2 * Copyright (c) 2008 Yahoo!, Inc.
4 * Written by: John Baldwin <jhb@FreeBSD.org>
6 * Redistribution and use in source and binary forms, with or without
7 * modification, are permitted provided that the following conditions
9 * 1. Redistributions of source code must retain the above copyright
10 * notice, this list of conditions and the following disclaimer.
11 * 2. Redistributions in binary form must reproduce the above copyright
12 * notice, this list of conditions and the following disclaimer in the
13 * documentation and/or other materials provided with the distribution.
14 * 3. Neither the name of the author nor the names of any co-contributors
15 * may be used to endorse or promote products derived from this software
16 * without specific prior written permission.
18 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
19 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
20 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
21 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
22 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
23 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
24 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
25 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
26 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
27 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
31 #include <sys/cdefs.h>
32 __FBSDID("$FreeBSD$");
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 * Allocate a scatter/gather list along with 'nsegs' segments. The
221 * 'mflags' parameters are the same as passed to malloc(9). The caller
222 * should use sglist_free() to free this list.
225 sglist_alloc(int nsegs, int mflags)
229 sg = malloc(sizeof(struct sglist) + nsegs * sizeof(struct sglist_seg),
233 sglist_init(sg, nsegs, (struct sglist_seg *)(sg + 1));
238 * Free a scatter/gather list allocated via sglist_allc().
241 sglist_free(struct sglist *sg)
247 if (refcount_release(&sg->sg_refs))
252 * Append the segments to describe a single kernel virtual address
253 * range to a scatter/gather list. If there are insufficient
254 * segments, then this fails with EFBIG.
257 sglist_append(struct sglist *sg, void *buf, size_t len)
262 if (sg->sg_maxseg == 0)
264 SGLIST_SAVE(sg, save);
265 error = _sglist_append_buf(sg, buf, len, NULL, NULL);
267 SGLIST_RESTORE(sg, save);
272 * Append the segments to describe a bio's data to a scatter/gather list.
273 * If there are insufficient segments, then this fails with EFBIG.
275 * NOTE: This function expects bio_bcount to be initialized.
278 sglist_append_bio(struct sglist *sg, struct bio *bp)
282 if ((bp->bio_flags & BIO_UNMAPPED) == 0)
283 error = sglist_append(sg, bp->bio_data, bp->bio_bcount);
285 error = sglist_append_vmpages(sg, bp->bio_ma,
286 bp->bio_ma_offset, bp->bio_bcount);
291 * Append a single physical address range to a scatter/gather list.
292 * If there are insufficient segments, then this fails with EFBIG.
295 sglist_append_phys(struct sglist *sg, vm_paddr_t paddr, size_t len)
297 struct sglist_seg *ss;
301 if (sg->sg_maxseg == 0)
306 if (sg->sg_nseg == 0) {
307 sg->sg_segs[0].ss_paddr = paddr;
308 sg->sg_segs[0].ss_len = len;
312 ss = &sg->sg_segs[sg->sg_nseg - 1];
313 SGLIST_SAVE(sg, save);
314 error = _sglist_append_range(sg, &ss, paddr, len);
316 SGLIST_RESTORE(sg, save);
321 * Append the segments that describe a single mbuf chain to a
322 * scatter/gather list. If there are insufficient segments, then this
326 sglist_append_mbuf(struct sglist *sg, struct mbuf *m0)
332 if (sg->sg_maxseg == 0)
336 SGLIST_SAVE(sg, save);
337 for (m = m0; m != NULL; m = m->m_next) {
339 error = sglist_append(sg, m->m_data, m->m_len);
341 SGLIST_RESTORE(sg, save);
350 * Append the segments that describe a buffer spanning an array of VM
351 * pages. The buffer begins at an offset of 'pgoff' in the first
355 sglist_append_vmpages(struct sglist *sg, vm_page_t *m, size_t pgoff,
359 struct sglist_seg *ss;
364 if (sg->sg_maxseg == 0)
369 SGLIST_SAVE(sg, save);
371 if (sg->sg_nseg == 0) {
372 seglen = min(PAGE_SIZE - pgoff, len);
373 sg->sg_segs[0].ss_paddr = VM_PAGE_TO_PHYS(m[0]) + pgoff;
374 sg->sg_segs[0].ss_len = seglen;
380 ss = &sg->sg_segs[sg->sg_nseg - 1];
381 for (; len > 0; i++, len -= seglen) {
382 seglen = min(PAGE_SIZE - pgoff, len);
383 paddr = VM_PAGE_TO_PHYS(m[i]) + pgoff;
384 error = _sglist_append_range(sg, &ss, paddr, seglen);
386 SGLIST_RESTORE(sg, save);
395 * Append the segments that describe a single user address range to a
396 * scatter/gather list. If there are insufficient segments, then this
400 sglist_append_user(struct sglist *sg, void *buf, size_t len, struct thread *td)
405 if (sg->sg_maxseg == 0)
407 SGLIST_SAVE(sg, save);
408 error = _sglist_append_buf(sg, buf, len,
409 vmspace_pmap(td->td_proc->p_vmspace), NULL);
411 SGLIST_RESTORE(sg, save);
416 * Append the segments that describe a single uio to a scatter/gather
417 * list. If there are insufficient segments, then this fails with
421 sglist_append_uio(struct sglist *sg, struct uio *uio)
425 size_t resid, minlen;
429 if (sg->sg_maxseg == 0)
432 resid = uio->uio_resid;
435 if (uio->uio_segflg == UIO_USERSPACE) {
436 KASSERT(uio->uio_td != NULL,
437 ("sglist_append_uio: USERSPACE but no thread"));
438 pmap = vmspace_pmap(uio->uio_td->td_proc->p_vmspace);
443 SGLIST_SAVE(sg, save);
444 for (i = 0; i < uio->uio_iovcnt && resid != 0; i++) {
446 * Now at the first iovec to load. Load each iovec
447 * until we have exhausted the residual count.
