Home | History | Annotate | Line # | Download | only in puffs
puffs_msgif.c revision 1.34
      1 /*	$NetBSD: puffs_msgif.c,v 1.34 2007/05/01 12:18:40 pooka Exp $	*/
      2 
      3 /*
      4  * Copyright (c) 2005, 2006, 2007  Antti Kantee.  All Rights Reserved.
      5  *
      6  * Development of this software was supported by the
      7  * Google Summer of Code program and the Ulla Tuominen Foundation.
      8  * The Google SoC project was mentored by Bill Studenmund.
      9  *
     10  * Redistribution and use in source and binary forms, with or without
     11  * modification, are permitted provided that the following conditions
     12  * are met:
     13  * 1. Redistributions of source code must retain the above copyright
     14  *    notice, this list of conditions and the following disclaimer.
     15  * 2. Redistributions in binary form must reproduce the above copyright
     16  *    notice, this list of conditions and the following disclaimer in the
     17  *    documentation and/or other materials provided with the distribution.
     18  * 3. The name of the company nor the name of the author may be used to
     19  *    endorse or promote products derived from this software without specific
     20  *    prior written permission.
     21  *
     22  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS
     23  * OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
     24  * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
     25  * DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
     26  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
     27  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
     28  * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
     29  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
     30  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
     31  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
     32  * SUCH DAMAGE.
     33  */
     34 
     35 #include <sys/cdefs.h>
     36 __KERNEL_RCSID(0, "$NetBSD: puffs_msgif.c,v 1.34 2007/05/01 12:18:40 pooka Exp $");
     37 
     38 #include <sys/param.h>
     39 #include <sys/fstrans.h>
     40 #include <sys/malloc.h>
     41 #include <sys/mount.h>
     42 #include <sys/vnode.h>
     43 #include <sys/lock.h>
     44 
     45 #include <fs/puffs/puffs_msgif.h>
     46 #include <fs/puffs/puffs_sys.h>
     47 
     48 /*
     49  * waitq data structures
     50  */
     51 
     52 /*
     53  * While a request is going to userspace, park the caller within the
     54  * kernel.  This is the kernel counterpart of "struct puffs_req".
     55  */
     56 struct puffs_park {
     57 	struct puffs_req	*park_preq;	/* req followed by buf	*/
     58 	uint64_t		park_id;	/* duplicate of preq_id */
     59 
     60 	size_t			park_copylen;	/* userspace copylength	*/
     61 	size_t			park_maxlen;	/* max size in comeback */
     62 
     63 	parkdone_fn		park_done;
     64 	void			*park_donearg;
     65 
     66 	int			park_flags;
     67 	int			park_refcount;
     68 
     69 	kcondvar_t		park_cv;
     70 	kmutex_t		park_mtx;
     71 
     72 	TAILQ_ENTRY(puffs_park) park_entries;
     73 };
     74 #define PARKFLAG_WAITERGONE	0x01
     75 #define PARKFLAG_DONE		0x02
     76 #define PARKFLAG_ONQUEUE1	0x04
     77 #define PARKFLAG_ONQUEUE2	0x08
     78 #define PARKFLAG_CALL		0x10
     79 #define PARKFLAG_WANTREPLY	0x20
     80 
     81 static struct pool_cache parkpc;
     82 static struct pool parkpool;
     83 
     84 static int
     85 makepark(void *arg, void *obj, int flags)
     86 {
     87 	struct puffs_park *park = obj;
     88 
     89 	mutex_init(&park->park_mtx, MUTEX_DEFAULT, IPL_NONE);
     90 	cv_init(&park->park_cv, "puffsrpl");
     91 
     92 	return 0;
     93 }
     94 
     95 static void
     96 nukepark(void *arg, void *obj)
     97 {
     98 	struct puffs_park *park = obj;
     99 
    100 	cv_destroy(&park->park_cv);
    101 	mutex_destroy(&park->park_mtx);
    102 }
    103 
    104 void
    105 puffs_msgif_init()
    106 {
    107 
    108 	pool_init(&parkpool, sizeof(struct puffs_park), 0, 0, 0,
    109 	    "puffprkl", &pool_allocator_nointr, IPL_NONE);
    110 	pool_cache_init(&parkpc, &parkpool, makepark, nukepark, NULL);
    111 }
    112 
    113 void
    114 puffs_msgif_destroy()
    115 {
    116 
    117 	pool_cache_destroy(&parkpc);
    118 	pool_destroy(&parkpool);
    119 }
    120 
