2 * SPDX-License-Identifier: BSD-3-Clause
4 * Copyright (c) 2008 Yahoo!, Inc.
6 * Written by: John Baldwin <jhb@FreeBSD.org>
8 * Redistribution and use in source and binary forms, with or without
9 * modification, are permitted provided that the following conditions
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in the
15 * documentation and/or other materials provided with the distribution.
16 * 3. Neither the name of the author nor the names of any co-contributors
17 * may be used to endorse or promote products derived from this software
18 * without specific prior written permission.
20 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
21 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
22 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
23 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
24 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
25 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
26 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
27 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
28 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
29 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
33 #include <sys/cdefs.h>
34 #include <sys/param.h>
35 #include <sys/kernel.h>
37 #include <sys/malloc.h>
40 #include <sys/sglist.h>
44 #include <vm/vm_page.h>
46 #include <vm/vm_map.h>
50 static MALLOC_DEFINE(M_SGLIST, "sglist", "scatter/gather lists");
53 * Convenience macros to save the state of an sglist so it can be restored
54 * if an append attempt fails. Since sglist's only grow we only need to
55 * save the current count of segments and the length of the ending segment.
56 * Earlier segments will not be changed by an append, and the only change
57 * that can occur to the ending segment is that it can be extended.
64 #define SGLIST_SAVE(sg, sgsave) do { \
65 (sgsave).sg_nseg = (sg)->sg_nseg; \
66 if ((sgsave).sg_nseg > 0) \
67 (sgsave).ss_len = (sg)->sg_segs[(sgsave).sg_nseg - 1].ss_len; \
69 (sgsave).ss_len = 0; \
72 #define SGLIST_RESTORE(sg, sgsave) do { \
73 (sg)->sg_nseg = (sgsave).sg_nseg; \
74 if ((sgsave).sg_nseg > 0) \
75 (sg)->sg_segs[(sgsave).sg_nseg - 1].ss_len = (sgsave).ss_len; \
79 * Append a single (paddr, len) to a sglist. sg is the list and ss is
80 * the current segment in the list. If we run out of segments then
81 * EFBIG will be returned.
84 _sglist_append_range(struct sglist *sg, struct sglist_seg **ssp,
85 vm_paddr_t paddr, size_t len)
87 struct sglist_seg *ss;
90 if (ss->ss_paddr + ss->ss_len == paddr)
93 if (sg->sg_nseg == sg->sg_maxseg)
105 * Worker routine to append a virtual address range (either kernel or
106 * user) to a scatter/gather list.
109 _sglist_append_buf(struct sglist *sg, void *buf, size_t len, pmap_t pmap,
112 struct sglist_seg *ss;
113 vm_offset_t vaddr, offset;
123 /* Do the first page. It may have an offset. */
124 vaddr = (vm_offset_t)buf;
125 offset = vaddr & PAGE_MASK;
127 paddr = pmap_extract(pmap, vaddr);
129 paddr = pmap_kextract(vaddr);
130 seglen = MIN(len, PAGE_SIZE - offset);
131 if (sg->sg_nseg == 0) {
133 ss->ss_paddr = paddr;
137 ss = &sg->sg_segs[sg->sg_nseg - 1];
138 error = _sglist_append_range(sg, &ss, paddr, seglen);
148 seglen = MIN(len, PAGE_SIZE);
150 paddr = pmap_extract(pmap, vaddr);
152 paddr = pmap_kextract(vaddr);
153 error = _sglist_append_range(sg, &ss, paddr, seglen);
166 * Determine the number of scatter/gather list elements needed to
167 * describe a kernel virtual address range.
170 sglist_count(void *buf, size_t len)
172 vm_offset_t vaddr, vendaddr;
173 vm_paddr_t lastaddr, paddr;
179 vaddr = trunc_page((vm_offset_t)buf);
180 vendaddr = (vm_offset_t)buf + len;
182 lastaddr = pmap_kextract(vaddr);
184 while (vaddr < vendaddr) {
185 paddr = pmap_kextract(vaddr);
186 if (lastaddr + PAGE_SIZE != paddr)
195 * Determine the number of scatter/gather list elements needed to
196 * describe a buffer backed by an array of VM pages.