449 minlen = MIN(resid, iov[i].iov_len);
451 error = _sglist_append_buf(sg, iov[i].iov_base, minlen,
454 SGLIST_RESTORE(sg, save);
464 * Append the segments that describe at most 'resid' bytes from a
465 * single uio to a scatter/gather list. If there are insufficient
466 * segments, then only the amount that fits is appended.
469 sglist_consume_uio(struct sglist *sg, struct uio *uio, size_t resid)
476 if (sg->sg_maxseg == 0)
479 if (uio->uio_segflg == UIO_USERSPACE) {
480 KASSERT(uio->uio_td != NULL,
481 ("sglist_consume_uio: USERSPACE but no thread"));
482 pmap = vmspace_pmap(uio->uio_td->td_proc->p_vmspace);
487 while (resid > 0 && uio->uio_resid) {
499 * Try to append this iovec. If we run out of room,
500 * then break out of the loop.
502 error = _sglist_append_buf(sg, iov->iov_base, len, pmap, &done);
503 iov->iov_base = (char *)iov->iov_base + done;
504 iov->iov_len -= done;
505 uio->uio_resid -= done;
506 uio->uio_offset += done;
515 * Allocate and populate a scatter/gather list to describe a single
516 * kernel virtual address range.
519 sglist_build(void *buf, size_t len, int mflags)
527 nsegs = sglist_count(buf, len);
528 sg = sglist_alloc(nsegs, mflags);
531 if (sglist_append(sg, buf, len) != 0) {
539 * Clone a new copy of a scatter/gather list.
542 sglist_clone(struct sglist *sg, int mflags)
548 new = sglist_alloc(sg->sg_maxseg, mflags);
551 new->sg_nseg = sg->sg_nseg;
552 bcopy(sg->sg_segs, new->sg_segs, sizeof(struct sglist_seg) *
558 * Calculate the total length of the segments described in a
559 * scatter/gather list.
562 sglist_length(struct sglist *sg)
568 for (i = 0; i < sg->sg_nseg; i++)
569 space += sg->sg_segs[i].ss_len;
574 * Split a scatter/gather list into two lists. The scatter/gather
575 * entries for the first 'length' bytes of the 'original' list are
576 * stored in the '*head' list and are removed from 'original'.
578 * If '*head' is NULL, then a new list will be allocated using
579 * 'mflags'. If M_NOWAIT is specified and the allocation fails,
580 * ENOMEM will be returned.
582 * If '*head' is not NULL, it should point to an empty sglist. If it
583 * does not have enough room for the remaining space, then EFBIG will
584 * be returned. If '*head' is not empty, then EINVAL will be
587 * If 'original' is shared (refcount > 1), then EDOOFUS will be
591 sglist_split(struct sglist *original, struct sglist **head, size_t length,
598 if (original->sg_refs > 1)
601 /* Figure out how big of a sglist '*head' has to hold. */
605 for (i = 0; i < original->sg_nseg; i++) {
606 space += original->sg_segs[i].ss_len;
608 if (space >= length) {
610 * If 'length' falls in the middle of a
611 * scatter/gather list entry, then 'split'
612 * holds how much of that entry will remain in
615 split = space - length;
620 /* Nothing to do, so leave head empty. */
625 sg = sglist_alloc(count, mflags);
631 if (sg->sg_maxseg < count)
633 if (sg->sg_nseg != 0)
637 /* Copy 'count' entries to 'sg' from 'original'. */
638 bcopy(original->sg_segs, sg->sg_segs, count *
639 sizeof(struct sglist_seg));
643 * If we had to split a list entry, fixup the last entry in
644 * 'sg' and the new first entry in 'original'. We also
645 * decrement 'count' by 1 since we will only be removing
646 * 'count - 1' segments from 'original' now.