    121 void *
    122 puffs_park_alloc(int waitok)
    123 {
    124 	struct puffs_park *park;
    125 
    126 	park = pool_cache_get(&parkpc, waitok ? PR_WAITOK : PR_NOWAIT);
    127 	if (park) {
    128 		park->park_refcount = 1;
    129 		mutex_enter(&park->park_mtx);
    130 	}
    131 
    132 	return park;
    133 }
    134 
    135 static void
    136 puffs_park_reference(struct puffs_park *park)
    137 {
    138 
    139 	mutex_enter(&park->park_mtx);
    140 	park->park_refcount++;
    141 }
    142 
    143 void
    144 puffs_park_release(void *arg, int fullnuke)
    145 {
    146 	struct puffs_park *park = arg;
    147 
    148 	KASSERT(mutex_owned(&park->park_mtx));
    149 	--park->park_refcount;
    150 
    151 	mutex_exit(&park->park_mtx);
    152 	if (park->park_refcount == 0 || fullnuke)
    153 		pool_cache_put(&parkpc, park);
    154 }
    155 
    156 #ifdef PUFFSDEBUG
    157 static void
    158 parkdump(struct puffs_park *park)
    159 {
    160 
    161 	DPRINTF(("park %p, preq %p, id %" PRIu64 "\n"
    162 	    "\tcopy %zu, max %zu - done: %p/%p\n"
    163 	    "\tflags 0x%08x, refcount %d, cv/mtx: %p/%p\n",
    164 	    park, park->park_preq, park->park_id,
    165 	    park->park_copylen, park->park_maxlen,
    166 	    park->park_done, park->park_donearg,
    167 	    park->park_flags, park->park_refcount,
    168 	    &park->park_cv, &park->park_mtx));
    169 }
    170 
    171 static void
    172 parkqdump(struct puffs_wq *q, int dumpall)
    173 {
    174 	struct puffs_park *park;
    175 	int total = 0;
    176 
    177 	TAILQ_FOREACH(park, q, park_entries) {
    178 		if (dumpall)
    179 			parkdump(park);
    180 		total++;
    181 	}
    182 	DPRINTF(("puffs waitqueue at %p dumped, %d total\n", q, total));
    183 
    184 }
    185 #endif /* PUFFSDEBUG */
    186 
    187 /*
    188  * Converts a non-FAF op to a FAF.  This simply involves making copies
    189  * of the park and request structures and tagging the request as a FAF.
    190  * It is safe to block here, since the original op is not a FAF.
    191  */
    192 static void
    193 puffs_reqtofaf(struct puffs_park *park)
    194 {
    195 	struct puffs_req *newpreq;
    196 
    197 	KASSERT((park->park_preq->preq_opclass & PUFFSOPFLAG_FAF) == 0);
    198 
    199 	MALLOC(newpreq, struct puffs_req *, park->park_copylen,
    200 	    M_PUFFS, M_ZERO | M_WAITOK);
    201 
    202 	memcpy(newpreq, park->park_preq, park->park_copylen);
    203 
    204 	park->park_preq = newpreq;
    205 	park->park_preq->preq_opclass |= PUFFSOPFLAG_FAF;
    206 }
    207 
    208 
    209 /*
    210  * kernel-user-kernel waitqueues
    211  */
    212 
    213 static int touser(struct puffs_mount *, struct puffs_park *, uint64_t,
    214 		  struct vnode *, struct vnode *);
    215 
    216 uint64_t
    217 puffs_getreqid(struct puffs_mount *pmp)
    218 {
    219 	uint64_t rv;
    220 
    221 	mutex_enter(&pmp->pmp_lock);
    222 	rv = pmp->pmp_nextreq++;
    223 	mutex_exit(&pmp->pmp_lock);
    224 
    225 	return rv;
    226 }
    227 
    228 /* vfs request */
    229 int
    230 puffs_vfstouser(struct puffs_mount *pmp, int optype, void *kbuf, size_t buflen)
    231 {
    232 	struct puffs_park *park;
    233 
    234 	park = puffs_park_alloc(1);
    235 	park->park_preq = kbuf;
    236 
    237 	park->park_preq->preq_opclass = PUFFSOP_VFS;
    238 	park->park_preq->preq_optype = optype;
    239 
    240 	park->park_maxlen = park->park_copylen = buflen;
    241 	park->park_flags = 0;
    242 
    243 	return touser(pmp, park, puffs_getreqid(pmp), NULL, NULL);
    244 }
    245 
    246 void
    247 puffs_suspendtouser(struct puffs_mount *pmp, int status)
    248 {
    249 	struct puffs_vfsreq_suspend *pvfsr_susp;
    250 	struct puffs_park *park;
    251 
    252 	pvfsr_susp = malloc(sizeof(struct puffs_vfsreq_suspend),
    253 	    M_PUFFS, M_WAITOK | M_ZERO);
    254 	park = puffs_park_alloc(1);
    255 
    256 	pvfsr_susp->pvfsr_status = status;
    257 	park->park_preq = (struct puffs_req *)pvfsr_susp;
    258 
    259 	park->park_preq->preq_opclass = PUFFSOP_VFS | PUFFSOPFLAG_FAF;
    260 	park->park_preq->preq_optype = PUFFS_VFS_SUSPEND;
    261 
    262 	park->park_maxlen = park->park_copylen
    263 	    = sizeof(struct puffs_vfsreq_suspend);
    264 	park->park_flags = 0;
    265 
    266 	(void)touser(pmp, park, 0, NULL, NULL);
    267 }
    268 
    269 /*
