xref: /linux/io_uring/zcrx.c (revision ca220141fa8ebae09765a242076b2b77338106b0)
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 
353 	if (unlikely(!atomic_read(uref)))
354 		return false;
355 	atomic_dec(uref);
356 	return true;
357 }
358 
359 static void io_zcrx_get_niov_uref(struct net_iov *niov)
360 {
361 	atomic_inc(io_get_user_counter(niov));
362 }
363 
364 static void io_fill_zcrx_offsets(struct io_uring_zcrx_offsets *offsets)
365 {
366 	offsets->head = offsetof(struct io_uring, head);
367 	offsets->tail = offsetof(struct io_uring, tail);
368 	offsets->rqes = ALIGN(sizeof(struct io_uring), L1_CACHE_BYTES);
369 }
370 
371 static int io_allocate_rbuf_ring(struct io_ring_ctx *ctx,
372 				 struct io_zcrx_ifq *ifq,
373 				 struct io_uring_zcrx_ifq_reg *reg,
374 				 struct io_uring_region_desc *rd,
375 				 u32 id)
376 {
377 	u64 mmap_offset;
378 	size_t off, size;
379 	void *ptr;
380 	int ret;
381 
382 	io_fill_zcrx_offsets(&reg->offsets);
383 	off = reg->offsets.rqes;
384 	size = off + sizeof(struct io_uring_zcrx_rqe) * reg->rq_entries;
385 	if (size > rd->size)
386 		return -EINVAL;
387 
388 	mmap_offset = IORING_MAP_OFF_ZCRX_REGION;
389 	mmap_offset += id << IORING_OFF_PBUF_SHIFT;
390 
391 	ret = io_create_region(ctx, &ifq->region, rd, mmap_offset);
392 	if (ret < 0)
393 		return ret;
394 
395 	ptr = io_region_get_ptr(&ifq->region);
396 	ifq->rq_ring = (struct io_uring *)ptr;
397 	ifq->rqes = (struct io_uring_zcrx_rqe *)(ptr + off);
398 
399 	return 0;
400 }
401 
402 static void io_free_rbuf_ring(struct io_zcrx_ifq *ifq)
403 {
404 	io_free_region(ifq->user, &ifq->region);
405 	ifq->rq_ring = NULL;
406 	ifq->rqes = NULL;
407 }
408 
409 static void io_zcrx_free_area(struct io_zcrx_ifq *ifq,
410 			      struct io_zcrx_area *area)
411 {
412 	io_zcrx_unmap_area(ifq, area);
413 	io_release_area_mem(&area->mem);
414 
415 	if (area->mem.account_pages)
416 		io_unaccount_mem(ifq->user, ifq->mm_account,
417 				 area->mem.account_pages);
418 
419 	kvfree(area->freelist);
420 	kvfree(area->nia.niovs);
421 	kvfree(area->user_refs);
422 	kfree(area);
423 }
424 
425 static int io_zcrx_append_area(struct io_zcrx_ifq *ifq,
426 				struct io_zcrx_area *area)
427 {
428 	if (ifq->area)
429 		return -EINVAL;
430 	ifq->area = area;
431 	return 0;
432 }
433 
434 static int io_zcrx_create_area(struct io_zcrx_ifq *ifq,
435 			       struct io_uring_zcrx_area_reg *area_reg,
436 			       struct io_uring_zcrx_ifq_reg *reg)
437 {
438 	int buf_size_shift = PAGE_SHIFT;
439 	struct io_zcrx_area *area;
440 	unsigned nr_iovs;
441 	int i, ret;
442 
443 	if (reg->rx_buf_len) {
444 		if (!is_power_of_2(reg->rx_buf_len) ||
445 		     reg->rx_buf_len < PAGE_SIZE)
446 			return -EINVAL;
447 		buf_size_shift = ilog2(reg->rx_buf_len);
448 	}
449 
450 	ret = -ENOMEM;
451 	area = kzalloc(sizeof(*area), GFP_KERNEL);
452 	if (!area)
453 		goto err;
454 	area->ifq = ifq;
455 
456 	ret = io_import_area(ifq, &area->mem, area_reg);
457 	if (ret)
458 		goto err;
459 
460 	if (buf_size_shift > io_area_max_shift(&area->mem)) {
461 		ret = -ERANGE;
462 		goto err;
463 	}
464 
465 	ifq->niov_shift = buf_size_shift;
466 	nr_iovs = area->mem.size >> ifq->niov_shift;
467 	area->nia.num_niovs = nr_iovs;
468 
469 	ret = -ENOMEM;
470 	area->nia.niovs = kvmalloc_array(nr_iovs, sizeof(area->nia.niovs[0]),
471 					 GFP_KERNEL_ACCOUNT | __GFP_ZERO);
472 	if (!area->nia.niovs)
473 		goto err;
474 
475 	area->freelist = kvmalloc_array(nr_iovs, sizeof(area->freelist[0]),
476 					GFP_KERNEL_ACCOUNT | __GFP_ZERO);
477 	if (!area->freelist)
478 		goto err;
479 
480 	area->user_refs = kvmalloc_array(nr_iovs, sizeof(area->user_refs[0]),
481 					GFP_KERNEL_ACCOUNT | __GFP_ZERO);
