xref: /titanic_50/usr/src/uts/common/xen/io/xdf.c (revision 13237b7e1e5bd293e466307b2e06f8e0e2321a0a)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 
22 /*
23  * Copyright 2009 Sun Microsystems, Inc.  All rights reserved.
24  * Use is subject to license terms.
25  */
26 
27 /*
28  * xdf.c - Xen Virtual Block Device Driver
29  * TODO:
30  *	- support alternate block size (currently only DEV_BSIZE supported)
31  *	- revalidate geometry for removable devices
32  *
33  * This driver export solaris disk device nodes, accepts IO requests from
34  * those nodes, and services those requests by talking to a backend device
35  * in another domain.
36  *
37  * Communication with the backend device is done via a ringbuffer (which is
38  * managed via xvdi interfaces) and dma memory (which is managed via ddi
39  * interfaces).
40  *
41  * Communication with the backend device is dependant upon establishing a
42  * connection to the backend device.  This connection process involves
43  * reading device configuration information from xenbus and publishing
44  * some frontend runtime configuration parameters via the xenbus (for
45  * consumption by the backend).  Once we've published runtime configuration
46  * information via the xenbus, the backend device can enter the connected
47  * state and we'll enter the XD_CONNECTED state.  But before we can allow
48  * random IO to begin, we need to do IO to the backend device to determine
49  * the device label and if flush operations are supported.  Once this is
50  * done we enter the XD_READY state and can process any IO operations.
51  *
52  * We recieve notifications of xenbus state changes for the backend device
53  * (aka, the "other end") via the xdf_oe_change() callback.  This callback
54  * is single threaded, meaning that we can't recieve new notification of
55  * other end state changes while we're processing an outstanding
56  * notification of an other end state change.  There for we can't do any
57  * blocking operations from the xdf_oe_change() callback.  This is why we
58  * have a seperate taskq (xdf_ready_tq) which exists to do the necessary
59  * IO to get us from the XD_CONNECTED to the XD_READY state.  All IO
60  * generated by the xdf_ready_tq thread (xdf_ready_tq_thread) will go
61  * throught xdf_lb_rdwr(), which is a synchronous IO interface.  IOs
62  * generated by the xdf_ready_tq_thread thread have priority over all
63  * other IO requests.
64  *
65  * We also communicate with the backend device via the xenbus "media-req"
66  * (XBP_MEDIA_REQ) property.  For more information on this see the
67  * comments in blkif.h.
68  */
69 
70 #include <io/xdf.h>
71 
72 #include <sys/conf.h>
73 #include <sys/dkio.h>
74 #include <sys/promif.h>
75 #include <sys/sysmacros.h>
76 #include <sys/kstat.h>
77 #include <sys/mach_mmu.h>
78 #ifdef XPV_HVM_DRIVER
79 #include <sys/xpv_support.h>
80 #include <sys/sunndi.h>
81 #else /* !XPV_HVM_DRIVER */
82 #include <sys/evtchn_impl.h>
83 #endif /* !XPV_HVM_DRIVER */
84 #include <public/io/xenbus.h>
85 #include <xen/sys/xenbus_impl.h>
86 #include <sys/scsi/generic/inquiry.h>
87 #include <xen/io/blkif_impl.h>
88 #include <sys/fdio.h>
89 #include <sys/cdio.h>
90 
91 /*
92  * DEBUG_EVAL can be used to include debug only statements without
93  * having to use '#ifdef DEBUG' statements
94  */
95 #ifdef DEBUG
96 #define	DEBUG_EVAL(x)	(x)
97 #else /* !DEBUG */
98 #define	DEBUG_EVAL(x)
99 #endif /* !DEBUG */
100 
101 #define	XDF_DRAIN_MSEC_DELAY		(50*1000)	/* 00.05 sec */
102 #define	XDF_DRAIN_RETRY_COUNT		200		/* 10.00 sec */
103 
104 #define	INVALID_DOMID	((domid_t)-1)
105 #define	FLUSH_DISKCACHE	0x1
106 #define	WRITE_BARRIER	0x2
107 #define	DEFAULT_FLUSH_BLOCK	156 /* block to write to cause a cache flush */
108 #define	USE_WRITE_BARRIER(vdp)						\
109 	((vdp)->xdf_feature_barrier && !(vdp)->xdf_flush_supported)
110 #define	USE_FLUSH_DISKCACHE(vdp)					\
111 	((vdp)->xdf_feature_barrier && (vdp)->xdf_flush_supported)
112 #define	IS_WRITE_BARRIER(vdp, bp)					\
113 	(!IS_READ(bp) && USE_WRITE_BARRIER(vdp) &&			\
114 	((bp)->b_un.b_addr == (vdp)->xdf_cache_flush_block))
115 #define	IS_FLUSH_DISKCACHE(bp)						\
116 	(!IS_READ(bp) && USE_FLUSH_DISKCACHE(vdp) && ((bp)->b_bcount == 0))
117 
118 #define	VREQ_DONE(vreq)							\
119 	VOID2BOOLEAN(((vreq)->v_status == VREQ_DMAWIN_DONE) &&		\
120 	    (((vreq)->v_flush_diskcache == FLUSH_DISKCACHE) ||		\
121 	    (((vreq)->v_dmaw + 1) == (vreq)->v_ndmaws)))
122 
123 #define	BP_VREQ(bp)		((v_req_t *)((bp)->av_back))
124 #define	BP_VREQ_SET(bp, vreq)	(((bp)->av_back = (buf_t *)(vreq)))
125 
126 extern int		do_polled_io;
127 
128 /* run-time tunables that we don't want the compiler to optimize away */
129 volatile int		xdf_debug = 0;
130 volatile boolean_t	xdf_barrier_flush_disable = B_FALSE;
131 
132 /* per module globals */
133 major_t			xdf_major;
134 static void		*xdf_ssp;
135 static kmem_cache_t	*xdf_vreq_cache;
136 static kmem_cache_t	*xdf_gs_cache;
137 static int		xdf_maxphys = XB_MAXPHYS;
138 static diskaddr_t	xdf_flush_block = DEFAULT_FLUSH_BLOCK;
139 static int		xdf_fbrewrites;	/* flush block re-write count */
140 
141 /* misc public functions (used by xdf_shell.c) */
142 int xdf_lb_rdwr(dev_info_t *, uchar_t, void *, diskaddr_t, size_t, void *);
143 int xdf_lb_getinfo(dev_info_t *, int, void *, void *);
144 
145 /*  misc private functions */
146 static void xdf_io_start(xdf_t *);
147 
148 /* callbacks from commmon label */
149 static cmlb_tg_ops_t xdf_lb_ops = {
150 	TG_DK_OPS_VERSION_1,
151 	xdf_lb_rdwr,
152 	xdf_lb_getinfo
153 };
154 
155 /*
156  * I/O buffer DMA attributes
157  * Make sure: one DMA window contains BLKIF_MAX_SEGMENTS_PER_REQUEST at most
158  */
159 static ddi_dma_attr_t xb_dma_attr = {
160 	DMA_ATTR_V0,
161 	(uint64_t)0,			/* lowest address */
162 	(uint64_t)0xffffffffffffffff,	/* highest usable address */
163 	(uint64_t)0xffffff,		/* DMA counter limit max */
164 	(uint64_t)XB_BSIZE,		/* alignment in bytes */
165 	XB_BSIZE - 1,			/* bitmap of burst sizes */
166 	XB_BSIZE,			/* min transfer */
167 	(uint64_t)XB_MAX_XFER, 		/* maximum transfer */
168 	(uint64_t)PAGEOFFSET,		/* 1 page segment length  */
169 	BLKIF_MAX_SEGMENTS_PER_REQUEST,	/* maximum number of segments */
170 	XB_BSIZE,			/* granularity */
171 	0,				/* flags (reserved) */
172 };
173 
174 static ddi_device_acc_attr_t xc_acc_attr = {
175 	DDI_DEVICE_ATTR_V0,
176 	DDI_NEVERSWAP_ACC,
177 	DDI_STRICTORDER_ACC
178 };
179 
180 static void
181 xdf_timeout_handler(void *arg)
182 {
183 	xdf_t *vdp = arg;
184 
185 	mutex_enter(&vdp->xdf_dev_lk);
186 	vdp->xdf_timeout_id = 0;
187 	mutex_exit(&vdp->xdf_dev_lk);
188 
189 	/* new timeout thread could be re-scheduled */
190 	xdf_io_start(vdp);
191 }
192 
193 /*
194  * callback func when DMA/GTE resources is available
195  *
196  * Note: we only register one callback function to grant table subsystem
197  * since we only have one 'struct gnttab_free_callback' in xdf_t.
198  */
199 static int
200 xdf_dmacallback(caddr_t arg)
201 {
202 	xdf_t *vdp = (xdf_t *)arg;
203 	ASSERT(vdp != NULL);
204 
205 	DPRINTF(DMA_DBG, ("xdf@%s: DMA callback started\n",
206 	    vdp->xdf_addr));
207 
208 	ddi_trigger_softintr(vdp->xdf_softintr_id);
209 	return (DDI_DMA_CALLBACK_DONE);
210 }
211 
212 static ge_slot_t *
213 gs_get(xdf_t *vdp, int isread)
214 {
215 	grant_ref_t gh;
216 	ge_slot_t *gs;
217 
218 	/* try to alloc GTEs needed in this slot, first */
219 	if (gnttab_alloc_grant_references(
220 	    BLKIF_MAX_SEGMENTS_PER_REQUEST, &gh) == -1) {
221 		if (vdp->xdf_gnt_callback.next == NULL) {
222 			SETDMACBON(vdp);
223 			gnttab_request_free_callback(
224 			    &vdp->xdf_gnt_callback,
225 			    (void (*)(void *))xdf_dmacallback,
226 			    (void *)vdp,
227 			    BLKIF_MAX_SEGMENTS_PER_REQUEST);
228 		}
229 		return (NULL);
230 	}
231 
232 	gs = kmem_cache_alloc(xdf_gs_cache, KM_NOSLEEP);
233 	if (gs == NULL) {
234 		gnttab_free_grant_references(gh);
235 		if (vdp->xdf_timeout_id == 0)
236 			/* restart I/O after one second */
237 			vdp->xdf_timeout_id =
238 			    timeout(xdf_timeout_handler, vdp, hz);
239 		return (NULL);
240 	}
241 
242 	/* init gs_slot */
243 	gs->gs_oeid = vdp->xdf_peer;
244 	gs->gs_isread = isread;
245 	gs->gs_ghead = gh;
246 	gs->gs_ngrefs = 0;
247 
248 	return (gs);
249 }
250 
251 static void
252 gs_free(ge_slot_t *gs)
253 {
254 	int		i;
255 
256 	/* release all grant table entry resources used in this slot */
257 	for (i = 0; i < gs->gs_ngrefs; i++)
258 		gnttab_end_foreign_access(gs->gs_ge[i], !gs->gs_isread, 0);
259 	gnttab_free_grant_references(gs->gs_ghead);
260 	list_remove(&gs->gs_vreq->v_gs, gs);
261 	kmem_cache_free(xdf_gs_cache, gs);
262 }
263 
264 static grant_ref_t
265 gs_grant(ge_slot_t *gs, mfn_t mfn)
266 {
267 	grant_ref_t gr = gnttab_claim_grant_reference(&gs->gs_ghead);
268 
269 	ASSERT(gr != -1);
270 	ASSERT(gs->gs_ngrefs < BLKIF_MAX_SEGMENTS_PER_REQUEST);
271 	gs->gs_ge[gs->gs_ngrefs++] = gr;
272 	gnttab_grant_foreign_access_ref(gr, gs->gs_oeid, mfn, !gs->gs_isread);
273 
274 	return (gr);
275 }
276 
277 /*
278  * Alloc a vreq for this bp
279  * bp->av_back contains the pointer to the vreq upon return
280  */
281 static v_req_t *
282 vreq_get(xdf_t *vdp, buf_t *bp)
283 {
284 	v_req_t *vreq = NULL;
285 
286 	ASSERT(BP_VREQ(bp) == NULL);
287 
288 	vreq = kmem_cache_alloc(xdf_vreq_cache, KM_NOSLEEP);
289 	if (vreq == NULL) {
290 		if (vdp->xdf_timeout_id == 0)
291 			/* restart I/O after one second */
292 			vdp->xdf_timeout_id =
293 			    timeout(xdf_timeout_handler, vdp, hz);
294 		return (NULL);
295 	}
296 	bzero(vreq, sizeof (v_req_t));
297 	list_create(&vreq->v_gs, sizeof (ge_slot_t),
298 	    offsetof(ge_slot_t, gs_vreq_link));
299 	vreq->v_buf = bp;
300 	vreq->v_status = VREQ_INIT;
301 	vreq->v_runq = B_FALSE;
302 	BP_VREQ_SET(bp, vreq);
303 	/* init of other fields in vreq is up to the caller */
304 
305 	list_insert_head(&vdp->xdf_vreq_act, (void *)vreq);
306 
307 	return (vreq);
308 }
309 
310 static void
311 vreq_free(xdf_t *vdp, v_req_t *vreq)
312 {
313 	buf_t	*bp = vreq->v_buf;
314 
315 	ASSERT(MUTEX_HELD(&vdp->xdf_dev_lk));
316 	ASSERT(BP_VREQ(bp) == vreq);
317 
318 	list_remove(&vdp->xdf_vreq_act, vreq);
319 
320 	if (vreq->v_flush_diskcache == FLUSH_DISKCACHE)
321 		goto done;
322 
323 	switch (vreq->v_status) {
324 	case VREQ_DMAWIN_DONE:
325 	case VREQ_GS_ALLOCED:
326 	case VREQ_DMABUF_BOUND:
327 		(void) ddi_dma_unbind_handle(vreq->v_dmahdl);
328 		/*FALLTHRU*/
329 	case VREQ_DMAMEM_ALLOCED:
330 		if (!ALIGNED_XFER(bp)) {
331 			ASSERT(vreq->v_abuf != NULL);
332 			if (!IS_ERROR(bp) && IS_READ(bp))
333 				bcopy(vreq->v_abuf, bp->b_un.b_addr,
334 				    bp->b_bcount);
335 			ddi_dma_mem_free(&vreq->v_align);
336 		}
337 		/*FALLTHRU*/
338 	case VREQ_MEMDMAHDL_ALLOCED:
339 		if (!ALIGNED_XFER(bp))
340 			ddi_dma_free_handle(&vreq->v_memdmahdl);
341 		/*FALLTHRU*/
342 	case VREQ_DMAHDL_ALLOCED:
343 		ddi_dma_free_handle(&vreq->v_dmahdl);
344 		break;
345 	default:
346 		break;
347 	}
348 done:
349 	ASSERT(!vreq->v_runq);
350 	list_destroy(&vreq->v_gs);
351 	kmem_cache_free(xdf_vreq_cache, vreq);
352 }
353 
354 /*
355  * Snarf new data if our flush block was re-written
356  */
357 static void
358 check_fbwrite(xdf_t *vdp, buf_t *bp, daddr_t blkno)
359 {
360 	int nblks;
361 	boolean_t mapin;
362 
363 	if (IS_WRITE_BARRIER(vdp, bp))
364 		return; /* write was a flush write */
365 
366 	mapin = B_FALSE;
367 	nblks = bp->b_bcount >> DEV_BSHIFT;
368 	if (xdf_flush_block >= blkno && xdf_flush_block < (blkno + nblks)) {
369 		xdf_fbrewrites++;
370 		if (bp->b_flags & (B_PAGEIO | B_PHYS)) {
371 			mapin = B_TRUE;
372 			bp_mapin(bp);
373 		}
374 		bcopy(bp->b_un.b_addr +
375 		    ((xdf_flush_block - blkno) << DEV_BSHIFT),
376 		    vdp->xdf_cache_flush_block, DEV_BSIZE);
377 		if (mapin)
378 			bp_mapout(bp);
379 	}
380 }
381 
382 /*
383  * Initalize the DMA and grant table resources for the buf
384  */
385 static int
386 vreq_setup(xdf_t *vdp, v_req_t *vreq)
387 {
388 	int rc;
389 	ddi_dma_attr_t dmaattr;
390 	uint_t ndcs, ndws;
391 	ddi_dma_handle_t dh;
392 	ddi_dma_handle_t mdh;