    270  * vnode level request
    271  */
    272 int
    273 puffs_vntouser(struct puffs_mount *pmp, int optype,
    274 	void *kbuf, size_t buflen, size_t maxdelta, void *cookie,
    275 	struct vnode *vp1, struct vnode *vp2)
    276 {
    277 	struct puffs_park *park;
    278 
    279 	park = puffs_park_alloc(1);
    280 	park->park_preq = kbuf;
    281 
    282 	park->park_preq->preq_opclass = PUFFSOP_VN;
    283 	park->park_preq->preq_optype = optype;
    284 	park->park_preq->preq_cookie = cookie;
    285 
    286 	park->park_copylen = buflen;
    287 	park->park_maxlen = buflen + maxdelta;
    288 	park->park_flags = 0;
    289 
    290 	return touser(pmp, park, puffs_getreqid(pmp), vp1, vp2);
    291 }
    292 
    293 /*
    294  * vnode level request, caller-controller req id
    295  */
    296 int
    297 puffs_vntouser_req(struct puffs_mount *pmp, int optype,
    298 	void *kbuf, size_t buflen, size_t maxdelta, void *cookie,
    299 	uint64_t reqid, struct vnode *vp1, struct vnode *vp2)
    300 {
    301 	struct puffs_park *park;
    302 
    303 	park = puffs_park_alloc(1);
    304 	park->park_preq = kbuf;
    305 
    306 	park->park_preq->preq_opclass = PUFFSOP_VN;
    307 	park->park_preq->preq_optype = optype;
    308 	park->park_preq->preq_cookie = cookie;
    309 
    310 	park->park_copylen = buflen;
    311 	park->park_maxlen = buflen + maxdelta;
    312 	park->park_flags = 0;
    313 
    314 	return touser(pmp, park, reqid, vp1, vp2);
    315 }
    316 
    317 void
    318 puffs_vntouser_call(struct puffs_mount *pmp, int optype,
    319 	void *kbuf, size_t buflen, size_t maxdelta, void *cookie,
    320 	parkdone_fn donefn, void *donearg,
    321 	struct vnode *vp1, struct vnode *vp2)
    322 {
    323 	struct puffs_park *park;
    324 
    325 	park = puffs_park_alloc(1);
    326 	park->park_preq = kbuf;
    327 
    328 	park->park_preq->preq_opclass = PUFFSOP_VN;
    329 	park->park_preq->preq_optype = optype;
    330 	park->park_preq->preq_cookie = cookie;
    331 
    332 	park->park_copylen = buflen;
    333 	park->park_maxlen = buflen + maxdelta;
    334 	park->park_done = donefn;
    335 	park->park_donearg = donearg;
    336 	park->park_flags = PARKFLAG_CALL;
    337 
    338 	(void) touser(pmp, park, puffs_getreqid(pmp), vp1, vp2);
    339 }
    340 
    341 /*
    342  * Notice: kbuf will be free'd later.  I must be allocated from the
    343  * kernel heap and it's ownership is shifted to this function from
    344  * now on, i.e. the caller is not allowed to use it anymore!
    345  */
    346 void
    347 puffs_vntouser_faf(struct puffs_mount *pmp, int optype,
    348 	void *kbuf, size_t buflen, void *cookie)
    349 {
    350 	struct puffs_park *park;
    351 
    352 	/* XXX: is it allowable to sleep here? */
    353 	park = puffs_park_alloc(0);
    354 	if (park == NULL)
    355 		return; /* 2bad */
    356 
    357 	park->park_preq = kbuf;
    358 
    359 	park->park_preq->preq_opclass = PUFFSOP_VN | PUFFSOPFLAG_FAF;
    360 	park->park_preq->preq_optype = optype;
    361 	park->park_preq->preq_cookie = cookie;
    362 
    363 	park->park_maxlen = park->park_copylen = buflen;
    364 	park->park_flags = 0;
    365 
    366 	(void)touser(pmp, park, 0, NULL, NULL);
    367 }
    368 
    369 void
    370 puffs_cacheop(struct puffs_mount *pmp, struct puffs_park *park,
    371 	struct puffs_cacheinfo *pcinfo, size_t pcilen, void *cookie)
    372 {
    373 
    374 	park->park_preq = (struct puffs_req *)pcinfo;
    375 	park->park_preq->preq_opclass = PUFFSOP_CACHE | PUFFSOPFLAG_FAF;
    376 	park->park_preq->preq_optype = PCACHE_TYPE_WRITE; /* XXX */
    377 	park->park_preq->preq_cookie = cookie;
    378 
    379 	park->park_maxlen = park->park_copylen = pcilen;
    380 	park->park_flags = 0;
    381 
    382 	(void)touser(pmp, park, 0, NULL, NULL);
    383 }
    384 
    385 /*
    386  * Wait for the userspace ping-pong game in calling process context.
    387  *
    388  * This unlocks vnodes if they are supplied.  vp1 is the vnode
    389  * before in the locking order, i.e. the one which must be locked
    390  * before accessing vp2.  This is done here so that operations are
    391  * already ordered in the queue when vnodes are unlocked (I'm not
    392  * sure if that's really necessary, but it can't hurt).  Okok, maybe
    393  * there's a slight ugly-factor also, but let's not worry about that.
