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
35 * This module implements the core syslink() system call and provides
36 * glue for kernel syslink frontends and backends, creating a intra-host
37 * communications infrastructure and DMA transport abstraction.
40 #include <sys/param.h>
41 #include <sys/systm.h>
42 #include <sys/kernel.h>
43 #include <sys/endian.h>
44 #include <sys/malloc.h>
45 #include <sys/alist.h>
51 #include <sys/objcache.h>
52 #include <sys/queue.h>
53 #include <sys/thread.h>
55 #include <sys/sysctl.h>
56 #include <sys/sysproto.h>
58 #include <sys/socket.h>
59 #include <sys/socketvar.h>
60 #include <sys/socketops.h>
61 #include <sys/sysref.h>
62 #include <sys/syslink.h>
63 #include <sys/syslink_msg.h>
64 #include <netinet/in.h>
66 #include <sys/thread2.h>
67 #include <sys/spinlock2.h>
69 #include <sys/mplock2.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, NULL, 64,
152 slmsg_ctor, slmsg_dtor,
154 sl_objcache_small= objcache_create_mbacked(M_SYSLINK, n, NULL, 64,
155 slmsg_ctor, slmsg_dtor,
157 sl_objcache_big = objcache_create_mbacked(M_SYSLINK, n, NULL, 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?");
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,
231 struct ucred *cred, struct sysmsg *msg);
232 static int slfileop_kqfilter(struct file *fp, struct knote *kn);
234 static struct fileops syslinkops = {
235 .fo_read = slfileop_read,
236 .fo_write = slfileop_write,
237 .fo_ioctl = slfileop_ioctl,
238 .fo_kqfilter = slfileop_kqfilter,
239 .fo_stat = slfileop_stat,
240 .fo_close = slfileop_close,
241 .fo_shutdown = slfileop_shutdown
244 /************************************************************************
245 * PRIMARY SYSTEM CALL INTERFACE *
246 ************************************************************************
248 * 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 = priv_check(curthread, PRIV_ROOT);
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);
286 * Process the command
289 case SYSLINK_CMD_NEW:
290 error = syslink_cmd_new(&info.cmd_new, &uap->sysmsg_result);
298 if (error == 0 && info.head.wbflag)
299 copyout(&info, uap->info, uap->bytes);
304 * Create a linked pair of descriptors, like a pipe.
308 syslink_cmd_new(struct syslink_info_new *info, int *result)
310 struct thread *td = curthread;
311 struct filedesc *fdp = td->td_proc->p_fd;
315 struct sldesc *slpeer;
319 error = falloc(td->td_lwp, &fp1, &fd1);
322 error = falloc(td->td_lwp, &fp2, &fd2);
324 fsetfd(fdp, NULL, fd1);
328 slpeer = allocsldesc(NULL);
329 slpeer->backend_wblocked = backend_wblocked_user;
330 slpeer->backend_write = backend_write_user;
331 slpeer->backend_reply = backend_reply_user;
332 slpeer->backend_dispose = backend_dispose_user;
333 sl = allocsldesc(slpeer->common);
335 sl->backend_wblocked = backend_wblocked_user;
336 sl->backend_write = backend_write_user;
337 sl->backend_reply = backend_reply_user;
338 sl->backend_dispose = backend_dispose_user;
341 setsldescfp(sl, fp1);
342 setsldescfp(slpeer, fp2);
344 fsetfd(fdp, fp1, fd1);
346 fsetfd(fdp, fp2, fd2);
349 info->head.wbflag = 1; /* write back */
356 /************************************************************************
357 * LOW LEVEL SLDESC SUPPORT *
358 ************************************************************************
364 allocsldesc(struct slcommon *common)
368 sl = kmalloc(sizeof(struct sldesc), M_SYSLINK, M_WAITOK|M_ZERO);