393 	ddi_dma_cookie_t dc;
394 	ddi_acc_handle_t abh;
395 	caddr_t	aba;
396 	ge_slot_t *gs;
397 	size_t bufsz;
398 	off_t off;
399 	size_t sz;
400 	buf_t *bp = vreq->v_buf;
401 	int dma_flags = (IS_READ(bp) ? DDI_DMA_READ : DDI_DMA_WRITE) |
402 	    DDI_DMA_STREAMING | DDI_DMA_PARTIAL;
403 
404 	switch (vreq->v_status) {
405 	case VREQ_INIT:
406 		if (IS_FLUSH_DISKCACHE(bp)) {
407 			if ((gs = gs_get(vdp, IS_READ(bp))) == NULL) {
408 				DPRINTF(DMA_DBG, ("xdf@%s: "
409 				    "get ge_slotfailed\n", vdp->xdf_addr));
410 				return (DDI_FAILURE);
411 			}
412 			vreq->v_blkno = 0;
413 			vreq->v_nslots = 1;
414 			vreq->v_flush_diskcache = FLUSH_DISKCACHE;
415 			vreq->v_status = VREQ_GS_ALLOCED;
416 			gs->gs_vreq = vreq;
417 			list_insert_head(&vreq->v_gs, gs);
418 			return (DDI_SUCCESS);
419 		}
420 
421 		if (IS_WRITE_BARRIER(vdp, bp))
422 			vreq->v_flush_diskcache = WRITE_BARRIER;
423 		vreq->v_blkno = bp->b_blkno +
424 		    (diskaddr_t)(uintptr_t)bp->b_private;
425 		/* See if we wrote new data to our flush block */
426 		if (!IS_READ(bp) && USE_WRITE_BARRIER(vdp))
427 			check_fbwrite(vdp, bp, vreq->v_blkno);
428 		vreq->v_status = VREQ_INIT_DONE;
429 		/*FALLTHRU*/
430 
431 	case VREQ_INIT_DONE:
432 		/*
433 		 * alloc DMA handle
434 		 */
435 		rc = ddi_dma_alloc_handle(vdp->xdf_dip, &xb_dma_attr,
436 		    xdf_dmacallback, (caddr_t)vdp, &dh);
437 		if (rc != DDI_SUCCESS) {
438 			SETDMACBON(vdp);
439 			DPRINTF(DMA_DBG, ("xdf@%s: DMA handle alloc failed\n",
440 			    vdp->xdf_addr));
441 			return (DDI_FAILURE);
442 		}
443 
444 		vreq->v_dmahdl = dh;
445 		vreq->v_status = VREQ_DMAHDL_ALLOCED;
446 		/*FALLTHRU*/
447 
448 	case VREQ_DMAHDL_ALLOCED:
449 		/*
450 		 * alloc dma handle for 512-byte aligned buf
451 		 */
452 		if (!ALIGNED_XFER(bp)) {
453 			/*
454 			 * XXPV: we need to temporarily enlarge the seg
455 			 * boundary and s/g length to work round CR6381968
456 			 */
457 			dmaattr = xb_dma_attr;
458 			dmaattr.dma_attr_seg = (uint64_t)-1;
459 			dmaattr.dma_attr_sgllen = INT_MAX;
460 			rc = ddi_dma_alloc_handle(vdp->xdf_dip, &dmaattr,
461 			    xdf_dmacallback, (caddr_t)vdp, &mdh);
462 			if (rc != DDI_SUCCESS) {
463 				SETDMACBON(vdp);
464 				DPRINTF(DMA_DBG, ("xdf@%s: "
465 				    "unaligned buf DMAhandle alloc failed\n",
466 				    vdp->xdf_addr));
467 				return (DDI_FAILURE);
468 			}
469 			vreq->v_memdmahdl = mdh;
470 			vreq->v_status = VREQ_MEMDMAHDL_ALLOCED;
471 		}
472 		/*FALLTHRU*/
473 
474 	case VREQ_MEMDMAHDL_ALLOCED:
475 		/*
476 		 * alloc 512-byte aligned buf
477 		 */
478 		if (!ALIGNED_XFER(bp)) {
479 			if (bp->b_flags & (B_PAGEIO | B_PHYS))
480 				bp_mapin(bp);
481 
482 			rc = ddi_dma_mem_alloc(vreq->v_memdmahdl,
483 			    roundup(bp->b_bcount, XB_BSIZE), &xc_acc_attr,
484 			    DDI_DMA_STREAMING, xdf_dmacallback, (caddr_t)vdp,
485 			    &aba, &bufsz, &abh);
486 			if (rc != DDI_SUCCESS) {
487 				SETDMACBON(vdp);
488 				DPRINTF(DMA_DBG, ("xdf@%s: "
489 				    "DMA mem allocation failed\n",
490 				    vdp->xdf_addr));
491 				return (DDI_FAILURE);
492 			}
493 
494 			vreq->v_abuf = aba;
495 			vreq->v_align = abh;
496 			vreq->v_status = VREQ_DMAMEM_ALLOCED;
497 
498 			ASSERT(bufsz >= bp->b_bcount);
499 			if (!IS_READ(bp))
500 				bcopy(bp->b_un.b_addr, vreq->v_abuf,
501 				    bp->b_bcount);
502 		}
503 		/*FALLTHRU*/
504 
505 	case VREQ_DMAMEM_ALLOCED:
506 		/*
507 		 * dma bind
508 		 */
509 		if (ALIGNED_XFER(bp)) {
510 			rc = ddi_dma_buf_bind_handle(vreq->v_dmahdl, bp,
511 			    dma_flags, xdf_dmacallback, (caddr_t)vdp,
512 			    &dc, &ndcs);
513 		} else {
514 			rc = ddi_dma_addr_bind_handle(vreq->v_dmahdl,
515 			    NULL, vreq->v_abuf, bp->b_bcount, dma_flags,
516 			    xdf_dmacallback, (caddr_t)vdp, &dc, &ndcs);
517 		}
518 		if (rc == DDI_DMA_MAPPED || rc == DDI_DMA_PARTIAL_MAP) {
519 			/* get num of dma windows */
520 			if (rc == DDI_DMA_PARTIAL_MAP) {
521 				rc = ddi_dma_numwin(vreq->v_dmahdl, &ndws);
522 				ASSERT(rc == DDI_SUCCESS);
523 			} else {
524 				ndws = 1;
525 			}
526 		} else {
527 			SETDMACBON(vdp);
528 			DPRINTF(DMA_DBG, ("xdf@%s: DMA bind failed\n",
529 			    vdp->xdf_addr));
530 			return (DDI_FAILURE);
531 		}
532 
533 		vreq->v_dmac = dc;
534 		vreq->v_dmaw = 0;
535 		vreq->v_ndmacs = ndcs;
536 		vreq->v_ndmaws = ndws;
537 		vreq->v_nslots = ndws;
538 		vreq->v_status = VREQ_DMABUF_BOUND;
539 		/*FALLTHRU*/
540 
541 	case VREQ_DMABUF_BOUND:
542 		/*
543 		 * get ge_slot, callback is set upon failure from gs_get(),
544 		 * if not set previously
545 		 */
546 		if ((gs = gs_get(vdp, IS_READ(bp))) == NULL) {
547 			DPRINTF(DMA_DBG, ("xdf@%s: get ge_slot failed\n",
548 			    vdp->xdf_addr));
549 			return (DDI_FAILURE);
550 		}
551 
552 		vreq->v_status = VREQ_GS_ALLOCED;
553 		gs->gs_vreq = vreq;
554 		list_insert_head(&vreq->v_gs, gs);
555 		break;
556 
557 	case VREQ_GS_ALLOCED:
558 		/* nothing need to be done */
559 		break;
560 
561 	case VREQ_DMAWIN_DONE:
562 		/*
563 		 * move to the next dma window
564 		 */
565 		ASSERT((vreq->v_dmaw + 1) < vreq->v_ndmaws);
566 
567 		/* get a ge_slot for this DMA window */
568 		if ((gs = gs_get(vdp, IS_READ(bp))) == NULL) {
569 			DPRINTF(DMA_DBG, ("xdf@%s: get ge_slot failed\n",
570 			    vdp->xdf_addr));
571 			return (DDI_FAILURE);
572 		}
573 
574 		vreq->v_dmaw++;
575 		VERIFY(ddi_dma_getwin(vreq->v_dmahdl, vreq->v_dmaw, &off, &sz,
576 		    &vreq->v_dmac, &vreq->v_ndmacs) == DDI_SUCCESS);
577 		vreq->v_status = VREQ_GS_ALLOCED;
578 		gs->gs_vreq = vreq;
579 		list_insert_head(&vreq->v_gs, gs);
580 		break;
581 
582 	default:
583 		return (DDI_FAILURE);
584 	}
585 
586 	return (DDI_SUCCESS);
587 }
588 
589 static int
590 xdf_cmlb_attach(xdf_t *vdp)
591 {
592 	dev_info_t	*dip = vdp->xdf_dip;
593 
594 	return (cmlb_attach(dip, &xdf_lb_ops,
595 	    XD_IS_CD(vdp) ? DTYPE_RODIRECT : DTYPE_DIRECT,
596 	    XD_IS_RM(vdp),
597 	    B_TRUE,
598 	    XD_IS_CD(vdp) ? DDI_NT_CD_XVMD : DDI_NT_BLOCK_XVMD,
599 #if defined(XPV_HVM_DRIVER)
600 	    (XD_IS_CD(vdp) ? 0 : CMLB_CREATE_ALTSLICE_VTOC_16_DTYPE_DIRECT) |
601 	    CMLB_INTERNAL_MINOR_NODES,
602 #else /* !XPV_HVM_DRIVER */
603 	    XD_IS_CD(vdp) ? 0 : CMLB_FAKE_LABEL_ONE_PARTITION,
604 #endif /* !XPV_HVM_DRIVER */
605 	    vdp->xdf_vd_lbl, NULL));
606 }
607 
608 static void
609 xdf_io_err(buf_t *bp, int err, size_t resid)
610 {
611 	bioerror(bp, err);
612 	if (resid == 0)
613 		bp->b_resid = bp->b_bcount;
614 	biodone(bp);
615 }
616 
617 static void
618 xdf_kstat_enter(xdf_t *vdp, buf_t *bp)
619 {
620 	v_req_t *vreq = BP_VREQ(bp);
621 
622 	ASSERT(MUTEX_HELD(&vdp->xdf_dev_lk));
623 
624 	if (vdp->xdf_xdev_iostat == NULL)
625 		return;
626 	if ((vreq != NULL) && vreq->v_runq) {
627 		kstat_runq_enter(KSTAT_IO_PTR(vdp->xdf_xdev_iostat));
628 	} else {
629 		kstat_waitq_enter(KSTAT_IO_PTR(vdp->xdf_xdev_iostat));
630 	}
631 }
632 
633 static void
634 xdf_kstat_exit(xdf_t *vdp, buf_t *bp)
635 {
636 	v_req_t *vreq = BP_VREQ(bp);
637 
638 	ASSERT(MUTEX_HELD(&vdp->xdf_dev_lk));
639 
640 	if (vdp->xdf_xdev_iostat == NULL)
641 		return;
642 	if ((vreq != NULL) && vreq->v_runq) {
643 		kstat_runq_exit(KSTAT_IO_PTR(vdp->xdf_xdev_iostat));
644 	} else {
645 		kstat_waitq_exit(KSTAT_IO_PTR(vdp->xdf_xdev_iostat));
646 	}
647 }
648 
649 static void
650 xdf_kstat_waitq_to_runq(xdf_t *vdp, buf_t *bp)
651 {
652 	v_req_t *vreq = BP_VREQ(bp);
653 
654 	ASSERT(MUTEX_HELD(&vdp->xdf_dev_lk));
655 	ASSERT(!vreq->v_runq);
656 
657 	vreq->v_runq = B_TRUE;
658 	if (vdp->xdf_xdev_iostat == NULL)
659 		return;
660 	kstat_waitq_to_runq(KSTAT_IO_PTR(vdp->xdf_xdev_iostat));
661 }
662 
663 static void
664 xdf_kstat_runq_to_waitq(xdf_t *vdp, buf_t *bp)
665 {
666 	v_req_t *vreq = BP_VREQ(bp);
667 
668 	ASSERT(MUTEX_HELD(&vdp->xdf_dev_lk));
669 	ASSERT(vreq->v_runq);
670 
671 	vreq->v_runq = B_FALSE;
672 	if (vdp->xdf_xdev_iostat == NULL)
673 		return;
674 	kstat_runq_back_to_waitq(KSTAT_IO_PTR(vdp->xdf_xdev_iostat));
675 }
676 
677 int
678 xdf_kstat_create(dev_info_t *dip, char *ks_module, int instance)
679 {
680 	xdf_t		*vdp = (xdf_t *)ddi_get_driver_private(dip);
681 	kstat_t		*kstat;
682 	buf_t		*bp;
683 
684 	if ((kstat = kstat_create(
685 	    ks_module, instance, NULL, "disk",
686 	    KSTAT_TYPE_IO, 1, KSTAT_FLAG_PERSISTENT)) == NULL)
687 		return (-1);
688 
689 	/* See comment about locking in xdf_kstat_delete(). */
690 	mutex_enter(&vdp->xdf_iostat_lk);
691 	mutex_enter(&vdp->xdf_dev_lk);
692 
693 	/* only one kstat can exist at a time */
694 	if (vdp->xdf_xdev_iostat != NULL) {
695 		mutex_exit(&vdp->xdf_dev_lk);
696 		mutex_exit(&vdp->xdf_iostat_lk);
697 		kstat_delete(kstat);
698 		return (-1);
699 	}
700 
701 	vdp->xdf_xdev_iostat = kstat;
702 	vdp->xdf_xdev_iostat->ks_lock = &vdp->xdf_dev_lk;
703 	kstat_install(vdp->xdf_xdev_iostat);
704 
705 	/*
706 	 * Now that we've created a kstat, we need to update the waitq and
707 	 * runq counts for the kstat to reflect our current state.
708 	 *
709 	 * For a buf_t structure to be on the runq, it must have a ring
710 	 * buffer slot associated with it.  To get a ring buffer slot the
711 	 * buf must first have a v_req_t and a ge_slot_t associated with it.
712 	 * Then when it is granted a ring buffer slot, v_runq will be set to
713 	 * true.
714 	 *
715 	 * For a buf_t structure to be on the waitq, it must not be on the
716 	 * runq.  So to find all the buf_t's that should be on waitq, we
717 	 * walk the active buf list and add any buf_t's which aren't on the
718 	 * runq to the waitq.
719 	 */
720 	bp = vdp->xdf_f_act;
721 	while (bp != NULL) {
722 		xdf_kstat_enter(vdp, bp);
723 		bp = bp->av_forw;
724 	}
725 	if (vdp->xdf_ready_tq_bp != NULL)
726 		xdf_kstat_enter(vdp, vdp->xdf_ready_tq_bp);
727 
728 	mutex_exit(&vdp->xdf_dev_lk);
729 	mutex_exit(&vdp->xdf_iostat_lk);
730 	return (0);
731 }
732 
733 void
734 xdf_kstat_delete(dev_info_t *dip)
735 {
736 	xdf_t		*vdp = (xdf_t *)ddi_get_driver_private(dip);
737 	kstat_t		*kstat;
738 	buf_t		*bp;
739 
740 	/*
741 	 * The locking order here is xdf_iostat_lk and then xdf_dev_lk.
742 	 * xdf_dev_lk is used to protect the xdf_xdev_iostat pointer
743 	 * and the contents of the our kstat.  xdf_iostat_lk is used
744 	 * to protect the allocation and freeing of the actual kstat.
745 	 * xdf_dev_lk can't be used for this purpose because kstat
746 	 * readers use it to access the contents of the kstat and
747 	 * hence it can't be held when calling kstat_delete().
748 	 */
749 	mutex_enter(&vdp->xdf_iostat_lk);
750 	mutex_enter(&vdp->xdf_dev_lk);
751 
752 	if (vdp->xdf_xdev_iostat == NULL) {
753 		mutex_exit(&vdp->xdf_dev_lk);
754 		mutex_exit(&vdp->xdf_iostat_lk);
755 		return;
756 	}
757 
758 	/*
759 	 * We're about to destroy the kstat structures, so it isn't really
760 	 * necessary to update the runq and waitq counts.  But, since this
761 	 * isn't a hot code path we can afford to be a little pedantic and
762 	 * go ahead and decrement the runq and waitq kstat counters to zero
763 	 * before free'ing them.  This helps us ensure that we've gotten all
764 	 * our accounting correct.
765 	 *
766 	 * For an explanation of how we determine which buffers go on the
767 	 * runq vs which go on the waitq, see the comments in
768 	 * xdf_kstat_create().
769 	 */
770 	bp = vdp->xdf_f_act;
771 	while (bp != NULL) {
772 		xdf_kstat_exit(vdp, bp);
773 		bp = bp->av_forw;
774 	}
775 	if (vdp->xdf_ready_tq_bp != NULL)
776 		xdf_kstat_exit(vdp, vdp->xdf_ready_tq_bp);
777 
778 	kstat = vdp->xdf_xdev_iostat;
779 	vdp->xdf_xdev_iostat = NULL;
780 	mutex_exit(&vdp->xdf_dev_lk);
781 	kstat_delete(kstat);
782 	mutex_exit(&vdp->xdf_iostat_lk);
783 }
784 
785 /*
786  * Add an IO requests onto the active queue.
787  *
788  * We have to detect IOs generated by xdf_ready_tq_thread.  These IOs
789  * are used to establish a connection to the backend, so they recieve
790  * priority over all other IOs.  Since xdf_ready_tq_thread only does
791  * synchronous IO, there can only be one xdf_ready_tq_thread request at any
792  * given time and we record the buf associated with that request in
793  * xdf_ready_tq_bp.