    394  */
    395 static int
    396 touser(struct puffs_mount *pmp, struct puffs_park *park, uint64_t reqid,
    397 	struct vnode *vp1, struct vnode *vp2)
    398 {
    399 	struct lwp *l = curlwp;
    400 	struct mount *mp;
    401 	struct puffs_req *preq;
    402 	int rv = 0;
    403 
    404 	mp = PMPTOMP(pmp);
    405 	preq = park->park_preq;
    406 	preq->preq_id = park->park_id = reqid;
    407 	preq->preq_buflen = ALIGN(park->park_maxlen);
    408 
    409 	if (PUFFSOP_WANTREPLY(preq->preq_opclass))
    410 		park->park_flags |= PARKFLAG_WANTREPLY;
    411 
    412 	/*
    413 	 * To support PCATCH, yet another movie: check if there are signals
    414 	 * pending and we are issueing a non-FAF.  If so, return an error
    415 	 * directly UNLESS we are issueing INACTIVE.  In that case, convert
    416 	 * it to a FAF, fire off to the file server and return an error.
    417 	 * Yes, this is bordering disgusting.  Barfbags are on me.
    418 	 */
    419 	if ((park->park_flags & PARKFLAG_WANTREPLY)
    420 	   && (park->park_flags & PARKFLAG_CALL) == 0
    421 	   && (l->l_flag & LW_PENDSIG) != 0 && sigispending(l, 0)) {
    422 		if (PUFFSOP_OPCLASS(preq->preq_opclass) == PUFFSOP_VN
    423 		    && preq->preq_optype == PUFFS_VN_INACTIVE) {
    424 			puffs_reqtofaf(park);
    425 			DPRINTF(("puffs touser: converted to FAF %p\n", park));
    426 			rv = EINTR;
    427 		} else {
    428 			puffs_park_release(park, 0);
    429 			return EINTR;
    430 		}
    431 	}
    432 
    433 	/*
    434 	 * test for suspension lock.
    435 	 *
    436 	 * Note that we *DO NOT* keep the lock, since that might block
    437 	 * lock acquiring PLUS it would give userlandia control over
    438 	 * the lock.  The operation queue enforces a strict ordering:
    439 	 * when the fs server gets in the op stream, it knows things
    440 	 * are in order.  The kernel locks can't guarantee that for
    441 	 * userspace, in any case.
    442 	 *
    443 	 * BUT: this presents a problem for ops which have a consistency
    444 	 * clause based on more than one operation.  Unfortunately such
    445 	 * operations (read, write) do not reliably work yet.
    446 	 *
    447 	 * Ya, Ya, it's wrong wong wrong, me be fixink this someday.
    448 	 *
    449 	 * XXX: and there is one more problem.  We sometimes need to
    450 	 * take a lazy lock in case the fs is suspending and we are
    451 	 * executing as the fs server context.  This might happen
    452 	 * e.g. in the case that the user server triggers a reclaim
    453 	 * in the kernel while the fs is suspending.  It's not a very
    454 	 * likely event, but it needs to be fixed some day.
    455 	 */
    456 
    457 	/*
    458 	 * MOREXXX: once PUFFS_WCACHEINFO is enabled, we can't take
    459 	 * the mutex here, since getpages() might be called locked.
    460 	 */
    461 	fstrans_start(mp, FSTRANS_NORMAL);
    462 	mutex_enter(&pmp->pmp_lock);
    463 	fstrans_done(mp);
    464 
    465 	if (pmp->pmp_status != PUFFSTAT_RUNNING) {
    466 		mutex_exit(&pmp->pmp_lock);
    467 		puffs_park_release(park, 0);
    468 		return ENXIO;
    469 	}
    470 
    471 #ifdef PUFFSDEBUG
    472 	parkqdump(&pmp->pmp_req_touser, puffsdebug > 1);
    473 	parkqdump(&pmp->pmp_req_replywait, puffsdebug > 1);
    474 #endif
    475 
    476 	TAILQ_INSERT_TAIL(&pmp->pmp_req_touser, park, park_entries);
    477 	park->park_flags |= PARKFLAG_ONQUEUE1;
    478 	puffs_mp_reference(pmp);
    479 	pmp->pmp_req_touser_count++;
    480 	mutex_exit(&pmp->pmp_lock);
    481 
    482 #if 0
    483 	/*
    484 	 * Don't do unlock-relock dance yet.  There are a couple of
    485 	 * unsolved issues with it.  If we don't unlock, we can have
    486 	 * processes wanting vn_lock in case userspace hangs.  But
    487 	 * that can be "solved" by killing the userspace process.  It
    488 	 * would of course be nicer to have antilocking in the userspace
    489 	 * interface protocol itself.. your patience will be rewarded.
