2 * Copyright (c) 2006-2007 The DragonFly Project. All rights reserved.
4 * This code is derived from software contributed to The DragonFly Project
5 * by Matthew Dillon <dillon@backplane.com>
7 * Redistribution and use in source and binary forms, with or without
8 * 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
15 * the documentation and/or other materials provided with the
17 * 3. Neither the name of The DragonFly Project nor the names of its
18 * contributors may be used to endorse or promote products derived
19 * from this software without specific, prior written permission.
21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
24 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
25 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
26 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
27 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
28 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
29 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
30 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
31 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
34 * $DragonFly: src/sys/kern/kern_syslink.c,v 1.15 2007/08/13 17:47:19 dillon Exp $
37 * This module implements the core syslink() system call and provides
38 * glue for kernel syslink frontends and backends, creating a intra-host
39 * communications infrastructure and DMA transport abstraction.
42 #include <sys/param.h>
43 #include <sys/systm.h>
44 #include <sys/kernel.h>
45 #include <sys/endian.h>
46 #include <sys/malloc.h>
47 #include <sys/alist.h>
52 #include <sys/objcache.h>
53 #include <sys/queue.h>
54 #include <sys/thread.h>
56 #include <sys/sysctl.h>
57 #include <sys/sysproto.h>
59 #include <sys/socket.h>
60 #include <sys/socketvar.h>
61 #include <sys/socketops.h>
62 #include <sys/sysref.h>
63 #include <sys/syslink.h>
64 #include <sys/syslink_msg.h>
65 #include <netinet/in.h>
67 #include <sys/thread2.h>
68 #include <sys/spinlock2.h>
71 #include "opt_syslink.h"
74 * Syslink Connection abstraction
83 struct slmsg_rb_tree reply_rb_root; /* replies to requests */
85 struct sldesc *peer; /* peer syslink, if any */
86 struct file *xfp; /* external file pointer */
87 struct slcommon *common;
89 int rwaiters; /* number of threads waiting */
90 int wblocked; /* blocked waiting for us to drain */
91 size_t cmdbytes; /* unreplied commands pending */
92 size_t repbytes; /* undrained replies pending */
93 int (*backend_wblocked)(struct sldesc *, int, sl_proto_t);
94 int (*backend_write)(struct sldesc *, struct slmsg *);
95 void (*backend_reply)(struct sldesc *,struct slmsg *,struct slmsg *);
96 void (*backend_dispose)(struct sldesc *, struct slmsg *);
99 #define SLF_RSHUTDOWN 0x0001
100 #define SLF_WSHUTDOWN 0x0002
102 static int syslink_cmd_new(struct syslink_info_new *info, int *result);
103 static struct sldesc *allocsldesc(struct slcommon *common);
104 static void setsldescfp(struct sldesc *sl, struct file *fp);
105 static void shutdownsldesc(struct sldesc *sl, int how);
106 static void shutdownsldesc2(struct sldesc *sl, int how);
107 static void sldrop(struct sldesc *sl);
108 static int syslink_validate_msg(struct syslink_msg *msg, int bytes);
109 static int syslink_validate_elm(struct syslink_elm *elm, sl_reclen_t bytes,
110 int swapit, int depth);
112 static int sl_local_mmap(struct slmsg *slmsg, char *base, size_t len);
113 static void sl_local_munmap(struct slmsg *slmsg);
115 static int backend_wblocked_user(struct sldesc *sl, int nbio, sl_proto_t proto);
116 static int backend_write_user(struct sldesc *sl, struct slmsg *slmsg);
117 static void backend_reply_user(struct sldesc *sl, struct slmsg *slcmd,
118 struct slmsg *slrep);
119 static void backend_dispose_user(struct sldesc *sl, struct slmsg *slmsg);
121 static int backend_wblocked_kern(struct sldesc *sl, int nbio, sl_proto_t proto);
122 static int backend_write_kern(struct sldesc *sl, struct slmsg *slmsg);
123 static void backend_reply_kern(struct sldesc *sl, struct slmsg *slcmd,
124 struct slmsg *slrep);
125 static void backend_dispose_kern(struct sldesc *sl, struct slmsg *slmsg);
126 static void slmsg_put(struct slmsg *slmsg);
129 * Objcache memory backend
131 * All three object caches return slmsg structures but each is optimized
132 * for syslink message buffers of varying sizes. We use the slightly
133 * more complex ctor/dtor API in order to provide ready-to-go slmsg's.