794  */
795 static void
796 xdf_bp_push(xdf_t *vdp, buf_t *bp)
797 {
798 	ASSERT(MUTEX_HELD(&vdp->xdf_dev_lk));
799 	ASSERT(bp->av_forw == NULL);
800 
801 	xdf_kstat_enter(vdp, bp);
802 
803 	if (curthread == vdp->xdf_ready_tq_thread) {
804 		/* new IO requests from the ready thread */
805 		ASSERT(vdp->xdf_ready_tq_bp == NULL);
806 		vdp->xdf_ready_tq_bp = bp;
807 		return;
808 	}
809 
810 	/* this is normal IO request */
811 	ASSERT(bp != vdp->xdf_ready_tq_bp);
812 
813 	if (vdp->xdf_f_act == NULL) {
814 		/* this is only only IO on the active queue */
815 		ASSERT(vdp->xdf_l_act == NULL);
816 		ASSERT(vdp->xdf_i_act == NULL);
817 		vdp->xdf_f_act = vdp->xdf_l_act = vdp->xdf_i_act = bp;
818 		return;
819 	}
820 
821 	/* add this IO to the tail of the active queue */
822 	vdp->xdf_l_act->av_forw = bp;
823 	vdp->xdf_l_act = bp;
824 	if (vdp->xdf_i_act == NULL)
825 		vdp->xdf_i_act = bp;
826 }
827 
828 static void
829 xdf_bp_pop(xdf_t *vdp, buf_t *bp)
830 {
831 	buf_t	*bp_iter;
832 
833 	ASSERT(MUTEX_HELD(&vdp->xdf_dev_lk));
834 	ASSERT(VREQ_DONE(BP_VREQ(bp)));
835 
836 	if (vdp->xdf_ready_tq_bp == bp) {
837 		/* we're done with a ready thread IO request */
838 		ASSERT(bp->av_forw == NULL);
839 		vdp->xdf_ready_tq_bp = NULL;
840 		return;
841 	}
842 
843 	/* we're done with a normal IO request */
844 	ASSERT((bp->av_forw != NULL) || (bp == vdp->xdf_l_act));
845 	ASSERT((bp->av_forw == NULL) || (bp != vdp->xdf_l_act));
846 	ASSERT(VREQ_DONE(BP_VREQ(vdp->xdf_f_act)));
847 	ASSERT(vdp->xdf_f_act != vdp->xdf_i_act);
848 
849 	if (bp == vdp->xdf_f_act) {
850 		/* This IO was at the head of our active queue. */
851 		vdp->xdf_f_act = bp->av_forw;
852 		if (bp == vdp->xdf_l_act)
853 			vdp->xdf_l_act = NULL;
854 	} else {
855 		/* There IO finished before some other pending IOs. */
856 		bp_iter = vdp->xdf_f_act;
857 		while (bp != bp_iter->av_forw) {
858 			bp_iter = bp_iter->av_forw;
859 			ASSERT(VREQ_DONE(BP_VREQ(bp_iter)));
860 			ASSERT(bp_iter != vdp->xdf_i_act);
861 		}
862 		bp_iter->av_forw = bp->av_forw;
863 		if (bp == vdp->xdf_l_act)
864 			vdp->xdf_l_act = bp_iter;
865 	}
866 	bp->av_forw = NULL;
867 }
868 
869 static buf_t *
870 xdf_bp_next(xdf_t *vdp)
871 {
872 	v_req_t	*vreq;
873 	buf_t	*bp;
874 
875 	if (vdp->xdf_state == XD_CONNECTED) {
876 		/*
877 		 * If we're in the XD_CONNECTED state, we only service IOs
878 		 * from the xdf_ready_tq_thread thread.
879 		 */
880 		if ((bp = vdp->xdf_ready_tq_bp) == NULL)
881 			return (NULL);
882 		if (((vreq = BP_VREQ(bp)) == NULL) || (!VREQ_DONE(vreq)))
883 			return (bp);
884 		return (NULL);
885 	}
886 
887 	/* if we're not in the XD_CONNECTED or XD_READY state we can't do IO */
888 	if (vdp->xdf_state != XD_READY)
889 		return (NULL);
890 
891 	ASSERT(vdp->xdf_ready_tq_bp == NULL);
892 	for (;;) {
893 		if ((bp = vdp->xdf_i_act) == NULL)
894 			return (NULL);
895 		if (((vreq = BP_VREQ(bp)) == NULL) || (!VREQ_DONE(vreq)))
896 			return (bp);
897 
898 		/* advance the active buf index pointer */
899 		vdp->xdf_i_act = bp->av_forw;
900 	}
901 }
902 
903 static void
904 xdf_io_fini(xdf_t *vdp, uint64_t id, int bioerr)
905 {
906 	ge_slot_t	*gs = (ge_slot_t *)(uintptr_t)id;
907 	v_req_t		*vreq = gs->gs_vreq;
908 	buf_t		*bp = vreq->v_buf;
909 
910 	ASSERT(MUTEX_HELD(&vdp->xdf_dev_lk));
911 	ASSERT(BP_VREQ(bp) == vreq);
912 
913 	gs_free(gs);
914 
915 	if (bioerr != 0)
916 		bioerror(bp, bioerr);
917 	ASSERT(vreq->v_nslots > 0);
918 	if (--vreq->v_nslots > 0)
919 		return;
920 
921 	/* remove this IO from our active queue */
922 	xdf_bp_pop(vdp, bp);
923 
924 	ASSERT(vreq->v_runq);
925 	xdf_kstat_exit(vdp, bp);
926 	vreq->v_runq = B_FALSE;
927 	vreq_free(vdp, vreq);
928 
929 	if (IS_ERROR(bp)) {
930 		xdf_io_err(bp, geterror(bp), 0);
931 	} else if (bp->b_resid != 0) {
932 		/* Partial transfers are an error */
933 		xdf_io_err(bp, EIO, bp->b_resid);
934 	} else {
935 		biodone(bp);
936 	}
937 }
938 
939 /*
940  * xdf interrupt handler
941  */
942 static uint_t
943 xdf_intr_locked(xdf_t *vdp)
944 {
945 	xendev_ring_t *xbr;
946 	blkif_response_t *resp;
947 	int bioerr;
948 	uint64_t id;
949 	uint8_t op;
950 	uint16_t status;
951 	ddi_acc_handle_t acchdl;
952 
953 	ASSERT(MUTEX_HELD(&vdp->xdf_dev_lk));
954 
955 	if ((xbr = vdp->xdf_xb_ring) == NULL)
956 		return (DDI_INTR_UNCLAIMED);
957 
958 	acchdl = vdp->xdf_xb_ring_hdl;
959 
960 	/*
961 	 * complete all requests which have a response
962 	 */
963 	while (resp = xvdi_ring_get_response(xbr)) {
964 		id = ddi_get64(acchdl, &resp->id);
965 		op = ddi_get8(acchdl, &resp->operation);
966 		status = ddi_get16(acchdl, (uint16_t *)&resp->status);
967 		DPRINTF(INTR_DBG, ("resp: op %d id %"PRIu64" status %d\n",
968 		    op, id, status));
969 
970 		if (status != BLKIF_RSP_OKAY) {
971 			DPRINTF(IO_DBG, ("xdf@%s: I/O error while %s",
972 			    vdp->xdf_addr,
973 			    (op == BLKIF_OP_READ) ? "reading" : "writing"));
974 			bioerr = EIO;
975 		} else {
976 			bioerr = 0;
977 		}
978 
979 		xdf_io_fini(vdp, id, bioerr);
980 	}
981 	return (DDI_INTR_CLAIMED);
982 }
983 
984 /*
985  * xdf_intr runs at PIL 5, so no one else can grab xdf_dev_lk and
986  * block at a lower pil.
987  */
988 static uint_t
989 xdf_intr(caddr_t arg)
990 {
991 	xdf_t *vdp = (xdf_t *)arg;
992 	int rv;
993 
994 	mutex_enter(&vdp->xdf_dev_lk);
995 	rv = xdf_intr_locked(vdp);
996 	mutex_exit(&vdp->xdf_dev_lk);
997 
998 	if (!do_polled_io)
999 		xdf_io_start(vdp);
1000 
1001 	return (rv);
1002 }
1003 
1004 static void
1005 xdf_ring_push(xdf_t *vdp)
1006 {
1007 	ASSERT(MUTEX_HELD(&vdp->xdf_dev_lk));
1008 
1009 	if (vdp->xdf_xb_ring == NULL)
1010 		return;
1011 
1012 	if (xvdi_ring_push_request(vdp->xdf_xb_ring)) {
1013 		DPRINTF(IO_DBG, (
1014 		    "xdf@%s: xdf_ring_push: sent request(s) to backend\n",
1015 		    vdp->xdf_addr));
1016 	}
1017 
1018 	if (xvdi_get_evtchn(vdp->xdf_dip) != INVALID_EVTCHN)
1019 		xvdi_notify_oe(vdp->xdf_dip);
1020 }
1021 
1022 static int
1023 xdf_ring_drain_locked(xdf_t *vdp)
1024 {
1025 	int		pollc, rv = 0;
1026 
1027 	ASSERT(MUTEX_HELD(&vdp->xdf_dev_lk));
1028 
1029 	if (xdf_debug & SUSRES_DBG)
1030 		xen_printf("xdf_ring_drain: start\n");
1031 
1032 	for (pollc = 0; pollc < XDF_DRAIN_RETRY_COUNT; pollc++) {
1033 		if (vdp->xdf_xb_ring == NULL)
1034 			goto out;
1035 
1036 		if (xvdi_ring_has_unconsumed_responses(vdp->xdf_xb_ring))
1037 			(void) xdf_intr_locked(vdp);
1038 		if (!xvdi_ring_has_incomp_request(vdp->xdf_xb_ring))
1039 			goto out;
1040 		xdf_ring_push(vdp);
1041 
1042 		/* file-backed devices can be slow */
1043 		mutex_exit(&vdp->xdf_dev_lk);
1044 #ifdef XPV_HVM_DRIVER
1045 		(void) HYPERVISOR_yield();
1046 #endif /* XPV_HVM_DRIVER */
1047 		delay(drv_usectohz(XDF_DRAIN_MSEC_DELAY));
1048 		mutex_enter(&vdp->xdf_dev_lk);
1049 	}
1050 	cmn_err(CE_WARN, "xdf@%s: xdf_ring_drain: timeout", vdp->xdf_addr);
1051 
1052 out:
1053 	if (vdp->xdf_xb_ring != NULL) {
1054 		if (xvdi_ring_has_incomp_request(vdp->xdf_xb_ring) ||
1055 		    xvdi_ring_has_unconsumed_responses(vdp->xdf_xb_ring))
1056 			rv = EIO;
1057 	}
1058 	if (xdf_debug & SUSRES_DBG)
1059 		xen_printf("xdf@%s: xdf_ring_drain: end, err=%d\n",
1060 		    vdp->xdf_addr, rv);
1061 	return (rv);
1062 }
1063 
1064 static int
1065 xdf_ring_drain(xdf_t *vdp)
1066 {
1067 	int rv;
1068 	mutex_enter(&vdp->xdf_dev_lk);
1069 	rv = xdf_ring_drain_locked(vdp);
1070 	mutex_exit(&vdp->xdf_dev_lk);
1071 	return (rv);
1072 }
1073 
1074 /*
1075  * Destroy all v_req_t, grant table entries, and our ring buffer.
1076  */
1077 static void
1078 xdf_ring_destroy(xdf_t *vdp)
1079 {
1080 	v_req_t		*vreq;
1081 	buf_t		*bp;
1082 	ge_slot_t	*gs;
1083 
1084 	ASSERT(MUTEX_HELD(&vdp->xdf_cb_lk));
1085 	ASSERT(MUTEX_HELD(&vdp->xdf_dev_lk));
1086 
1087 	if ((vdp->xdf_state != XD_INIT) &&
1088 	    (vdp->xdf_state != XD_CONNECTED) &&
1089 	    (vdp->xdf_state != XD_READY)) {
1090 		ASSERT(vdp->xdf_xb_ring == NULL);
1091 		ASSERT(vdp->xdf_xb_ring_hdl == NULL);
1092 		ASSERT(vdp->xdf_peer == INVALID_DOMID);
1093 		ASSERT(vdp->xdf_evtchn == INVALID_EVTCHN);
1094 		ASSERT(list_is_empty(&vdp->xdf_vreq_act));
1095 		return;
1096 	}
1097 
1098 	/*
1099 	 * We don't want to recieve async notifications from the backend
1100 	 * when it finishes processing ring entries.
1101 	 */
1102 #ifdef XPV_HVM_DRIVER
1103 	ec_unbind_evtchn(vdp->xdf_evtchn);
1104 #else /* !XPV_HVM_DRIVER */
1105 	(void) ddi_remove_intr(vdp->xdf_dip, 0, NULL);
1106 #endif /* !XPV_HVM_DRIVER */
1107 
1108 	/*
1109 	 * Drain any requests in the ring.  We need to do this before we
1110 	 * can free grant table entries, because if active ring entries
1111 	 * point to grants, then the backend could be trying to access
1112 	 * those grants.
1113 	 */
1114 	(void) xdf_ring_drain_locked(vdp);
1115 
1116 	/* We're done talking to the backend so free up our event channel */
1117 	xvdi_free_evtchn(vdp->xdf_dip);
1118 	vdp->xdf_evtchn = INVALID_EVTCHN;
1119 
1120 	while ((vreq = list_head(&vdp->xdf_vreq_act)) != NULL) {
1121 		bp = vreq->v_buf;
1122 		ASSERT(BP_VREQ(bp) == vreq);
1123 
1124 		/* Free up any grant table entries associaed with this IO */
1125 		while ((gs = list_head(&vreq->v_gs)) != NULL)
1126 			gs_free(gs);
1127 
1128 		/* If this IO was on the runq, move it back to the waitq. */
1129 		if (vreq->v_runq)
1130 			xdf_kstat_runq_to_waitq(vdp, bp);
1131 
1132 		/*
1133 		 * Reset any buf IO state since we're going to re-issue the
1134 		 * IO when we reconnect.
1135 		 */
1136 		vreq_free(vdp, vreq);
1137 		BP_VREQ_SET(bp, NULL);
1138 		bioerror(bp, 0);
1139 	}
1140 
1141 	/* reset the active queue index pointer */
1142 	vdp->xdf_i_act = vdp->xdf_f_act;
1143 
1144 	/* Destroy the ring */
1145 	xvdi_free_ring(vdp->xdf_xb_ring);
1146 	vdp->xdf_xb_ring = NULL;
1147 	vdp->xdf_xb_ring_hdl = NULL;
1148 	vdp->xdf_peer = INVALID_DOMID;
1149 }
1150 
1151 void
1152 xdfmin(struct buf *bp)
1153 {
1154 	if (bp->b_bcount > xdf_maxphys)
1155 		bp->b_bcount = xdf_maxphys;
1156 }
1157 
1158 /*
1159  * Check if we have a pending "eject" media request.
1160  */
1161 static int
1162 xdf_eject_pending(xdf_t *vdp)
1163 {
1164 	dev_info_t	*dip = vdp->xdf_dip;
1165 	char		*xsname, *str;
1166 
1167 	if (!vdp->xdf_media_req_supported)
1168 		return (B_FALSE);
1169 
1170 	if (((xsname = xvdi_get_xsname(dip)) == NULL) ||
1171 	    (xenbus_read_str(xsname, XBP_MEDIA_REQ, &str) != 0))
1172 		return (B_FALSE);
1173 
1174 	if (strcmp(str, XBV_MEDIA_REQ_EJECT) != 0) {
1175 		strfree(str);
1176 		return (B_FALSE);
1177 	}
1178 	strfree(str);
1179 	return (B_TRUE);
1180 }
1181 
1182 /*
1183  * Generate a media request.
1184  */
1185 static int
1186 xdf_media_req(xdf_t *vdp, char *req, boolean_t media_required)
1187 {
1188 	dev_info_t	*dip = vdp->xdf_dip;
1189 	char		*xsname;
1190 
1191 	/*
1192 	 * we can't be holding xdf_dev_lk because xenbus_printf() can
1193 	 * block while waiting for a PIL 1 interrupt message.  this
1194 	 * would cause a deadlock with xdf_intr() which needs to grab
1195 	 * xdf_dev_lk as well and runs at PIL 5.