370 common = kmalloc(sizeof(*common), M_SYSLINK, M_WAITOK|M_ZERO);
371 TAILQ_INIT(&sl->inq); /* incoming requests */
372 RB_INIT(&sl->reply_rb_root); /* match incoming replies */
373 spin_init(&sl->spin);
381 setsldescfp(struct sldesc *sl, struct file *fp)
384 fp->f_type = DTYPE_SYSLINK;
385 fp->f_flag = FREAD | FWRITE;
386 fp->f_ops = &syslinkops;
391 * Red-black tree compare function
395 rb_slmsg_compare(struct slmsg *msg1, struct slmsg *msg2)
397 if (msg1->msg->sm_msgid < msg2->msg->sm_msgid)
399 if (msg1->msg->sm_msgid == msg2->msg->sm_msgid)
406 shutdownsldesc(struct sldesc *sl, int how)
411 shutdownsldesc2(sl, how);
414 * Return unread and unreplied messages
416 spin_lock(&sl->spin);
417 while ((slmsg = TAILQ_FIRST(&sl->inq)) != NULL) {
418 TAILQ_REMOVE(&sl->inq, slmsg, tqnode);
419 spin_unlock(&sl->spin);
420 if (slmsg->msg->sm_proto & SM_PROTO_REPLY) {
421 sl->repbytes -= slmsg->maxsize;
422 slmsg->flags &= ~SLMSGF_ONINQ;
423 sl->peer->backend_dispose(sl->peer, slmsg);
425 /* leave ONINQ set for commands, it will cleared below */
426 spin_lock(&sl->spin);
428 while ((slmsg = RB_ROOT(&sl->reply_rb_root)) != NULL) {
429 RB_REMOVE(slmsg_rb_tree, &sl->reply_rb_root, slmsg);
430 sl->cmdbytes -= slmsg->maxsize;
431 spin_unlock(&sl->spin);
432 slmsg->flags &= ~SLMSGF_ONINQ;
433 sl->peer->backend_reply(sl->peer, slmsg, NULL);
434 spin_lock(&sl->spin);
436 spin_unlock(&sl->spin);
439 * Call shutdown on the peer with the opposite flags
453 shutdownsldesc2(sl->peer, rhow);
458 shutdownsldesc2(struct sldesc *sl, int how)
460 spin_lock(&sl->spin);
463 sl->flags |= SLF_RSHUTDOWN;
466 sl->flags |= SLF_WSHUTDOWN;
469 sl->flags |= SLF_RSHUTDOWN | SLF_WSHUTDOWN;
472 spin_unlock(&sl->spin);
475 * Handle signaling on the user side
479 wakeup(&sl->rwaiters);
483 sl->wblocked = 0; /* race ok */
484 wakeup(&sl->wblocked);
491 sldrop(struct sldesc *sl)
493 struct sldesc *slpeer;
495 spin_lock(&sl->common->spin);
496 if (--sl->common->refs == 0) {
497 spin_unlock(&sl->common->spin);
498 if ((slpeer = sl->peer) != NULL) {
501 slpeer->common = NULL;
502 KKASSERT(slpeer->xfp == NULL);
503 KKASSERT(TAILQ_EMPTY(&slpeer->inq));
504 KKASSERT(RB_EMPTY(&slpeer->reply_rb_root));
505 kfree(slpeer, M_SYSLINK);
507 KKASSERT(sl->xfp == NULL);
508 KKASSERT(TAILQ_EMPTY(&sl->inq));
509 KKASSERT(RB_EMPTY(&sl->reply_rb_root));
510 kfree(sl->common, M_SYSLINK);
512 kfree(sl, M_SYSLINK);
514 spin_unlock(&sl->common->spin);
520 slmsg_put(struct slmsg *slmsg)
522 if (slmsg->flags & SLMSGF_HASXIO) {
523 slmsg->flags &= ~SLMSGF_HASXIO;
525 xio_release(&slmsg->xio);
528 slmsg->flags &= ~SLMSGF_LINMAP;
529 objcache_put(slmsg->oc, slmsg);
532 /************************************************************************
534 ************************************************************************
536 * Implement userland fileops.
542 slfileop_read(struct file *fp, struct uio *uio, struct ucred *cred, int flags)
544 struct sldesc *sl = fp->f_data; /* fp refed on call */
548 struct syslink_msg *wmsg;
553 * Kinda messy. Figure out the non-blocking state
555 if (flags & O_FBLOCKING)
557 else if (flags & O_FNONBLOCKING)
559 else if (fp->f_flag & O_NONBLOCK)
567 * iov0 - message buffer
568 * iov1 - DMA buffer or backup buffer
570 if (uio->uio_iovcnt < 1) {
574 iov0 = &uio->uio_iov[0];
575 if (uio->uio_iovcnt > 2) {
581 * Get a message, blocking if necessary.