136 static struct objcache *sl_objcache_big;
137 static struct objcache *sl_objcache_small;
138 static struct objcache *sl_objcache_none;
140 MALLOC_DEFINE(M_SYSLINK, "syslink", "syslink manager");
142 static boolean_t slmsg_ctor(void *data, void *private, int ocflags);
143 static void slmsg_dtor(void *data, void *private);
147 syslinkinit(void *dummy __unused)
149 size_t n = sizeof(struct slmsg);
151 sl_objcache_none = objcache_create_mbacked(M_SYSLINK, n, 0, 64,
152 slmsg_ctor, slmsg_dtor,
154 sl_objcache_small= objcache_create_mbacked(M_SYSLINK, n, 0, 64,
155 slmsg_ctor, slmsg_dtor,
157 sl_objcache_big = objcache_create_mbacked(M_SYSLINK, n, 0, 16,
158 slmsg_ctor, slmsg_dtor,
164 slmsg_ctor(void *data, void *private, int ocflags)
166 struct slmsg *slmsg = data;
168 bzero(slmsg, sizeof(*slmsg));
170 slmsg->oc = *(struct objcache **)private;
171 if (slmsg->oc == sl_objcache_none) {
173 } else if (slmsg->oc == sl_objcache_small) {
174 slmsg->maxsize = SLMSG_SMALL;
175 } else if (slmsg->oc == sl_objcache_big) {
176 slmsg->maxsize = SLMSG_BIG;
178 panic("slmsg_ctor: bad objcache?\n");
180 if (slmsg->maxsize) {
181 slmsg->msg = kmalloc(slmsg->maxsize,
182 M_SYSLINK, M_WAITOK|M_ZERO);
184 xio_init(&slmsg->xio);
190 slmsg_dtor(void *data, void *private)
192 struct slmsg *slmsg = data;
194 if (slmsg->maxsize && slmsg->msg) {
195 kfree(slmsg->msg, M_SYSLINK);
201 SYSINIT(syslink, SI_BOOT2_MACHDEP, SI_ORDER_ANY, syslinkinit, NULL)
203 static int rb_slmsg_compare(struct slmsg *msg1, struct slmsg *msg2);
204 RB_GENERATE2(slmsg_rb_tree, slmsg, rbnode, rb_slmsg_compare,
205 sysid_t, msg->sm_msgid);
210 static int syslink_enabled;
211 SYSCTL_NODE(_kern, OID_AUTO, syslink, CTLFLAG_RW, 0, "Pipe operation");
212 SYSCTL_INT(_kern_syslink, OID_AUTO, enabled,
213 CTLFLAG_RW, &syslink_enabled, 0, "Enable SYSLINK");
214 static size_t syslink_bufsize = 65536;
215 SYSCTL_UINT(_kern_syslink, OID_AUTO, bufsize,
216 CTLFLAG_RW, &syslink_bufsize, 0, "Maximum buffer size");
219 * Fileops API - typically used to glue a userland frontend with a
223 static int slfileop_read(struct file *fp, struct uio *uio,
224 struct ucred *cred, int flags);
225 static int slfileop_write(struct file *fp, struct uio *uio,
226 struct ucred *cred, int flags);
227 static int slfileop_close(struct file *fp);
228 static int slfileop_stat(struct file *fp, struct stat *sb, struct ucred *cred);
229 static int slfileop_shutdown(struct file *fp, int how);
230 static int slfileop_ioctl(struct file *fp, u_long cmd, caddr_t data,
232 static int slfileop_poll(struct file *fp, int events, struct ucred *cred);
233 static int slfileop_kqfilter(struct file *fp, struct knote *kn);
235 static struct fileops syslinkops = {
236 .fo_read = slfileop_read,
237 .fo_write = slfileop_write,
238 .fo_ioctl = slfileop_ioctl,
239 .fo_poll = slfileop_poll,
240 .fo_kqfilter = slfileop_kqfilter,
241 .fo_stat = slfileop_stat,
242 .fo_close = slfileop_close,
243 .fo_shutdown = slfileop_shutdown
246 /************************************************************************
247 * PRIMARY SYSTEM CALL INTERFACE *
248 ************************************************************************
250 * syslink(int cmd, struct syslink_info *info, size_t bytes)
253 sys_syslink(struct syslink_args *uap)
255 union syslink_info_all info;
259 * System call is under construction and disabled by default.
260 * Superuser access is also required for now, but eventually
261 * will not be needed.
263 if (syslink_enabled == 0)
265 error = suser(curthread);
270 * Load and validate the info structure. Unloaded bytes are zerod
271 * out. The label field must always be 0-filled, even if not used
274 bzero(&info, sizeof(info));
275 if ((unsigned)uap->bytes <= sizeof(info)) {
277 error = copyin(uap->info, &info, uap->bytes);
285 * Process the command
288 case SYSLINK_CMD_NEW:
289 error = syslink_cmd_new(&info.cmd_new, &uap->sysmsg_result);
295 if (error == 0 && info.head.wbflag)
296 copyout(&info, uap->info, uap->bytes);
301 * Create a linked pair of descriptors, like a pipe.
