2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
4 * Copyright (c) 2007 Seccuris Inc.
7 * This software was developed by Robert N. M. Watson under contract to
10 * Redistribution and use in source and binary forms, with or without
11 * modification, are permitted provided that the following conditions
13 * 1. Redistributions of source code must retain the above copyright
14 * notice, this list of conditions and the following disclaimer.
15 * 2. Redistributions in binary form must reproduce the above copyright
16 * notice, this list of conditions and the following disclaimer in the
17 * documentation and/or other materials provided with the distribution.
19 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
20 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
21 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
22 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
23 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
24 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
25 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
26 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
27 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
28 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32 #include <sys/cdefs.h>
33 __FBSDID("$FreeBSD$");
37 #include <sys/param.h>
39 #include <sys/malloc.h>
41 #include <sys/mutex.h>
43 #include <sys/sf_buf.h>
44 #include <sys/socket.h>
47 #include <machine/atomic.h>
51 #include <net/bpf_zerocopy.h>
52 #include <net/bpfdesc.h>
55 #include <vm/vm_param.h>
57 #include <vm/vm_extern.h>
58 #include <vm/vm_map.h>
59 #include <vm/vm_page.h>
62 * Zero-copy buffer scheme for BPF: user space "donates" two buffers, which
63 * are mapped into the kernel address space using sf_bufs and used directly
64 * by BPF. Memory is wired since page faults cannot be tolerated in the
65 * contexts where the buffers are copied to (locks held, interrupt context,
66 * etc). Access to shared memory buffers is synchronized using a header on
67 * each buffer, allowing the number of system calls to go to zero as BPF
68 * reaches saturation (buffers filled as fast as they can be drained by the
69 * user process). Full details of the protocol for communicating between the
70 * user process and BPF may be found in bpf(4).
74 * Maximum number of pages per buffer. Since all BPF devices use two, the
75 * maximum per device is 2*BPF_MAX_PAGES. Resource limits on the number of
76 * sf_bufs may be an issue, so do not set this too high. On older systems,
77 * kernel address space limits may also be an issue.
79 #define BPF_MAX_PAGES 512
82 * struct zbuf describes a memory buffer loaned by a user process to the
83 * kernel. We represent this as a series of pages managed using an array of
84 * sf_bufs. Even though the memory is contiguous in user space, it may not
85 * be mapped contiguously in the kernel (i.e., a set of physically
86 * non-contiguous pages in the direct map region) so we must implement
87 * scatter-gather copying. One significant mitigating factor is that on
88 * systems with a direct memory map, we can avoid TLB misses.
90 * At the front of the shared memory region is a bpf_zbuf_header, which
91 * contains shared control data to allow user space and the kernel to
92 * synchronize; this is included in zb_size, but not bpf_bufsize, so that BPF
93 * knows that the space is not available.
96 vm_offset_t zb_uaddr; /* User address at time of setup. */
97 size_t zb_size; /* Size of buffer, incl. header. */
98 u_int zb_numpages; /* Number of pages. */
99 int zb_flags; /* Flags on zbuf. */
100 struct sf_buf **zb_pages; /* Pages themselves. */
101 struct bpf_zbuf_header *zb_header; /* Shared header. */
105 * When a buffer has been assigned to userspace, flag it as such, as the
106 * buffer may remain in the store position as a result of the user process
107 * not yet having acknowledged the buffer in the hold position yet.
109 #define ZBUF_FLAG_ASSIGNED 0x00000001 /* Set when owned by user. */
112 * Release a page we've previously wired.
115 zbuf_page_free(vm_page_t pp)
118 vm_page_unwire(pp, PQ_INACTIVE);
122 * Free an sf_buf with attached page.
125 zbuf_sfbuf_free(struct sf_buf *sf)
129 pp = sf_buf_page(sf);
135 * Free a zbuf, including its page array, sbufs, and pages. Allow partially
136 * allocated zbufs to be freed so that it may be used even during a zbuf
140 zbuf_free(struct zbuf *zb)
144 for (i = 0; i < zb->zb_numpages; i++) {
145 if (zb->zb_pages[i] != NULL)
146 zbuf_sfbuf_free(zb->zb_pages[i]);
148 free(zb->zb_pages, M_BPF);
153 * Given a user pointer to a page of user memory, return an sf_buf for the
154 * page. Because we may be requesting quite a few sf_bufs, prefer failure to
155 * deadlock and use SFB_NOWAIT.
