp, nest);
1079 
1080 	return 0;
1081 }
1082 
1083 static void io_pp_uninstall(void *mp_priv, struct netdev_rx_queue *rxq)
1084 {
1085 	struct pp_memory_provider_params *p = &rxq->mp_params;
1086 	struct io_zcrx_ifq *ifq = mp_priv;
1087 
1088 	io_zcrx_drop_netdev(ifq);
1089 	if (ifq->area)
1090 		io_zcrx_unmap_area(ifq, ifq->area);
1091 
1092 	p->mp_ops = NULL;
1093 	p->mp_priv = NULL;
1094 }
1095 
1096 static const struct memory_provider_ops io_uring_pp_zc_ops = {
1097 	.alloc_netmems		= io_pp_zc_alloc_netmems,
1098 	.release_netmem		= io_pp_zc_release_netmem,
1099 	.init			= io_pp_zc_init,
1100 	.destroy		= io_pp_zc_destroy,
1101 	.nl_fill		= io_pp_nl_fill,
1102 	.uninstall		= io_pp_uninstall,
1103 };
1104 
1105 static unsigned zcrx_parse_rq(netmem_ref *netmem_array, unsigned nr,
1106 			      struct io_zcrx_ifq *zcrx)
1107 {
1108 	unsigned int mask = zcrx->rq_entries - 1;
1109 	unsigned int i;
1110 
1111 	nr = min(nr, io_zcrx_rqring_entries(zcrx));
1112 	for (i = 0; i < nr; i++) {
1113 		struct io_uring_zcrx_rqe *rqe = io_zcrx_get_rqe(zcrx, mask);
1114 		struct net_iov *niov;
1115 
1116 		if (!io_parse_rqe(rqe, zcrx, &niov))
1117 			break;
1118 		netmem_array[i] = net_iov_to_netmem(niov);
1119 	}
1120 
1121 	smp_store_release(&zcrx->rq_ring->head, zcrx->cached_rq_head);
1122 	return i;
1123 }
1124 
1125 #define ZCRX_FLUSH_BATCH 32
1126 
1127 static void zcrx_return_buffers(netmem_ref *netmems, unsigned nr)
1128 {
1129 	unsigned i;
1130 
1131 	for (i = 0; i < nr; i++) {
1132 		netmem_ref netmem = netmems[i];
1133 		struct net_iov *niov = netmem_to_net_iov(netmem);
1134 
1135 		if (!io_zcrx_put_niov_uref(niov))
1136 			continue;
1137 		if (!page_pool_unref_and_test(netmem))
1138 			continue;
1139 		io_zcrx_return_niov(niov);
1140 	}
1141 }
1142 
1143 static int zcrx_flush_rq(struct io_ring_ctx *ctx, struct io_zcrx_ifq *zcrx,
1144 			 struct zcrx_ctrl *ctrl)
1145 {
1146 	struct zcrx_ctrl_flush_rq *frq = &ctrl->zc_flush;
1147 	netmem_ref netmems[ZCRX_FLUSH_BATCH];
1148 	unsigned total = 0;
1149 	unsigned nr;
1150 
1151 	if (!mem_is_zero(&frq->__resv, sizeof(frq->__resv)))
1152 		return -EINVAL;
1153 
1154 	do {
1155 		scoped_guard(spinlock_bh, &zcrx->rq_lock) {
1156 			nr = zcrx_parse_rq(netmems, ZCRX_FLUSH_BATCH, zcrx);
1157 			zcrx_return_buffers(netmems, nr);
1158 		}
1159 
1160 		total += nr;
1161 
1162 		if (fatal_signal_pending(current))
1163 			break;
1164 		cond_resched();
1165 	} while (nr == ZCRX_FLUSH_BATCH && total < zcrx->rq_entries);
1166 
1167 	return 0;
1168 }
1169 
1170 int io_zcrx_ctrl(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
1171 {
1172 	struct zcrx_ctrl ctrl;
1173 	struct io_zcrx_ifq *zcrx;
1174 
1175 	if (nr_args)
1176 		return -EINVAL;
1177 	if (copy_from_user(&ctrl, arg, sizeof(ctrl)))
1178 		return -EFAULT;
1179 	if (!mem_is_zero(&ctrl.__resv, sizeof(ctrl.__resv)))
1180 		return -EFAULT;
1181 