305 syslink_cmd_new(struct syslink_info_new *info, int *result)
307 struct proc *p = curproc;
311 struct sldesc *slpeer;
315 error = falloc(p, &fp1, &fd1);
318 error = falloc(p, &fp2, &fd2);
320 fsetfd(p, NULL, fd1);
324 slpeer = allocsldesc(NULL);
325 slpeer->backend_wblocked = backend_wblocked_user;
326 slpeer->backend_write = backend_write_user;
327 slpeer->backend_reply = backend_reply_user;
328 slpeer->backend_dispose = backend_dispose_user;
329 sl = allocsldesc(slpeer->common);
331 sl->backend_wblocked = backend_wblocked_user;
332 sl->backend_write = backend_write_user;
333 sl->backend_reply = backend_reply_user;
334 sl->backend_dispose = backend_dispose_user;
337 setsldescfp(sl, fp1);
338 setsldescfp(slpeer, fp2);
345 info->head.wbflag = 1; /* write back */
352 /************************************************************************
353 * LOW LEVEL SLDESC SUPPORT *
354 ************************************************************************
360 allocsldesc(struct slcommon *common)
364 sl = kmalloc(sizeof(struct sldesc), M_SYSLINK, M_WAITOK|M_ZERO);
366 common = kmalloc(sizeof(*common), M_SYSLINK, M_WAITOK|M_ZERO);
367 TAILQ_INIT(&sl->inq); /* incoming requests */
368 RB_INIT(&sl->reply_rb_root); /* match incoming replies */
369 spin_init(&sl->spin);
377 setsldescfp(struct sldesc *sl, struct file *fp)
380 fp->f_type = DTYPE_SYSLINK;
381 fp->f_flag = FREAD | FWRITE;
382 fp->f_ops = &syslinkops;
387 * Red-black tree compare function
391 rb_slmsg_compare(struct slmsg *msg1, struct slmsg *msg2)
393 if (msg1->msg->sm_msgid < msg2->msg->sm_msgid)
395 if (msg1->msg->sm_msgid == msg2->msg->sm_msgid)
402 shutdownsldesc(struct sldesc *sl, int how)
407 shutdownsldesc2(sl, how);
410 * Return unread and unreplied messages
412 spin_lock_wr(&sl->spin);
413 while ((slmsg = TAILQ_FIRST(&sl->inq)) != NULL) {
414 TAILQ_REMOVE(&sl->inq, slmsg, tqnode);
415 spin_unlock_wr(&sl->spin);
416 if (slmsg->msg->sm_proto & SM_PROTO_REPLY) {
417 sl->repbytes -= slmsg->maxsize;
418 slmsg->flags &= ~SLMSGF_ONINQ;
419 sl->peer->backend_dispose(sl->peer, slmsg);
421 /* leave ONINQ set for commands, it will cleared below */
422 spin_lock_wr(&sl->spin);
424 while ((slmsg = RB_ROOT(&sl->reply_rb_root)) != NULL) {
425 RB_REMOVE(slmsg_rb_tree, &sl->reply_rb_root, slmsg);
426 sl->cmdbytes -= slmsg->maxsize;
427 spin_unlock_wr(&sl->spin);
428 slmsg->flags &= ~SLMSGF_ONINQ;
429 sl->peer->backend_reply(sl->peer, slmsg, NULL);
430 spin_lock_wr(&sl->spin);
432 spin_unlock_wr(&sl->spin);
435 * Call shutdown on the peer with the opposite flags
449 shutdownsldesc2(sl->peer, rhow);
454 shutdownsldesc2(struct sldesc *sl, int how)
456 spin_lock_wr(&sl->spin);
459 sl->flags |= SLF_RSHUTDOWN;
462 sl->flags |= SLF_WSHUTDOWN;
465 sl->flags |= SLF_RSHUTDOWN | SLF_WSHUTDOWN;
468 spin_unlock_wr(&sl->spin);
471 * Handle signaling on the user side
475 wakeup(&sl->rwaiters);
479 sl->wblocked = 0; /* race ok */
480 wakeup(&sl->wblocked);
487 sldrop(struct sldesc *sl)
489 struct sldesc *slpeer;
491 spin_lock_wr(&sl->common->spin);
492 if (--sl->common->refs == 0) {
493 spin_unlock_wr(&sl->common->spin);
494 if ((slpeer = sl->peer) != NULL) {
497 slpeer->common = NULL;
498 KKASSERT(slpeer->xfp == NULL);
499 KKASSERT(TAILQ_EMPTY(&slpeer->inq));
500 KKASSERT(RB_EMPTY(&slpeer->reply_rb_root));
501 kfree(slpeer, M_SYSLINK);
503 KKASSERT(sl->xfp == NULL);
504 KKASSERT(TAILQ_EMPTY(&sl->inq));
505 KKASSERT(RB_EMPTY(&sl->reply_rb_root));
506 kfree(sl->common, M_SYSLINK);
508 kfree(sl, M_SYSLINK);
510 spin_unlock_wr(&sl->common->spin);
516 slmsg_put(struct slmsg *slmsg)
518 if (slmsg->flags & SLMSGF_HASXIO) {
519 slmsg->flags &= ~SLMSGF_HASXIO;
521 xio_release(&slmsg->xio);
524 slmsg->flags &= ~SLMSGF_LINMAP;
525 objcache_put(slmsg->oc, slmsg);
528 /************************************************************************
530 ************************************************************************
532 * Implement userland fileops.
538 slfileop_read(struct file *fp, struct uio *uio, struct ucred *cred, int flags)
540 struct sldesc *sl = fp->f_data; /* fp refed on call */
544 struct syslink_msg *wmsg;
549 * Kinda messy. Figure out the non-blocking state
551 if (flags & O_FBLOCKING)
553 else if (flags & O_FNONBLOCKING)
555 else if (fp->f_flag & O_NONBLOCK)
563 * iov0 - message buffer
564 * iov1 - DMA buffer or backup buffer
566 if (uio->uio_iovcnt < 1) {
570 iov0 = &uio->uio_iov[0];
571 if (uio->uio_iovcnt > 2) {
577 * Get a message, blocking if necessary.