157 static struct sf_buf *
158 zbuf_sfbuf_get(struct vm_map *map, vm_offset_t uaddr)
163 if (vm_fault_quick_hold_pages(map, uaddr, PAGE_SIZE, VM_PROT_READ |
164 VM_PROT_WRITE, &pp, 1) < 0)
166 sf = sf_buf_alloc(pp, SFB_NOWAIT);
175 * Create a zbuf describing a range of user address space memory. Validate
176 * page alignment, size requirements, etc.
179 zbuf_setup(struct thread *td, vm_offset_t uaddr, size_t len,
189 * User address must be page-aligned.
191 if (uaddr & PAGE_MASK)
195 * Length must be an integer number of full pages.
201 * Length must not exceed per-buffer resource limit.
203 if ((len / PAGE_SIZE) > BPF_MAX_PAGES)
207 * Allocate the buffer and set up each page with is own sf_buf.
210 zb = malloc(sizeof(*zb), M_BPF, M_ZERO | M_WAITOK);
211 zb->zb_uaddr = uaddr;
213 zb->zb_numpages = len / PAGE_SIZE;
214 zb->zb_pages = malloc(sizeof(struct sf_buf *) *
215 zb->zb_numpages, M_BPF, M_ZERO | M_WAITOK);
216 map = &td->td_proc->p_vmspace->vm_map;
217 for (i = 0; i < zb->zb_numpages; i++) {
218 zb->zb_pages[i] = zbuf_sfbuf_get(map,
219 uaddr + (i * PAGE_SIZE));
220 if (zb->zb_pages[i] == NULL) {
226 (struct bpf_zbuf_header *)sf_buf_kva(zb->zb_pages[0]);
227 bzero(zb->zb_header, sizeof(*zb->zb_header));
237 * Copy bytes from a source into the specified zbuf. The caller is
238 * responsible for performing bounds checking, etc.
241 bpf_zerocopy_append_bytes(struct bpf_d *d, caddr_t buf, u_int offset,
242 void *src, u_int len)
244 u_int count, page, poffset;
248 KASSERT(d->bd_bufmode == BPF_BUFMODE_ZBUF,
249 ("bpf_zerocopy_append_bytes: not in zbuf mode"));
250 KASSERT(buf != NULL, ("bpf_zerocopy_append_bytes: NULL buf"));
252 src_bytes = (u_char *)src;
253 zb = (struct zbuf *)buf;
255 KASSERT((zb->zb_flags & ZBUF_FLAG_ASSIGNED) == 0,
256 ("bpf_zerocopy_append_bytes: ZBUF_FLAG_ASSIGNED"));
259 * Scatter-gather copy to user pages mapped into kernel address space
260 * using sf_bufs: copy up to a page at a time.
262 offset += sizeof(struct bpf_zbuf_header);
263 page = offset / PAGE_SIZE;
264 poffset = offset % PAGE_SIZE;
266 KASSERT(page < zb->zb_numpages, ("bpf_zerocopy_append_bytes:"
267 " page overflow (%d p %d np)\n", page, zb->zb_numpages));
269 count = min(len, PAGE_SIZE - poffset);
270 bcopy(src_bytes, ((u_char *)sf_buf_kva(zb->zb_pages[page])) +
273 if (poffset == PAGE_SIZE) {
277 KASSERT(poffset < PAGE_SIZE,
278 ("bpf_zerocopy_append_bytes: page offset overflow (%d)",
286 * Copy bytes from an mbuf chain to the specified zbuf: copying will be
287 * scatter-gather both from mbufs, which may be fragmented over memory, and
288 * to pages, which may not be contiguously mapped in kernel address space.
289 * As with bpf_zerocopy_append_bytes(), the caller is responsible for
290 * checking that this will not exceed the buffer limit.