482 	if (!area->user_refs)
483 		goto err;
484 
485 	for (i = 0; i < nr_iovs; i++) {
486 		struct net_iov *niov = &area->nia.niovs[i];
487 
488 		niov->owner = &area->nia;
489 		area->freelist[i] = i;
490 		atomic_set(&area->user_refs[i], 0);
491 		niov->type = NET_IOV_IOURING;
492 	}
493 
494 	area->free_count = nr_iovs;
495 	/* we're only supporting one area per ifq for now */
496 	area->area_id = 0;
497 	area_reg->rq_area_token = (u64)area->area_id << IORING_ZCRX_AREA_SHIFT;
498 	spin_lock_init(&area->freelist_lock);
499 
500 	ret = io_zcrx_append_area(ifq, area);
501 	if (!ret)
502 		return 0;
503 err:
504 	if (area)
505 		io_zcrx_free_area(ifq, area);
506 	return ret;
507 }
508 
509 static struct io_zcrx_ifq *io_zcrx_ifq_alloc(struct io_ring_ctx *ctx)
510 {
511 	struct io_zcrx_ifq *ifq;
512 
513 	ifq = kzalloc(sizeof(*ifq), GFP_KERNEL);
514 	if (!ifq)
515 		return NULL;
516 
517 	ifq->if_rxq = -1;
518 	spin_lock_init(&ifq->rq_lock);
519 	mutex_init(&ifq->pp_lock);
520 	refcount_set(&ifq->refs, 1);
521 	refcount_set(&ifq->user_refs, 1);
522 	return ifq;
523 }
524 
525 static void io_zcrx_drop_netdev(struct io_zcrx_ifq *ifq)
526 {
527 	guard(mutex)(&ifq->pp_lock);
528 
529 	if (!ifq->netdev)
530 		return;
531 	netdev_put(ifq->netdev, &ifq->netdev_tracker);
532 	ifq->netdev = NULL;
533 }
534 
535 static void io_close_queue(struct io_zcrx_ifq *ifq)
536 {
537 	struct net_device *netdev;
538 	netdevice_tracker netdev_tracker;
539 	struct pp_memory_provider_params p = {
540 		.mp_ops = &io_uring_pp_zc_ops,
541 		.mp_priv = ifq,
542 	};
543 
544 	scoped_guard(mutex, &ifq->pp_lock) {
545 		netdev = ifq->netdev;
546 		netdev_tracker = ifq->netdev_tracker;
547 		ifq->netdev = NULL;
548 	}
549 
550 	if (netdev) {
551 		if (ifq->if_rxq != -1)
552 			net_mp_close_rxq(netdev, ifq->if_rxq, &p);
553 		netdev_put(netdev, &netdev_tracker);
554 	}
555 	ifq->if_rxq = -1;
556 }
557 
558 static void io_zcrx_ifq_free(struct io_zcrx_ifq *ifq)
559 {
560 	io_close_queue(ifq);
561 
562 	if (ifq->area)
563 		io_zcrx_free_area(ifq, ifq->area);
564 	free_uid(ifq->user);
565 	if (ifq->mm_account)
566 		mmdrop(ifq->mm_account);
567 	if (ifq->dev)
568 		put_device(ifq->dev);
569 
570 	io_free_rbuf_ring(ifq);
571 	mutex_destroy(&ifq->pp_lock);
572 	kfree(ifq);
573 }
574 
575 static void io_put_zcrx_ifq(struct io_zcrx_ifq *ifq)
576 {
577 	if (refcount_dec_and_test(&ifq->refs))
578 		io_zcrx_ifq_free(ifq);
579 }
580 
581 static void io_zcrx_return_niov_freelist(struct net_iov *niov)
582 {
583 	struct io_zcrx_area *area = io_zcrx_iov_to_area(niov);
584 
585 	spin_lock_bh(&area->freelist_lock);
586 	area->freelist[area->free_count++] = net_iov_idx(niov);
587 	spin_unlock_bh(&area->freelist_lock);
588 }
589 
590 static void io_zcrx_return_niov(struct net_iov *niov)
591 {
592 	netmem_ref netmem = net_iov_to_netmem(niov);
593 
594 	if (!niov->desc.pp) {
595 		/* copy fallback allocated niovs */
596 		io_zcrx_return_niov_freelist(niov);
597 		return;
598 	}
599 	page_pool_put_unrefed_netmem(niov->desc.pp, netmem, -1, false);
600 }
601 
602 static void io_zcrx_scrub(struct io_zcrx_ifq *ifq)
603 {
604 	struct io_zcrx_area *area = ifq->area;
605 	int i;
606 
607 	if (!area)
608 		return;
609 
610 	/* Reclaim back all buffers given to the user space. */
611 	for (i = 0; i < area->nia.num_niovs; i++) {
612 		struct net_iov *niov = &area->nia.niovs[i];
613 		int nr;
614 
615 		if (!atomic_read(io_get_user_counter(niov)))
616 			continue;
617 		nr = atomic_xchg(io_get_user_counter(niov), 0);
618 		if (nr && !page_pool_unref_netmem(net_iov_to_netmem(niov), nr))
619 			io_zcrx_return_niov(niov);
620 	}
621 }
622 
623 static void zcrx_unregister(struct io_zcrx_ifq *ifq)