579 spin_lock_wr(&sl->spin);
580 while ((slmsg = TAILQ_FIRST(&sl->inq)) == NULL) {
581 if (sl->flags & SLF_RSHUTDOWN) {
590 error = msleep(&sl->rwaiters, &sl->spin, PCATCH, "slrmsg", 0);
598 * We have a message and still hold the spinlock. Make sure the
599 * uio has enough room to hold the message.
601 * Note that replies do not have XIOs.
603 if (slmsg->msgsize > iov0->iov_len) {
607 if (slmsg->xio.xio_bytes) {
608 if (uio->uio_iovcnt != 2) {
612 iov1 = &uio->uio_iov[1];
613 if (slmsg->xio.xio_bytes > iov1->iov_len) {
622 * Dequeue the message. Adjust repbytes immediately. cmdbytes
623 * are adjusted when the command is replied to, not here.
625 TAILQ_REMOVE(&sl->inq, slmsg, tqnode);
626 if (slmsg->msg->sm_proto & SM_PROTO_REPLY)
627 sl->repbytes -= slmsg->maxsize;
628 spin_unlock_wr(&sl->spin);
631 * Load the message data into the user buffer.
633 * If receiving a command an XIO may exist specifying a DMA buffer.
634 * For commands, if DMAW is set we have to copy or map the buffer
635 * so the caller can access the data being written. If DMAR is set
636 * we do not have to copy but we still must map the buffer so the
637 * caller can directly fill in the data being requested.
639 error = uiomove((void *)slmsg->msg, slmsg->msgsize, uio);
640 if (error == 0 && slmsg->xio.xio_bytes &&
641 (wmsg->sm_head.se_cmd & SE_CMDF_REPLY) == 0) {
642 if (wmsg->sm_head.se_cmd & SE_CMDF_DMAW) {
644 * Data being passed to caller or being passed in both
645 * directions, copy or map.
648 if ((flags & O_MAPONREAD) &&
649 (slmsg->xio.xio_flags & XIOF_VMLINEAR)) {
650 error = sl_local_mmap(slmsg,
654 error = xio_copy_xtou(&slmsg->xio, 0,
656 slmsg->xio.xio_bytes);
658 error = xio_copy_xtou(&slmsg->xio, 0,
660 slmsg->xio.xio_bytes);
663 } else if (wmsg->sm_head.se_cmd & SE_CMDF_DMAR) {
665 * Data will be passed back to originator, map
666 * the buffer if we can, else use the backup
667 * buffer at the same VA supplied by the caller.
670 if ((flags & O_MAPONREAD) &&
671 (slmsg->xio.xio_flags & XIOF_VMLINEAR)) {
672 error = sl_local_mmap(slmsg,
675 error = 0; /* ignore errors */
686 * Requeue the message if we could not read it successfully
688 spin_lock_wr(&sl->spin);
689 TAILQ_INSERT_HEAD(&sl->inq, slmsg, tqnode);
690 slmsg->flags |= SLMSGF_ONINQ;
691 spin_unlock_wr(&sl->spin);
692 } else if (slmsg->msg->sm_proto & SM_PROTO_REPLY) {
694 * Dispose of any received reply after we've copied it
695 * to userland. We don't need the slmsg any more.
697 slmsg->flags &= ~SLMSGF_ONINQ;
698 sl->peer->backend_dispose(sl->peer, slmsg);
699 if (sl->wblocked && sl->repbytes < syslink_bufsize) {
700 sl->wblocked = 0; /* MP race ok here */
701 wakeup(&sl->wblocked);
705 * Leave the command in the RB tree but clear ONINQ now
706 * that we have returned it to userland so userland can
709 slmsg->flags &= ~SLMSGF_ONINQ;
713 spin_unlock_wr(&sl->spin);
719 * Userland writes syslink message (optionally with DMA buffer in iov[1]).
723 slfileop_write(struct file *fp, struct uio *uio, struct ucred *cred, int flags)
725 struct sldesc *sl = fp->f_data;
728 struct syslink_msg sltmp;
729 struct syslink_msg *wmsg; /* wire message */
738 * Kinda messy. Figure out the non-blocking state
740 if (flags & O_FBLOCKING)
742 else if (flags & O_FNONBLOCKING)
744 else if (fp->f_flag & O_NONBLOCK)
752 if (uio->uio_iovcnt < 1) {
756 iov0 = &uio->uio_iov[0];
757 if (iov0->iov_len > SLMSG_BIG) {
761 if (uio->uio_iovcnt > 2) {
765 if (uio->uio_iovcnt > 1) {
766 iov1 = &uio->uio_iov[1];
767 if (iov1->iov_len > XIO_INTERNAL_SIZE) {
771 if ((intptr_t)iov1->iov_base & PAGE_MASK) {
780 * Handle the buffer-full case. slpeer cmdbytes is managed
781 * by the backend function, not us so if the callback just
782 * directly implements the message and never adjusts cmdbytes,
783 * we will never sleep here.
785 if (sl->flags & SLF_WSHUTDOWN) {
791 * Only commands can block the pipe, not replies. Otherwise a
792 * deadlock is possible.