    490 	 */
    491 	/* unlock */
    492 	if (vp2)
    493 		VOP_UNLOCK(vp2, 0);
    494 	if (vp1)
    495 		VOP_UNLOCK(vp1, 0);
    496 #endif
    497 
    498 	DPRINTF(("touser: req %" PRIu64 ", preq: %p, park: %p, "
    499 	    "c/t: 0x%x/0x%x, f: 0x%x\n", preq->preq_id, preq, park,
    500 	    preq->preq_opclass, preq->preq_optype, park->park_flags));
    501 
    502 	cv_broadcast(&pmp->pmp_req_waiter_cv);
    503 	selnotify(pmp->pmp_sel, 0);
    504 
    505 	if ((park->park_flags & PARKFLAG_WANTREPLY)
    506 	    && (park->park_flags & PARKFLAG_CALL) == 0) {
    507 		int error;
    508 
    509 		error = cv_wait_sig(&park->park_cv, &park->park_mtx);
    510 		if (error) {
    511 			park->park_flags |= PARKFLAG_WAITERGONE;
    512 			if (park->park_flags & PARKFLAG_DONE) {
    513 				rv = preq->preq_rv;
    514 				puffs_park_release(park, 0);
    515 			} else {
    516 				/*
    517 				 * ok, we marked it as going away, but
    518 				 * still need to do queue ops.  take locks
    519 				 * in correct order.
    520 				 *
    521 				 * We don't want to release our reference
    522 				 * if it's on replywait queue to avoid error
    523 				 * to file server.  putop() code will DTRT.
    524 				 */
    525 				KASSERT(park->park_flags &
    526 				    (PARKFLAG_ONQUEUE1 | PARKFLAG_ONQUEUE2));
    527 				mutex_exit(&park->park_mtx);
    528 
    529 				mutex_enter(&pmp->pmp_lock);
    530 				mutex_enter(&park->park_mtx);
    531 				if (park->park_flags & PARKFLAG_ONQUEUE1)
    532 					TAILQ_REMOVE(&pmp->pmp_req_touser,
    533 					    park, park_entries);
    534 				park->park_flags &= ~PARKFLAG_ONQUEUE1;
    535 				pmp->pmp_req_touser_count--;
    536 				if ((park->park_flags & PARKFLAG_ONQUEUE2) == 0)
    537 					puffs_park_release(park, 0);
    538 				else
    539 					mutex_exit(&park->park_mtx);
    540 				mutex_exit(&pmp->pmp_lock);
    541 
    542 				rv = error;
    543 			}
    544 		} else {
    545 			rv = preq->preq_rv;
    546 			puffs_park_release(park, 0);
    547 		}
    548 
    549 		/*
    550 		 * retake the lock and release.  This makes sure (haha,
    551 		 * I'm humorous) that we don't process the same vnode in
    552 		 * multiple threads due to the locks hacks we have in
    553 		 * puffs_lock().  In reality this is well protected by
    554 		 * the biglock, but once that's gone, well, hopefully
    555 		 * this will be fixed for real.  (and when you read this
    556 		 * comment in 2017 and subsequently barf, my condolences ;).
    557 		 */
    558 		if (rv == 0 && !fstrans_is_owner(mp)) {
    559 			fstrans_start(mp, FSTRANS_NORMAL);
    560 			fstrans_done(mp);
    561 		}
    562 	} else {
    563 		mutex_exit(&park->park_mtx);
    564 	}
    565 
    566 #if 0
    567 	/* relock */
    568 	if (vp1)
    569 		KASSERT(vn_lock(vp1, LK_EXCLUSIVE | LK_RETRY) == 0);
    570 	if (vp2)
    571 		KASSERT(vn_lock(vp2, LK_EXCLUSIVE | LK_RETRY) == 0);
    572 #endif
    573 
    574 	mutex_enter(&pmp->pmp_lock);
    575 	puffs_mp_release(pmp);
    576 	mutex_exit(&pmp->pmp_lock);
    577 
    578 	return rv;
    579 }
    580 
    581 
    582 /*
    583  * getop: scan through queued requests until:
    584  *  1) max number of requests satisfied
    585  *     OR
    586  *  2) buffer runs out of space
    587  *     OR
    588  *  3) nonblocking is set AND there are no operations available
    589  *     OR
    590  *  4) at least one operation was transferred AND there are no more waiting
    591  */
    592 int
    593 puffs_getop(struct puffs_mount *pmp, struct puffs_reqh_get *phg, int nonblock)
    594 {
    595 	struct puffs_park *park;
    596 	struct puffs_req *preq;
    597 	uint8_t *bufpos;
    598 	int error, donesome;
    599 
    600 	donesome = error = 0;
    601 	bufpos = phg->phg_buf;
    602 
    603 	mutex_enter(&pmp->pmp_lock);
    604 	while (phg->phg_nops == 0 || donesome != phg->phg_nops) {
    605  again:
    606 		if (pmp->pmp_status != PUFFSTAT_RUNNING) {
    607 			/* if we got some, they don't really matter anymore */
    608 			error = ENXIO;
    609 			goto out;
    610 		}
    611 		if (TAILQ_EMPTY(&pmp->pmp_req_touser)) {
    612 			if (donesome)
    613 				goto out;
    614 
    615 			if (nonblock) {
    616 				error = EWOULDBLOCK;
    617 				goto out;
    618 			}
    619 
    620 			error = cv_wait_sig(&pmp->pmp_req_waiter_cv,
    621 			    &pmp->pmp_lock);
    622 			if (error)
    623 				goto out;
    624 			else
    625 				goto again;
    626 		}
    627 
    628 		park = TAILQ_FIRST(&pmp->pmp_req_touser);
    629 		puffs_park_reference(park);