583 spin_lock(&sl->spin);
584 while ((slmsg = TAILQ_FIRST(&sl->inq)) == NULL) {
585 if (sl->flags & SLF_RSHUTDOWN) {
594 error = ssleep(&sl->rwaiters, &sl->spin, PCATCH, "slrmsg", 0);
602 * We have a message and still hold the spinlock. Make sure the
603 * uio has enough room to hold the message.
605 * Note that replies do not have XIOs.
607 if (slmsg->msgsize > iov0->iov_len) {
611 if (slmsg->xio.xio_bytes) {
612 if (uio->uio_iovcnt != 2) {
616 iov1 = &uio->uio_iov[1];
617 if (slmsg->xio.xio_bytes > iov1->iov_len) {
626 * Dequeue the message. Adjust repbytes immediately. cmdbytes
627 * are adjusted when the command is replied to, not here.
629 TAILQ_REMOVE(&sl->inq, slmsg, tqnode);
630 if (slmsg->msg->sm_proto & SM_PROTO_REPLY)
631 sl->repbytes -= slmsg->maxsize;
632 spin_unlock(&sl->spin);
635 * Load the message data into the user buffer.
637 * If receiving a command an XIO may exist specifying a DMA buffer.
638 * For commands, if DMAW is set we have to copy or map the buffer
639 * so the caller can access the data being written. If DMAR is set
640 * we do not have to copy but we still must map the buffer so the
641 * caller can directly fill in the data being requested.
643 error = uiomove((void *)slmsg->msg, slmsg->msgsize, uio);
644 if (error == 0 && slmsg->xio.xio_bytes &&
645 (wmsg->sm_head.se_cmd & SE_CMDF_REPLY) == 0) {
646 if (wmsg->sm_head.se_cmd & SE_CMDF_DMAW) {
648 * Data being passed to caller or being passed in both
649 * directions, copy or map.
652 if ((flags & O_MAPONREAD) &&
653 (slmsg->xio.xio_flags & XIOF_VMLINEAR)) {
654 error = sl_local_mmap(slmsg,
658 error = xio_copy_xtou(&slmsg->xio, 0,
660 slmsg->xio.xio_bytes);
662 error = xio_copy_xtou(&slmsg->xio, 0,
664 slmsg->xio.xio_bytes);
667 } else if (wmsg->sm_head.se_cmd & SE_CMDF_DMAR) {
669 * Data will be passed back to originator, map
670 * the buffer if we can, else use the backup
671 * buffer at the same VA supplied by the caller.
674 if ((flags & O_MAPONREAD) &&
675 (slmsg->xio.xio_flags & XIOF_VMLINEAR)) {
676 error = sl_local_mmap(slmsg,
679 error = 0; /* ignore errors */
690 * Requeue the message if we could not read it successfully
692 spin_lock(&sl->spin);
693 TAILQ_INSERT_HEAD(&sl->inq, slmsg, tqnode);
694 slmsg->flags |= SLMSGF_ONINQ;
695 spin_unlock(&sl->spin);
696 } else if (slmsg->msg->sm_proto & SM_PROTO_REPLY) {
698 * Dispose of any received reply after we've copied it
699 * to userland. We don't need the slmsg any more.
701 slmsg->flags &= ~SLMSGF_ONINQ;
702 sl->peer->backend_dispose(sl->peer, slmsg);
703 if (sl->wblocked && sl->repbytes < syslink_bufsize) {
704 sl->wblocked = 0; /* MP race ok here */
705 wakeup(&sl->wblocked);
709 * Leave the command in the RB tree but clear ONINQ now
710 * that we have returned it to userland so userland can
713 slmsg->flags &= ~SLMSGF_ONINQ;
717 spin_unlock(&sl->spin);
723 * Userland writes syslink message (optionally with DMA buffer in iov[1]).
727 slfileop_write(struct file *fp, struct uio *uio, struct ucred *cred, int flags)
729 struct sldesc *sl = fp->f_data;
732 struct syslink_msg sltmp;
733 struct syslink_msg *wmsg; /* wire message */
742 * Kinda messy. Figure out the non-blocking state
744 if (flags & O_FBLOCKING)
746 else if (flags & O_FNONBLOCKING)
748 else if (fp->f_flag & O_NONBLOCK)
756 if (uio->uio_iovcnt < 1) {
760 iov0 = &uio->uio_iov[0];
761 if (iov0->iov_len > SLMSG_BIG) {
765 if (uio->uio_iovcnt > 2) {
769 if (uio->uio_iovcnt > 1) {
770 iov1 = &uio->uio_iov[1];
771 if (iov1->iov_len > XIO_INTERNAL_SIZE) {
775 if ((intptr_t)iov1->iov_base & PAGE_MASK) {
784 * Handle the buffer-full case. slpeer cmdbytes is managed
785 * by the backend function, not us so if the callback just
786 * directly implements the message and never adjusts cmdbytes,
787 * we will never sleep here.