1196 	 */
1197 	ASSERT(MUTEX_HELD(&vdp->xdf_cb_lk));
1198 	ASSERT(MUTEX_NOT_HELD(&vdp->xdf_dev_lk));
1199 
1200 	if ((xsname = xvdi_get_xsname(dip)) == NULL)
1201 		return (ENXIO);
1202 
1203 	/* Check if we support media requests */
1204 	if (!XD_IS_CD(vdp) || !vdp->xdf_media_req_supported)
1205 		return (ENOTTY);
1206 
1207 	/* If an eject is pending then don't allow any new requests */
1208 	if (xdf_eject_pending(vdp))
1209 		return (ENXIO);
1210 
1211 	/* Make sure that there is media present */
1212 	if (media_required && (vdp->xdf_xdev_nblocks == 0))
1213 		return (ENXIO);
1214 
1215 	/* We only allow operations when the device is ready and connected */
1216 	if (vdp->xdf_state != XD_READY)
1217 		return (EIO);
1218 
1219 	if (xenbus_printf(XBT_NULL, xsname, XBP_MEDIA_REQ, "%s", req) != 0)
1220 		return (EIO);
1221 
1222 	return (0);
1223 }
1224 
1225 /*
1226  * populate a single blkif_request_t w/ a buf
1227  */
1228 static void
1229 xdf_process_rreq(xdf_t *vdp, struct buf *bp, blkif_request_t *rreq)
1230 {
1231 	grant_ref_t	gr;
1232 	uint8_t		fsect, lsect;
1233 	size_t		bcnt;
1234 	paddr_t		dma_addr;
1235 	off_t		blk_off;
1236 	dev_info_t	*dip = vdp->xdf_dip;
1237 	blkif_vdev_t	vdev = xvdi_get_vdevnum(dip);
1238 	v_req_t		*vreq = BP_VREQ(bp);
1239 	uint64_t	blkno = vreq->v_blkno;
1240 	uint_t		ndmacs = vreq->v_ndmacs;
1241 	ddi_acc_handle_t acchdl = vdp->xdf_xb_ring_hdl;
1242 	int		seg = 0;
1243 	int		isread = IS_READ(bp);
1244 	ge_slot_t	*gs = list_head(&vreq->v_gs);
1245 
1246 	ASSERT(MUTEX_HELD(&vdp->xdf_dev_lk));
1247 	ASSERT(vreq->v_status == VREQ_GS_ALLOCED);
1248 
1249 	if (isread)
1250 		ddi_put8(acchdl, &rreq->operation, BLKIF_OP_READ);
1251 	else {
1252 		switch (vreq->v_flush_diskcache) {
1253 		case FLUSH_DISKCACHE:
1254 			ddi_put8(acchdl, &rreq->operation,
1255 			    BLKIF_OP_FLUSH_DISKCACHE);
1256 			ddi_put16(acchdl, &rreq->handle, vdev);
1257 			ddi_put64(acchdl, &rreq->id,
1258 			    (uint64_t)(uintptr_t)(gs));
1259 			ddi_put8(acchdl, &rreq->nr_segments, 0);
1260 			vreq->v_status = VREQ_DMAWIN_DONE;
1261 			return;
1262 		case WRITE_BARRIER:
1263 			ddi_put8(acchdl, &rreq->operation,
1264 			    BLKIF_OP_WRITE_BARRIER);
1265 			break;
1266 		default:
1267 			if (!vdp->xdf_wce)
1268 				ddi_put8(acchdl, &rreq->operation,
1269 				    BLKIF_OP_WRITE_BARRIER);
1270 			else
1271 				ddi_put8(acchdl, &rreq->operation,
1272 				    BLKIF_OP_WRITE);
1273 			break;
1274 		}
1275 	}
1276 
1277 	ddi_put16(acchdl, &rreq->handle, vdev);
1278 	ddi_put64(acchdl, &rreq->sector_number, blkno);
1279 	ddi_put64(acchdl, &rreq->id, (uint64_t)(uintptr_t)(gs));
1280 
1281 	/*
1282 	 * loop until all segments are populated or no more dma cookie in buf
1283 	 */
1284 	for (;;) {
1285 		/*
1286 		 * Each segment of a blkif request can transfer up to
1287 		 * one 4K page of data.
1288 		 */
1289 		bcnt = vreq->v_dmac.dmac_size;
1290 		dma_addr = vreq->v_dmac.dmac_laddress;
1291 		blk_off = (uint_t)((paddr_t)XB_SEGOFFSET & dma_addr);
1292 		fsect = blk_off >> XB_BSHIFT;
1293 		lsect = fsect + (bcnt >> XB_BSHIFT) - 1;
1294 
1295 		ASSERT(bcnt <= PAGESIZE);
1296 		ASSERT((bcnt % XB_BSIZE) == 0);
1297 		ASSERT((blk_off & XB_BMASK) == 0);
1298 		ASSERT(fsect < XB_MAX_SEGLEN / XB_BSIZE &&
1299 		    lsect < XB_MAX_SEGLEN / XB_BSIZE);
1300 
1301 		gr = gs_grant(gs, PATOMA(dma_addr) >> PAGESHIFT);
1302 		ddi_put32(acchdl, &rreq->seg[seg].gref, gr);
1303 		ddi_put8(acchdl, &rreq->seg[seg].first_sect, fsect);
1304 		ddi_put8(acchdl, &rreq->seg[seg].last_sect, lsect);
1305 
1306 		DPRINTF(IO_DBG, (
1307 		    "xdf@%s: seg%d: dmacS %lu blk_off %ld\n",
1308 		    vdp->xdf_addr, seg, vreq->v_dmac.dmac_size, blk_off));
1309 		DPRINTF(IO_DBG, (
1310 		    "xdf@%s: seg%d: fs %d ls %d gr %d dma 0x%"PRIx64"\n",
1311 		    vdp->xdf_addr, seg, fsect, lsect, gr, dma_addr));
1312 
1313 		blkno += (bcnt >> XB_BSHIFT);
1314 		seg++;
1315 		ASSERT(seg <= BLKIF_MAX_SEGMENTS_PER_REQUEST);
1316 		if (--ndmacs) {
1317 			ddi_dma_nextcookie(vreq->v_dmahdl, &vreq->v_dmac);
1318 			continue;
1319 		}
1320 
1321 		vreq->v_status = VREQ_DMAWIN_DONE;
1322 		vreq->v_blkno = blkno;
1323 		break;
1324 	}
1325 	ddi_put8(acchdl,  &rreq->nr_segments, seg);
1326 	DPRINTF(IO_DBG, (
1327 	    "xdf@%s: xdf_process_rreq: request id=%"PRIx64" ready\n",
1328 	    vdp->xdf_addr, rreq->id));
1329 }
1330 
1331 static void
1332 xdf_io_start(xdf_t *vdp)
1333 {
1334 	struct buf	*bp;
1335 	v_req_t		*vreq;
1336 	blkif_request_t	*rreq;
1337 	boolean_t	rreqready = B_FALSE;
1338 
1339 	mutex_enter(&vdp->xdf_dev_lk);
1340 
1341 	/*
1342 	 * Populate the ring request(s).  Loop until there is no buf to
1343 	 * transfer or no free slot available in I/O ring.
1344 	 */
1345 	for (;;) {
1346 		/* don't start any new IO if we're suspending */
1347 		if (vdp->xdf_suspending)
1348 			break;
1349 		if ((bp = xdf_bp_next(vdp)) == NULL)
1350 			break;
1351 
1352 		/* if the buf doesn't already have a vreq, allocate one */
1353 		if (((vreq = BP_VREQ(bp)) == NULL) &&
1354 		    ((vreq = vreq_get(vdp, bp)) == NULL))
1355 			break;
1356 
1357 		/* alloc DMA/GTE resources */
1358 		if (vreq_setup(vdp, vreq) != DDI_SUCCESS)
1359 			break;
1360 
1361 		/* get next blkif_request in the ring */
1362 		if ((rreq = xvdi_ring_get_request(vdp->xdf_xb_ring)) == NULL)
1363 			break;
1364 		bzero(rreq, sizeof (blkif_request_t));
1365 		rreqready = B_TRUE;
1366 
1367 		/* populate blkif_request with this buf */
1368 		xdf_process_rreq(vdp, bp, rreq);
1369 
1370 		/*
1371 		 * This buffer/vreq pair is has been allocated a ring buffer
1372 		 * resources, so if it isn't already in our runq, add it.
1373 		 */
1374 		if (!vreq->v_runq)
1375 			xdf_kstat_waitq_to_runq(vdp, bp);
1376 	}
1377 
1378 	/* Send the request(s) to the backend */
1379 	if (rreqready)
1380 		xdf_ring_push(vdp);
1381 
1382 	mutex_exit(&vdp->xdf_dev_lk);
1383 }
1384 
1385 
1386 /* check if partition is open, -1 - check all partitions on the disk */
1387 static boolean_t
1388 xdf_isopen(xdf_t *vdp, int partition)
1389 {
1390 	int i;
1391 	ulong_t parbit;
1392 	boolean_t rval = B_FALSE;
1393 
1394 	ASSERT((partition == -1) ||
1395 	    ((partition >= 0) || (partition < XDF_PEXT)));
1396 
1397 	if (partition == -1)
1398 		parbit = (ulong_t)-1;
1399 	else
1400 		parbit = 1 << partition;
1401 
1402 	for (i = 0; i < OTYPCNT; i++) {
1403 		if (vdp->xdf_vd_open[i] & parbit)
1404 			rval = B_TRUE;
1405 	}
1406 
1407 	return (rval);
1408 }
1409 
1410 /*
1411  * The connection should never be closed as long as someone is holding
1412  * us open, there is pending IO, or someone is waiting waiting for a
1413  * connection.
1414  */
1415 static boolean_t
1416 xdf_busy(xdf_t *vdp)
1417 {
1418 	ASSERT(MUTEX_HELD(&vdp->xdf_dev_lk));
1419 
1420 	if ((vdp->xdf_xb_ring != NULL) &&
1421 	    xvdi_ring_has_unconsumed_responses(vdp->xdf_xb_ring)) {
1422 		ASSERT(vdp->xdf_state != XD_CLOSED);
1423 		return (B_TRUE);
1424 	}
1425 
1426 	if (!list_is_empty(&vdp->xdf_vreq_act) || (vdp->xdf_f_act != NULL)) {
1427 		ASSERT(vdp->xdf_state != XD_CLOSED);
1428 		return (B_TRUE);
1429 	}
1430 
1431 	if (xdf_isopen(vdp, -1)) {
1432 		ASSERT(vdp->xdf_state != XD_CLOSED);
1433 		return (B_TRUE);
1434 	}
1435 
1436 	if (vdp->xdf_connect_req > 0) {
1437 		ASSERT(vdp->xdf_state != XD_CLOSED);
1438 		return (B_TRUE);
1439 	}
1440 
1441 	return (B_FALSE);
1442 }
1443 
1444 static void
1445 xdf_set_state(xdf_t *vdp, xdf_state_t new_state)
1446 {
1447 	ASSERT(MUTEX_HELD(&vdp->xdf_cb_lk));
1448 	ASSERT(MUTEX_HELD(&vdp->xdf_dev_lk));
1449 	DPRINTF(DDI_DBG, ("xdf@%s: state change %d -> %d\n",
1450 	    vdp->xdf_addr, vdp->xdf_state, new_state));
1451 	vdp->xdf_state = new_state;
1452 	cv_broadcast(&vdp->xdf_dev_cv);
1453 }
1454 
1455 static void
1456 xdf_disconnect(xdf_t *vdp, xdf_state_t new_state, boolean_t quiet)
1457 {
1458 	dev_info_t	*dip = vdp->xdf_dip;
1459 	boolean_t	busy;
1460 
1461 	ASSERT(MUTEX_HELD(&vdp->xdf_cb_lk));
1462 	ASSERT(MUTEX_NOT_HELD(&vdp->xdf_dev_lk));
1463 	ASSERT((new_state == XD_UNKNOWN) || (new_state == XD_CLOSED));
1464 
1465 	/* Check if we're already there. */
1466 	if (vdp->xdf_state == new_state)
1467 		return;
1468 
1469 	mutex_enter(&vdp->xdf_dev_lk);
1470 	busy = xdf_busy(vdp);
1471 
1472 	/* If we're already closed then there's nothing todo. */
1473 	if (vdp->xdf_state == XD_CLOSED) {
1474 		ASSERT(!busy);
1475 		xdf_set_state(vdp, new_state);
1476 		mutex_exit(&vdp->xdf_dev_lk);
1477 		return;
1478 	}
1479 
1480 #ifdef DEBUG
1481 	/* UhOh.  Warn the user that something bad has happened. */
1482 	if (!quiet && busy && (vdp->xdf_state == XD_READY) &&
1483 	    (vdp->xdf_xdev_nblocks != 0)) {
1484 		cmn_err(CE_WARN, "xdf@%s: disconnected while in use",
1485 		    vdp->xdf_addr);
1486 	}
1487 #endif /* DEBUG */
1488 
1489 	xdf_ring_destroy(vdp);
1490 
1491 	/* If we're busy then we can only go into the unknown state */
1492 	xdf_set_state(vdp, (busy) ? XD_UNKNOWN : new_state);
1493 	mutex_exit(&vdp->xdf_dev_lk);
1494 
1495 	/* if we're closed now, let the other end know */
1496 	if (vdp->xdf_state == XD_CLOSED)
1497 		(void) xvdi_switch_state(dip, XBT_NULL, XenbusStateClosed);
1498 }
1499 
1500 
1501 /*
1502  * Kick-off connect process
1503  * Status should be XD_UNKNOWN or XD_CLOSED
1504  * On success, status will be changed to XD_INIT
1505  * On error, it will be changed to XD_UNKNOWN
1506  */
1507 static int
1508 xdf_setstate_init(xdf_t *vdp)
1509 {
1510 	dev_info_t		*dip = vdp->xdf_dip;
1511 	xenbus_transaction_t	xbt;
1512 	grant_ref_t		gref;
1513 	char			*xsname, *str;
1514 	int 			rv;
1515 
1516 	ASSERT(MUTEX_HELD(&vdp->xdf_cb_lk));
1517 	ASSERT(MUTEX_NOT_HELD(&vdp->xdf_dev_lk));
1518 	ASSERT((vdp->xdf_state == XD_UNKNOWN) ||
1519 	    (vdp->xdf_state == XD_CLOSED));
1520 
1521 	DPRINTF(DDI_DBG,
1522 	    ("xdf@%s: starting connection process\n", vdp->xdf_addr));
1523 
1524 	/*
1525 	 * If an eject is pending then don't allow a new connection.
1526 	 * (Only the backend can clear media request eject request.)
1527 	 */
1528 	if (xdf_eject_pending(vdp))
1529 		return (DDI_FAILURE);
1530 
1531 	if ((xsname = xvdi_get_xsname(dip)) == NULL)
1532 		goto errout;
1533 
1534 	if ((vdp->xdf_peer = xvdi_get_oeid(dip)) == INVALID_DOMID)
1535 		goto errout;
1536 
1537 	(void) xvdi_switch_state(dip, XBT_NULL, XenbusStateInitialising);
1538 
1539 	/*
1540 	 * Sanity check for the existance of the xenbus device-type property.
1541 	 * This property might not exist if we our xenbus device nodes was
1542 	 * force destroyed while we were still connected to the backend.
1543 	 */
1544 	if (xenbus_read_str(xsname, XBP_DEV_TYPE, &str) != 0)
1545 		goto errout;
1546 	strfree(str);
1547 
1548 	if (xvdi_alloc_evtchn(dip) != DDI_SUCCESS)
1549 		goto errout;
1550 
1551 	vdp->xdf_evtchn = xvdi_get_evtchn(dip);
1552 #ifdef XPV_HVM_DRIVER
1553 	ec_bind_evtchn_to_handler(vdp->xdf_evtchn, IPL_VBD, xdf_intr, vdp);
1554 #else /* !XPV_HVM_DRIVER */
1555 	if (ddi_add_intr(dip, 0, NULL, NULL, xdf_intr, (caddr_t)vdp) !=
1556 	    DDI_SUCCESS) {
1557 		cmn_err(CE_WARN, "xdf@%s: xdf_setstate_init: "
1558 		    "failed to add intr handler", vdp->xdf_addr);
1559 		goto errout1;
1560 	}
1561 #endif /* !XPV_HVM_DRIVER */
1562 
1563 	if (xvdi_alloc_ring(dip, BLKIF_RING_SIZE,
1564 	    sizeof (union blkif_sring_entry), &gref, &vdp->xdf_xb_ring) !=
1565 	    DDI_SUCCESS) {
1566 		cmn_err(CE_WARN, "xdf@%s: failed to alloc comm ring",
1567 		    vdp->xdf_addr);
1568 		goto errout2;
1569 	}
1570 	vdp->xdf_xb_ring_hdl = vdp->xdf_xb_ring->xr_acc_hdl; /* ugly!! */
1571 
1572 	/*
1573 	 * Write into xenstore the info needed by backend
1574 	 */
1575 trans_retry:
1576 	if (xenbus_transaction_start(&xbt)) {
1577 		cmn_err(CE_WARN, "xdf@%s: failed to start transaction",
1578 		    vdp->xdf_addr);
1579 		xvdi_fatal_error(dip, EIO, "connect transaction init");
1580 		goto fail_trans;
1581 	}
1582 
1583 	/*
1584 	 * XBP_PROTOCOL is written by the domain builder in the case of PV
1585 	 * domains. However, it is not written for HVM domains, so let's
1586 	 * write it here.