    630 
    631 		/* If it's a goner, don't process any furher */
    632 		if (park->park_flags & PARKFLAG_WAITERGONE) {
    633 			puffs_park_release(park, 0);
    634 			continue;
    635 		}
    636 
    637 		preq = park->park_preq;
    638 		if (phg->phg_buflen < preq->preq_buflen) {
    639 			if (!donesome)
    640 				error = E2BIG;
    641 			puffs_park_release(park, 0);
    642 			goto out;
    643 		}
    644 
    645 		TAILQ_REMOVE(&pmp->pmp_req_touser, park, park_entries);
    646 		KASSERT(park->park_flags & PARKFLAG_ONQUEUE1);
    647 		park->park_flags &= ~PARKFLAG_ONQUEUE1;
    648 		pmp->pmp_req_touser_count--;
    649 		KASSERT(pmp->pmp_req_touser_count >= 0);
    650 		mutex_exit(&pmp->pmp_lock);
    651 
    652 		DPRINTF(("puffsgetop: get op %" PRIu64 " (%d.), from %p "
    653 		    "len %zu (buflen %zu), target %p\n", preq->preq_id,
    654 		    donesome, preq, park->park_copylen, preq->preq_buflen,
    655 		    bufpos));
    656 
    657 		if ((error = copyout(preq, bufpos, park->park_copylen)) != 0) {
    658 			DPRINTF(("puffs_getop: copyout failed: %d\n", error));
    659 			/*
    660 			 * ok, user server is probably trying to cheat.
    661 			 * stuff op back & return error to user.  We need
    662 			 * to take locks in the correct order.
    663 			 */
    664 			mutex_exit(&park->park_mtx);
    665 
    666 			/*
    667 			 * XXX: ONQUEUE1 | ONQUEUE2 invariant doesn't
    668 			 * hold here
    669 			 */
    670 
    671 			mutex_enter(&pmp->pmp_lock);
    672 			mutex_enter(&park->park_mtx);
    673 			if ((park->park_flags & PARKFLAG_WAITERGONE) == 0) {
    674 				 TAILQ_INSERT_HEAD(&pmp->pmp_req_touser, park,
    675 				     park_entries);
    676 				 park->park_flags |= PARKFLAG_ONQUEUE1;
    677 				 pmp->pmp_req_touser_count++;
    678 			}
    679 
    680 			if (donesome)
    681 				error = 0;
    682 			puffs_park_release(park, 0);
    683 			goto out;
    684 		}
    685 		bufpos += preq->preq_buflen;
    686 		phg->phg_buflen -= preq->preq_buflen;
    687 		donesome++;
    688 
    689 		/* XXXfixme: taking this lock in the wrong order */
    690 		mutex_enter(&pmp->pmp_lock);
    691 
    692 		if (park->park_flags & PARKFLAG_WANTREPLY) {
    693 			TAILQ_INSERT_TAIL(&pmp->pmp_req_replywait, park,
    694 			    park_entries);
    695 			park->park_flags |= PARKFLAG_ONQUEUE2;
    696 			puffs_park_release(park, 0);
    697 		} else {
    698 			free(preq, M_PUFFS);
    699 			puffs_park_release(park, 1);
    700 		}
    701 	}
    702 
    703  out:
    704 	phg->phg_more = pmp->pmp_req_touser_count;
    705 	mutex_exit(&pmp->pmp_lock);
    706 
    707 	phg->phg_nops = donesome;
    708 
    709 	return error;
    710 }
    711 
    712 int
    713 puffs_putop(struct puffs_mount *pmp, struct puffs_reqh_put *php)
    714 {
    715 	struct puffs_park *park;
    716 	struct puffs_req tmpreq;
    717 	struct puffs_req *nextpreq;
    718 	void *userbuf;
    719 	uint64_t id;
    720 	size_t reqlen;
    721 	int donesome, error, wgone, release;
    722 
    723 	donesome = error = wgone = 0;
    724 
    725 	id = php->php_id;
    726 	userbuf = php->php_buf;
    727 	reqlen = php->php_buflen;
    728 
    729 	mutex_enter(&pmp->pmp_lock);
    730 	while (donesome != php->php_nops) {
    731 		release = 0;
    732 #ifdef PUFFSDEBUG
    733 		DPRINTF(("puffsputop: searching for %" PRIu64 ", ubuf: %p, "
    734 		    "len %zu\n", id, userbuf, reqlen));
    735 #endif
    736 		TAILQ_FOREACH(park, &pmp->pmp_req_replywait, park_entries) {
    737 			if (park->park_id == id)
    738 				break;
    739 		}
    740 
    741 		if (park == NULL) {
    742 			DPRINTF(("puffsputop: no request: %" PRIu64 "\n", id));
    743 			error = EINVAL;
    744 			break;
    745 		}
    746 
    747 		puffs_park_reference(park);
    748 		if (reqlen == 0 || reqlen > park->park_maxlen) {
    749 			DPRINTF(("puffsputop: invalid buffer length: "
    750 			    "%zu\n", reqlen));
    751 			error = E2BIG;
    752 			puffs_park_release(park, 0);
    753 			break;
    754 		}
    755 		wgone = park->park_flags & PARKFLAG_WAITERGONE;
    756 
    757 		/* check if it's still on the queue after acquiring lock */
    758 		if (park->park_flags & PARKFLAG_ONQUEUE2) {
    759 			TAILQ_REMOVE(&pmp->pmp_req_replywait, park,
    760 			    park_entries);
    761 			park->park_flags &= ~PARKFLAG_ONQUEUE2;
    762 		}
    763 
    764 		mutex_exit(&pmp->pmp_lock);
    765 
    766 		/*
    767 		 * If the caller has gone south, go to next, collect
    768 		 * $200 and free the structure there instead of wakeup.