794 error = copyin(iov0->iov_base, &sltmp, sizeof(sltmp));
797 if ((proto = sltmp.sm_proto) & SM_PROTO_ENDIAN_REV)
798 proto = bswap16(proto);
799 error = sl->peer->backend_wblocked(sl->peer, nbio, proto);
804 * Allocate a slmsg and load the message. Note that the bytes
805 * returned to userland only reflects the primary syslink message
806 * and does not include any DMA buffers.
808 if (iov0->iov_len <= SLMSG_SMALL)
809 slmsg = objcache_get(sl_objcache_small, M_WAITOK);
811 slmsg = objcache_get(sl_objcache_big, M_WAITOK);
812 slmsg->msgsize = iov0->iov_len;
815 error = uiomove((void *)wmsg, iov0->iov_len, uio);
818 error = syslink_validate_msg(wmsg, slmsg->msgsize);
822 if ((wmsg->sm_head.se_cmd & SE_CMDF_REPLY) == 0) {
824 * Install the XIO for commands if any DMA flags are set.
826 * XIOF_VMLINEAR requires that the XIO represent a
827 * contiguous set of pages associated with a single VM
828 * object (so the reader side can mmap it easily).
830 * XIOF_VMLINEAR might not be set when the kernel sends
831 * commands to userland so the reader side backs off to
832 * a backup buffer if it isn't set, but we require it
833 * for userland writes.
835 xflags = XIOF_VMLINEAR;
836 if (wmsg->sm_head.se_cmd & SE_CMDF_DMAR)
837 xflags |= XIOF_READ | XIOF_WRITE;
838 else if (wmsg->sm_head.se_cmd & SE_CMDF_DMAW)
840 if (xflags && iov1) {
842 error = xio_init_ubuf(&slmsg->xio, iov1->iov_base,
843 iov1->iov_len, xflags);
847 slmsg->flags |= SLMSGF_HASXIO;
849 error = sl->peer->backend_write(sl->peer, slmsg);
852 * Replies have to be matched up against received commands.
854 spin_lock_wr(&sl->spin);
855 slcmd = slmsg_rb_tree_RB_LOOKUP(&sl->reply_rb_root,
856 slmsg->msg->sm_msgid);
857 if (slcmd == NULL || (slcmd->flags & SLMSGF_ONINQ)) {
859 spin_unlock_wr(&sl->spin);
862 RB_REMOVE(slmsg_rb_tree, &sl->reply_rb_root, slcmd);
863 sl->cmdbytes -= slcmd->maxsize;
864 spin_unlock_wr(&sl->spin);
867 * If the original command specified DMAR, has an xio, and
868 * our write specifies a DMA buffer, then we can do a
869 * copyback. But if we are linearly mapped and the caller
870 * is using the map base address, then the caller filled in
871 * the data via the direct memory map and no copyback is
874 if ((slcmd->msg->sm_head.se_cmd & SE_CMDF_DMAR) && iov1 &&
875 (slcmd->flags & SLMSGF_HASXIO) &&
876 ((slcmd->flags & SLMSGF_LINMAP) == 0 ||
877 iov1->iov_base != slcmd->vmbase)
880 if (iov1->iov_len > slcmd->xio.xio_bytes)
881 count = slcmd->xio.xio_bytes;
883 count = iov1->iov_len;
885 error = xio_copy_utox(&slcmd->xio, 0, iov1->iov_base,
891 * If we had mapped a DMA buffer, remove it
893 if (slcmd->flags & SLMSGF_LINMAP) {
895 sl_local_munmap(slcmd);
900 * Reply and handle unblocking
902 sl->peer->backend_reply(sl->peer, slcmd, slmsg);
903 if (sl->wblocked && sl->cmdbytes < syslink_bufsize) {
904 sl->wblocked = 0; /* MP race ok here */
905 wakeup(&sl->wblocked);
909 * slmsg has already been dealt with, make sure error is
910 * 0 so we do not double-free it.
924 * Close a syslink descriptor.
926 * Disassociate the syslink from the file descriptor and disconnect from
931 slfileop_close(struct file *fp)
936 * Disassociate the file pointer. Take ownership of the ref on the
941 fp->f_ops = &badfileops;
945 * Shutdown both directions. The other side will not issue API
946 * calls to us after we've shutdown both directions.
948 shutdownsldesc(sl, SHUT_RDWR);
953 KKASSERT(sl->cmdbytes == 0);
954 KKASSERT(sl->repbytes == 0);
961 slfileop_stat (struct file *fp, struct stat *sb, struct ucred *cred)
968 slfileop_shutdown (struct file *fp, int how)
970 shutdownsldesc((struct sldesc *)fp->f_data, how);
976 slfileop_ioctl (struct file *fp, u_long cmd, caddr_t data, struct ucred *cred)
983 slfileop_poll (struct file *fp, int events, struct ucred *cred)
990 slfileop_kqfilter(struct file *fp, struct knote *kn)
995 /************************************************************************
996 * LOCAL MEMORY MAPPING *
997 ************************************************************************
999 * This feature is currently not implemented
1005 sl_local_mmap(struct slmsg *slmsg, char *base, size_t len)
1007 return (EOPNOTSUPP);
1012 sl_local_munmap(struct slmsg *slmsg)
1021 sl_local_mmap(struct slmsg *slmsg, char *base, size_t len)
1023 struct vmspace *vms = curproc->p_vmspace;
1024 vm_offset_t addr = (vm_offset_t)base;
1026 /* XXX check user address range */
1027 error = vm_map_replace(
1029 (vm_offset_t)base, (vm_offset_t)base + len,
1030 slmsg->xio.xio_pages[0]->object,
1031 slmsg->xio.xio_pages[0]->pindex << PAGE_SHIFT,
1032 VM_PROT_READ|VM_PROT_WRITE,
1033 VM_PROT_READ|VM_PROT_WRITE,
1034 MAP_DISABLE_SYNCER);
1037 slmsg->flags |= SLMSGF_LINMAP;
1038 slmsg->vmbase = base;
1039 slmsg->vmsize = len;
1046 sl_local_munmap(struct slmsg *slmsg)
1048 if (slmsg->flags & SLMSGF_LINMAP) {
1049 vm_map_remove(&curproc->p_vmspace->vm_map,
1051 slmsg->vmbase + slcmd->vmsize);
1052 slmsg->flags &= ~SLMSGF_LINMAP;
1058 /************************************************************************
1059 * MESSAGE VALIDATION *
1060 ************************************************************************
1062 * Validate that the syslink message. Check that all headers and elements
1063 * conform. Correct the endian if necessary.