199 sglist_count_vmpages(vm_page_t *m, size_t pgoff, size_t len)
201 vm_paddr_t lastaddr, paddr;
209 lastaddr = VM_PAGE_TO_PHYS(m[0]);
210 for (i = 1; len > PAGE_SIZE; len -= PAGE_SIZE, i++) {
211 paddr = VM_PAGE_TO_PHYS(m[i]);
212 if (lastaddr + PAGE_SIZE != paddr)
220 * Determine the number of scatter/gather list elements needed to
221 * describe an M_EXTPG mbuf.
224 sglist_count_mbuf_epg(struct mbuf *m, size_t off, size_t len)
226 vm_paddr_t nextaddr, paddr;
227 size_t seglen, segoff;
228 int i, nsegs, pglen, pgoff;
234 if (m->m_epg_hdrlen != 0) {
235 if (off >= m->m_epg_hdrlen) {
236 off -= m->m_epg_hdrlen;
238 seglen = m->m_epg_hdrlen - off;
240 seglen = MIN(seglen, len);
243 nsegs += sglist_count(&m->m_epg_hdr[segoff],
248 pgoff = m->m_epg_1st_off;
249 for (i = 0; i < m->m_epg_npgs && len > 0; i++) {
250 pglen = m_epg_pagelen(m, i, pgoff);
256 seglen = pglen - off;
257 segoff = pgoff + off;
259 seglen = MIN(seglen, len);
261 paddr = m->m_epg_pa[i] + segoff;
262 if (paddr != nextaddr)
264 nextaddr = paddr + seglen;
268 seglen = MIN(len, m->m_epg_trllen - off);
270 nsegs += sglist_count(&m->m_epg_trail[off], seglen);
272 KASSERT(len == 0, ("len != 0"));
277 * Allocate a scatter/gather list along with 'nsegs' segments. The
278 * 'mflags' parameters are the same as passed to malloc(9). The caller
279 * should use sglist_free() to free this list.
282 sglist_alloc(int nsegs, int mflags)
286 sg = malloc(sizeof(struct sglist) + nsegs * sizeof(struct sglist_seg),
290 sglist_init(sg, nsegs, (struct sglist_seg *)(sg + 1));
295 * Free a scatter/gather list allocated via sglist_allc().
298 sglist_free(struct sglist *sg)
304 if (refcount_release(&sg->sg_refs))
309 * Append the segments to describe a single kernel virtual address
310 * range to a scatter/gather list. If there are insufficient
311 * segments, then this fails with EFBIG.
314 sglist_append(struct sglist *sg, void *buf, size_t len)
319 if (sg->sg_maxseg == 0)
321 SGLIST_SAVE(sg, save);
322 error = _sglist_append_buf(sg, buf, len, NULL, NULL);
324 SGLIST_RESTORE(sg, save);
329 * Append the segments to describe a bio's data to a scatter/gather list.
330 * If there are insufficient segments, then this fails with EFBIG.
332 * NOTE: This function expects bio_bcount to be initialized.
335 sglist_append_bio(struct sglist *sg, struct bio *bp)
339 if ((bp->bio_flags & BIO_UNMAPPED) == 0)
340 error = sglist_append(sg, bp->bio_data, bp->bio_bcount);
342 error = sglist_append_vmpages(sg, bp->bio_ma,
343 bp->bio_ma_offset, bp->bio_bcount);
348 * Append a single physical address range to a scatter/gather list.
349 * If there are insufficient segments, then this fails with EFBIG.
352 sglist_append_phys(struct sglist *sg, vm_paddr_t paddr, size_t len)
354 struct sglist_seg *ss;
358 if (sg->sg_maxseg == 0)
363 if (sg->sg_nseg == 0) {
364 sg->sg_segs[0].ss_paddr = paddr;
365 sg->sg_segs[0].ss_len = len;
369 ss = &sg->sg_segs[sg->sg_nseg - 1];
370 SGLIST_SAVE(sg, save);
371 error = _sglist_append_range(sg, &ss, paddr, len);
373 SGLIST_RESTORE(sg, save);
378 * Append the segments of single multi-page mbuf.
379 * If there are insufficient segments, then this fails with EFBIG.