293 bpf_zerocopy_append_mbuf(struct bpf_d *d, caddr_t buf, u_int offset,
294 void *src, u_int len)
296 u_int count, moffset, page, poffset;
297 const struct mbuf *m;
300 KASSERT(d->bd_bufmode == BPF_BUFMODE_ZBUF,
301 ("bpf_zerocopy_append_mbuf not in zbuf mode"));
302 KASSERT(buf != NULL, ("bpf_zerocopy_append_mbuf: NULL buf"));
304 m = (struct mbuf *)src;
305 zb = (struct zbuf *)buf;
307 KASSERT((zb->zb_flags & ZBUF_FLAG_ASSIGNED) == 0,
308 ("bpf_zerocopy_append_mbuf: ZBUF_FLAG_ASSIGNED"));
311 * Scatter gather both from an mbuf chain and to a user page set
312 * mapped into kernel address space using sf_bufs. If we're lucky,
313 * each mbuf requires one copy operation, but if page alignment and
314 * mbuf alignment work out less well, we'll be doing two copies per
317 offset += sizeof(struct bpf_zbuf_header);
318 page = offset / PAGE_SIZE;
319 poffset = offset % PAGE_SIZE;
322 KASSERT(page < zb->zb_numpages,
323 ("bpf_zerocopy_append_mbuf: page overflow (%d p %d "
324 "np)\n", page, zb->zb_numpages));
326 ("bpf_zerocopy_append_mbuf: end of mbuf chain"));
328 count = min(m->m_len - moffset, len);
329 count = min(count, PAGE_SIZE - poffset);
330 bcopy(mtod(m, u_char *) + moffset,
331 ((u_char *)sf_buf_kva(zb->zb_pages[page])) + poffset,
334 if (poffset == PAGE_SIZE) {
338 KASSERT(poffset < PAGE_SIZE,
339 ("bpf_zerocopy_append_mbuf: page offset overflow (%d)",
342 if (moffset == m->m_len) {
351 * Notification from the BPF framework that a buffer in the store position is
352 * rejecting packets and may be considered full. We mark the buffer as
353 * immutable and assign to userspace so that it is immediately available for
354 * the user process to access.
357 bpf_zerocopy_buffull(struct bpf_d *d)
361 KASSERT(d->bd_bufmode == BPF_BUFMODE_ZBUF,
362 ("bpf_zerocopy_buffull: not in zbuf mode"));
364 zb = (struct zbuf *)d->bd_sbuf;
365 KASSERT(zb != NULL, ("bpf_zerocopy_buffull: zb == NULL"));
367 if ((zb->zb_flags & ZBUF_FLAG_ASSIGNED) == 0) {
368 zb->zb_flags |= ZBUF_FLAG_ASSIGNED;
369 zb->zb_header->bzh_kernel_len = d->bd_slen;
370 atomic_add_rel_int(&zb->zb_header->bzh_kernel_gen, 1);
375 * Notification from the BPF framework that a buffer has moved into the held
376 * slot on a descriptor. Zero-copy BPF will update the shared page to let
377 * the user process know and flag the buffer as assigned if it hasn't already
378 * been marked assigned due to filling while it was in the store position.
380 * Note: identical logic as in bpf_zerocopy_buffull(), except that we operate
381 * on bd_hbuf and bd_hlen.
384 bpf_zerocopy_bufheld(struct bpf_d *d)
388 KASSERT(d->bd_bufmode == BPF_BUFMODE_ZBUF,
389 ("bpf_zerocopy_bufheld: not in zbuf mode"));
391 zb = (struct zbuf *)d->bd_hbuf;
392 KASSERT(zb != NULL, ("bpf_zerocopy_bufheld: zb == NULL"));
394 if ((zb->zb_flags & ZBUF_FLAG_ASSIGNED) == 0) {
395 zb->zb_flags |= ZBUF_FLAG_ASSIGNED;
396 zb->zb_header->bzh_kernel_len = d->bd_hlen;
397 atomic_add_rel_int(&zb->zb_header->bzh_kernel_gen, 1);
402 * Notification from the BPF framework that the free buffer has been been
403 * rotated out of the held position to the free position. This happens when
404 * the user acknowledges the held buffer.