1065 * NOTE: If reverse endian needs to be corrected, SE_CMDF_UNTRANSLATED
1066 * is recursively flipped on all syslink_elm's in the message. As the
1067 * message traverses the mesh, multiple flips may occur. It is
1068 * up to the RPC protocol layer to correct opaque data payloads and
1069 * SE_CMDF_UNTRANSLATED prevents the protocol layer from misinterpreting
1070 * a command or reply element which has not been endian-corrected.
1074 syslink_validate_msg(struct syslink_msg *msg, int bytes)
1081 * The raw message must be properly-aligned.
1083 if (bytes & SL_ALIGNMASK)
1088 * The message must at least contain the msgid, bytes, and
1091 if (bytes < SL_MIN_PAD_SIZE)
1095 * Fix the endian if it is reversed.
1097 if (msg->sm_proto & SM_PROTO_ENDIAN_REV) {
1098 msg->sm_msgid = bswap64(msg->sm_msgid);
1099 msg->sm_sessid = bswap64(msg->sm_sessid);
1100 msg->sm_bytes = bswap16(msg->sm_bytes);
1101 msg->sm_proto = bswap16(msg->sm_proto);
1102 msg->sm_rlabel = bswap32(msg->sm_rlabel);
1103 if (msg->sm_proto & SM_PROTO_ENDIAN_REV)
1111 * Validate the contents. For PADs, the entire payload is
1112 * ignored and the minimum message size can be as small as
1115 if (msg->sm_proto == SMPROTO_PAD) {
1116 if (msg->sm_bytes < SL_MIN_PAD_SIZE ||
1117 msg->sm_bytes > bytes) {
1120 /* ignore the entire payload, it can be garbage */
1122 if (msg->sm_bytes < SL_MIN_MSG_SIZE ||
1123 msg->sm_bytes > bytes) {
1126 error = syslink_validate_elm(
1129 offsetof(struct syslink_msg,
1131 swapit, SL_MAXDEPTH);
1137 * The aligned payload size must be used to locate the
1138 * next syslink_msg in the buffer.
1140 aligned_reclen = SL_MSG_ALIGN(msg->sm_bytes);
1141 bytes -= aligned_reclen;
1142 msg = (void *)((char *)msg + aligned_reclen);
1149 syslink_validate_elm(struct syslink_elm *elm, sl_reclen_t bytes,
1150 int swapit, int depth)
1155 * If the buffer isn't big enough to fit the header, stop now!
1157 if (bytes < SL_MIN_ELM_SIZE)
1160 * All syslink_elm headers are recursively endian-adjusted. Opaque
1161 * data payloads are not.
1164 elm->se_cmd = bswap16(elm->se_cmd) ^ SE_CMDF_UNTRANSLATED;
1165 elm->se_bytes = bswap16(elm->se_bytes);
1166 elm->se_aux = bswap32(elm->se_aux);
1170 * Check element size requirements.
1172 if (elm->se_bytes < SL_MIN_ELM_SIZE || elm->se_bytes > bytes)
1176 * Recursively check structured payloads. A structured payload may
1177 * contain as few as 0 recursive elements.
1179 if (elm->se_cmd & SE_CMDF_STRUCTURED) {
1182 bytes -= SL_MIN_ELM_SIZE;
1185 if (syslink_validate_elm(elm, bytes, swapit, depth - 1))
1187 aligned_reclen = SL_MSG_ALIGN(elm->se_bytes);
1188 elm = (void *)((char *)elm + aligned_reclen);
1189 bytes -= aligned_reclen;
1195 /************************************************************************
1196 * BACKEND FUNCTIONS - USER DESCRIPTOR *
1197 ************************************************************************
1199 * Peer backend links are primarily used when userland creates a pair
1200 * of linked descriptors.
1204 * Do any required blocking / nbio handling for attempts to write to
1205 * a sldesc associated with a user descriptor.
1209 backend_wblocked_user(struct sldesc *sl, int nbio, sl_proto_t proto)
1212 int *bytesp = (proto & SM_PROTO_REPLY) ? &sl->repbytes : &sl->cmdbytes;
1215 * Block until sufficient data is drained by the target. It is
1216 * ok to have a MP race against cmdbytes.