624 {
625 	if (refcount_dec_and_test(&ifq->user_refs)) {
626 		io_close_queue(ifq);
627 		io_zcrx_scrub(ifq);
628 	}
629 	io_put_zcrx_ifq(ifq);
630 }
631 
632 struct io_mapped_region *io_zcrx_get_region(struct io_ring_ctx *ctx,
633 					    unsigned int id)
634 {
635 	struct io_zcrx_ifq *ifq = xa_load(&ctx->zcrx_ctxs, id);
636 
637 	lockdep_assert_held(&ctx->mmap_lock);
638 
639 	return ifq ? &ifq->region : NULL;
640 }
641 
642 static int zcrx_box_release(struct inode *inode, struct file *file)
643 {
644 	struct io_zcrx_ifq *ifq = file->private_data;
645 
646 	if (WARN_ON_ONCE(!ifq))
647 		return -EFAULT;
648 	zcrx_unregister(ifq);
649 	return 0;
650 }
651 
652 static const struct file_operations zcrx_box_fops = {
653 	.owner		= THIS_MODULE,
654 	.release	= zcrx_box_release,
655 };
656 
657 static int zcrx_export(struct io_ring_ctx *ctx, struct io_zcrx_ifq *ifq,
658 		       struct zcrx_ctrl *ctrl, void __user *arg)
659 {
660 	struct zcrx_ctrl_export *ce = &ctrl->zc_export;
661 	struct file *file;
662 	int fd = -1;
663 
664 	if (!mem_is_zero(ce, sizeof(*ce)))
665 		return -EINVAL;
666 	fd = get_unused_fd_flags(O_CLOEXEC);
667 	if (fd < 0)
668 		return fd;
669 
670 	ce->zcrx_fd = fd;
671 	if (copy_to_user(arg, ctrl, sizeof(*ctrl))) {
672 		put_unused_fd(fd);
673 		return -EFAULT;
674 	}
675 
676 	refcount_inc(&ifq->refs);
677 	refcount_inc(&ifq->user_refs);
678 
679 	file = anon_inode_create_getfile("[zcrx]", &zcrx_box_fops,
680 					 ifq, O_CLOEXEC, NULL);
681 	if (IS_ERR(file)) {
682 		put_unused_fd(fd);
683 		zcrx_unregister(ifq);
684 		return PTR_ERR(file);
685 	}
686 
687 	fd_install(fd, file);
688 	return 0;
689 }
690 
691 static int import_zcrx(struct io_ring_ctx *ctx,
692 		       struct io_uring_zcrx_ifq_reg __user *arg,
693 		       struct io_uring_zcrx_ifq_reg *reg)
694 {
695 	struct io_zcrx_ifq *ifq;
696 	struct file *file;
697 	int fd, ret;
698 	u32 id;
699 
700 	if (!(ctx->flags & IORING_SETUP_DEFER_TASKRUN))
701 		return -EINVAL;
702 	if (!(ctx->flags & (IORING_SETUP_CQE32|IORING_SETUP_CQE_MIXED)))
703 		return -EINVAL;
704 	if (reg->if_rxq || reg->rq_entries || reg->area_ptr || reg->region_ptr)
705 		return -EINVAL;
706 	if (reg->flags & ~ZCRX_REG_IMPORT)
707 		return -EINVAL;
708 
709 	fd = reg->if_idx;
710 	CLASS(fd, f)(fd);
711 	if (fd_empty(f))
712 		return -EBADF;
713 
714 	file = fd_file(f);
715 	if (file->f_op != &zcrx_box_fops || !file->private_data)
716 		return -EBADF;
717 
718 	ifq = file->private_data;
719 	refcount_inc(&ifq->refs);
720 	refcount_inc(&ifq->user_refs);
721 
722 	scoped_guard(mutex, &ctx->mmap_lock) {
723 		ret = xa_alloc(&ctx->zcrx_ctxs, &id, NULL, xa_limit_31b, GFP_KERNEL);
724 		if (ret)
725 			goto err;
726 	}
727 
728 	reg->zcrx_id = id;
729 	io_fill_zcrx_offsets(&reg->offsets);
730 	if (copy_to_user(arg, reg, sizeof(*reg))) {
731 		ret = -EFAULT;
732 		goto err_xa_erase;
733 	}
734 
735 	scoped_guard(mutex, &ctx->mmap_lock) {
736 		ret = -ENOMEM;
737 		if (xa_store(&ctx->zcrx_ctxs, id, ifq, GFP_KERNEL))
738 			goto err_xa_erase;
739 	}
740 
741 	return 0;
742 err_xa_erase:
743 	scoped_guard(mutex, &ctx->mmap_lock)
744 		xa_erase(&ctx->zcrx_ctxs, id);
745 err:
746 	zcrx_unregister(ifq);
747 	return ret;
748 }
749 
750 int io_register_zcrx_ifq(struct io_ring_ctx *ctx,
751 			  struct io_uring_zcrx_ifq_reg __user *arg)
752 {
753 	struct pp_memory_provider_params mp_param = {};
754 	struct io_uring_zcrx_area_reg area;
755 	struct io_uring_zcrx_ifq_reg reg;
756 	struct io_uring_region_desc rd;
757 	struct io_zcrx_ifq *ifq;
758 	int ret;
759 	u32 id;
760 
761 	/*
762 	 * 1. Interface queue allocation.
763 	 * 2. It can observe data destined for sockets of other tasks.