1587 	 */
1588 	if (((rv = xenbus_printf(xbt, xsname,
1589 	    XBP_MEDIA_REQ, "%s", XBV_MEDIA_REQ_NONE)) != 0) ||
1590 	    ((rv = xenbus_printf(xbt, xsname,
1591 	    XBP_RING_REF, "%u", gref)) != 0) ||
1592 	    ((rv = xenbus_printf(xbt, xsname,
1593 	    XBP_EVENT_CHAN, "%u", vdp->xdf_evtchn)) != 0) ||
1594 	    ((rv = xenbus_printf(xbt, xsname,
1595 	    XBP_PROTOCOL, "%s", XEN_IO_PROTO_ABI_NATIVE)) != 0) ||
1596 	    ((rv = xvdi_switch_state(dip, xbt, XenbusStateInitialised)) > 0)) {
1597 		(void) xenbus_transaction_end(xbt, 1);
1598 		xvdi_fatal_error(dip, rv, "connect transaction setup");
1599 		goto fail_trans;
1600 	}
1601 
1602 	/* kick-off connect process */
1603 	if (rv = xenbus_transaction_end(xbt, 0)) {
1604 		if (rv == EAGAIN)
1605 			goto trans_retry;
1606 		xvdi_fatal_error(dip, rv, "connect transaction commit");
1607 		goto fail_trans;
1608 	}
1609 
1610 	ASSERT(MUTEX_HELD(&vdp->xdf_cb_lk));
1611 	mutex_enter(&vdp->xdf_dev_lk);
1612 	xdf_set_state(vdp, XD_INIT);
1613 	mutex_exit(&vdp->xdf_dev_lk);
1614 
1615 	return (DDI_SUCCESS);
1616 
1617 fail_trans:
1618 	xvdi_free_ring(vdp->xdf_xb_ring);
1619 errout2:
1620 #ifdef XPV_HVM_DRIVER
1621 	ec_unbind_evtchn(vdp->xdf_evtchn);
1622 #else /* !XPV_HVM_DRIVER */
1623 	(void) ddi_remove_intr(vdp->xdf_dip, 0, NULL);
1624 #endif /* !XPV_HVM_DRIVER */
1625 errout1:
1626 	xvdi_free_evtchn(dip);
1627 	vdp->xdf_evtchn = INVALID_EVTCHN;
1628 errout:
1629 	xdf_disconnect(vdp, XD_UNKNOWN, B_FALSE);
1630 	cmn_err(CE_WARN, "xdf@%s: failed to start connection to backend",
1631 	    vdp->xdf_addr);
1632 	return (DDI_FAILURE);
1633 }
1634 
1635 int
1636 xdf_get_flush_block(xdf_t *vdp)
1637 {
1638 	/*
1639 	 * Get a DEV_BSIZE aligned bufer
1640 	 */
1641 	vdp->xdf_flush_mem = kmem_alloc(DEV_BSIZE * 2, KM_SLEEP);
1642 	vdp->xdf_cache_flush_block =
1643 	    (char *)P2ROUNDUP((uintptr_t)(vdp->xdf_flush_mem), DEV_BSIZE);
1644 	if (xdf_lb_rdwr(vdp->xdf_dip, TG_READ, vdp->xdf_cache_flush_block,
1645 	    xdf_flush_block, DEV_BSIZE, NULL) != 0)
1646 		return (DDI_FAILURE);
1647 	return (DDI_SUCCESS);
1648 }
1649 
1650 static void
1651 xdf_setstate_ready(void *arg)
1652 {
1653 	xdf_t	*vdp = (xdf_t *)arg;
1654 
1655 	vdp->xdf_ready_tq_thread = curthread;
1656 
1657 	/*
1658 	 * We've created all the minor nodes via cmlb_attach() using default
1659 	 * value in xdf_attach() to make it possible to block in xdf_open(),
1660 	 * in case there's anyone (say, booting thread) ever trying to open
1661 	 * it before connected to backend. We will refresh all those minor
1662 	 * nodes w/ latest info we've got now when we are almost connected.
1663 	 */
1664 	mutex_enter(&vdp->xdf_dev_lk);
1665 	if (vdp->xdf_cmbl_reattach) {
1666 		vdp->xdf_cmbl_reattach = B_FALSE;
1667 
1668 		mutex_exit(&vdp->xdf_dev_lk);
1669 		if (xdf_cmlb_attach(vdp) != 0) {
1670 			xdf_disconnect(vdp, XD_UNKNOWN, B_FALSE);
1671 			return;
1672 		}
1673 		mutex_enter(&vdp->xdf_dev_lk);
1674 	}
1675 
1676 	/* If we're not still trying to get to the ready state, then bail. */
1677 	if (vdp->xdf_state != XD_CONNECTED) {
1678 		mutex_exit(&vdp->xdf_dev_lk);
1679 		return;
1680 	}
1681 	mutex_exit(&vdp->xdf_dev_lk);
1682 
1683 	/*
1684 	 * If backend has feature-barrier, see if it supports disk
1685 	 * cache flush op.
1686 	 */
1687 	vdp->xdf_flush_supported = B_FALSE;
1688 	if (vdp->xdf_feature_barrier) {
1689 		/*
1690 		 * Pretend we already know flush is supported so probe
1691 		 * will attempt the correct op.
1692 		 */
1693 		vdp->xdf_flush_supported = B_TRUE;
1694 		if (xdf_lb_rdwr(vdp->xdf_dip, TG_WRITE, NULL, 0, 0, 0) == 0) {
1695 			vdp->xdf_flush_supported = B_TRUE;
1696 		} else {
1697 			vdp->xdf_flush_supported = B_FALSE;
1698 			/*
1699 			 * If the other end does not support the cache flush op
1700 			 * then we must use a barrier-write to force disk
1701 			 * cache flushing.  Barrier writes require that a data
1702 			 * block actually be written.
1703 			 * Cache a block to barrier-write when we are
1704 			 * asked to perform a flush.
1705 			 * XXX - would it be better to just copy 1 block
1706 			 * (512 bytes) from whatever write we did last
1707 			 * and rewrite that block?
1708 			 */
1709 			if (xdf_get_flush_block(vdp) != DDI_SUCCESS) {
1710 				xdf_disconnect(vdp, XD_UNKNOWN, B_FALSE);
1711 				return;
1712 			}
1713 		}
1714 	}
1715 
1716 	mutex_enter(&vdp->xdf_cb_lk);
1717 	mutex_enter(&vdp->xdf_dev_lk);
1718 	if (vdp->xdf_state == XD_CONNECTED)
1719 		xdf_set_state(vdp, XD_READY);
1720 	mutex_exit(&vdp->xdf_dev_lk);
1721 
1722 	/* Restart any currently queued up io */
1723 	xdf_io_start(vdp);
1724 
1725 	mutex_exit(&vdp->xdf_cb_lk);
1726 }
1727 
1728 /*
1729  * synthetic geometry
1730  */
1731 #define	XDF_NSECTS	256
1732 #define	XDF_NHEADS	16
1733 
1734 static void
1735 xdf_synthetic_pgeom(dev_info_t *dip, cmlb_geom_t *geomp)
1736 {
1737 	xdf_t *vdp;
1738 	uint_t ncyl;
1739 
1740 	vdp = ddi_get_soft_state(xdf_ssp, ddi_get_instance(dip));
1741 
1742 	ncyl = vdp->xdf_xdev_nblocks / (XDF_NHEADS * XDF_NSECTS);
1743 
1744 	bzero(geomp, sizeof (*geomp));
1745 	geomp->g_ncyl = ncyl == 0 ? 1 : ncyl;
1746 	geomp->g_acyl = 0;
1747 	geomp->g_nhead = XDF_NHEADS;
1748 	geomp->g_nsect = XDF_NSECTS;
1749 	geomp->g_secsize = XB_BSIZE;
1750 	geomp->g_capacity = vdp->xdf_xdev_nblocks;
1751 	geomp->g_intrlv = 0;
1752 	geomp->g_rpm = 7200;
1753 }
1754 
1755 /*
1756  * Finish other initialization after we've connected to backend
1757  * Status should be XD_INIT before calling this routine
1758  * On success, status should be changed to XD_CONNECTED.
1759  * On error, status should stay XD_INIT
1760  */
1761 static int
1762 xdf_setstate_connected(xdf_t *vdp)
1763 {
1764 	dev_info_t	*dip = vdp->xdf_dip;
1765 	cmlb_geom_t	pgeom;
1766 	diskaddr_t	nblocks = 0;
1767 	char		*oename, *xsname, *str;
1768 	uint_t		dinfo;
1769 
1770 	ASSERT(MUTEX_HELD(&vdp->xdf_cb_lk));
1771 	ASSERT(MUTEX_NOT_HELD(&vdp->xdf_dev_lk));
1772 	ASSERT(vdp->xdf_state == XD_INIT);
1773 
1774 	if (((xsname = xvdi_get_xsname(dip)) == NULL) ||
1775 	    ((oename = xvdi_get_oename(dip)) == NULL))
1776 		return (DDI_FAILURE);
1777 
1778 	/* Make sure the other end is XenbusStateConnected */
1779 	if (xenbus_read_driver_state(oename) != XenbusStateConnected)
1780 		return (DDI_FAILURE);
1781 
1782 	/* Determine if feature barrier is supported by backend */
1783 	if (!(vdp->xdf_feature_barrier = xenbus_exists(oename, XBP_FB)))
1784 		cmn_err(CE_NOTE, "xdf@%s: failed to read feature-barrier",
1785 		    vdp->xdf_addr);
1786 
1787 	/*
1788 	 * Probe backend.  Read the device size into xdf_xdev_nblocks
1789 	 * and set the VDISK_READONLY, VDISK_CDROM, and VDISK_REMOVABLE
1790 	 * flags in xdf_dinfo.  If the emulated device type is "cdrom",
1791 	 * we always set VDISK_CDROM, regardless of if it's present in
1792 	 * the xenbus info parameter.
1793 	 */
1794 	if (xenbus_gather(XBT_NULL, oename,
1795 	    XBP_SECTORS, "%"SCNu64, &nblocks,
1796 	    XBP_INFO, "%u", &dinfo,
1797 	    NULL) != 0) {
1798 		cmn_err(CE_WARN, "xdf@%s: xdf_setstate_connected: "
1799 		    "cannot read backend info", vdp->xdf_addr);
1800 		return (DDI_FAILURE);
1801 	}
1802 	if (xenbus_read_str(xsname, XBP_DEV_TYPE, &str) != 0) {
1803 		cmn_err(CE_WARN, "xdf@%s: cannot read device-type",
1804 		    vdp->xdf_addr);
1805 		return (DDI_FAILURE);
1806 	}
1807 	if (strcmp(str, XBV_DEV_TYPE_CD) == 0)
1808 		dinfo |= VDISK_CDROM;
1809 	strfree(str);
1810 
1811 	vdp->xdf_xdev_nblocks = nblocks;
1812 #ifdef _ILP32
1813 	if (vdp->xdf_xdev_nblocks > DK_MAX_BLOCKS) {
1814 		cmn_err(CE_WARN, "xdf@%s: xdf_setstate_connected: "
1815 		    "backend disk device too large with %llu blocks for"
1816 		    " 32-bit kernel", vdp->xdf_addr, vdp->xdf_xdev_nblocks);
1817 		xvdi_fatal_error(dip, EFBIG, "reading backend info");
1818 		return (DDI_FAILURE);
1819 	}
1820 #endif
1821 
1822 	/*
1823 	 * If the physical geometry for a fixed disk has been explicity
1824 	 * set then make sure that the specified physical geometry isn't
1825 	 * larger than the device we connected to.
1826 	 */
1827 	if (vdp->xdf_pgeom_fixed &&
1828 	    (vdp->xdf_pgeom.g_capacity > vdp->xdf_xdev_nblocks)) {
1829 		cmn_err(CE_WARN,
1830 		    "xdf@%s: connect failed, fixed geometry too large",
1831 		    vdp->xdf_addr);
1832 		return (DDI_FAILURE);
1833 	}
1834 
1835 	vdp->xdf_media_req_supported = xenbus_exists(oename, XBP_MEDIA_REQ_SUP);
1836 
1837 	/* mark vbd is ready for I/O */
1838 	mutex_enter(&vdp->xdf_dev_lk);
1839 	xdf_set_state(vdp, XD_CONNECTED);
1840 
1841 	/* check if the cmlb label should be updated */
1842 	xdf_synthetic_pgeom(dip, &pgeom);
1843 	if ((vdp->xdf_dinfo != dinfo) ||
1844 	    (!vdp->xdf_pgeom_fixed &&
1845 	    (memcmp(&vdp->xdf_pgeom, &pgeom, sizeof (pgeom)) != 0))) {
1846 		vdp->xdf_cmbl_reattach = B_TRUE;
1847 
1848 		vdp->xdf_dinfo = dinfo;
1849 		if (!vdp->xdf_pgeom_fixed)
1850 			vdp->xdf_pgeom = pgeom;
1851 	}
1852 
1853 	if (XD_IS_CD(vdp) || XD_IS_RM(vdp)) {
1854 		if (vdp->xdf_xdev_nblocks == 0) {
1855 			vdp->xdf_mstate = DKIO_EJECTED;
1856 			cv_broadcast(&vdp->xdf_mstate_cv);
1857 		} else {
1858 			vdp->xdf_mstate = DKIO_INSERTED;
1859 			cv_broadcast(&vdp->xdf_mstate_cv);
1860 		}
1861 	} else {
1862 		if (vdp->xdf_mstate != DKIO_NONE) {
1863 			vdp->xdf_mstate = DKIO_NONE;
1864 			cv_broadcast(&vdp->xdf_mstate_cv);
1865 		}
1866 	}
1867 
1868 	mutex_exit(&vdp->xdf_dev_lk);
1869 
1870 	cmn_err(CE_CONT, "?xdf@%s: %"PRIu64" blocks", vdp->xdf_addr,
1871 	    (uint64_t)vdp->xdf_xdev_nblocks);
1872 
1873 	/* Restart any currently queued up io */
1874 	xdf_io_start(vdp);
1875 
1876 	/*
1877 	 * To get to the ready state we have to do IO to the backend device,
1878 	 * but we can't initiate IO from the other end change callback thread
1879 	 * (which is the current context we're executing in.)  This is because
1880 	 * if the other end disconnects while we're doing IO from the callback
1881 	 * thread, then we can't recieve that disconnect event and we hang
1882 	 * waiting for an IO that can never complete.
1883 	 */
1884 	(void) ddi_taskq_dispatch(vdp->xdf_ready_tq, xdf_setstate_ready, vdp,
1885 	    DDI_SLEEP);
1886 
1887 	(void) xvdi_switch_state(dip, XBT_NULL, XenbusStateConnected);
1888 	return (DDI_SUCCESS);
1889 }
1890 
1891 /*ARGSUSED*/
1892 static void
1893 xdf_oe_change(dev_info_t *dip, ddi_eventcookie_t id, void *arg, void *impl_data)
1894 {
1895 	XenbusState new_state = *(XenbusState *)impl_data;
1896 	xdf_t *vdp = (xdf_t *)ddi_get_driver_private(dip);
1897 
1898 	DPRINTF(DDI_DBG, ("xdf@%s: otherend state change to %d!\n",
1899 	    vdp->xdf_addr, new_state));
1900 
1901 	mutex_enter(&vdp->xdf_cb_lk);
1902 
1903 	/* We assume that this callback is single threaded */
1904 	ASSERT(vdp->xdf_oe_change_thread == NULL);
1905 	DEBUG_EVAL(vdp->xdf_oe_change_thread = curthread);
1906 
1907 	/* ignore any backend state changes if we're suspending/suspended */
1908 	if (vdp->xdf_suspending || (vdp->xdf_state == XD_SUSPEND)) {
1909 		DEBUG_EVAL(vdp->xdf_oe_change_thread = NULL);
1910 		mutex_exit(&vdp->xdf_cb_lk);
1911 		return;
1912 	}
1913 
1914 	switch (new_state) {
1915 	case XenbusStateUnknown:
1916 	case XenbusStateInitialising:
1917 	case XenbusStateInitWait:
1918 	case XenbusStateInitialised:
1919 		if (vdp->xdf_state == XD_INIT)
1920 			break;
1921 
1922 		xdf_disconnect(vdp, XD_UNKNOWN, B_FALSE);
1923 		if (xdf_setstate_init(vdp) != DDI_SUCCESS)
1924 			break;
1925 		ASSERT(vdp->xdf_state == XD_INIT);
1926 		break;
1927 
1928 	case XenbusStateConnected:
1929 		if ((vdp->xdf_state == XD_CONNECTED) ||
1930 		    (vdp->xdf_state == XD_READY))
1931 			break;
1932 
1933 		if (vdp->xdf_state != XD_INIT) {
1934 			xdf_disconnect(vdp, XD_UNKNOWN, B_FALSE);
1935 			if (xdf_setstate_init(vdp) != DDI_SUCCESS)
1936 				break;
1937 			ASSERT(vdp->xdf_state == XD_INIT);
1938 		}
1939 
1940 		if (xdf_setstate_connected(vdp) != DDI_SUCCESS) {
1941 			xdf_disconnect(vdp, XD_UNKNOWN, B_FALSE);
1942 			break;
1943 		}
1944 		ASSERT(vdp->xdf_state == XD_CONNECTED);
1945 		break;
1946 
1947 	case XenbusStateClosing:
1948 		if (xdf_isopen(vdp, -1)) {
1949 			cmn_err(CE_NOTE,
1950 			    "xdf@%s: hot-unplug failed, still in use",
1951 			    vdp->xdf_addr);
1952 			break;
1953 		}
1954 		/*FALLTHROUGH*/
1955 	case XenbusStateClosed:
1956 		xdf_disconnect(vdp, XD_CLOSED, B_FALSE);
1957 		break;
1958 	}
1959 
1960 	/* notify anybody waiting for oe state change */
1961 	cv_broadcast(&vdp->xdf_dev_cv);
1962 	DEBUG_EVAL(vdp->xdf_oe_change_thread = NULL);
1963 	mutex_exit(&vdp->xdf_cb_lk);
1964 }
1965 
1966 static int
1967 xdf_connect_locked(xdf_t *vdp, boolean_t wait)
1968 {
1969 	int	rv, timeouts = 0, reset = 20;
1970 
1971 	ASSERT(MUTEX_HELD(&vdp->xdf_cb_lk));
1972 	ASSERT(MUTEX_HELD(&vdp->xdf_dev_lk));
1973 
1974 	/* we can't connect once we're in the closed state */
1975 	if (vdp->xdf_state == XD_CLOSED)
1976 		return (XD_CLOSED);
1977 
1978 	vdp->xdf_connect_req++;
1979 	while (vdp->xdf_state != XD_READY) {
1980 		mutex_exit(&vdp->xdf_dev_lk);
1981 
1982 		/* only one thread at a time can be the connection thread */
1983 		if (vdp->xdf_connect_thread == NULL)
1984 			vdp->xdf_connect_thread = curthread;
1985 
1986 		if (vdp->xdf_connect_thread == curthread) {
1987 			if ((timeouts > 0) && ((timeouts % reset) == 0)) {
1988 				/*
1989 				 * If we haven't establised a connection
1990 				 * within the reset time, then disconnect
1991 				 * so we can try again, and double the reset
1992 				 * time.  The reset time starts at 2 sec.