1182 	zcrx = xa_load(&ctx->zcrx_ctxs, ctrl.zcrx_id);
1183 	if (!zcrx)
1184 		return -ENXIO;
1185 
1186 	switch (ctrl.op) {
1187 	case ZCRX_CTRL_FLUSH_RQ:
1188 		return zcrx_flush_rq(ctx, zcrx, &ctrl);
1189 	case ZCRX_CTRL_EXPORT:
1190 		return zcrx_export(ctx, zcrx, &ctrl, arg);
1191 	}
1192 
1193 	return -EOPNOTSUPP;
1194 }
1195 
1196 static bool io_zcrx_queue_cqe(struct io_kiocb *req, struct net_iov *niov,
1197 			      struct io_zcrx_ifq *ifq, int off, int len)
1198 {
1199 	struct io_ring_ctx *ctx = req->ctx;
1200 	struct io_uring_zcrx_cqe *rcqe;
1201 	struct io_zcrx_area *area;
1202 	struct io_uring_cqe *cqe;
1203 	u64 offset;
1204 
1205 	if (!io_defer_get_uncommited_cqe(ctx, &cqe))
1206 		return false;
1207 
1208 	cqe->user_data = req->cqe.user_data;
1209 	cqe->res = len;
1210 	cqe->flags = IORING_CQE_F_MORE;
1211 	if (ctx->flags & IORING_SETUP_CQE_MIXED)
1212 		cqe->flags |= IORING_CQE_F_32;
1213 
1214 	area = io_zcrx_iov_to_area(niov);
1215 	offset = off + (net_iov_idx(niov) << ifq->niov_shift);
1216 	rcqe = (struct io_uring_zcrx_cqe *)(cqe + 1);
1217 	rcqe->off = offset + ((u64)area->area_id << IORING_ZCRX_AREA_SHIFT);
1218 	rcqe->__pad = 0;
1219 	return true;
1220 }
1221 
1222 static struct net_iov *io_alloc_fallback_niov(struct io_zcrx_ifq *ifq)
1223 {
1224 	struct io_zcrx_area *area = ifq->area;
1225 	struct net_iov *niov = NULL;
1226 
1227 	if (area->mem.is_dmabuf)
1228 		return NULL;
1229 
1230 	spin_lock_bh(&area->freelist_lock);
1231 	if (area->free_count)
1232 		niov = __io_zcrx_get_free_niov(area);
1233 	spin_unlock_bh(&area->freelist_lock);
1234 
1235 	if (niov)
1236 		page_pool_fragment_netmem(net_iov_to_netmem(niov), 1);
1237 	return niov;
1238 }
1239 
1240 struct io_copy_cache {
1241 	struct page		*page;
1242 	unsigned long		offset;
1243 	size_t			size;
1244 };
1245 
1246 static ssize_t io_copy_page(struct io_copy_cache *cc, struct page *src_page,
1247 			    unsigned int src_offset, size_t len)
1248 {
1249 	size_t copied = 0;
1250 
1251 	len = min(len, cc->size);
1252 
1253 	while (len) {
1254 		void *src_addr, *dst_addr;
1255 		struct page *dst_page = cc->page;
1256 		unsigned dst_offset = cc->offset;
1257 		size_t n = len;
1258 
1259 		if (folio_test_partial_kmap(page_folio(dst_page)) ||
1260 		    folio_test_partial_kmap(page_folio(src_page))) {
1261 			dst_page += dst_offset / PAGE_SIZE;
1262 			dst_offset = offset_in_page(dst_offset);
1263 			src_page += src_offset / PAGE_SIZE;
1264 			src_offset = offset_in_page(src_offset);
1265 			n = min(PAGE_SIZE - src_offset, PAGE_SIZE - dst_offset);
1266 			n = min(n, len);
1267 		}
1268 
1269 		dst_addr = kmap_local_page(dst_page) + dst_offset;
1270 		src_addr = kmap_local_page(src_page) + src_offset;
1271 
1272 		memcpy(dst_addr, src_addr, n);
1273 
1274 		kunmap_local(src_addr);
1275 		kunmap_local(dst_addr);
1276 
1277 		cc->size -= n;
1278 		cc->offset += n;