789 if (sl->flags & SLF_WSHUTDOWN) {
795 * Only commands can block the pipe, not replies. Otherwise a
796 * deadlock is possible.
798 error = copyin(iov0->iov_base, &sltmp, sizeof(sltmp));
801 if ((proto = sltmp.sm_proto) & SM_PROTO_ENDIAN_REV)
802 proto = bswap16(proto);
803 error = sl->peer->backend_wblocked(sl->peer, nbio, proto);
808 * Allocate a slmsg and load the message. Note that the bytes
809 * returned to userland only reflects the primary syslink message
810 * and does not include any DMA buffers.
812 if (iov0->iov_len <= SLMSG_SMALL)
813 slmsg = objcache_get(sl_objcache_small, M_WAITOK);
815 slmsg = objcache_get(sl_objcache_big, M_WAITOK);
816 slmsg->msgsize = iov0->iov_len;
819 error = uiomove((void *)wmsg, iov0->iov_len, uio);
822 error = syslink_validate_msg(wmsg, slmsg->msgsize);
826 if ((wmsg->sm_head.se_cmd & SE_CMDF_REPLY) == 0) {
828 * Install the XIO for commands if any DMA flags are set.
830 * XIOF_VMLINEAR requires that the XIO represent a
831 * contiguous set of pages associated with a single VM
832 * object (so the reader side can mmap it easily).
834 * XIOF_VMLINEAR might not be set when the kernel sends
835 * commands to userland so the reader side backs off to
836 * a backup buffer if it isn't set, but we require it
837 * for userland writes.
839 xflags = XIOF_VMLINEAR;
840 if (wmsg->sm_head.se_cmd & SE_CMDF_DMAR)
841 xflags |= XIOF_READ | XIOF_WRITE;
842 else if (wmsg->sm_head.se_cmd & SE_CMDF_DMAW)
844 if (xflags && iov1) {
846 error = xio_init_ubuf(&slmsg->xio, iov1->iov_base,
847 iov1->iov_len, xflags);
851 slmsg->flags |= SLMSGF_HASXIO;
853 error = sl->peer->backend_write(sl->peer, slmsg);
856 * Replies have to be matched up against received commands.
858 spin_lock(&sl->spin);
859 slcmd = slmsg_rb_tree_RB_LOOKUP(&sl->reply_rb_root,
860 slmsg->msg->sm_msgid);
861 if (slcmd == NULL || (slcmd->flags & SLMSGF_ONINQ)) {
863 spin_unlock(&sl->spin);
866 RB_REMOVE(slmsg_rb_tree, &sl->reply_rb_root, slcmd);
867 sl->cmdbytes -= slcmd->maxsize;
868 spin_unlock(&sl->spin);
871 * If the original command specified DMAR, has an xio, and
872 * our write specifies a DMA buffer, then we can do a
873 * copyback. But if we are linearly mapped and the caller
874 * is using the map base address, then the caller filled in
875 * the data via the direct memory map and no copyback is
878 if ((slcmd->msg->sm_head.se_cmd & SE_CMDF_DMAR) && iov1 &&
879 (slcmd->flags & SLMSGF_HASXIO) &&
880 ((slcmd->flags & SLMSGF_LINMAP) == 0 ||
881 iov1->iov_base != slcmd->vmbase)
884 if (iov1->iov_len > slcmd->xio.xio_bytes)
885 count = slcmd->xio.xio_bytes;
887 count = iov1->iov_len;
889 error = xio_copy_utox(&slcmd->xio, 0, iov1->iov_base,
895 * If we had mapped a DMA buffer, remove it
897 if (slcmd->flags & SLMSGF_LINMAP) {
899 sl_local_munmap(slcmd);
904 * Reply and handle unblocking
906 sl->peer->backend_reply(sl->peer, slcmd, slmsg);
907 if (sl->wblocked && sl->cmdbytes < syslink_bufsize) {
908 sl->wblocked = 0; /* MP race ok here */
909 wakeup(&sl->wblocked);
913 * slmsg has already been dealt with, make sure error is
914 * 0 so we do not double-free it.