382 sglist_append_mbuf_epg(struct sglist *sg, struct mbuf *m, size_t off,
385 size_t seglen, segoff;
387 int error, i, pglen, pgoff;
392 if (m->m_epg_hdrlen != 0) {
393 if (off >= m->m_epg_hdrlen) {
394 off -= m->m_epg_hdrlen;
396 seglen = m->m_epg_hdrlen - off;
398 seglen = MIN(seglen, len);
401 error = sglist_append(sg,
402 &m->m_epg_hdr[segoff], seglen);
405 pgoff = m->m_epg_1st_off;
406 for (i = 0; i < m->m_epg_npgs && error == 0 && len > 0; i++) {
407 pglen = m_epg_pagelen(m, i, pgoff);
413 seglen = pglen - off;
414 segoff = pgoff + off;
416 seglen = MIN(seglen, len);
418 paddr = m->m_epg_pa[i] + segoff;
419 error = sglist_append_phys(sg, paddr, seglen);
422 if (error == 0 && len > 0) {
423 seglen = MIN(len, m->m_epg_trllen - off);
425 error = sglist_append(sg,
426 &m->m_epg_trail[off], seglen);
429 KASSERT(len == 0, ("len != 0"));
434 * Append the segments that describe a single mbuf chain to a
435 * scatter/gather list. If there are insufficient segments, then this
439 sglist_append_mbuf(struct sglist *sg, struct mbuf *m0)
445 if (sg->sg_maxseg == 0)
449 SGLIST_SAVE(sg, save);
450 for (m = m0; m != NULL; m = m->m_next) {
452 if ((m->m_flags & M_EXTPG) != 0)
453 error = sglist_append_mbuf_epg(sg, m,
454 mtod(m, vm_offset_t), m->m_len);
456 error = sglist_append(sg, m->m_data,
459 SGLIST_RESTORE(sg, save);
468 * Append the segments that describe a single mbuf to a scatter/gather
469 * list. If there are insufficient segments, then this fails with
473 sglist_append_single_mbuf(struct sglist *sg, struct mbuf *m)
475 if ((m->m_flags & M_EXTPG) != 0)
476 return (sglist_append_mbuf_epg(sg, m,
477 mtod(m, vm_offset_t), m->m_len));
479 return (sglist_append(sg, m->m_data, m->m_len));
483 * Append the segments that describe a buffer spanning an array of VM
484 * pages. The buffer begins at an offset of 'pgoff' in the first
488 sglist_append_vmpages(struct sglist *sg, vm_page_t *m, size_t pgoff,
492 struct sglist_seg *ss;
497 if (sg->sg_maxseg == 0)
502 SGLIST_SAVE(sg, save);
504 if (sg->sg_nseg == 0) {
505 seglen = min(PAGE_SIZE - pgoff, len);
506 sg->sg_segs[0].ss_paddr = VM_PAGE_TO_PHYS(m[0]) + pgoff;
507 sg->sg_segs[0].ss_len = seglen;
513 ss = &sg->sg_segs[sg->sg_nseg - 1];
514 for (; len > 0; i++, len -= seglen) {
515 seglen = min(PAGE_SIZE - pgoff, len);
516 paddr = VM_PAGE_TO_PHYS(m[i]) + pgoff;
517 error = _sglist_append_range(sg, &ss, paddr, seglen);
519 SGLIST_RESTORE(sg, save);
528 * Append the segments that describe a single user address range to a
529 * scatter/gather list. If there are insufficient segments, then this
533 sglist_append_user(struct sglist *sg, void *buf, size_t len, struct thread *td)
538 if (sg->sg_maxseg == 0)
540 SGLIST_SAVE(sg, save);
541 error = _sglist_append_buf(sg, buf, len,
542 vmspace_pmap(td->td_proc->p_vmspace), NULL);
544 SGLIST_RESTORE(sg, save);
549 * Append a subset of an existing scatter/gather list 'source' to a
550 * the scatter/gather list 'sg'. If there are insufficient segments,
551 * then this fails with EFBIG.
554 sglist_append_sglist(struct sglist *sg, struct sglist *source, size_t offset,
558 struct sglist_seg *ss;
562 if (sg->sg_maxseg == 0 || length == 0)
564 SGLIST_SAVE(sg, save);
566 ss = &sg->sg_segs[sg->sg_nseg - 1];
567 for (i = 0; i < source->sg_nseg; i++) {
568 if (offset >= source->sg_segs[i].ss_len) {
569 offset -= source->sg_segs[i].ss_len;
572 seglen = source->sg_segs[i].ss_len - offset;
575 error = _sglist_append_range(sg, &ss,
576 source->sg_segs[i].ss_paddr + offset, seglen);
587 SGLIST_RESTORE(sg, save);
592 * Append the segments that describe a single uio to a scatter/gather
593 * list. If there are insufficient segments, then this fails with
597 sglist_append_uio(struct sglist *sg, struct uio *uio)
601 size_t resid, minlen;
605 if (sg->sg_maxseg == 0)
608 resid = uio->uio_resid;
611 if (uio->uio_segflg == UIO_USERSPACE) {
612 KASSERT(uio->uio_td != NULL,
613 ("sglist_append_uio: USERSPACE but no thread"));
614 pmap = vmspace_pmap(uio->uio_td->td_proc->p_vmspace);
619 SGLIST_SAVE(sg, save);
620 for (i = 0; i < uio->uio_iovcnt && resid != 0; i++) {
622 * Now at the first iovec to load. Load each iovec
623 * until we have exhausted the residual count.