407 bpf_zerocopy_buf_reclaimed(struct bpf_d *d)
411 KASSERT(d->bd_bufmode == BPF_BUFMODE_ZBUF,
412 ("bpf_zerocopy_reclaim_buf: not in zbuf mode"));
414 KASSERT(d->bd_fbuf != NULL,
415 ("bpf_zerocopy_buf_reclaimed: NULL free buf"));
416 zb = (struct zbuf *)d->bd_fbuf;
417 zb->zb_flags &= ~ZBUF_FLAG_ASSIGNED;
421 * Query from the BPF framework regarding whether the buffer currently in the
422 * held position can be moved to the free position, which can be indicated by
423 * the user process making their generation number equal to the kernel
427 bpf_zerocopy_canfreebuf(struct bpf_d *d)
431 KASSERT(d->bd_bufmode == BPF_BUFMODE_ZBUF,
432 ("bpf_zerocopy_canfreebuf: not in zbuf mode"));
434 zb = (struct zbuf *)d->bd_hbuf;
437 if (zb->zb_header->bzh_kernel_gen ==
438 atomic_load_acq_int(&zb->zb_header->bzh_user_gen))
444 * Query from the BPF framework as to whether or not the buffer current in
445 * the store position can actually be written to. This may return false if
446 * the store buffer is assigned to userspace before the hold buffer is
450 bpf_zerocopy_canwritebuf(struct bpf_d *d)
454 KASSERT(d->bd_bufmode == BPF_BUFMODE_ZBUF,
455 ("bpf_zerocopy_canwritebuf: not in zbuf mode"));
457 zb = (struct zbuf *)d->bd_sbuf;
458 KASSERT(zb != NULL, ("bpf_zerocopy_canwritebuf: bd_sbuf NULL"));
460 if (zb->zb_flags & ZBUF_FLAG_ASSIGNED)
466 * Free zero copy buffers at request of descriptor.
469 bpf_zerocopy_free(struct bpf_d *d)
473 KASSERT(d->bd_bufmode == BPF_BUFMODE_ZBUF,
474 ("bpf_zerocopy_free: not in zbuf mode"));
476 zb = (struct zbuf *)d->bd_sbuf;
479 zb = (struct zbuf *)d->bd_hbuf;
482 zb = (struct zbuf *)d->bd_fbuf;
488 * Ioctl to return the maximum buffer size.
491 bpf_zerocopy_ioctl_getzmax(struct thread *td, struct bpf_d *d, size_t *i)
494 KASSERT(d->bd_bufmode == BPF_BUFMODE_ZBUF,
495 ("bpf_zerocopy_ioctl_getzmax: not in zbuf mode"));
497 *i = BPF_MAX_PAGES * PAGE_SIZE;
502 * Ioctl to force rotation of the two buffers, if there's any data available.
503 * This can be used by user space to implement timeouts when waiting for a
507 bpf_zerocopy_ioctl_rotzbuf(struct thread *td, struct bpf_d *d,
512 bzero(bz, sizeof(*bz));
514 if (d->bd_hbuf == NULL && d->bd_slen != 0) {
516 bzh = (struct zbuf *)d->bd_hbuf;
517 bz->bz_bufa = (void *)bzh->zb_uaddr;
518 bz->bz_buflen = d->bd_hlen;
525 * Ioctl to configure zero-copy buffers -- may be done only once.
528 bpf_zerocopy_ioctl_setzbuf(struct thread *td, struct bpf_d *d,
531 struct zbuf *zba, *zbb;
534 KASSERT(d->bd_bufmode == BPF_BUFMODE_ZBUF,
535 ("bpf_zerocopy_ioctl_setzbuf: not in zbuf mode"));
538 * Must set both buffers. Cannot clear them.
540 if (bz->bz_bufa == NULL || bz->bz_bufb == NULL)
544 * Buffers must have a size greater than 0. Alignment and other size
545 * validity checking is done in zbuf_setup().
547 if (bz->bz_buflen == 0)
551 * Allocate new buffers.
553 error = zbuf_setup(td, (vm_offset_t)bz->bz_bufa, bz->bz_buflen,
557 error = zbuf_setup(td, (vm_offset_t)bz->bz_bufb, bz->bz_buflen,
565 * We only allow buffers to be installed once, so atomically check
566 * that no buffers are currently installed and install new buffers.
569 if (d->bd_hbuf != NULL || d->bd_sbuf != NULL || d->bd_fbuf != NULL ||
578 * Point BPF descriptor at buffers; initialize sbuf as zba so that
579 * it is always filled first in the sequence, per bpf(4).
581 d->bd_fbuf = (caddr_t)zbb;
582 d->bd_sbuf = (caddr_t)zba;
587 * We expose only the space left in the buffer after the size of the
588 * shared management region.
590 d->bd_bufsize = bz->bz_buflen - sizeof(struct bpf_zbuf_header);