928 * Close a syslink descriptor.
930 * Disassociate the syslink from the file descriptor and disconnect from
935 slfileop_close(struct file *fp)
940 * Disassociate the file pointer. Take ownership of the ref on the
945 fp->f_ops = &badfileops;
949 * Shutdown both directions. The other side will not issue API
950 * calls to us after we've shutdown both directions.
952 shutdownsldesc(sl, SHUT_RDWR);
957 KKASSERT(sl->cmdbytes == 0);
958 KKASSERT(sl->repbytes == 0);
968 slfileop_stat (struct file *fp, struct stat *sb, struct ucred *cred)
975 slfileop_shutdown (struct file *fp, int how)
977 shutdownsldesc((struct sldesc *)fp->f_data, how);
983 slfileop_ioctl (struct file *fp, u_long cmd, caddr_t data,
984 struct ucred *cred, struct sysmsg *msg)
991 slfileop_kqfilter(struct file *fp, struct knote *kn)
996 /************************************************************************
997 * LOCAL MEMORY MAPPING *
998 ************************************************************************
1000 * This feature is currently not implemented
1006 sl_local_mmap(struct slmsg *slmsg, char *base, size_t len)
1008 return (EOPNOTSUPP);
1013 sl_local_munmap(struct slmsg *slmsg)
1022 sl_local_mmap(struct slmsg *slmsg, char *base, size_t len)
1024 struct vmspace *vms = curproc->p_vmspace;
1025 vm_offset_t addr = (vm_offset_t)base;
1027 /* XXX check user address range */
1028 error = vm_map_replace(
1030 (vm_offset_t)base, (vm_offset_t)base + len,
1031 slmsg->xio.xio_pages[0]->object,
1032 slmsg->xio.xio_pages[0]->pindex << PAGE_SHIFT,
1033 VM_PROT_READ|VM_PROT_WRITE,
1034 VM_PROT_READ|VM_PROT_WRITE,
1035 MAP_DISABLE_SYNCER);
1038 slmsg->flags |= SLMSGF_LINMAP;
1039 slmsg->vmbase = base;
1040 slmsg->vmsize = len;
1047 sl_local_munmap(struct slmsg *slmsg)
1049 if (slmsg->flags & SLMSGF_LINMAP) {
1050 vm_map_remove(&curproc->p_vmspace->vm_map,
1052 slmsg->vmbase + slcmd->vmsize);
1053 slmsg->flags &= ~SLMSGF_LINMAP;
1059 /************************************************************************
1060 * MESSAGE VALIDATION *
1061 ************************************************************************
1063 * Validate that the syslink message. Check that all headers and elements
1064 * conform. Correct the endian if necessary.
1066 * NOTE: If reverse endian needs to be corrected, SE_CMDF_UNTRANSLATED
1067 * is recursively flipped on all syslink_elm's in the message. As the
1068 * message traverses the mesh, multiple flips may occur. It is
1069 * up to the RPC protocol layer to correct opaque data payloads and
1070 * SE_CMDF_UNTRANSLATED prevents the protocol layer from misinterpreting
1071 * a command or reply element which has not been endian-corrected.
1075 syslink_validate_msg(struct syslink_msg *msg, int bytes)
1082 * The raw message must be properly-aligned.
1084 if (bytes & SL_ALIGNMASK)
1089 * The message must at least contain the msgid, bytes, and
1092 if (bytes < SL_MIN_PAD_SIZE)
1096 * Fix the endian if it is reversed.
1098 if (msg->sm_proto & SM_PROTO_ENDIAN_REV) {
1099 msg->sm_msgid = bswap64(msg->sm_msgid);
1100 msg->sm_sessid = bswap64(msg->sm_sessid);
1101 msg->sm_bytes = bswap16(msg->sm_bytes);
1102 msg->sm_proto = bswap16(msg->sm_proto);
1103 msg->sm_rlabel = bswap32(msg->sm_rlabel);
1104 if (msg->sm_proto & SM_PROTO_ENDIAN_REV)
1112 * Validate the contents. For PADs, the entire payload is
1113 * ignored and the minimum message size can be as small as
1116 if (msg->sm_proto == SMPROTO_PAD) {
1117 if (msg->sm_bytes < SL_MIN_PAD_SIZE ||
1118 msg->sm_bytes > bytes) {
1121 /* ignore the entire payload, it can be garbage */
1123 if (msg->sm_bytes < SL_MIN_MSG_SIZE ||
1124 msg->sm_bytes > bytes) {
1127 error = syslink_validate_elm(
1130 offsetof(struct syslink_msg,
1132 swapit, SL_MAXDEPTH);
1138 * The aligned payload size must be used to locate the
1139 * next syslink_msg in the buffer.