625 minlen = MIN(resid, iov[i].iov_len);
627 error = _sglist_append_buf(sg, iov[i].iov_base, minlen,
630 SGLIST_RESTORE(sg, save);
640 * Append the segments that describe at most 'resid' bytes from a
641 * single uio to a scatter/gather list. If there are insufficient
642 * segments, then only the amount that fits is appended.
645 sglist_consume_uio(struct sglist *sg, struct uio *uio, size_t resid)
652 if (sg->sg_maxseg == 0)
655 if (uio->uio_segflg == UIO_USERSPACE) {
656 KASSERT(uio->uio_td != NULL,
657 ("sglist_consume_uio: USERSPACE but no thread"));
658 pmap = vmspace_pmap(uio->uio_td->td_proc->p_vmspace);
663 while (resid > 0 && uio->uio_resid) {
675 * Try to append this iovec. If we run out of room,
676 * then break out of the loop.
678 error = _sglist_append_buf(sg, iov->iov_base, len, pmap, &done);
679 iov->iov_base = (char *)iov->iov_base + done;
680 iov->iov_len -= done;
681 uio->uio_resid -= done;
682 uio->uio_offset += done;
691 * Allocate and populate a scatter/gather list to describe a single
692 * kernel virtual address range.
695 sglist_build(void *buf, size_t len, int mflags)
703 nsegs = sglist_count(buf, len);
704 sg = sglist_alloc(nsegs, mflags);
707 if (sglist_append(sg, buf, len) != 0) {
715 * Clone a new copy of a scatter/gather list.
718 sglist_clone(struct sglist *sg, int mflags)
724 new = sglist_alloc(sg->sg_maxseg, mflags);
727 new->sg_nseg = sg->sg_nseg;
728 bcopy(sg->sg_segs, new->sg_segs, sizeof(struct sglist_seg) *
734 * Calculate the total length of the segments described in a
735 * scatter/gather list.
738 sglist_length(struct sglist *sg)
744 for (i = 0; i < sg->sg_nseg; i++)
745 space += sg->sg_segs[i].ss_len;
750 * Split a scatter/gather list into two lists. The scatter/gather
751 * entries for the first 'length' bytes of the 'original' list are
752 * stored in the '*head' list and are removed from 'original'.
754 * If '*head' is NULL, then a new list will be allocated using
755 * 'mflags'. If M_NOWAIT is specified and the allocation fails,
756 * ENOMEM will be returned.
758 * If '*head' is not NULL, it should point to an empty sglist. If it
759 * does not have enough room for the remaining space, then EFBIG will
760 * be returned. If '*head' is not empty, then EINVAL will be
763 * If 'original' is shared (refcount > 1), then EDOOFUS will be
767 sglist_split(struct sglist *original, struct sglist **head, size_t length,
774 if (original->sg_refs > 1)
777 /* Figure out how big of a sglist '*head' has to hold. */
781 for (i = 0; i < original->sg_nseg; i++) {
782 space += original->sg_segs[i].ss_len;
784 if (space >= length) {
786 * If 'length' falls in the middle of a
787 * scatter/gather list entry, then 'split'
788 * holds how much of that entry will remain in
791 split = space - length;
796 /* Nothing to do, so leave head empty. */
801 sg = sglist_alloc(count, mflags);
807 if (sg->sg_maxseg < count)
809 if (sg->sg_nseg != 0)
813 /* Copy 'count' entries to 'sg' from 'original'. */
814 bcopy(original->sg_segs, sg->sg_segs, count *
815 sizeof(struct sglist_seg));
819 * If we had to split a list entry, fixup the last entry in
820 * 'sg' and the new first entry in 'original'. We also
821 * decrement 'count' by 1 since we will only be removing
822 * 'count - 1' segments from 'original' now.