1218 if (*bytesp >= syslink_bufsize) {
1219 spin_lock_wr(&sl->spin);
1220 while (*bytesp >= syslink_bufsize) {
1221 if (sl->flags & SLF_WSHUTDOWN) {
1230 error = msleep(&sl->wblocked, &sl->spin,
1231 PCATCH, "slwmsg", 0);
1235 spin_unlock_wr(&sl->spin);
1241 * Unconditionally write a syslink message to the sldesc associated with
1242 * a user descriptor. Command messages are also placed in a red-black
1243 * tree so their DMA tag (if any) can be accessed and so they can be
1244 * linked to any reply message.
1248 backend_write_user(struct sldesc *sl, struct slmsg *slmsg)
1252 spin_lock_wr(&sl->spin);
1253 if (sl->flags & SLF_RSHUTDOWN) {
1255 * Not accepting new messages
1258 } else if (slmsg->msg->sm_proto & SM_PROTO_REPLY) {
1262 TAILQ_INSERT_TAIL(&sl->inq, slmsg, tqnode);
1263 sl->repbytes += slmsg->maxsize;
1264 slmsg->flags |= SLMSGF_ONINQ;
1266 } else if (RB_INSERT(slmsg_rb_tree, &sl->reply_rb_root, slmsg)) {
1268 * Write a command, but there was a msgid collision when
1269 * we tried to insert it into the RB tree.
1274 * Write a command, successful insertion into the RB tree.
1276 TAILQ_INSERT_TAIL(&sl->inq, slmsg, tqnode);
1277 sl->cmdbytes += slmsg->maxsize;
1278 slmsg->flags |= SLMSGF_ONINQ;
1281 spin_unlock_wr(&sl->spin);
1283 wakeup(&sl->rwaiters);
1288 * Our peer is replying a command we previously sent it back to us, along
1289 * with the reply message (if not NULL). We just queue the reply to
1290 * userland and free of the command.
1294 backend_reply_user(struct sldesc *sl, struct slmsg *slcmd, struct slmsg *slrep)
1300 spin_lock_wr(&sl->spin);
1301 if ((sl->flags & SLF_RSHUTDOWN) == 0) {
1302 TAILQ_INSERT_TAIL(&sl->inq, slrep, tqnode);
1303 sl->repbytes += slrep->maxsize;
1308 spin_unlock_wr(&sl->spin);
1310 sl->peer->backend_dispose(sl->peer, slrep);
1311 else if (sl->rwaiters)
1312 wakeup(&sl->rwaiters);
1318 backend_dispose_user(struct sldesc *sl, struct slmsg *slmsg)
1323 /************************************************************************
1324 * KERNEL DRIVER OR FILESYSTEM API *
1325 ************************************************************************
1330 * Create a user<->kernel link, returning the user descriptor in *fdp
1331 * and the kernel descriptor in *kslp. 0 is returned on success, and an
1332 * error code is returned on failure.
1335 syslink_ukbackend(int *fdp, struct sldesc **kslp)
1337 struct proc *p = curproc;
1347 error = falloc(p, &fp, &fd);
1350 usl = allocsldesc(NULL);
1351 usl->backend_wblocked = backend_wblocked_user;
1352 usl->backend_write = backend_write_user;
1353 usl->backend_reply = backend_reply_user;
1354 usl->backend_dispose = backend_dispose_user;
1356 ksl = allocsldesc(usl->common);
1358 ksl->backend_wblocked = backend_wblocked_kern;
1359 ksl->backend_write = backend_write_kern;
1360 ksl->backend_reply = backend_reply_kern;
1361 ksl->backend_dispose = backend_dispose_kern;
1365 setsldescfp(usl, fp);
1375 * Assign a unique message id, issue a syslink message to userland,
1376 * and wait for a reply.
1379 syslink_kdomsg(struct sldesc *ksl, struct slmsg *slmsg)
1381 struct syslink_msg *msg;
1385 * Finish initializing slmsg and post it to the red-black tree for
1386 * reply matching. If the message id is already in use we return
1387 * EEXIST, giving the originator the chance to roll a new msgid.
1390 slmsg->msgsize = msg->sm_bytes;
1391 if ((error = syslink_validate_msg(msg, msg->sm_bytes)) != 0)
1393 msg->sm_msgid = allocsysid();
1396 * Issue the request and wait for a matching reply or failure,
1397 * then remove the message from the matching tree and return.
1399 error = ksl->peer->backend_write(ksl->peer, slmsg);
1400 spin_lock_wr(&ksl->spin);
1402 while (slmsg->rep == NULL) {
1403 error = msleep(slmsg, &ksl->spin, 0, "kwtmsg", 0);
1404 /* XXX ignore error for now */
1406 if (slmsg->rep == (struct slmsg *)-1) {
1410 error = slmsg->rep->msg->sm_head.se_aux;
1413 spin_unlock_wr(&ksl->spin);
1418 * Similar to syslink_kdomsg but return immediately instead of
1419 * waiting for a reply. The kernel must supply a callback function
1420 * which will be made in the context of the user process replying
1424 syslink_ksendmsg(struct sldesc *ksl, struct slmsg *slmsg,
1425 void (*func)(struct slmsg *, void *, int), void *arg)
1427 struct syslink_msg *msg;
1431 * Finish initializing slmsg and post it to the red-black tree for
1432 * reply matching. If the message id is already in use we return
1433 * EEXIST, giving the originator the chance to roll a new msgid.