1141 aligned_reclen = SL_MSG_ALIGN(msg->sm_bytes);
1142 bytes -= aligned_reclen;
1143 msg = (void *)((char *)msg + aligned_reclen);
1150 syslink_validate_elm(struct syslink_elm *elm, sl_reclen_t bytes,
1151 int swapit, int depth)
1156 * If the buffer isn't big enough to fit the header, stop now!
1158 if (bytes < SL_MIN_ELM_SIZE)
1161 * All syslink_elm headers are recursively endian-adjusted. Opaque
1162 * data payloads are not.
1165 elm->se_cmd = bswap16(elm->se_cmd) ^ SE_CMDF_UNTRANSLATED;
1166 elm->se_bytes = bswap16(elm->se_bytes);
1167 elm->se_aux = bswap32(elm->se_aux);
1171 * Check element size requirements.
1173 if (elm->se_bytes < SL_MIN_ELM_SIZE || elm->se_bytes > bytes)
1177 * Recursively check structured payloads. A structured payload may
1178 * contain as few as 0 recursive elements.
1180 if (elm->se_cmd & SE_CMDF_STRUCTURED) {
1183 bytes -= SL_MIN_ELM_SIZE;
1186 if (syslink_validate_elm(elm, bytes, swapit, depth - 1))
1188 aligned_reclen = SL_MSG_ALIGN(elm->se_bytes);
1189 elm = (void *)((char *)elm + aligned_reclen);
1190 bytes -= aligned_reclen;
1196 /************************************************************************
1197 * BACKEND FUNCTIONS - USER DESCRIPTOR *
1198 ************************************************************************
1200 * Peer backend links are primarily used when userland creates a pair
1201 * of linked descriptors.
1205 * Do any required blocking / nbio handling for attempts to write to
1206 * a sldesc associated with a user descriptor.
1210 backend_wblocked_user(struct sldesc *sl, int nbio, sl_proto_t proto)
1213 int *bytesp = (proto & SM_PROTO_REPLY) ? &sl->repbytes : &sl->cmdbytes;
1216 * Block until sufficient data is drained by the target. It is
1217 * ok to have a MP race against cmdbytes.
1219 if (*bytesp >= syslink_bufsize) {
1220 spin_lock(&sl->spin);
1221 while (*bytesp >= syslink_bufsize) {
1222 if (sl->flags & SLF_WSHUTDOWN) {
1231 error = ssleep(&sl->wblocked, &sl->spin,
1232 PCATCH, "slwmsg", 0);
1236 spin_unlock(&sl->spin);
1242 * Unconditionally write a syslink message to the sldesc associated with
1243 * a user descriptor. Command messages are also placed in a red-black
1244 * tree so their DMA tag (if any) can be accessed and so they can be
1245 * linked to any reply message.
1249 backend_write_user(struct sldesc *sl, struct slmsg *slmsg)
1253 spin_lock(&sl->spin);
1254 if (sl->flags & SLF_RSHUTDOWN) {
1256 * Not accepting new messages
1259 } else if (slmsg->msg->sm_proto & SM_PROTO_REPLY) {
1263 TAILQ_INSERT_TAIL(&sl->inq, slmsg, tqnode);
1264 sl->repbytes += slmsg->maxsize;
1265 slmsg->flags |= SLMSGF_ONINQ;
1267 } else if (RB_INSERT(slmsg_rb_tree, &sl->reply_rb_root, slmsg)) {
1269 * Write a command, but there was a msgid collision when
1270 * we tried to insert it into the RB tree.
1275 * Write a command, successful insertion into the RB tree.
1277 TAILQ_INSERT_TAIL(&sl->inq, slmsg, tqnode);
1278 sl->cmdbytes += slmsg->maxsize;
1279 slmsg->flags |= SLMSGF_ONINQ;
1282 spin_unlock(&sl->spin);
1284 wakeup(&sl->rwaiters);
1289 * Our peer is replying a command we previously sent it back to us, along
1290 * with the reply message (if not NULL). We just queue the reply to
1291 * userland and free of the command.