    769 		 * We also need to copyin the header info.  Flag structure
    770 		 * release to mode total and utter destruction.
    771 		 */
    772 		if (wgone) {
    773 			DPRINTF(("puffs_putop: bad service - waiter gone for "
    774 			    "park %p\n", park));
    775 			error = copyin(userbuf, &tmpreq,
    776 			    sizeof(struct puffs_req));
    777 			release = 1;
    778 			if (error)
    779 				goto loopout;
    780 			nextpreq = &tmpreq;
    781 			goto next;
    782 		}
    783 
    784 		DPRINTF(("puffsputpop: copyin from %p to %p, len %zu\n",
    785 		    userbuf, park->park_preq, reqlen));
    786 		error = copyin(userbuf, park->park_preq, reqlen);
    787 		if (error)
    788 			goto loopout;
    789 		nextpreq = park->park_preq;
    790 
    791  next:
    792 		/* all's well, prepare for next op */
    793 		id = nextpreq->preq_id;
    794 		reqlen = nextpreq->preq_buflen;
    795 		userbuf = nextpreq->preq_nextbuf;
    796 		donesome++;
    797 
    798  loopout:
    799 		if (error && !wgone)
    800 			park->park_preq->preq_rv = error;
    801 
    802 		if (park->park_flags & PARKFLAG_CALL) {
    803 			park->park_done(park->park_preq, park->park_donearg);
    804 			release = 1;
    805 		}
    806 
    807 		if (!wgone) {
    808 			DPRINTF(("puffs_putop: flagging done for "
    809 			    "park %p\n", park));
    810 
    811 			cv_signal(&park->park_cv);
    812 		}
    813 		park->park_flags |= PARKFLAG_DONE;
    814 		puffs_park_release(park, release);
    815 
    816 		mutex_enter(&pmp->pmp_lock);
    817 		if (error)
    818 			break;
    819 		wgone = 0;
    820 	}
    821 
    822 	mutex_exit(&pmp->pmp_lock);
    823 	php->php_nops -= donesome;
    824 
    825 	return error;
    826 }
    827 
    828 /*
    829  * We're dead, kaput, RIP, slightly more than merely pining for the
    830  * fjords, belly-up, fallen, lifeless, finished, expired, gone to meet
    831  * our maker, ceased to be, etcetc.  YASD.  It's a dead FS!
    832  *
    833  * Caller must hold puffs mutex.
    834  */
    835 void
    836 puffs_userdead(struct puffs_mount *pmp)
    837 {
    838 	struct puffs_park *park, *park_next;
    839 
    840 	/*
    841 	 * Mark filesystem status as dying so that operations don't
    842 	 * attempt to march to userspace any longer.
    843 	 */
    844 	pmp->pmp_status = PUFFSTAT_DYING;
    845 
    846 	/* signal waiters on REQUEST TO file server queue */
    847 	for (park = TAILQ_FIRST(&pmp->pmp_req_touser); park; park = park_next) {
    848 		uint8_t opclass;
    849 
    850 		puffs_park_reference(park);
    851 		park_next = TAILQ_NEXT(park, park_entries);
    852 
    853 		KASSERT(park->park_flags & PARKFLAG_ONQUEUE1);
    854 		TAILQ_REMOVE(&pmp->pmp_req_touser, park, park_entries);
    855 		park->park_flags &= ~PARKFLAG_ONQUEUE1;
    856 		pmp->pmp_req_touser_count--;
    857 
    858 		/*
    859 		 * If the waiter is gone, we may *NOT* access preq anymore.