1279 		src_offset += n;
1280 		len -= n;
1281 		copied += n;
1282 	}
1283 	return copied;
1284 }
1285 
1286 static ssize_t io_zcrx_copy_chunk(struct io_kiocb *req, struct io_zcrx_ifq *ifq,
1287 				  struct page *src_page, unsigned int src_offset,
1288 				  size_t len)
1289 {
1290 	size_t copied = 0;
1291 	int ret = 0;
1292 
1293 	while (len) {
1294 		struct io_copy_cache cc;
1295 		struct net_iov *niov;
1296 		size_t n;
1297 
1298 		niov = io_alloc_fallback_niov(ifq);
1299 		if (!niov) {
1300 			ret = -ENOMEM;
1301 			break;
1302 		}
1303 
1304 		cc.page = io_zcrx_iov_page(niov);
1305 		cc.offset = 0;
1306 		cc.size = PAGE_SIZE;
1307 
1308 		n = io_copy_page(&cc, src_page, src_offset, len);
1309 
1310 		if (!io_zcrx_queue_cqe(req, niov, ifq, 0, n)) {
1311 			io_zcrx_return_niov(niov);
1312 			ret = -ENOSPC;
1313 			break;
1314 		}
1315 
1316 		io_zcrx_get_niov_uref(niov);
1317 		src_offset += n;
1318 		len -= n;
1319 		copied += n;
1320 	}
1321 
1322 	return copied ? copied : ret;
1323 }
1324 
1325 static int io_zcrx_copy_frag(struct io_kiocb *req, struct io_zcrx_ifq *ifq,
1326 			     const skb_frag_t *frag, int off, int len)
1327 {
1328 	struct page *page = skb_frag_page(frag);
1329 
1330 	return io_zcrx_copy_chunk(req, ifq, page, off + skb_frag_off(frag), len);
1331 }
1332 
1333 static int io_zcrx_recv_frag(struct io_kiocb *req, struct io_zcrx_ifq *ifq,
1334 			     const skb_frag_t *frag, int off, int len)
1335 {
1336 	struct net_iov *niov;
1337 	struct page_pool *pp;
1338 
1339 	if (unlikely(!skb_frag_is_net_iov(frag)))
1340 		return io_zcrx_copy_frag(req, ifq, frag, off, len);
1341 
1342 	niov = netmem_to_net_iov(frag->netmem);
1343 	pp = niov->desc.pp;
1344 
1345 	if (!pp || pp->mp_ops != &io_uring_pp_zc_ops || io_pp_to_ifq(pp) != ifq)
1346 		return -EFAULT;
1347 
1348 	if (!io_zcrx_queue_cqe(req, niov, ifq, off + skb_frag_off(frag), len))
1349 		return -ENOSPC;
1350 
1351 	/*
1352 	 * Prevent it from being recycled while user is accessing it.
1353 	 * It has to be done before grabbing a user reference.
1354 	 */
1355 	page_pool_ref_netmem(net_iov_to_netmem(niov));
1356 	io_zcrx_get_niov_uref(niov);
1357 	return len;
1358 }
1359 
1360 static int
1361 io_zcrx_recv_skb(read_descriptor_t *desc, struct sk_buff *skb,
1362 		 unsigned int offset, size_t len)
1363 {
1364 	struct io_zcrx_args *args = desc->arg.data;
1365 	struct io_zcrx_ifq *ifq = args->ifq;
1366 	struct io_kiocb *req = args->req;
1367 	struct sk_buff *frag_iter;
1368 	unsigned start, start_off = offset;
1369 	int i, copy, end, off;
1370 	int ret = 0;
1371 
1372 	len = min_t(size_t, len, desc->count);
1373 	/*
1374 	 * __tcp_read_sock() always calls io_zcrx_recv_skb one last time, even
1375 	 * if desc->count is already 0. This is caused by the if (offset + 1 !=
1376 	 * skb->len) check. Return early in this case to break out of
1377 	 * __tcp_read_sock().