1295 backend_reply_user(struct sldesc *sl, struct slmsg *slcmd, struct slmsg *slrep)
1301 spin_lock(&sl->spin);
1302 if ((sl->flags & SLF_RSHUTDOWN) == 0) {
1303 TAILQ_INSERT_TAIL(&sl->inq, slrep, tqnode);
1304 sl->repbytes += slrep->maxsize;
1309 spin_unlock(&sl->spin);
1311 sl->peer->backend_dispose(sl->peer, slrep);
1312 else if (sl->rwaiters)
1313 wakeup(&sl->rwaiters);
1319 backend_dispose_user(struct sldesc *sl, struct slmsg *slmsg)
1324 /************************************************************************
1325 * KERNEL DRIVER OR FILESYSTEM API *
1326 ************************************************************************
1331 * Create a user<->kernel link, returning the user descriptor in *fdp
1332 * and the kernel descriptor in *kslp. 0 is returned on success, and an
1333 * error code is returned on failure.
1336 syslink_ukbackend(int *pfd, struct sldesc **kslp)
1338 struct thread *td = curthread;
1339 struct filedesc *fdp = td->td_proc->p_fd;
1349 error = falloc(td->td_lwp, &fp, &fd);
1352 usl = allocsldesc(NULL);
1353 usl->backend_wblocked = backend_wblocked_user;
1354 usl->backend_write = backend_write_user;
1355 usl->backend_reply = backend_reply_user;
1356 usl->backend_dispose = backend_dispose_user;
1358 ksl = allocsldesc(usl->common);
1360 ksl->backend_wblocked = backend_wblocked_kern;
1361 ksl->backend_write = backend_write_kern;
1362 ksl->backend_reply = backend_reply_kern;
1363 ksl->backend_dispose = backend_dispose_kern;
1367 setsldescfp(usl, fp);
1368 fsetfd(fdp, fp, fd);
1377 * Assign a unique message id, issue a syslink message to userland,
1378 * and wait for a reply.
1381 syslink_kdomsg(struct sldesc *ksl, struct slmsg *slmsg)
1383 struct syslink_msg *msg;
1387 * Finish initializing slmsg and post it to the red-black tree for
1388 * reply matching. If the message id is already in use we return
1389 * EEXIST, giving the originator the chance to roll a new msgid.
1392 slmsg->msgsize = msg->sm_bytes;
1393 if ((error = syslink_validate_msg(msg, msg->sm_bytes)) != 0)
1395 msg->sm_msgid = allocsysid();
1398 * Issue the request and wait for a matching reply or failure,
1399 * then remove the message from the matching tree and return.
1401 error = ksl->peer->backend_write(ksl->peer, slmsg);
1402 spin_lock(&ksl->spin);
1404 while (slmsg->rep == NULL) {
1405 error = ssleep(slmsg, &ksl->spin, 0, "kwtmsg", 0);
1406 /* XXX ignore error for now */
1408 if (slmsg->rep == (struct slmsg *)-1) {
1412 error = slmsg->rep->msg->sm_head.se_aux;
1415 spin_unlock(&ksl->spin);
1420 * Similar to syslink_kdomsg but return immediately instead of
1421 * waiting for a reply. The kernel must supply a callback function
1422 * which will be made in the context of the user process replying
1426 syslink_ksendmsg(struct sldesc *ksl, struct slmsg *slmsg,
1427 void (*func)(struct slmsg *, void *, int), void *arg)
1429 struct syslink_msg *msg;
1433 * Finish initializing slmsg and post it to the red-black tree for
1434 * reply matching. If the message id is already in use we return
1435 * EEXIST, giving the originator the chance to roll a new msgid.
1438 slmsg->msgsize = msg->sm_bytes;
1439 slmsg->callback_func = func;
1440 slmsg->callback_data = arg;
1441 if ((error = syslink_validate_msg(msg, msg->sm_bytes)) != 0)
1443 msg->sm_msgid = allocsysid();
1446 * Issue the request. If no error occured the operation will be
1447 * in progress, otherwise the operation is considered to have failed
1448 * and the caller can deallocate the slmsg.