650 sg->sg_segs[count].ss_len -= split;
651 original->sg_segs[count].ss_paddr =
652 sg->sg_segs[count].ss_paddr + split;
653 original->sg_segs[count].ss_len = split;
656 /* Trim 'count' entries from the front of 'original'. */
657 original->sg_nseg -= count;
658 bcopy(original->sg_segs + count, original->sg_segs, count *
659 sizeof(struct sglist_seg));
664 * Append the scatter/gather list elements in 'second' to the
665 * scatter/gather list 'first'. If there is not enough space in
666 * 'first', EFBIG is returned.
669 sglist_join(struct sglist *first, struct sglist *second)
671 struct sglist_seg *flast, *sfirst;
674 /* If 'second' is empty, there is nothing to do. */
675 if (second->sg_nseg == 0)
679 * If the first entry in 'second' can be appended to the last entry
680 * in 'first' then set append to '1'.
683 flast = &first->sg_segs[first->sg_nseg - 1];
684 sfirst = &second->sg_segs[0];
685 if (first->sg_nseg != 0 &&
686 flast->ss_paddr + flast->ss_len == sfirst->ss_paddr)
689 /* Make sure 'first' has enough room. */
690 if (first->sg_nseg + second->sg_nseg - append > first->sg_maxseg)
693 /* Merge last in 'first' and first in 'second' if needed. */
695 flast->ss_len += sfirst->ss_len;
697 /* Append new segments from 'second' to 'first'. */
698 bcopy(first->sg_segs + first->sg_nseg, second->sg_segs + append,
699 (second->sg_nseg - append) * sizeof(struct sglist_seg));
700 first->sg_nseg += second->sg_nseg - append;
701 sglist_reset(second);
706 * Generate a new scatter/gather list from a range of an existing
707 * scatter/gather list. The 'offset' and 'length' parameters specify
708 * the logical range of the 'original' list to extract. If that range
709 * is not a subset of the length of 'original', then EINVAL is
710 * returned. The new scatter/gather list is stored in '*slice'.
712 * If '*slice' is NULL, then a new list will be allocated using
713 * 'mflags'. If M_NOWAIT is specified and the allocation fails,
714 * ENOMEM will be returned.
716 * If '*slice' is not NULL, it should point to an empty sglist. If it
717 * does not have enough room for the remaining space, then EFBIG will
718 * be returned. If '*slice' is not empty, then EINVAL will be
722 sglist_slice(struct sglist *original, struct sglist **slice, size_t offset,
723 size_t length, int mflags)
726 size_t space, end, foffs, loffs;
733 /* Figure out how many segments '*slice' needs to have. */
734 end = offset + length;
739 for (i = 0; i < original->sg_nseg; i++) {
740 space += original->sg_segs[i].ss_len;
741 if (space > offset) {
743 * When we hit the first segment, store its index
744 * in 'fseg' and the offset into the first segment
745 * of 'offset' in 'foffs'.
749 foffs = offset - (space -
750 original->sg_segs[i].ss_len);
751 CTR1(KTR_DEV, "sglist_slice: foffs = %08lx",
757 * When we hit the last segment, break out of
758 * the loop. Store the amount of extra space
759 * at the end of this segment in 'loffs'.
763 CTR1(KTR_DEV, "sglist_slice: loffs = %08lx",
770 /* If we never hit 'end', then 'length' ran off the end, so fail. */
774 if (*slice == NULL) {
775 sg = sglist_alloc(count, mflags);
781 if (sg->sg_maxseg < count)
783 if (sg->sg_nseg != 0)
788 * Copy over 'count' segments from 'original' starting at
791 bcopy(original->sg_segs + fseg, sg->sg_segs,
792 count * sizeof(struct sglist_seg));
795 /* Fixup first and last segments if needed. */
797 sg->sg_segs[0].ss_paddr += foffs;
798 sg->sg_segs[0].ss_len -= foffs;
799 CTR2(KTR_DEV, "sglist_slice seg[0]: %08lx:%08lx",
800 (long)sg->sg_segs[0].ss_paddr, sg->sg_segs[0].ss_len);
803 sg->sg_segs[count - 1].ss_len -= loffs;
804 CTR2(KTR_DEV, "sglist_slice seg[%d]: len %08x", count - 1,
805 sg->sg_segs[count - 1].ss_len);