1378 	 */
1379 	if (!len)
1380 		return 0;
1381 	if (unlikely(args->nr_skbs++ > IO_SKBS_PER_CALL_LIMIT))
1382 		return -EAGAIN;
1383 
1384 	if (unlikely(offset < skb_headlen(skb))) {
1385 		ssize_t copied;
1386 		size_t to_copy;
1387 
1388 		to_copy = min_t(size_t, skb_headlen(skb) - offset, len);
1389 		copied = io_zcrx_copy_chunk(req, ifq, virt_to_page(skb->data),
1390 					    offset_in_page(skb->data) + offset,
1391 					    to_copy);
1392 		if (copied < 0) {
1393 			ret = copied;
1394 			goto out;
1395 		}
1396 		offset += copied;
1397 		len -= copied;
1398 		if (!len)
1399 			goto out;
1400 		if (offset != skb_headlen(skb))
1401 			goto out;
1402 	}
1403 
1404 	start = skb_headlen(skb);
1405 
1406 	for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
1407 		const skb_frag_t *frag;
1408 
1409 		if (WARN_ON(start > offset + len))
1410 			return -EFAULT;
1411 
1412 		frag = &skb_shinfo(skb)->frags[i];
1413 		end = start + skb_frag_size(frag);
1414 
1415 		if (offset < end) {
1416 			copy = end - offset;
1417 			if (copy > len)
1418 				copy = len;
1419 
1420 			off = offset - start;
1421 			ret = io_zcrx_recv_frag(req, ifq, frag, off, copy);
1422 			if (ret < 0)
1423 				goto out;
1424 
1425 			offset += ret;
1426 			len -= ret;
1427 			if (len == 0 || ret != copy)
1428 				goto out;
1429 		}
1430 		start = end;
1431 	}
1432 
1433 	skb_walk_frags(skb, frag_iter) {
1434 		if (WARN_ON(start > offset + len))
1435 			return -EFAULT;
1436 
1437 		end = start + frag_iter->len;
1438 		if (offset < end) {
1439 			size_t count;
1440 
1441 			copy = end - offset;
1442 			if (copy > len)
1443 				copy = len;
1444 
1445 			off = offset - start;
1446 			count = desc->count;
1447 			ret = io_zcrx_recv_skb(desc, frag_iter, off, copy);
1448 			desc->count = count;
1449 			if (ret < 0)
1450 				goto out;
1451 
1452 			offset += ret;
1453 			len -= ret;
1454 			if (len == 0 || ret != copy)
1455 				goto out;
1456 		}
1457 		start = end;
1458 	}
1459 
1460 out:
1461 	if (offset == start_off)
1462 		return ret;
1463 	desc->count -= (offset - start_off);
1464 	return offset - start_off;
1465 }
1466 
1467 static int io_zcrx_tcp_recvmsg(struct io_kiocb *req, struct io_zcrx_ifq *ifq,
1468 				struct sock *sk, int flags,
1469 				unsigned issue_flags, unsigned int *outlen)
1470 {
1471 	unsigned int len = *outlen;
1472 	struct io_zcrx_args args = {
1473 		.req = req,
1474 		.ifq = ifq,
1475 		.sock = sk->sk_socket,
1476 	};
1477 	read_descriptor_t rd_desc = {
1478 		.count = len ? len : UINT_MAX,
1479 		.arg.data = &args,
1480 	};
1481 	int ret;
1482 
1483 	lock_sock(sk);
1484 	ret = tcp_read_sock(sk, &rd_desc, io_zcrx_recv_skb);
1485 	if (len && ret > 0)
1486 		*outlen = len - ret;
1487 	if (ret <= 0) {
1488 		if (ret < 0 || sock_flag(sk, SOCK_DONE))
1489 			goto out;
1490 		if (sk->sk_err)
1491 			ret = sock_error(sk);
1492 		else if (sk->sk_shutdown & RCV_SHUTDOWN)
1493 			goto out;
1494 		else if (sk->sk_state == TCP_CLOSE)
1495 			ret = -ENOTCONN;
1496 		else
1497 			ret = -EAGAIN;
1498 	} else if (unlikely(args.nr_skbs > IO_SKBS_PER_CALL_LIMIT) &&
1499 		   (issue_flags & IO_URING_F_MULTISHOT)) {
1500 		ret = IOU_REQUEUE;
1501 	} else if (sock_flag(sk, SOCK_DONE)) {
1502 		/* Make it to retry until it finally gets 0. */
1503 		if (issue_flags & IO_URING_F_MULTISHOT)
1504 			ret = IOU_REQUEUE;
1505 		else
1506 			ret = -EAGAIN;
1507 	}
1508 out:
1509 	release_sock(sk);
1510 	return ret;
1511 }
1512 
1513 int io_zcrx_recv(struct io_kiocb *req, struct io_zcrx_ifq *ifq,
1514 		 struct socket *sock, unsigned int flags,
1515 		 unsigned issue_flags, unsigned int *len)
1516 {
1517 	struct sock *sk = sock->sk;
1518 	const struct proto *prot = READ_ONCE(sk->sk_prot);
1519 
1520 	if (prot->recvmsg != tcp_recvmsg)
1521 		return -EPROTONOSUPPORT;
1522 
1523 	sock_rps_record_flow(sk);
1524 	return io_zcrx_tcp_recvmsg(req, ifq, sk, flags, issue_flags, len);
1525 }
1526