764 	 */
765 	if (!capable(CAP_NET_ADMIN))
766 		return -EPERM;
767 
768 	/* mandatory io_uring features for zc rx */
769 	if (!(ctx->flags & IORING_SETUP_DEFER_TASKRUN))
770 		return -EINVAL;
771 	if (!(ctx->flags & (IORING_SETUP_CQE32|IORING_SETUP_CQE_MIXED)))
772 		return -EINVAL;
773 	if (copy_from_user(&reg, arg, sizeof(reg)))
774 		return -EFAULT;
775 	if (!mem_is_zero(&reg.__resv, sizeof(reg.__resv)) || reg.zcrx_id)
776 		return -EINVAL;
777 	if (reg.flags & ZCRX_REG_IMPORT)
778 		return import_zcrx(ctx, arg, &reg);
779 	if (copy_from_user(&rd, u64_to_user_ptr(reg.region_ptr), sizeof(rd)))
780 		return -EFAULT;
781 	if (reg.if_rxq == -1 || !reg.rq_entries || reg.flags)
782 		return -EINVAL;
783 	if (reg.rq_entries > IO_RQ_MAX_ENTRIES) {
784 		if (!(ctx->flags & IORING_SETUP_CLAMP))
785 			return -EINVAL;
786 		reg.rq_entries = IO_RQ_MAX_ENTRIES;
787 	}
788 	reg.rq_entries = roundup_pow_of_two(reg.rq_entries);
789 
790 	if (copy_from_user(&area, u64_to_user_ptr(reg.area_ptr), sizeof(area)))
791 		return -EFAULT;
792 
793 	ifq = io_zcrx_ifq_alloc(ctx);
794 	if (!ifq)
795 		return -ENOMEM;
796 
797 	if (ctx->user) {
798 		get_uid(ctx->user);
799 		ifq->user = ctx->user;
800 	}
801 	if (ctx->mm_account) {
802 		mmgrab(ctx->mm_account);
803 		ifq->mm_account = ctx->mm_account;
804 	}
805 	ifq->rq_entries = reg.rq_entries;
806 
807 	scoped_guard(mutex, &ctx->mmap_lock) {
808 		/* preallocate id */
809 		ret = xa_alloc(&ctx->zcrx_ctxs, &id, NULL, xa_limit_31b, GFP_KERNEL);
810 		if (ret)
811 			goto ifq_free;
812 	}
813 
814 	ret = io_allocate_rbuf_ring(ctx, ifq, &reg, &rd, id);
815 	if (ret)
816 		goto err;
817 
818 	ifq->netdev = netdev_get_by_index_lock(current->nsproxy->net_ns, reg.if_idx);
819 	if (!ifq->netdev) {
820 		ret = -ENODEV;
821 		goto err;
822 	}
823 	netdev_hold(ifq->netdev, &ifq->netdev_tracker, GFP_KERNEL);
824 
825 	ifq->dev = netdev_queue_get_dma_dev(ifq->netdev, reg.if_rxq);
826 	if (!ifq->dev) {
827 		ret = -EOPNOTSUPP;
828 		goto netdev_put_unlock;
829 	}
830 	get_device(ifq->dev);
831 
832 	ret = io_zcrx_create_area(ifq, &area, &reg);
833 	if (ret)
834 		goto netdev_put_unlock;
835 
836 	mp_param.rx_page_size = 1U << ifq->niov_shift;
837 	mp_param.mp_ops = &io_uring_pp_zc_ops;
838 	mp_param.mp_priv = ifq;
839 	ret = __net_mp_open_rxq(ifq->netdev, reg.if_rxq, &mp_param, NULL);
840 	if (ret)
841 		goto netdev_put_unlock;
842 	netdev_unlock(ifq->netdev);
843 	ifq->if_rxq = reg.if_rxq;
844 
845 	reg.zcrx_id = id;
846 
847 	scoped_guard(mutex, &ctx->mmap_lock) {
848 		/* publish ifq */
849 		ret = -ENOMEM;
850 		if (xa_store(&ctx->zcrx_ctxs, id, ifq, GFP_KERNEL))
851 			goto err;
852 	}
853 
854 	reg.rx_buf_len = 1U << ifq->niov_shift;
855 
856 	if (copy_to_user(arg, &reg, sizeof(reg)) ||
857 	    copy_to_user(u64_to_user_ptr(reg.region_ptr), &rd, sizeof(rd)) ||
858 	    copy_to_user(u64_to_user_ptr(reg.area_ptr), &area, sizeof(area))) {
859 		ret = -EFAULT;
860 		goto err;
861 	}
862 	return 0;
863 netdev_put_unlock:
864 	netdev_unlock(ifq->netdev);
865 err:
866 	scoped_guard(mutex, &ctx->mmap_lock)
867 		xa_erase(&ctx->zcrx_ctxs, id);
868 ifq_free:
869 	zcrx_unregister(ifq);
870 	return ret;
871 }
872 
873 static struct net_iov *__io_zcrx_get_free_niov(struct io_zcrx_area *area)
874 {
875 	unsigned niov_idx;
876 
877 	lockdep_assert_held(&area->freelist_lock);
878 
879 	niov_idx = area->freelist[--area->free_count];
880 	return &area->nia.niovs[niov_idx];
881 }
882 
883 void io_unregister_zcrx_ifqs(struct io_ring_ctx *ctx)
884 {
885 	struct io_zcrx_ifq *ifq;
886 
887 	lockdep_assert_held(&ctx->uring_lock);
888 
889 	while (1) {
890 		scoped_guard(mutex, &ctx->mmap_lock) {
891 			unsigned long id = 0;
892 
893 			ifq = xa_find(&ctx->zcrx_ctxs, &id, ULONG_MAX, XA_PRESENT);
894 			if (ifq)
895 				xa_erase(&ctx->zcrx_ctxs, id);
896 		}
897 		if (!ifq)
898 			break;
899 		zcrx_unregister(ifq);
900 	}
901 
902 	xa_destroy(&ctx->zcrx_ctxs);
903 }
904 
905 static inline u32 io_zcrx_rqring_entries(struct io_zcrx_ifq *ifq)
906 {
907 	u32 entries;
908 
909 	entries = smp_load_acquire(&ifq->rq_ring->tail) - ifq->cached_rq_head;
910 	return min(entries, ifq->rq_entries);
911 }
912 
913 static struct io_uring_zcrx_rqe *io_zcrx_get_rqe(struct io_zcrx_ifq *ifq,
914 						 unsigned mask)
915 {