1436 slmsg->msgsize = msg->sm_bytes;
1437 slmsg->callback_func = func;
1438 slmsg->callback_data = arg;
1439 if ((error = syslink_validate_msg(msg, msg->sm_bytes)) != 0)
1441 msg->sm_msgid = allocsysid();
1444 * Issue the request. If no error occured the operation will be
1445 * in progress, otherwise the operation is considered to have failed
1446 * and the caller can deallocate the slmsg.
1448 error = ksl->peer->backend_write(ksl->peer, slmsg);
1453 syslink_kwaitmsg(struct sldesc *ksl, struct slmsg *slmsg)
1457 spin_lock_wr(&ksl->spin);
1458 while (slmsg->rep == NULL) {
1459 error = msleep(slmsg, &ksl->spin, 0, "kwtmsg", 0);
1460 /* XXX ignore error for now */
1462 if (slmsg->rep == (struct slmsg *)-1) {
1466 error = slmsg->rep->msg->sm_head.se_aux;
1468 spin_unlock_wr(&ksl->spin);
1473 syslink_kallocmsg(void)
1475 return(objcache_get(sl_objcache_small, M_WAITOK));
1479 syslink_kfreemsg(struct sldesc *ksl, struct slmsg *slmsg)
1483 if ((rep = slmsg->rep) != NULL) {
1485 ksl->peer->backend_dispose(ksl->peer, rep);
1487 slmsg->callback_func = NULL;
1492 syslink_kshutdown(struct sldesc *ksl, int how)
1494 shutdownsldesc(ksl, how);
1498 syslink_kclose(struct sldesc *ksl)
1500 shutdownsldesc(ksl, SHUT_RDWR);
1505 * Associate a DMA buffer with a kernel syslink message prior to it
1506 * being sent to userland. The DMA buffer is set up from the point
1507 * of view of the target.
1510 syslink_kdmabuf_pages(struct slmsg *slmsg, struct vm_page **mbase, int npages)
1515 xflags = XIOF_VMLINEAR;
1516 if (slmsg->msg->sm_head.se_cmd & SE_CMDF_DMAR)
1517 xflags |= XIOF_READ | XIOF_WRITE;
1518 else if (slmsg->msg->sm_head.se_cmd & SE_CMDF_DMAW)
1519 xflags |= XIOF_READ;
1520 error = xio_init_pages(&slmsg->xio, mbase, npages, xflags);
1521 slmsg->flags |= SLMSGF_HASXIO;
1526 * Associate a DMA buffer with a kernel syslink message prior to it
1527 * being sent to userland. The DMA buffer is set up from the point
1528 * of view of the target.
1531 syslink_kdmabuf_data(struct slmsg *slmsg, char *base, int bytes)
1535 xflags = XIOF_VMLINEAR;
1536 if (slmsg->msg->sm_head.se_cmd & SE_CMDF_DMAR)
1537 xflags |= XIOF_READ | XIOF_WRITE;
1538 else if (slmsg->msg->sm_head.se_cmd & SE_CMDF_DMAW)
1539 xflags |= XIOF_READ;
1540 xio_init_kbuf(&slmsg->xio, base, bytes);
1541 slmsg->xio.xio_flags |= xflags;
1542 slmsg->flags |= SLMSGF_HASXIO;
1546 /************************************************************************
1547 * BACKEND FUNCTIONS FOR KERNEL API *
1548 ************************************************************************
1550 * These are the backend functions for a sldesc associated with a kernel
1555 * Our peer wants to write a syslink message to us and is asking us to
1556 * block if our input queue is full. We don't implement command reception
1557 * so don't block right now.
1561 backend_wblocked_kern(struct sldesc *ksl, int nbio, sl_proto_t proto)
1568 * Our peer is writing a request to the kernel. At the moment we do not
1573 backend_write_kern(struct sldesc *ksl, struct slmsg *slmsg)
1579 * Our peer wants to reply to a syslink message we sent it earlier. The
1580 * original command (that we passed to our peer), and the peer's reply
1581 * is specified. If the peer has failed slrep will be NULL.
1585 backend_reply_kern(struct sldesc *ksl, struct slmsg *slcmd, struct slmsg *slrep)
1589 spin_lock_wr(&ksl->spin);
1590 if (slrep == NULL) {
1591 slcmd->rep = (struct slmsg *)-1;
1595 error = slrep->msg->sm_head.se_aux;
1597 spin_unlock_wr(&ksl->spin);
1600 * Issue callback or wakeup a synchronous waiter.
1602 if (slcmd->callback_func) {
1603 slcmd->callback_func(slcmd, slcmd->callback_data, error);
1610 * Any reply messages we sent to our peer are returned to us for disposal.
1611 * Since we do not currently accept commands from our peer, there will not
1612 * be any replies returned to the peer to dispose of.
1616 backend_dispose_kern(struct sldesc *ksl, struct slmsg *slmsg)
1618 panic("backend_dispose_kern: kernel can't accept commands so it "
1619 "certainly did not reply to one!");