1993 				 */
1994 				(void) xdf_disconnect(vdp, XD_UNKNOWN, B_TRUE);
1995 				reset *= 2;
1996 			}
1997 			if (vdp->xdf_state == XD_UNKNOWN)
1998 				(void) xdf_setstate_init(vdp);
1999 			if (vdp->xdf_state == XD_INIT)
2000 				(void) xdf_setstate_connected(vdp);
2001 		}
2002 
2003 		mutex_enter(&vdp->xdf_dev_lk);
2004 		if (!wait || (vdp->xdf_state == XD_READY))
2005 			goto out;
2006 
2007 		mutex_exit((&vdp->xdf_cb_lk));
2008 		if (vdp->xdf_connect_thread != curthread) {
2009 			rv = cv_wait_sig(&vdp->xdf_dev_cv, &vdp->xdf_dev_lk);
2010 		} else {
2011 			/* delay for 0.1 sec */
2012 			rv = cv_timedwait_sig(&vdp->xdf_dev_cv,
2013 			    &vdp->xdf_dev_lk, lbolt + drv_usectohz(100*1000));
2014 			if (rv == -1)
2015 				timeouts++;
2016 		}
2017 		mutex_exit((&vdp->xdf_dev_lk));
2018 		mutex_enter((&vdp->xdf_cb_lk));
2019 		mutex_enter((&vdp->xdf_dev_lk));
2020 		if (rv == 0)
2021 			goto out;
2022 	}
2023 
2024 out:
2025 	ASSERT(MUTEX_HELD(&vdp->xdf_cb_lk));
2026 	ASSERT(MUTEX_HELD(&vdp->xdf_dev_lk));
2027 
2028 	if (vdp->xdf_connect_thread == curthread) {
2029 		/*
2030 		 * wake up someone else so they can become the connection
2031 		 * thread.
2032 		 */
2033 		cv_signal(&vdp->xdf_dev_cv);
2034 		vdp->xdf_connect_thread = NULL;
2035 	}
2036 
2037 	/* Try to lock the media */
2038 	mutex_exit((&vdp->xdf_dev_lk));
2039 	(void) xdf_media_req(vdp, XBV_MEDIA_REQ_LOCK, B_TRUE);
2040 	mutex_enter((&vdp->xdf_dev_lk));
2041 
2042 	vdp->xdf_connect_req--;
2043 	return (vdp->xdf_state);
2044 }
2045 
2046 static uint_t
2047 xdf_iorestart(caddr_t arg)
2048 {
2049 	xdf_t *vdp = (xdf_t *)arg;
2050 
2051 	ASSERT(vdp != NULL);
2052 
2053 	mutex_enter(&vdp->xdf_dev_lk);
2054 	ASSERT(ISDMACBON(vdp));
2055 	SETDMACBOFF(vdp);
2056 	mutex_exit(&vdp->xdf_dev_lk);
2057 
2058 	xdf_io_start(vdp);
2059 
2060 	return (DDI_INTR_CLAIMED);
2061 }
2062 
2063 #if defined(XPV_HVM_DRIVER)
2064 
2065 typedef struct xdf_hvm_entry {
2066 	list_node_t	xdf_he_list;
2067 	char		*xdf_he_path;
2068 	dev_info_t	*xdf_he_dip;
2069 } xdf_hvm_entry_t;
2070 
2071 static list_t xdf_hvm_list;
2072 static kmutex_t xdf_hvm_list_lock;
2073 
2074 static xdf_hvm_entry_t *
2075 i_xdf_hvm_find(const char *path, dev_info_t *dip)
2076 {
2077 	xdf_hvm_entry_t	*i;
2078 
2079 	ASSERT((path != NULL) || (dip != NULL));
2080 	ASSERT(MUTEX_HELD(&xdf_hvm_list_lock));
2081 
2082 	i = list_head(&xdf_hvm_list);
2083 	while (i != NULL) {
2084 		if ((path != NULL) && strcmp(i->xdf_he_path, path) != 0) {
2085 			i = list_next(&xdf_hvm_list, i);
2086 			continue;
2087 		}
2088 		if ((dip != NULL) && (i->xdf_he_dip != dip)) {
2089 			i = list_next(&xdf_hvm_list, i);
2090 			continue;
2091 		}
2092 		break;
2093 	}
2094 	return (i);
2095 }
2096 
2097 dev_info_t *
2098 xdf_hvm_hold(const char *path)
2099 {
2100 	xdf_hvm_entry_t	*i;
2101 	dev_info_t	*dip;
2102 
2103 	mutex_enter(&xdf_hvm_list_lock);
2104 	i = i_xdf_hvm_find(path, NULL);
2105 	if (i == NULL) {
2106 		mutex_exit(&xdf_hvm_list_lock);
2107 		return (B_FALSE);
2108 	}
2109 	ndi_hold_devi(dip = i->xdf_he_dip);
2110 	mutex_exit(&xdf_hvm_list_lock);
2111 	return (dip);
2112 }
2113 
2114 static void
2115 xdf_hvm_add(dev_info_t *dip)
2116 {
2117 	xdf_hvm_entry_t	*i;
2118 	char		*path;
2119 
2120 	/* figure out the path for the dip */
2121 	path = kmem_zalloc(MAXPATHLEN, KM_SLEEP);
2122 	(void) ddi_pathname(dip, path);
2123 
2124 	i = kmem_alloc(sizeof (*i), KM_SLEEP);
2125 	i->xdf_he_dip = dip;
2126 	i->xdf_he_path = i_ddi_strdup(path, KM_SLEEP);
2127 
2128 	mutex_enter(&xdf_hvm_list_lock);
2129 	ASSERT(i_xdf_hvm_find(path, NULL) == NULL);
2130 	ASSERT(i_xdf_hvm_find(NULL, dip) == NULL);
2131 	list_insert_head(&xdf_hvm_list, i);
2132 	mutex_exit(&xdf_hvm_list_lock);
2133 
2134 	kmem_free(path, MAXPATHLEN);
2135 }
2136 
2137 static void
2138 xdf_hvm_rm(dev_info_t *dip)
2139 {
2140 	xdf_hvm_entry_t	*i;
2141 
2142 	mutex_enter(&xdf_hvm_list_lock);
2143 	VERIFY((i = i_xdf_hvm_find(NULL, dip)) != NULL);
2144 	list_remove(&xdf_hvm_list, i);
2145 	mutex_exit(&xdf_hvm_list_lock);
2146 
2147 	kmem_free(i->xdf_he_path, strlen(i->xdf_he_path) + 1);
2148 	kmem_free(i, sizeof (*i));
2149 }
2150 
2151 static void
2152 xdf_hvm_init(void)
2153 {
2154 	list_create(&xdf_hvm_list, sizeof (xdf_hvm_entry_t),
2155 	    offsetof(xdf_hvm_entry_t, xdf_he_list));
2156 	mutex_init(&xdf_hvm_list_lock, NULL, MUTEX_DEFAULT, NULL);
2157 }
2158 
2159 static void
2160 xdf_hvm_fini(void)
2161 {
2162 	ASSERT(list_head(&xdf_hvm_list) == NULL);
2163 	list_destroy(&xdf_hvm_list);
2164 	mutex_destroy(&xdf_hvm_list_lock);
2165 }
2166 
2167 boolean_t
2168 xdf_hvm_connect(dev_info_t *dip)
2169 {
2170 	xdf_t	*vdp = (xdf_t *)ddi_get_driver_private(dip);
2171 	char	*oename, *str;
2172 	int	rv;
2173 
2174 	mutex_enter(&vdp->xdf_cb_lk);
2175 
2176 	/*
2177 	 * Before try to establish a connection we need to wait for the
2178 	 * backend hotplug scripts to have run.  Once they are run the
2179 	 * "<oename>/hotplug-status" property will be set to "connected".
2180 	 */
2181 	for (;;) {
2182 		ASSERT(MUTEX_HELD(&vdp->xdf_cb_lk));
2183 
2184 		/*
2185 		 * Get the xenbus path to the backend device.  Note that
2186 		 * we can't cache this path (and we look it up on each pass
2187 		 * through this loop) because it could change during
2188 		 * suspend, resume, and migration operations.
2189 		 */
2190 		if ((oename = xvdi_get_oename(dip)) == NULL) {
2191 			mutex_exit(&vdp->xdf_cb_lk);
2192 			return (B_FALSE);
2193 		}
2194 
2195 		str = NULL;
2196 		if ((xenbus_read_str(oename, XBP_HP_STATUS, &str) == 0) &&
2197 		    (strcmp(str, XBV_HP_STATUS_CONN) == 0))
2198 			break;
2199 
2200 		if (str != NULL)
2201 			strfree(str);
2202 
2203 		/* wait for an update to "<oename>/hotplug-status" */
2204 		if (cv_wait_sig(&vdp->xdf_hp_status_cv, &vdp->xdf_cb_lk) == 0) {
2205 			/* we got interrupted by a signal */
2206 			mutex_exit(&vdp->xdf_cb_lk);
2207 			return (B_FALSE);
2208 		}
2209 	}
2210 
2211 	/* Good news.  The backend hotplug scripts have been run. */
2212 	ASSERT(MUTEX_HELD(&vdp->xdf_cb_lk));
2213 	ASSERT(strcmp(str, XBV_HP_STATUS_CONN) == 0);
2214 	strfree(str);
2215 
2216 	/*
2217 	 * If we're emulating a cd device and if the backend doesn't support
2218 	 * media request opreations, then we're not going to bother trying
2219 	 * to establish a connection for a couple reasons.  First off, media
2220 	 * requests support is required to support operations like eject and
2221 	 * media locking.  Second, other backend platforms like Linux don't
2222 	 * support hvm pv cdrom access.  They don't even have a backend pv
2223 	 * driver for cdrom device nodes, so we don't want to block forever
2224 	 * waiting for a connection to a backend driver that doesn't exist.
2225 	 */
2226 	if (XD_IS_CD(vdp) && !xenbus_exists(oename, XBP_MEDIA_REQ_SUP)) {
2227 		mutex_exit(&vdp->xdf_cb_lk);
2228 		return (B_FALSE);
2229 	}
2230 
2231 	mutex_enter(&vdp->xdf_dev_lk);
2232 	rv = xdf_connect_locked(vdp, B_TRUE);
2233 	mutex_exit(&vdp->xdf_dev_lk);
2234 	mutex_exit(&vdp->xdf_cb_lk);
2235 
2236 	return ((rv == XD_READY) ? B_TRUE : B_FALSE);
2237 }
2238 
2239 int
2240 xdf_hvm_setpgeom(dev_info_t *dip, cmlb_geom_t *geomp)
2241 {
2242 	xdf_t	*vdp = (xdf_t *)ddi_get_driver_private(dip);
2243 
2244 	/* sanity check the requested physical geometry */
2245 	mutex_enter(&vdp->xdf_dev_lk);
2246 	if ((geomp->g_secsize != XB_BSIZE) ||
2247 	    (geomp->g_capacity == 0)) {
2248 		mutex_exit(&vdp->xdf_dev_lk);
2249 		return (EINVAL);
2250 	}
2251 
2252 	/*
2253 	 * If we've already connected to the backend device then make sure
2254 	 * we're not defining a physical geometry larger than our backend
2255 	 * device.
2256 	 */
2257 	if ((vdp->xdf_xdev_nblocks != 0) &&
2258 	    (geomp->g_capacity > vdp->xdf_xdev_nblocks)) {
2259 		mutex_exit(&vdp->xdf_dev_lk);
2260 		return (EINVAL);
2261 	}
2262 
2263 	bzero(&vdp->xdf_pgeom, sizeof (vdp->xdf_pgeom));
2264 	vdp->xdf_pgeom.g_ncyl = geomp->g_ncyl;
2265 	vdp->xdf_pgeom.g_acyl = geomp->g_acyl;
2266 	vdp->xdf_pgeom.g_nhead = geomp->g_nhead;
2267 	vdp->xdf_pgeom.g_nsect = geomp->g_nsect;
2268 	vdp->xdf_pgeom.g_secsize = geomp->g_secsize;
2269 	vdp->xdf_pgeom.g_capacity = geomp->g_capacity;
2270 	vdp->xdf_pgeom.g_intrlv = geomp->g_intrlv;
2271 	vdp->xdf_pgeom.g_rpm = geomp->g_rpm;
2272 
2273 	vdp->xdf_pgeom_fixed = B_TRUE;
2274 	mutex_exit(&vdp->xdf_dev_lk);
2275 
2276 	/* force a re-validation */
2277 	cmlb_invalidate(vdp->xdf_vd_lbl, NULL);
2278 
2279 	return (0);
2280 }
2281 
2282 boolean_t
2283 xdf_is_cd(dev_info_t *dip)
2284 {
2285 	xdf_t		*vdp = (xdf_t *)ddi_get_driver_private(dip);
2286 	boolean_t	rv;
2287 
2288 	mutex_enter(&vdp->xdf_cb_lk);
2289 	rv = XD_IS_CD(vdp);
2290 	mutex_exit(&vdp->xdf_cb_lk);
2291 	return (rv);
2292 }
2293 
2294 boolean_t
2295 xdf_is_rm(dev_info_t *dip)
2296 {
2297 	xdf_t		*vdp = (xdf_t *)ddi_get_driver_private(dip);
2298 	boolean_t	rv;
2299 
2300 	mutex_enter(&vdp->xdf_cb_lk);
2301 	rv = XD_IS_RM(vdp);
2302 	mutex_exit(&vdp->xdf_cb_lk);
2303 	return (rv);
2304 }
2305 
2306 boolean_t
2307 xdf_media_req_supported(dev_info_t *dip)
2308 {
2309 	xdf_t		*vdp = (xdf_t *)ddi_get_driver_private(dip);
2310 	boolean_t	rv;
2311 
2312 	mutex_enter(&vdp->xdf_cb_lk);
2313 	rv = vdp->xdf_media_req_supported;
2314 	mutex_exit(&vdp->xdf_cb_lk);
2315 	return (rv);
2316 }
2317 
2318 #endif /* XPV_HVM_DRIVER */
2319 
2320 static int
2321 xdf_lb_getcap(dev_info_t *dip, diskaddr_t *capp)
2322 {
2323 	xdf_t *vdp;
2324 	vdp = ddi_get_soft_state(xdf_ssp, ddi_get_instance(dip));
2325 
2326 	if (vdp == NULL)
2327 		return (ENXIO);
2328 
2329 	mutex_enter(&vdp->xdf_dev_lk);
2330 	*capp = vdp->xdf_pgeom.g_capacity;
2331 	DPRINTF(LBL_DBG, ("xdf@%s:capacity %llu\n", vdp->xdf_addr, *capp));
2332 	mutex_exit(&vdp->xdf_dev_lk);
2333 	return (0);
2334 }
2335 
2336 static int
2337 xdf_lb_getpgeom(dev_info_t *dip, cmlb_geom_t *geomp)
2338 {
2339 	xdf_t *vdp;
2340 
2341 	if ((vdp = ddi_get_soft_state(xdf_ssp, ddi_get_instance(dip))) == NULL)
2342 		return (ENXIO);
2343 	*geomp = vdp->xdf_pgeom;
2344 	return (0);
2345 }
2346 
2347 /*
2348  * No real HBA, no geometry available from it
2349  */
2350 /*ARGSUSED*/
2351 static int
2352 xdf_lb_getvgeom(dev_info_t *dip, cmlb_geom_t *geomp)
2353 {
2354 	return (EINVAL);
2355 }
2356 
2357 static int
2358 xdf_lb_getattribute(dev_info_t *dip, tg_attribute_t *tgattributep)
2359 {
2360 	xdf_t *vdp;
2361 
2362 	if (!(vdp = ddi_get_soft_state(xdf_ssp, ddi_get_instance(dip))))
2363 		return (ENXIO);
2364 
2365 	if (XD_IS_RO(vdp))
2366 		tgattributep->media_is_writable = 0;
2367 	else
2368 		tgattributep->media_is_writable = 1;
2369 	return (0);
2370 }
2371 
2372 /* ARGSUSED3 */
2373 int
2374 xdf_lb_getinfo(dev_info_t *dip, int cmd, void *arg, void *tg_cookie)
2375 {
2376 	switch (cmd) {
2377 	case TG_GETPHYGEOM:
2378 		return (xdf_lb_getpgeom(dip, (cmlb_geom_t *)arg));
2379 	case TG_GETVIRTGEOM:
2380 		return (xdf_lb_getvgeom(dip, (cmlb_geom_t *)arg));
2381 	case TG_GETCAPACITY:
2382 		return (xdf_lb_getcap(dip, (diskaddr_t *)arg));
2383 	case TG_GETBLOCKSIZE:
2384 		*(uint32_t *)arg = XB_BSIZE;
2385 		return (0);
2386 	case TG_GETATTR:
2387 		return (xdf_lb_getattribute(dip, (tg_attribute_t *)arg));
2388 	default:
2389 		return (ENOTTY);
2390 	}
2391 }
2392 
2393 /* ARGSUSED5 */
2394 int
2395 xdf_lb_rdwr(dev_info_t *dip, uchar_t cmd, void *bufp,
2396     diskaddr_t start, size_t reqlen, void *tg_cookie)
2397 {
2398 	xdf_t *vdp;
2399 	struct buf *bp;
2400 	int err = 0;
2401 
2402 	vdp = ddi_get_soft_state(xdf_ssp, ddi_get_instance(dip));
2403 
2404 	/* We don't allow IO from the oe_change callback thread */
2405 	ASSERT(curthread != vdp->xdf_oe_change_thread);
2406 
2407 	if ((start + (reqlen >> DEV_BSHIFT)) > vdp->xdf_pgeom.g_capacity)
2408 		return (EINVAL);
2409 
2410 	bp = getrbuf(KM_SLEEP);
2411 	if (cmd == TG_READ)
2412 		bp->b_flags = B_BUSY | B_READ;
2413 	else
2414 		bp->b_flags = B_BUSY | B_WRITE;
2415 	bp->b_un.b_addr = bufp;
2416 	bp->b_bcount = reqlen;
2417 	bp->b_blkno = start;
2418 	bp->b_edev = DDI_DEV_T_NONE; /* don't have dev_t */
2419 
2420 	mutex_enter(&vdp->xdf_dev_lk);
2421 	xdf_bp_push(vdp, bp);
2422 	mutex_exit(&vdp->xdf_dev_lk);
2423 	xdf_io_start(vdp);
2424 	if (curthread == vdp->xdf_ready_tq_thread)
2425 		(void) xdf_ring_drain(vdp);
2426 	err = biowait(bp);
2427 	ASSERT(bp->b_flags & B_DONE);
2428 	freerbuf(bp);
2429 	return (err);
2430 }
2431 
2432 /*
2433  * Lock the current media.  Set the media state to "lock".