1450 error = ksl->peer->backend_write(ksl->peer, slmsg);
1455 syslink_kwaitmsg(struct sldesc *ksl, struct slmsg *slmsg)
1459 spin_lock(&ksl->spin);
1460 while (slmsg->rep == NULL) {
1461 error = ssleep(slmsg, &ksl->spin, 0, "kwtmsg", 0);
1462 /* XXX ignore error for now */
1464 if (slmsg->rep == (struct slmsg *)-1) {
1468 error = slmsg->rep->msg->sm_head.se_aux;
1470 spin_unlock(&ksl->spin);
1475 syslink_kallocmsg(void)
1477 return(objcache_get(sl_objcache_small, M_WAITOK));
1481 syslink_kfreemsg(struct sldesc *ksl, struct slmsg *slmsg)
1485 if ((rep = slmsg->rep) != NULL) {
1487 ksl->peer->backend_dispose(ksl->peer, rep);
1489 slmsg->callback_func = NULL;
1494 syslink_kshutdown(struct sldesc *ksl, int how)
1496 shutdownsldesc(ksl, how);
1500 syslink_kclose(struct sldesc *ksl)
1502 shutdownsldesc(ksl, SHUT_RDWR);
1507 * Associate a DMA buffer with a kernel syslink message prior to it
1508 * being sent to userland. The DMA buffer is set up from the point
1509 * of view of the target.
1512 syslink_kdmabuf_pages(struct slmsg *slmsg, struct vm_page **mbase, int npages)
1517 xflags = XIOF_VMLINEAR;
1518 if (slmsg->msg->sm_head.se_cmd & SE_CMDF_DMAR)
1519 xflags |= XIOF_READ | XIOF_WRITE;
1520 else if (slmsg->msg->sm_head.se_cmd & SE_CMDF_DMAW)
1521 xflags |= XIOF_READ;
1522 error = xio_init_pages(&slmsg->xio, mbase, npages, xflags);
1523 slmsg->flags |= SLMSGF_HASXIO;
1528 * Associate a DMA buffer with a kernel syslink message prior to it
1529 * being sent to userland. The DMA buffer is set up from the point
1530 * of view of the target.
1533 syslink_kdmabuf_data(struct slmsg *slmsg, char *base, int bytes)
1537 xflags = XIOF_VMLINEAR;
1538 if (slmsg->msg->sm_head.se_cmd & SE_CMDF_DMAR)
1539 xflags |= XIOF_READ | XIOF_WRITE;
1540 else if (slmsg->msg->sm_head.se_cmd & SE_CMDF_DMAW)
1541 xflags |= XIOF_READ;
1542 xio_init_kbuf(&slmsg->xio, base, bytes);
1543 slmsg->xio.xio_flags |= xflags;
1544 slmsg->flags |= SLMSGF_HASXIO;
1548 /************************************************************************
1549 * BACKEND FUNCTIONS FOR KERNEL API *
1550 ************************************************************************
1552 * These are the backend functions for a sldesc associated with a kernel
1557 * Our peer wants to write a syslink message to us and is asking us to
1558 * block if our input queue is full. We don't implement command reception
1559 * so don't block right now.
1563 backend_wblocked_kern(struct sldesc *ksl, int nbio, sl_proto_t proto)
1570 * Our peer is writing a request to the kernel. At the moment we do not
1575 backend_write_kern(struct sldesc *ksl, struct slmsg *slmsg)
1581 * Our peer wants to reply to a syslink message we sent it earlier. The
1582 * original command (that we passed to our peer), and the peer's reply
1583 * is specified. If the peer has failed slrep will be NULL.
1587 backend_reply_kern(struct sldesc *ksl, struct slmsg *slcmd, struct slmsg *slrep)
1591 spin_lock(&ksl->spin);
1592 if (slrep == NULL) {
1593 slcmd->rep = (struct slmsg *)-1;
1597 error = slrep->msg->sm_head.se_aux;
1599 spin_unlock(&ksl->spin);
1602 * Issue callback or wakeup a synchronous waiter.
1604 if (slcmd->callback_func) {
1605 slcmd->callback_func(slcmd, slcmd->callback_data, error);
1612 * Any reply messages we sent to our peer are returned to us for disposal.
1613 * Since we do not currently accept commands from our peer, there will not
1614 * be any replies returned to the peer to dispose of.
1618 backend_dispose_kern(struct sldesc *ksl, struct slmsg *slmsg)
1620 panic("backend_dispose_kern: kernel can't accept commands so it "
1621 "certainly did not reply to one!");