916 	unsigned int idx = ifq->cached_rq_head++ & mask;
917 
918 	return &ifq->rqes[idx];
919 }
920 
921 static inline bool io_parse_rqe(struct io_uring_zcrx_rqe *rqe,
922 				struct io_zcrx_ifq *ifq,
923 				struct net_iov **ret_niov)
924 {
925 	unsigned niov_idx, area_idx;
926 	struct io_zcrx_area *area;
927 
928 	area_idx = rqe->off >> IORING_ZCRX_AREA_SHIFT;
929 	niov_idx = (rqe->off & ~IORING_ZCRX_AREA_MASK) >> ifq->niov_shift;
930 
931 	if (unlikely(rqe->__pad || area_idx))
932 		return false;
933 	area = ifq->area;
934 
935 	if (unlikely(niov_idx >= area->nia.num_niovs))
936 		return false;
937 	niov_idx = array_index_nospec(niov_idx, area->nia.num_niovs);
938 
939 	*ret_niov = &area->nia.niovs[niov_idx];
940 	return true;
941 }
942 
943 static void io_zcrx_ring_refill(struct page_pool *pp,
944 				struct io_zcrx_ifq *ifq)
945 {
946 	unsigned int mask = ifq->rq_entries - 1;
947 	unsigned int entries;
948 
949 	guard(spinlock_bh)(&ifq->rq_lock);
950 
951 	entries = io_zcrx_rqring_entries(ifq);
952 	entries = min_t(unsigned, entries, PP_ALLOC_CACHE_REFILL);
953 	if (unlikely(!entries))
954 		return;
955 
956 	do {
957 		struct io_uring_zcrx_rqe *rqe = io_zcrx_get_rqe(ifq, mask);
958 		struct net_iov *niov;
959 		netmem_ref netmem;
960 
961 		if (!io_parse_rqe(rqe, ifq, &niov))
962 			continue;
963 		if (!io_zcrx_put_niov_uref(niov))
964 			continue;
965 
966 		netmem = net_iov_to_netmem(niov);
967 		if (!page_pool_unref_and_test(netmem))
968 			continue;
969 
970 		if (unlikely(niov->desc.pp != pp)) {
971 			io_zcrx_return_niov(niov);
972 			continue;
973 		}
974 
975 		io_zcrx_sync_for_device(pp, niov);
976 		net_mp_netmem_place_in_cache(pp, netmem);
977 	} while (--entries);
978 
979 	smp_store_release(&ifq->rq_ring->head, ifq->cached_rq_head);
980 }
981 
982 static void io_zcrx_refill_slow(struct page_pool *pp, struct io_zcrx_ifq *ifq)
983 {
984 	struct io_zcrx_area *area = ifq->area;
985 
986 	spin_lock_bh(&area->freelist_lock);
987 	while (area->free_count && pp->alloc.count < PP_ALLOC_CACHE_REFILL) {
988 		struct net_iov *niov = __io_zcrx_get_free_niov(area);
989 		netmem_ref netmem = net_iov_to_netmem(niov);
990 
991 		net_mp_niov_set_page_pool(pp, niov);
992 		io_zcrx_sync_for_device(pp, niov);
993 		net_mp_netmem_place_in_cache(pp, netmem);
994 	}
995 	spin_unlock_bh(&area->freelist_lock);
996 }
997 
998 static netmem_ref io_pp_zc_alloc_netmems(struct page_pool *pp, gfp_t gfp)
999 {
1000 	struct io_zcrx_ifq *ifq = io_pp_to_ifq(pp);
1001 
1002 	/* pp should already be ensuring that */
1003 	if (unlikely(pp->alloc.count))
1004 		goto out_return;
1005 
1006 	io_zcrx_ring_refill(pp, ifq);
1007 	if (likely(pp->alloc.count))
1008 		goto out_return;
1009 
1010 	io_zcrx_refill_slow(pp, ifq);
1011 	if (!pp->alloc.count)
1012 		return 0;
1013 out_return:
1014 	return pp->alloc.cache[--pp->alloc.count];
1015 }
1016 
1017 static bool io_pp_zc_release_netmem(struct page_pool *pp, netmem_ref netmem)
1018 {
1019 	struct net_iov *niov;
1020 
1021 	if (WARN_ON_ONCE(!netmem_is_net_iov(netmem)))
1022 		return false;
1023 
1024 	niov = netmem_to_net_iov(netmem);
1025 	net_mp_niov_clear_page_pool(niov);
1026 	io_zcrx_return_niov_freelist(niov);
1027 	return false;
1028 }
1029 
1030 static int io_pp_zc_init(struct page_pool *pp)
1031 {
1032 	struct io_zcrx_ifq *ifq = io_pp_to_ifq(pp);
1033 	int ret;
1034 
1035 	if (WARN_ON_ONCE(!ifq))
1036 		return -EINVAL;
1037 	if (WARN_ON_ONCE(ifq->dev != pp->p.dev))
1038 		return -EINVAL;
1039 	if (WARN_ON_ONCE(!pp->dma_map))
1040 		return -EOPNOTSUPP;
1041 	if (pp->p.order + PAGE_SHIFT != ifq->niov_shift)
1042 		return -EINVAL;
1043 	if (pp->p.dma_dir != DMA_FROM_DEVICE)
1044 		return -EOPNOTSUPP;
1045 
1046 	ret = io_zcrx_map_area(ifq, ifq->area);
1047 	if (ret)
1048 		return ret;
1049 
1050 	refcount_inc(&ifq->refs);
1051 	return 0;
1052 }
1053 
1054 static void io_pp_zc_destroy(struct page_pool *pp)
1055 {
1056 	io_put_zcrx_ifq(io_pp_to_ifq(pp));
1057 }
1058 
1059 static int io_pp_nl_fill(void *mp_priv, struct sk_buff *rsp,
1060 			 struct netdev_rx_queue *rxq)
1061 {
1062 	struct nlattr *nest;
1063 	int type;
1064 
1065 	type = rxq ? NETDEV_A_QUEUE_IO_URING : NETDEV_A_PAGE_POOL_IO_URING;
1066 	nest = nla_nest_start(rsp, type);
1067 	if (!nest)
1068 		return -EMSGSIZE;
1069 	nla_nest_end(rsp, nest);
1070 
1071 	return 0;