826 sg->sg_segs[count].ss_len -= split;
827 original->sg_segs[count].ss_paddr =
828 sg->sg_segs[count].ss_paddr + split;
829 original->sg_segs[count].ss_len = split;
832 /* Trim 'count' entries from the front of 'original'. */
833 original->sg_nseg -= count;
834 bcopy(original->sg_segs + count, original->sg_segs, count *
835 sizeof(struct sglist_seg));
840 * Append the scatter/gather list elements in 'second' to the
841 * scatter/gather list 'first'. If there is not enough space in
842 * 'first', EFBIG is returned.
845 sglist_join(struct sglist *first, struct sglist *second)
847 struct sglist_seg *flast, *sfirst;
850 /* If 'second' is empty, there is nothing to do. */
851 if (second->sg_nseg == 0)
855 * If the first entry in 'second' can be appended to the last entry
856 * in 'first' then set append to '1'.
859 flast = &first->sg_segs[first->sg_nseg - 1];
860 sfirst = &second->sg_segs[0];
861 if (first->sg_nseg != 0 &&
862 flast->ss_paddr + flast->ss_len == sfirst->ss_paddr)
865 /* Make sure 'first' has enough room. */
866 if (first->sg_nseg + second->sg_nseg - append > first->sg_maxseg)
869 /* Merge last in 'first' and first in 'second' if needed. */
871 flast->ss_len += sfirst->ss_len;
873 /* Append new segments from 'second' to 'first'. */
874 bcopy(first->sg_segs + first->sg_nseg, second->sg_segs + append,
875 (second->sg_nseg - append) * sizeof(struct sglist_seg));
876 first->sg_nseg += second->sg_nseg - append;
877 sglist_reset(second);
882 * Generate a new scatter/gather list from a range of an existing
883 * scatter/gather list. The 'offset' and 'length' parameters specify
884 * the logical range of the 'original' list to extract. If that range
885 * is not a subset of the length of 'original', then EINVAL is
886 * returned. The new scatter/gather list is stored in '*slice'.
888 * If '*slice' is NULL, then a new list will be allocated using
889 * 'mflags'. If M_NOWAIT is specified and the allocation fails,
890 * ENOMEM will be returned.
892 * If '*slice' is not NULL, it should point to an empty sglist. If it
893 * does not have enough room for the remaining space, then EFBIG will
894 * be returned. If '*slice' is not empty, then EINVAL will be
898 sglist_slice(struct sglist *original, struct sglist **slice, size_t offset,
899 size_t length, int mflags)
902 size_t space, end, foffs, loffs;
909 /* Figure out how many segments '*slice' needs to have. */
910 end = offset + length;
915 for (i = 0; i < original->sg_nseg; i++) {
916 space += original->sg_segs[i].ss_len;
917 if (space > offset) {
919 * When we hit the first segment, store its index
920 * in 'fseg' and the offset into the first segment
921 * of 'offset' in 'foffs'.
925 foffs = offset - (space -
926 original->sg_segs[i].ss_len);
927 CTR1(KTR_DEV, "sglist_slice: foffs = %08lx",
933 * When we hit the last segment, break out of
934 * the loop. Store the amount of extra space
935 * at the end of this segment in 'loffs'.
939 CTR1(KTR_DEV, "sglist_slice: loffs = %08lx",
946 /* If we never hit 'end', then 'length' ran off the end, so fail. */
950 if (*slice == NULL) {
951 sg = sglist_alloc(count, mflags);
957 if (sg->sg_maxseg < count)
959 if (sg->sg_nseg != 0)
964 * Copy over 'count' segments from 'original' starting at
967 bcopy(original->sg_segs + fseg, sg->sg_segs,
968 count * sizeof(struct sglist_seg));
971 /* Fixup first and last segments if needed. */
973 sg->sg_segs[0].ss_paddr += foffs;
974 sg->sg_segs[0].ss_len -= foffs;
975 CTR2(KTR_DEV, "sglist_slice seg[0]: %08lx:%08lx",
976 (long)sg->sg_segs[0].ss_paddr, sg->sg_segs[0].ss_len);
979 sg->sg_segs[count - 1].ss_len -= loffs;
980 CTR2(KTR_DEV, "sglist_slice seg[%d]: len %08x", count - 1,
981 sg->sg_segs[count - 1].ss_len);