    860 		 */
    861 		if (park->park_flags & PARKFLAG_WAITERGONE) {
    862 			KASSERT((park->park_flags & PARKFLAG_CALL) == 0);
    863 			KASSERT(park->park_flags & PARKFLAG_WANTREPLY);
    864 			puffs_park_release(park, 0);
    865 		} else {
    866 			opclass = park->park_preq->preq_opclass;
    867 			park->park_preq->preq_rv = ENXIO;
    868 
    869 			if (park->park_flags & PARKFLAG_CALL) {
    870 				park->park_done(park->park_preq,
    871 				    park->park_donearg);
    872 				puffs_park_release(park, 1);
    873 			} else if ((park->park_flags & PARKFLAG_WANTREPLY)==0) {
    874 				free(park->park_preq, M_PUFFS);
    875 				puffs_park_release(park, 1);
    876 			} else {
    877 				park->park_preq->preq_rv = ENXIO;
    878 				cv_signal(&park->park_cv);
    879 				puffs_park_release(park, 0);
    880 			}
    881 		}
    882 	}
    883 
    884 	/* signal waiters on RESPONSE FROM file server queue */
    885 	for (park=TAILQ_FIRST(&pmp->pmp_req_replywait); park; park=park_next) {
    886 		puffs_park_reference(park);
    887 		park_next = TAILQ_NEXT(park, park_entries);
    888 
    889 		KASSERT(park->park_flags & PARKFLAG_ONQUEUE2);
    890 		KASSERT(park->park_flags & PARKFLAG_WANTREPLY);
    891 
    892 		TAILQ_REMOVE(&pmp->pmp_req_replywait, park, park_entries);
    893 		park->park_flags &= ~PARKFLAG_ONQUEUE2;
    894 
    895 		/*
    896 		 * If the waiter is gone, we may *NOT* access preq anymore.
    897 		 */
    898 		if (park->park_flags & PARKFLAG_WAITERGONE) {
    899 			KASSERT((park->park_flags & PARKFLAG_CALL) == 0);
    900 			puffs_park_release(park, 0);
    901 		} else {
    902 			park->park_preq->preq_rv = ENXIO;
    903 			if (park->park_flags & PARKFLAG_CALL) {
    904 				park->park_done(park->park_preq,
    905 				    park->park_donearg);
    906 				puffs_park_release(park, 1);
    907 			} else {
    908 				cv_signal(&park->park_cv);
    909 				puffs_park_release(park, 0);
    910 			}
    911 		}
    912 	}
    913 }
    914 
    915 /* this is probably going to die away at some point? */
    916 /*
    917  * XXX: currently bitrotted
    918  */
    919 #if 0
    920 static int
    921 puffssizeop(struct puffs_mount *pmp, struct puffs_sizeop *psop_user)
    922 {
    923 	struct puffs_sizepark *pspark;
    924 	void *kernbuf;
    925 	size_t copylen;
    926 	int error;
    927 
    928 	/* locate correct op */
    929 	mutex_enter(&pmp->pmp_lock);
    930 	TAILQ_FOREACH(pspark, &pmp->pmp_req_sizepark, pkso_entries) {
    931 		if (pspark->pkso_reqid == psop_user->pso_reqid) {
    932 			TAILQ_REMOVE(&pmp->pmp_req_sizepark, pspark,
    933 			    pkso_entries);
    934 			break;
    935 		}
    936 	}
    937 	mutex_exit(&pmp->pmp_lock);
    938 
    939 	if (pspark == NULL)
    940 		return EINVAL;
    941 
    942 	error = 0;
    943 	copylen = MIN(pspark->pkso_bufsize, psop_user->pso_bufsize);
    944 
    945 	/*
    946 	 * XXX: uvm stuff to avoid bouncy-bouncy copying?
    947 	 */
    948 	if (PUFFS_SIZEOP_UIO(pspark->pkso_reqtype)) {
    949 		kernbuf = malloc(copylen, M_PUFFS, M_WAITOK | M_ZERO);
    950 		if (pspark->pkso_reqtype == PUFFS_SIZEOPREQ_UIO_IN) {
    951 			error = copyin(psop_user->pso_userbuf,
    952 			    kernbuf, copylen);
    953 			if (error) {
    954 				printf("psop ERROR1 %d\n", error);
    955 				goto escape;
    956 			}
    957 		}
    958 		error = uiomove(kernbuf, copylen, pspark->pkso_uio);
    959 		if (error) {
    960 			printf("uiomove from kernel %p, len %d failed: %d\n",
    961 			    kernbuf, (int)copylen, error);
    962 			goto escape;
    963 		}
    964 
    965 		if (pspark->pkso_reqtype == PUFFS_SIZEOPREQ_UIO_OUT) {
    966 			error = copyout(kernbuf,
    967 			    psop_user->pso_userbuf, copylen);
    968 			if (error) {
    969 				printf("psop ERROR2 %d\n", error);
    970 				goto escape;
    971 			}
    972 		}
    973  escape:
    974 		free(kernbuf, M_PUFFS);
    975 	} else if (PUFFS_SIZEOP_BUF(pspark->pkso_reqtype)) {
    976 		copylen = MAX(pspark->pkso_bufsize, psop_user->pso_bufsize);
    977 		if (pspark->pkso_reqtype == PUFFS_SIZEOPREQ_BUF_IN) {
    978 			error = copyin(psop_user->pso_userbuf,
    979 			pspark->pkso_copybuf, copylen);
    980 		} else {
    981 			error = copyout(pspark->pkso_copybuf,
    982 			    psop_user->pso_userbuf, copylen);
    983 		}
    984 	}
    985 #ifdef DIAGNOSTIC
    986 	else
    987 		panic("puffssizeop: invalid reqtype %d\n",
    988 		    pspark->pkso_reqtype);
    989 #endif /* DIAGNOSTIC */
    990 
    991 	return error;
    992 }
    993 #endif
    994