1072 }
1073 
1074 static void io_pp_uninstall(void *mp_priv, struct netdev_rx_queue *rxq)
1075 {
1076 	struct pp_memory_provider_params *p = &rxq->mp_params;
1077 	struct io_zcrx_ifq *ifq = mp_priv;
1078 
1079 	io_zcrx_drop_netdev(ifq);
1080 	if (ifq->area)
1081 		io_zcrx_unmap_area(ifq, ifq->area);
1082 
1083 	p->mp_ops = NULL;
1084 	p->mp_priv = NULL;
1085 }
1086 
1087 static const struct memory_provider_ops io_uring_pp_zc_ops = {
1088 	.alloc_netmems		= io_pp_zc_alloc_netmems,
1089 	.release_netmem		= io_pp_zc_release_netmem,
1090 	.init			= io_pp_zc_init,
1091 	.destroy		= io_pp_zc_destroy,
1092 	.nl_fill		= io_pp_nl_fill,
1093 	.uninstall		= io_pp_uninstall,
1094 };
1095 
1096 static unsigned zcrx_parse_rq(netmem_ref *netmem_array, unsigned nr,
1097 			      struct io_zcrx_ifq *zcrx)
1098 {
1099 	unsigned int mask = zcrx->rq_entries - 1;
1100 	unsigned int i;
1101 
1102 	nr = min(nr, io_zcrx_rqring_entries(zcrx));
1103 	for (i = 0; i < nr; i++) {
1104 		struct io_uring_zcrx_rqe *rqe = io_zcrx_get_rqe(zcrx, mask);
1105 		struct net_iov *niov;
1106 
1107 		if (!io_parse_rqe(rqe, zcrx, &niov))
1108 			break;
1109 		netmem_array[i] = net_iov_to_netmem(niov);
1110 	}
1111 
1112 	smp_store_release(&zcrx->rq_ring->head, zcrx->cached_rq_head);
1113 	return i;
1114 }
1115 
1116 #define ZCRX_FLUSH_BATCH 32
1117 
1118 static void zcrx_return_buffers(netmem_ref *netmems, unsigned nr)
1119 {
1120 	unsigned i;
1121 
1122 	for (i = 0; i < nr; i++) {
1123 		netmem_ref netmem = netmems[i];
1124 		struct net_iov *niov = netmem_to_net_iov(netmem);
1125 
1126 		if (!io_zcrx_put_niov_uref(niov))
1127 			continue;
1128 		if (!page_pool_unref_and_test(netmem))
1129 			continue;
1130 		io_zcrx_return_niov(niov);
1131 	}
1132 }
1133 
1134 static int zcrx_flush_rq(struct io_ring_ctx *ctx, struct io_zcrx_ifq *zcrx,
1135 			 struct zcrx_ctrl *ctrl)
1136 {
1137 	struct zcrx_ctrl_flush_rq *frq = &ctrl->zc_flush;
1138 	netmem_ref netmems[ZCRX_FLUSH_BATCH];
1139 	unsigned total = 0;
1140 	unsigned nr;
1141 
1142 	if (!mem_is_zero(&frq->__resv, sizeof(frq->__resv)))
1143 		return -EINVAL;
1144 
1145 	do {
1146 		scoped_guard(spinlock_bh, &zcrx->rq_lock) {
1147 			nr = zcrx_parse_rq(netmems, ZCRX_FLUSH_BATCH, zcrx);
1148 			zcrx_return_buffers(netmems, nr);
1149 		}
1150 
1151 		total += nr;
1152 
1153 		if (fatal_signal_pending(current))
1154 			break;
1155 		cond_resched();
1156 	} while (nr == ZCRX_FLUSH_BATCH && total < zcrx->rq_entries);
1157 
1158 	return 0;
1159 }
1160 
1161 int io_zcrx_ctrl(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
1162 {
1163 	struct zcrx_ctrl ctrl;
1164 	struct io_zcrx_ifq *zcrx;
1165 
1166 	if (nr_args)
1167 		return -EINVAL;
1168 	if (copy_from_user(&ctrl, arg, sizeof(ctrl)))
1169 		return -EFAULT;
1170 	if (!mem_is_zero(&ctrl.__resv, sizeof(ctrl.__resv)))
1171 		return -EFAULT;
1172 
1173 	zcrx = xa_load(&ctx->zcrx_ctxs, ctrl.zcrx_id);
1174 	if (!zcrx)
1175 		return -ENXIO;
1176 
1177 	switch (ctrl.op) {
1178 	case ZCRX_CTRL_FLUSH_RQ:
1179 		return zcrx_flush_rq(ctx, zcrx, &ctrl);
1180 	case ZCRX_CTRL_EXPORT:
1181 		return zcrx_export(ctx, zcrx, &ctrl, arg);
1182 	}
1183 
1184 	return -EOPNOTSUPP;
1185 }
1186 
1187 static bool io_zcrx_queue_cqe(struct io_kiocb *req, struct net_iov *niov,
1188 			      struct io_zcrx_ifq *ifq, int off, int len)
1189 {
1190 	struct io_ring_ctx *ctx = req->ctx;
1191 	struct io_uring_zcrx_cqe *rcqe;
1192 	struct io_zcrx_area *area;
1193 	struct io_uring_cqe *cqe;
1194 	u64 offset;
1195 
1196 	if (!io_defer_get_uncommited_cqe(ctx, &cqe))
1197 		return false;
1198 
1199 	cqe->user_data = req->cqe.user_data;
1200 	cqe->res = len;
1201 	cqe->flags = IORING_CQE_F_MORE;
1202 	if (ctx->flags & IORING_SETUP_CQE_MIXED)
1203 		cqe->flags |= IORING_CQE_F_32;
1204 
1205 	area = io_zcrx_iov_to_area(niov);
1206 	offset = off + (net_iov_idx(niov) << ifq->niov_shift);
1207 	rcqe = (struct io_uring_zcrx_cqe *)(cqe + 1);
1208 	rcqe->off = offset + ((u64)area->area_id << IORING_ZCRX_AREA_SHIFT);
1209 	rcqe->__pad = 0;
1210 	return true;
1211 }
1212 
1213 static struct net_iov *io_alloc_fallback_niov(struct io_zcrx_ifq *ifq)
1214 {
1215 	struct io_zcrx_area *area = ifq->area;
1216 	struct net_iov *niov = NULL;
1217 
1218 	if (area->mem.is_dmabuf)
1219 		return NULL;
1220 
1221 	spin_lock_bh(&area->freelist_lock);