2434  * (Media locks are only respected by the backend driver.)
2435  */
2436 static int
2437 xdf_ioctl_mlock(xdf_t *vdp)
2438 {
2439 	int rv;
2440 	mutex_enter(&vdp->xdf_cb_lk);
2441 	rv = xdf_media_req(vdp, XBV_MEDIA_REQ_LOCK, B_TRUE);
2442 	mutex_exit(&vdp->xdf_cb_lk);
2443 	return (rv);
2444 }
2445 
2446 /*
2447  * Release a media lock.  Set the media state to "none".
2448  */
2449 static int
2450 xdf_ioctl_munlock(xdf_t *vdp)
2451 {
2452 	int rv;
2453 	mutex_enter(&vdp->xdf_cb_lk);
2454 	rv = xdf_media_req(vdp, XBV_MEDIA_REQ_NONE, B_TRUE);
2455 	mutex_exit(&vdp->xdf_cb_lk);
2456 	return (rv);
2457 }
2458 
2459 /*
2460  * Eject the current media.  Ignores any media locks.  (Media locks
2461  * are only for benifit of the the backend.)
2462  */
2463 static int
2464 xdf_ioctl_eject(xdf_t *vdp)
2465 {
2466 	int rv;
2467 
2468 	mutex_enter(&vdp->xdf_cb_lk);
2469 	if ((rv = xdf_media_req(vdp, XBV_MEDIA_REQ_EJECT, B_FALSE)) != 0) {
2470 		mutex_exit(&vdp->xdf_cb_lk);
2471 		return (rv);
2472 	}
2473 
2474 	/*
2475 	 * We've set the media requests xenbus parameter to eject, so now
2476 	 * disconnect from the backend, wait for the backend to clear
2477 	 * the media requets xenbus paramter, and then we can reconnect
2478 	 * to the backend.
2479 	 */
2480 	(void) xdf_disconnect(vdp, XD_UNKNOWN, B_TRUE);
2481 	mutex_enter(&vdp->xdf_dev_lk);
2482 	if (xdf_connect_locked(vdp, B_TRUE) != XD_READY) {
2483 		mutex_exit(&vdp->xdf_dev_lk);
2484 		mutex_exit(&vdp->xdf_cb_lk);
2485 		return (EIO);
2486 	}
2487 	mutex_exit(&vdp->xdf_dev_lk);
2488 	mutex_exit(&vdp->xdf_cb_lk);
2489 	return (0);
2490 }
2491 
2492 /*
2493  * Watch for media state changes.  This can be an insertion of a device
2494  * (triggered by a 'xm block-configure' request in another domain) or
2495  * the ejection of a device (triggered by a local "eject" operation).
2496  * For a full description of the DKIOCSTATE ioctl behavior see dkio(7I).
2497  */
2498 static int
2499 xdf_dkstate(xdf_t *vdp, enum dkio_state mstate)
2500 {
2501 	enum dkio_state		prev_state;
2502 
2503 	mutex_enter(&vdp->xdf_cb_lk);
2504 	prev_state = vdp->xdf_mstate;
2505 
2506 	if (vdp->xdf_mstate == mstate) {
2507 		while (vdp->xdf_mstate == prev_state) {
2508 			if (cv_wait_sig(&vdp->xdf_mstate_cv,
2509 			    &vdp->xdf_cb_lk) == 0) {
2510 				mutex_exit(&vdp->xdf_cb_lk);
2511 				return (EINTR);
2512 			}
2513 		}
2514 	}
2515 
2516 	if ((prev_state != DKIO_INSERTED) &&
2517 	    (vdp->xdf_mstate == DKIO_INSERTED)) {
2518 		(void) xdf_media_req(vdp, XBV_MEDIA_REQ_LOCK, B_TRUE);
2519 		mutex_exit(&vdp->xdf_cb_lk);
2520 		return (0);
2521 	}
2522 
2523 	mutex_exit(&vdp->xdf_cb_lk);
2524 	return (0);
2525 }
2526 
2527 /*ARGSUSED*/
2528 static int
2529 xdf_ioctl(dev_t dev, int cmd, intptr_t arg, int mode, cred_t *credp,
2530     int *rvalp)
2531 {
2532 	minor_t		minor = getminor(dev);
2533 	int		part = XDF_PART(minor);
2534 	xdf_t		*vdp;
2535 	int		rv;
2536 
2537 	if (((vdp = ddi_get_soft_state(xdf_ssp, XDF_INST(minor))) == NULL) ||
2538 	    (!xdf_isopen(vdp, part)))
2539 		return (ENXIO);
2540 
2541 	DPRINTF(IOCTL_DBG, ("xdf@%s:ioctl: cmd %d (0x%x)\n",
2542 	    vdp->xdf_addr, cmd, cmd));
2543 
2544 	switch (cmd) {
2545 	default:
2546 		return (ENOTTY);
2547 	case DKIOCG_PHYGEOM:
2548 	case DKIOCG_VIRTGEOM:
2549 	case DKIOCGGEOM:
2550 	case DKIOCSGEOM:
2551 	case DKIOCGAPART:
2552 	case DKIOCSAPART:
2553 	case DKIOCGVTOC:
2554 	case DKIOCSVTOC:
2555 	case DKIOCPARTINFO:
2556 	case DKIOCGEXTVTOC:
2557 	case DKIOCSEXTVTOC:
2558 	case DKIOCEXTPARTINFO:
2559 	case DKIOCGMBOOT:
2560 	case DKIOCSMBOOT:
2561 	case DKIOCGETEFI:
2562 	case DKIOCSETEFI:
2563 	case DKIOCPARTITION:
2564 		return (cmlb_ioctl(vdp->xdf_vd_lbl, dev, cmd, arg, mode, credp,
2565 		    rvalp, NULL));
2566 	case FDEJECT:
2567 	case DKIOCEJECT:
2568 	case CDROMEJECT:
2569 		return (xdf_ioctl_eject(vdp));
2570 	case DKIOCLOCK:
2571 		return (xdf_ioctl_mlock(vdp));
2572 	case DKIOCUNLOCK:
2573 		return (xdf_ioctl_munlock(vdp));
2574 	case CDROMREADOFFSET: {
2575 		int offset = 0;
2576 		if (!XD_IS_CD(vdp))
2577 			return (ENOTTY);
2578 		if (ddi_copyout(&offset, (void *)arg, sizeof (int), mode))
2579 			return (EFAULT);
2580 		return (0);
2581 	}
2582 	case DKIOCGMEDIAINFO: {
2583 		struct dk_minfo media_info;
2584 
2585 		media_info.dki_lbsize = DEV_BSIZE;
2586 		media_info.dki_capacity = vdp->xdf_pgeom.g_capacity;
2587 		if (XD_IS_CD(vdp))
2588 			media_info.dki_media_type = DK_CDROM;
2589 		else
2590 			media_info.dki_media_type = DK_FIXED_DISK;
2591 
2592 		if (ddi_copyout(&media_info, (void *)arg,
2593 		    sizeof (struct dk_minfo), mode))
2594 			return (EFAULT);
2595 		return (0);
2596 	}
2597 	case DKIOCINFO: {
2598 		struct dk_cinfo info;
2599 
2600 		/* controller information */
2601 		if (XD_IS_CD(vdp))
2602 			info.dki_ctype = DKC_CDROM;
2603 		else
2604 			info.dki_ctype = DKC_VBD;
2605 
2606 		info.dki_cnum = 0;
2607 		(void) strncpy((char *)(&info.dki_cname), "xdf", 8);
2608 
2609 		/* unit information */
2610 		info.dki_unit = ddi_get_instance(vdp->xdf_dip);
2611 		(void) strncpy((char *)(&info.dki_dname), "xdf", 8);
2612 		info.dki_flags = DKI_FMTVOL;
2613 		info.dki_partition = part;
2614 		info.dki_maxtransfer = maxphys / DEV_BSIZE;
2615 		info.dki_addr = 0;
2616 		info.dki_space = 0;
2617 		info.dki_prio = 0;
2618 		info.dki_vec = 0;
2619 
2620 		if (ddi_copyout(&info, (void *)arg, sizeof (info), mode))
2621 			return (EFAULT);
2622 		return (0);
2623 	}
2624 	case DKIOCSTATE: {
2625 		enum dkio_state mstate;
2626 
2627 		if (ddi_copyin((void *)arg, &mstate,
2628 		    sizeof (mstate), mode) != 0)
2629 			return (EFAULT);
2630 		if ((rv = xdf_dkstate(vdp, mstate)) != 0)
2631 			return (rv);
2632 		mstate = vdp->xdf_mstate;
2633 		if (ddi_copyout(&mstate, (void *)arg,
2634 		    sizeof (mstate), mode) != 0)
2635 			return (EFAULT);
2636 		return (0);
2637 	}
2638 	case DKIOCREMOVABLE: {
2639 		int i = BOOLEAN2VOID(XD_IS_RM(vdp));
2640 		if (ddi_copyout(&i, (caddr_t)arg, sizeof (i), mode))
2641 			return (EFAULT);
2642 		return (0);
2643 	}
2644 	case DKIOCGETWCE: {
2645 		int i = BOOLEAN2VOID(XD_IS_RM(vdp));
2646 		if (ddi_copyout(&i, (void *)arg, sizeof (i), mode))
2647 			return (EFAULT);
2648 		return (0);
2649 	}
2650 	case DKIOCSETWCE: {
2651 		int i;
2652 		if (ddi_copyin((void *)arg, &i, sizeof (i), mode))
2653 			return (EFAULT);
2654 		vdp->xdf_wce = VOID2BOOLEAN(i);
2655 		return (0);
2656 	}
2657 	case DKIOCFLUSHWRITECACHE: {
2658 		struct dk_callback *dkc = (struct dk_callback *)arg;
2659 
2660 		if (vdp->xdf_flush_supported) {
2661 			rv = xdf_lb_rdwr(vdp->xdf_dip, TG_WRITE,
2662 			    NULL, 0, 0, (void *)dev);
2663 		} else if (vdp->xdf_feature_barrier &&
2664 		    !xdf_barrier_flush_disable) {
2665 			rv = xdf_lb_rdwr(vdp->xdf_dip, TG_WRITE,
2666 			    vdp->xdf_cache_flush_block, xdf_flush_block,
2667 			    DEV_BSIZE, (void *)dev);
2668 		} else {
2669 			return (ENOTTY);
2670 		}
2671 		if ((mode & FKIOCTL) && (dkc != NULL) &&
2672 		    (dkc->dkc_callback != NULL)) {
2673 			(*dkc->dkc_callback)(dkc->dkc_cookie, rv);
2674 			/* need to return 0 after calling callback */
2675 			rv = 0;
2676 		}
2677 		return (rv);
2678 	}
2679 	}
2680 	/*NOTREACHED*/
2681 }
2682 
2683 static int
2684 xdf_strategy(struct buf *bp)
2685 {
2686 	xdf_t	*vdp;
2687 	minor_t minor;
2688 	diskaddr_t p_blkct, p_blkst;
2689 	ulong_t nblks;
2690 	int part;
2691 
2692 	minor = getminor(bp->b_edev);
2693 	part = XDF_PART(minor);
2694 	vdp = ddi_get_soft_state(xdf_ssp, XDF_INST(minor));
2695 
2696 	mutex_enter(&vdp->xdf_dev_lk);
2697 	if (!xdf_isopen(vdp, part)) {
2698 		mutex_exit(&vdp->xdf_dev_lk);
2699 		xdf_io_err(bp, ENXIO, 0);
2700 		return (0);
2701 	}
2702 
2703 	/* We don't allow IO from the oe_change callback thread */
2704 	ASSERT(curthread != vdp->xdf_oe_change_thread);
2705 
2706 	/* Check for writes to a read only device */
2707 	if (!IS_READ(bp) && XD_IS_RO(vdp)) {
2708 		mutex_exit(&vdp->xdf_dev_lk);
2709 		xdf_io_err(bp, EROFS, 0);
2710 		return (0);
2711 	}
2712 
2713 	/* Check if this I/O is accessing a partition or the entire disk */
2714 	if ((long)bp->b_private == XB_SLICE_NONE) {
2715 		/* This I/O is using an absolute offset */
2716 		p_blkct = vdp->xdf_xdev_nblocks;
2717 		p_blkst = 0;
2718 	} else {
2719 		/* This I/O is using a partition relative offset */
2720 		mutex_exit(&vdp->xdf_dev_lk);
2721 		if (cmlb_partinfo(vdp->xdf_vd_lbl, part, &p_blkct,
2722 		    &p_blkst, NULL, NULL, NULL)) {
2723 			xdf_io_err(bp, ENXIO, 0);
2724 			return (0);
2725 		}
2726 		mutex_enter(&vdp->xdf_dev_lk);
2727 	}
2728 
2729 	/* check for a starting block beyond the disk or partition limit */
2730 	if (bp->b_blkno > p_blkct) {
2731 		DPRINTF(IO_DBG, ("xdf@%s: block %lld exceeds VBD size %"PRIu64,
2732 		    vdp->xdf_addr, (longlong_t)bp->b_blkno, (uint64_t)p_blkct));
2733 		xdf_io_err(bp, EINVAL, 0);
2734 		return (0);
2735 	}
2736 
2737 	/* Legacy: don't set error flag at this case */
2738 	if (bp->b_blkno == p_blkct) {
2739 		bp->b_resid = bp->b_bcount;
2740 		biodone(bp);
2741 		return (0);
2742 	}
2743 
2744 	/* sanitize the input buf */
2745 	bioerror(bp, 0);
2746 	bp->b_resid = 0;
2747 	bp->av_back = bp->av_forw = NULL;
2748 
2749 	/* Adjust for partial transfer, this will result in an error later */
2750 	nblks = bp->b_bcount >> XB_BSHIFT;
2751 	if ((bp->b_blkno + nblks) > p_blkct) {
2752 		bp->b_resid = ((bp->b_blkno + nblks) - p_blkct) << XB_BSHIFT;
2753 		bp->b_bcount -= bp->b_resid;
2754 	}
2755 
2756 	DPRINTF(IO_DBG, ("xdf@%s: strategy blk %lld len %lu\n",
2757 	    vdp->xdf_addr, (longlong_t)bp->b_blkno, (ulong_t)bp->b_bcount));
2758 
2759 	/* Fix up the buf struct */
2760 	bp->b_flags |= B_BUSY;
2761 	bp->b_private = (void *)(uintptr_t)p_blkst;
2762 
2763 	xdf_bp_push(vdp, bp);
2764 	mutex_exit(&vdp->xdf_dev_lk);
2765 	xdf_io_start(vdp);
2766 	if (do_polled_io)
2767 		(void) xdf_ring_drain(vdp);
2768 	return (0);
2769 }
2770 
2771 /*ARGSUSED*/
2772 static int
2773 xdf_read(dev_t dev, struct uio *uiop, cred_t *credp)
2774 {
2775 	xdf_t	*vdp;
2776 	minor_t minor;
2777 	diskaddr_t p_blkcnt;
2778 	int part;
2779 