1222 	if (area->free_count)
1223 		niov = __io_zcrx_get_free_niov(area);
1224 	spin_unlock_bh(&area->freelist_lock);
1225 
1226 	if (niov)
1227 		page_pool_fragment_netmem(net_iov_to_netmem(niov), 1);
1228 	return niov;
1229 }
1230 
1231 struct io_copy_cache {
1232 	struct page		*page;
1233 	unsigned long		offset;
1234 	size_t			size;
1235 };
1236 
1237 static ssize_t io_copy_page(struct io_copy_cache *cc, struct page *src_page,
1238 			    unsigned int src_offset, size_t len)
1239 {
1240 	size_t copied = 0;
1241 
1242 	len = min(len, cc->size);
1243 
1244 	while (len) {
1245 		void *src_addr, *dst_addr;
1246 		struct page *dst_page = cc->page;
1247 		unsigned dst_offset = cc->offset;
1248 		size_t n = len;
1249 
1250 		if (folio_test_partial_kmap(page_folio(dst_page)) ||
1251 		    folio_test_partial_kmap(page_folio(src_page))) {
1252 			dst_page += dst_offset / PAGE_SIZE;
1253 			dst_offset = offset_in_page(dst_offset);
1254 			src_page += src_offset / PAGE_SIZE;
1255 			src_offset = offset_in_page(src_offset);
1256 			n = min(PAGE_SIZE - src_offset, PAGE_SIZE - dst_offset);
1257 			n = min(n, len);
1258 		}
1259 
1260 		dst_addr = kmap_local_page(dst_page) + dst_offset;
1261 		src_addr = kmap_local_page(src_page) + src_offset;
1262 
1263 		memcpy(dst_addr, src_addr, n);
1264 
1265 		kunmap_local(src_addr);
1266 		kunmap_local(dst_addr);
1267 
1268 		cc->size -= n;
1269 		cc->offset += n;
1270 		src_offset += n;
1271 		len -= n;
1272 		copied += n;
1273 	}
1274 	return copied;
1275 }
1276 
1277 static ssize_t io_zcrx_copy_chunk(struct io_kiocb *req, struct io_zcrx_ifq *ifq,
1278 				  struct page *src_page, unsigned int src_offset,
1279 				  size_t len)
1280 {
1281 	size_t copied = 0;
1282 	int ret = 0;
1283 
1284 	while (len) {
1285 		struct io_copy_cache cc;
1286 		struct net_iov *niov;
1287 		size_t n;
1288 
1289 		niov = io_alloc_fallback_niov(ifq);
1290 		if (!niov) {
1291 			ret = -ENOMEM;
1292 			break;
1293 		}
1294 
1295 		cc.page = io_zcrx_iov_page(niov);
1296 		cc.offset = 0;
1297 		cc.size = PAGE_SIZE;
1298 
1299 		n = io_copy_page(&cc, src_page, src_offset, len);
1300 
1301 		if (!io_zcrx_queue_cqe(req, niov, ifq, 0, n)) {
1302 			io_zcrx_return_niov(niov);
1303 			ret = -ENOSPC;
1304 			break;
1305 		}
1306 
1307 		io_zcrx_get_niov_uref(niov);
1308 		src_offset += n;
1309 		len -= n;
1310 		copied += n;
1311 	}
1312 
1313 	return copied ? copied : ret;
1314 }
1315 
1316 static int io_zcrx_copy_frag(struct io_kiocb *req, struct io_zcrx_ifq *ifq,
1317 			     const skb_frag_t *frag, int off, int len)
1318 {
1319 	struct page *page = skb_frag_page(frag);
1320 
1321 	return io_zcrx_copy_chunk(req, ifq, page, off + skb_frag_off(frag), len);
1322 }
1323 
1324 static int io_zcrx_recv_frag(struct io_kiocb *req, struct io_zcrx_ifq *ifq,
1325 			     const skb_frag_t *frag, int off, int len)
1326 {
1327 	struct net_iov *niov;
1328 	struct page_pool *pp;
1329 
1330 	if (unlikely(!skb_frag_is_net_iov(frag)))
1331 		return io_zcrx_copy_frag(req, ifq, frag, off, len);
1332 
1333 	niov = netmem_to_net_iov(frag->netmem);
1334 	pp = niov->desc.pp;
1335 
1336 	if (!pp || pp->mp_ops != &io_uring_pp_zc_ops || io_pp_to_ifq(pp) != ifq)
1337 		return -EFAULT;
1338 
1339 	if (!io_zcrx_queue_cqe(req, niov, ifq, off + skb_frag_off(frag), len))
1340 		return -ENOSPC;
1341 
1342 	/*
1343 	 * Prevent it from being recycled while user is accessing it.
1344 	 * It has to be done before grabbing a user reference.
1345 	 */
1346 	page_pool_ref_netmem(net_iov_to_netmem(niov));
1347 	io_zcrx_get_niov_uref(niov);
1348 	return len;
1349 }
1350 
1351 static int
1352 io_zcrx_recv_skb(read_descriptor_t *desc, struct sk_buff *skb,
1353 		 unsigned int offset, size_t len)
1354 {
1355 	struct io_zcrx_args *args = desc->arg.data;
1356 	struct io_zcrx_ifq *ifq = args->ifq;
1357 	struct io_kiocb *req = args->req;
1358 	struct sk_buff *frag_iter;
1359 	unsigned start, start_off = offset;
1360 	int i, copy, end, off;
1361 	int ret = 0;
1362 
1363 	len = min_t(size_t, len, desc->count);
1364 	/*
1365 	 * __tcp_read_sock() always calls io_zcrx_recv_skb one last time, even
1366 	 * if desc->count is already 0. This is caused by the if (offset + 1 !=
1367 	 * skb->len) check. Return early in this case to break out of
1368 	 * __tcp_read_sock().
1369 	 */
1370 	if (!len)
1371 		return 0;
1372 	if (unlikely(args->nr_skbs++ > IO_SKBS_PER_CALL_LIMIT))
1373 		return -EAGAIN;
1374 
1375 	if (unlikely(offset < skb_headlen(skb))) {
1376 		ssize_t copied;
1377 		size_t to_copy;
1378 
1379 		to_copy = min_t(size_t, skb_headlen(skb) - offset, len);
1380 		copied = io_zcrx_copy_chunk(req, ifq, virt_to_page(skb->data),
1381 					    offset_in_page(skb->data) + offset,
1382 					    to_copy);
1383 		if (copied < 0) {
1384 			ret = copied;
1385 			goto out;
1386 		}
1387 		offset += copied;
1388 		len -= copied;
1389 		if (!len)
1390 			goto out;
1391 		if (offset != skb_headlen(skb))
1392 			goto out;
1393 	}
1394 
1395 	start = skb_headlen(skb);
1396 
1397 	for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
1398 		const skb_frag_t *frag;
1399 
1400 		if (WARN_ON(start > offset + len))
1401 			return -EFAULT;
1402 
1403 		frag = &skb_shinfo(skb)->frags[i];
1404 		end = start + skb_frag_size(frag);
1405 
1406 		if (offset < end) {
1407 			copy = end - offset;
1408 			if (copy > len)
1409 				copy = len;
1410 
1411 			off = offset - start;
1412 			ret = io_zcrx_recv_frag(req, ifq, frag, off, copy);
1413 			if (ret < 0)
1414 				goto out;
1415 
1416 			offset += ret;
1417 			len -= ret;
1418 			if (len == 0 || ret != copy)
1419 				goto out;
1420 		}
1421 		start = end;
1422 	}
1423 
1424 	skb_walk_frags(skb, frag_iter) {
1425 		if (WARN_ON(start > offset + len))
1426 			return -EFAULT;
1427 
1428 		end = start + frag_iter->len;
1429 		if (offset < end) {
1430 			size_t count;
1431 
1432 			copy = end - offset;
1433 			if (copy > len)
1434 				copy = len;
1435 
1436 			off = offset - start;
1437 			count = desc->count;
1438 			ret = io_zcrx_recv_skb(desc, frag_iter, off, copy);
1439 			desc->count = count;
1440 			if (ret < 0)
1441 				goto out;
1442 
1443 			offset += ret;
1444 			len -= ret;
1445 			if (len == 0 || ret != copy)
1446 				goto out;
1447 		}
1448 		start = end;
1449 	}
1450 
1451 out:
1452 	if (offset == start_off)
1453 		return ret;
1454 	desc->count -= (offset - start_off);
1455 	return offset - start_off;
1456 }
1457 
1458 static int io_zcrx_tcp_recvmsg(struct io_kiocb *req, struct io_zcrx_ifq *ifq,
1459 				struct sock *sk, int flags,
1460 				unsigned issue_flags, unsigned int *outlen)
1461 {
1462 	unsigned int len = *outlen;
1463 	struct io_zcrx_args args = {
1464 		.req = req,
1465 		.ifq = ifq,
1466 		.sock = sk->sk_socket,
1467 	};
1468 	read_descriptor_t rd_desc = {
1469 		.count = len ? len : UINT_MAX,
1470 		.arg.data = &args,
1471 	};
1472 	int ret;
1473 
1474 	lock_sock(sk);
1475 	ret = tcp_read_sock(sk, &rd_desc, io_zcrx_recv_skb);
1476 	if (len && ret > 0)
1477 		*outlen = len - ret;
1478 	if (ret <= 0) {
1479 		if (ret < 0 || sock_flag(sk, SOCK_DONE))
1480 			goto out;
1481 		if (sk->sk_err)
1482 			ret = sock_error(sk);
1483 		else if (sk->sk_shutdown & RCV_SHUTDOWN)
1484 			goto out;
1485 		else if (sk->sk_state == TCP_CLOSE)
1486 			ret = -ENOTCONN;
1487 		else
1488 			ret = -EAGAIN;
1489 	} else if (unlikely(args.nr_skbs > IO_SKBS_PER_CALL_LIMIT) &&
1490 		   (issue_flags & IO_URING_F_MULTISHOT)) {
1491 		ret = IOU_REQUEUE;
1492 	} else if (sock_flag(sk, SOCK_DONE)) {
1493 		/* Make it to retry until it finally gets 0. */
1494 		if (issue_flags & IO_URING_F_MULTISHOT)
1495 			ret = IOU_REQUEUE;
1496 		else
1497 			ret = -EAGAIN;
1498 	}
1499 out:
1500 	release_sock(sk);
1501 	return ret;
1502 }
1503 
1504 int io_zcrx_recv(struct io_kiocb *req, struct io_zcrx_ifq *ifq,
1505 		 struct socket *sock, unsigned int flags,
1506 		 unsigned issue_flags, unsigned int *len)
1507 {
1508 	struct sock *sk = sock->sk;
1509 	const struct proto *prot = READ_ONCE(sk->sk_prot);
1510 
1511 	if (prot->recvmsg != tcp_recvmsg)
1512 		return -EPROTONOSUPPORT;
1513 
1514 	sock_rps_record_flow(sk);
1515 	return io_zcrx_tcp_recvmsg(req, ifq, sk, flags, issue_flags, len);
1516 }
1517