2780 	minor = getminor(dev);
2781 	if ((vdp = ddi_get_soft_state(xdf_ssp, XDF_INST(minor))) == NULL)
2782 		return (ENXIO);
2783 
2784 	DPRINTF(IO_DBG, ("xdf@%s: read offset 0x%"PRIx64"\n",
2785 	    vdp->xdf_addr, (int64_t)uiop->uio_offset));
2786 
2787 	part = XDF_PART(minor);
2788 	if (!xdf_isopen(vdp, part))
2789 		return (ENXIO);
2790 
2791 	if (cmlb_partinfo(vdp->xdf_vd_lbl, part, &p_blkcnt,
2792 	    NULL, NULL, NULL, NULL))
2793 		return (ENXIO);
2794 
2795 	if (U_INVAL(uiop))
2796 		return (EINVAL);
2797 
2798 	return (physio(xdf_strategy, NULL, dev, B_READ, xdfmin, uiop));
2799 }
2800 
2801 /*ARGSUSED*/
2802 static int
2803 xdf_write(dev_t dev, struct uio *uiop, cred_t *credp)
2804 {
2805 	xdf_t *vdp;
2806 	minor_t minor;
2807 	diskaddr_t p_blkcnt;
2808 	int part;
2809 
2810 	minor = getminor(dev);
2811 	if ((vdp = ddi_get_soft_state(xdf_ssp, XDF_INST(minor))) == NULL)
2812 		return (ENXIO);
2813 
2814 	DPRINTF(IO_DBG, ("xdf@%s: write offset 0x%"PRIx64"\n",
2815 	    vdp->xdf_addr, (int64_t)uiop->uio_offset));
2816 
2817 	part = XDF_PART(minor);
2818 	if (!xdf_isopen(vdp, part))
2819 		return (ENXIO);
2820 
2821 	if (cmlb_partinfo(vdp->xdf_vd_lbl, part, &p_blkcnt,
2822 	    NULL, NULL, NULL, NULL))
2823 		return (ENXIO);
2824 
2825 	if (uiop->uio_loffset >= XB_DTOB(p_blkcnt))
2826 		return (ENOSPC);
2827 
2828 	if (U_INVAL(uiop))
2829 		return (EINVAL);
2830 
2831 	return (physio(xdf_strategy, NULL, dev, B_WRITE, xdfmin, uiop));
2832 }
2833 
2834 /*ARGSUSED*/
2835 static int
2836 xdf_aread(dev_t dev, struct aio_req *aiop, cred_t *credp)
2837 {
2838 	xdf_t	*vdp;
2839 	minor_t minor;
2840 	struct uio *uiop = aiop->aio_uio;
2841 	diskaddr_t p_blkcnt;
2842 	int part;
2843 
2844 	minor = getminor(dev);
2845 	if ((vdp = ddi_get_soft_state(xdf_ssp, XDF_INST(minor))) == NULL)
2846 		return (ENXIO);
2847 
2848 	part = XDF_PART(minor);
2849 	if (!xdf_isopen(vdp, part))
2850 		return (ENXIO);
2851 
2852 	if (cmlb_partinfo(vdp->xdf_vd_lbl, part, &p_blkcnt,
2853 	    NULL, NULL, NULL, NULL))
2854 		return (ENXIO);
2855 
2856 	if (uiop->uio_loffset >= XB_DTOB(p_blkcnt))
2857 		return (ENOSPC);
2858 
2859 	if (U_INVAL(uiop))
2860 		return (EINVAL);
2861 
2862 	return (aphysio(xdf_strategy, anocancel, dev, B_READ, xdfmin, aiop));
2863 }
2864 
2865 /*ARGSUSED*/
2866 static int
2867 xdf_awrite(dev_t dev, struct aio_req *aiop, cred_t *credp)
2868 {
2869 	xdf_t *vdp;
2870 	minor_t minor;
2871 	struct uio *uiop = aiop->aio_uio;
2872 	diskaddr_t p_blkcnt;
2873 	int part;
2874 
2875 	minor = getminor(dev);
2876 	if ((vdp = ddi_get_soft_state(xdf_ssp, XDF_INST(minor))) == NULL)
2877 		return (ENXIO);
2878 
2879 	part = XDF_PART(minor);
2880 	if (!xdf_isopen(vdp, part))
2881 		return (ENXIO);
2882 
2883 	if (cmlb_partinfo(vdp->xdf_vd_lbl, part, &p_blkcnt,
2884 	    NULL, NULL, NULL, NULL))
2885 		return (ENXIO);
2886 
2887 	if (uiop->uio_loffset >= XB_DTOB(p_blkcnt))
2888 		return (ENOSPC);
2889 
2890 	if (U_INVAL(uiop))
2891 		return (EINVAL);
2892 
2893 	return (aphysio(xdf_strategy, anocancel, dev, B_WRITE, xdfmin, aiop));
2894 }
2895 
2896 static int
2897 xdf_dump(dev_t dev, caddr_t addr, daddr_t blkno, int nblk)
2898 {
2899 	struct buf dumpbuf, *dbp = &dumpbuf;
2900 	xdf_t	*vdp;
2901 	minor_t minor;
2902 	int err = 0;
2903 	int part;
2904 	diskaddr_t p_blkcnt, p_blkst;
2905 
2906 	minor = getminor(dev);
2907 	if ((vdp = ddi_get_soft_state(xdf_ssp, XDF_INST(minor))) == NULL)
2908 		return (ENXIO);
2909 
2910 	DPRINTF(IO_DBG, ("xdf@%s: dump addr (0x%p) blk (%ld) nblks (%d)\n",
2911 	    vdp->xdf_addr, (void *)addr, blkno, nblk));
2912 
2913 	/* We don't allow IO from the oe_change callback thread */
2914 	ASSERT(curthread != vdp->xdf_oe_change_thread);
2915 
2916 	part = XDF_PART(minor);
2917 	if (!xdf_isopen(vdp, part))
2918 		return (ENXIO);
2919 
2920 	if (cmlb_partinfo(vdp->xdf_vd_lbl, part, &p_blkcnt, &p_blkst,
2921 	    NULL, NULL, NULL))
2922 		return (ENXIO);
2923 
2924 	if ((blkno + nblk) > p_blkcnt) {
2925 		cmn_err(CE_WARN, "xdf@%s: block %ld exceeds VBD size %"PRIu64,
2926 		    vdp->xdf_addr, blkno + nblk, (uint64_t)p_blkcnt);
2927 		return (EINVAL);
2928 	}
2929 
2930 	bioinit(dbp);
2931 	dbp->b_flags = B_BUSY;
2932 	dbp->b_un.b_addr = addr;
2933 	dbp->b_bcount = nblk << DEV_BSHIFT;
2934 	dbp->b_blkno = blkno;
2935 	dbp->b_edev = dev;
2936 	dbp->b_private = (void *)(uintptr_t)p_blkst;
2937 
2938 	mutex_enter(&vdp->xdf_dev_lk);
2939 	xdf_bp_push(vdp, dbp);
2940 	mutex_exit(&vdp->xdf_dev_lk);
2941 	xdf_io_start(vdp);
2942 	err = xdf_ring_drain(vdp);
2943 	biofini(dbp);
2944 	return (err);
2945 }
2946 
2947 /*ARGSUSED*/
2948 static int
2949 xdf_close(dev_t dev, int flag, int otyp, struct cred *credp)
2950 {
2951 	minor_t	minor;
2952 	xdf_t	*vdp;
2953 	int part;
2954 	ulong_t parbit;
2955 
2956 	minor = getminor(dev);
2957 	if ((vdp = ddi_get_soft_state(xdf_ssp, XDF_INST(minor))) == NULL)
2958 		return (ENXIO);
2959 
2960 	mutex_enter(&vdp->xdf_dev_lk);
2961 	part = XDF_PART(minor);
2962 	if (!xdf_isopen(vdp, part)) {
2963 		mutex_exit(&vdp->xdf_dev_lk);
2964 		return (ENXIO);
2965 	}
2966 	parbit = 1 << part;
2967 
2968 	ASSERT((vdp->xdf_vd_open[otyp] & parbit) != 0);
2969 	if (otyp == OTYP_LYR) {
2970 		ASSERT(vdp->xdf_vd_lyropen[part] > 0);
2971 		if (--vdp->xdf_vd_lyropen[part] == 0)
2972 			vdp->xdf_vd_open[otyp] &= ~parbit;
2973 	} else {
2974 		vdp->xdf_vd_open[otyp] &= ~parbit;
2975 	}
2976 	vdp->xdf_vd_exclopen &= ~parbit;
2977 
2978 	mutex_exit(&vdp->xdf_dev_lk);
2979 	return (0);
2980 }
2981 
2982 static int
2983 xdf_open(dev_t *devp, int flag, int otyp, cred_t *credp)
2984 {
2985 	minor_t	minor;
2986 	xdf_t	*vdp;
2987 	int part;
2988 	ulong_t parbit;
2989 	diskaddr_t p_blkct = 0;
2990 	boolean_t firstopen;
2991 	boolean_t nodelay;
2992 
2993 	minor = getminor(*devp);
2994 	if ((vdp = ddi_get_soft_state(xdf_ssp, XDF_INST(minor))) == NULL)
2995 		return (ENXIO);
2996 
2997 	nodelay = (flag & (FNDELAY | FNONBLOCK));
2998 
2999 	DPRINTF(DDI_DBG, ("xdf@%s: opening\n", vdp->xdf_addr));
3000 
3001 	/* do cv_wait until connected or failed */
3002 	mutex_enter(&vdp->xdf_cb_lk);
3003 	mutex_enter(&vdp->xdf_dev_lk);
3004 	if (!nodelay && (xdf_connect_locked(vdp, B_TRUE) != XD_READY)) {
3005 		mutex_exit(&vdp->xdf_dev_lk);
3006 		mutex_exit(&vdp->xdf_cb_lk);
3007 		return (ENXIO);
3008 	}
3009 	mutex_exit(&vdp->xdf_cb_lk);
3010 
3011 	if ((flag & FWRITE) && XD_IS_RO(vdp)) {
3012 		mutex_exit(&vdp->xdf_dev_lk);
3013 		return (EROFS);
3014 	}
3015 
3016 	part = XDF_PART(minor);
3017 	parbit = 1 << part;
3018 	if ((vdp->xdf_vd_exclopen & parbit) ||
3019 	    ((flag & FEXCL) && xdf_isopen(vdp, part))) {
3020 		mutex_exit(&vdp->xdf_dev_lk);
3021 		return (EBUSY);
3022 	}
3023 
3024 	/* are we the first one to open this node? */
3025 	firstopen = !xdf_isopen(vdp, -1);
3026 
3027 	if (otyp == OTYP_LYR)
3028 		vdp->xdf_vd_lyropen[part]++;
3029 
3030 	vdp->xdf_vd_open[otyp] |= parbit;
3031 
3032 	if (flag & FEXCL)
3033 		vdp->xdf_vd_exclopen |= parbit;
3034 
3035 	mutex_exit(&vdp->xdf_dev_lk);
3036 
3037 	/* force a re-validation */
3038 	if (firstopen)
3039 		cmlb_invalidate(vdp->xdf_vd_lbl, NULL);
3040 
3041 	/* If this is a non-blocking open then we're done */
3042 	if (nodelay)
3043 		return (0);
3044 
3045 	/*
3046 	 * This is a blocking open, so we require:
3047 	 * - that the disk have a valid label on it
3048 	 * - that the size of the partition that we're opening is non-zero
3049 	 */
3050 	if ((cmlb_partinfo(vdp->xdf_vd_lbl, part, &p_blkct,
3051 	    NULL, NULL, NULL, NULL) != 0) || (p_blkct == 0)) {
3052 		(void) xdf_close(*devp, flag, otyp, credp);
3053 		return (ENXIO);
3054 	}
3055 
3056 	return (0);
3057 }
3058 
3059 /*ARGSUSED*/
3060 static void
3061 xdf_watch_hp_status_cb(dev_info_t *dip, const char *path, void *arg)
3062 {
3063 	xdf_t *vdp = (xdf_t *)ddi_get_driver_private(dip);
3064 	cv_broadcast(&vdp->xdf_hp_status_cv);
3065 }
3066 
3067 static int
3068 xdf_prop_op(dev_t dev, dev_info_t *dip, ddi_prop_op_t prop_op, int flags,
3069 	char *name, caddr_t valuep, int *lengthp)
3070 {
3071 	xdf_t	*vdp = ddi_get_soft_state(xdf_ssp, ddi_get_instance(dip));
3072 
3073 	/*
3074 	 * Sanity check that if a dev_t or dip were specified that they
3075 	 * correspond to this device driver.  On debug kernels we'll
3076 	 * panic and on non-debug kernels we'll return failure.
3077 	 */
3078 	ASSERT(ddi_driver_major(dip) == xdf_major);
3079 	ASSERT((dev == DDI_DEV_T_ANY) || (getmajor(dev) == xdf_major));
3080 	if ((ddi_driver_major(dip) != xdf_major) ||
3081 	    ((dev != DDI_DEV_T_ANY) && (getmajor(dev) != xdf_major)))
3082 		return (DDI_PROP_NOT_FOUND);
3083 
3084 	if (vdp == NULL)
3085 		return (ddi_prop_op(dev, dip, prop_op, flags,
3086 		    name, valuep, lengthp));
3087 
3088 	return (cmlb_prop_op(vdp->xdf_vd_lbl,
3089 	    dev, dip, prop_op, flags, name, valuep, lengthp,
3090 	    XDF_PART(getminor(dev)), NULL));
3091 }
3092 
3093 /*ARGSUSED*/
3094 static int
3095 xdf_getinfo(dev_info_t *dip, ddi_info_cmd_t cmd, void *arg, void **rp)
3096 {
3097 	int	instance = XDF_INST(getminor((dev_t)arg));
3098 	xdf_t	*vbdp;
3099 
3100 	switch (cmd) {
3101 	case DDI_INFO_DEVT2DEVINFO:
3102 		if ((vbdp = ddi_get_soft_state(xdf_ssp, instance)) == NULL) {
3103 			*rp = NULL;
3104 			return (DDI_FAILURE);
3105 		}
3106 		*rp = vbdp->xdf_dip;
3107 		return (DDI_SUCCESS);
3108 
3109 	case DDI_INFO_DEVT2INSTANCE:
3110 		*rp = (void *)(uintptr_t)instance;
3111 		return (DDI_SUCCESS);
3112 
3113 	default:
3114 		return (DDI_FAILURE);
3115 	}
3116 }
3117 
3118 /*ARGSUSED*/
3119 static int
3120 xdf_resume(dev_info_t *dip)
3121 {
3122 	xdf_t	*vdp;
3123 	char	*oename;
3124 
3125 	if ((vdp = ddi_get_soft_state(xdf_ssp, ddi_get_instance(dip))) == NULL)
3126 		goto err;
3127 
3128 	if (xdf_debug & SUSRES_DBG)
3129 		xen_printf("xdf@%s: xdf_resume\n", vdp->xdf_addr);
3130 
3131 	mutex_enter(&vdp->xdf_cb_lk);
3132 
3133 	if (xvdi_resume(dip) != DDI_SUCCESS) {
3134 		mutex_exit(&vdp->xdf_cb_lk);
3135 		goto err;
3136 	}
3137 
3138 	if (((oename = xvdi_get_oename(dip)) == NULL) ||
3139 	    (xvdi_add_xb_watch_handler(dip, oename, XBP_HP_STATUS,
3140 	    xdf_watch_hp_status_cb, NULL) != DDI_SUCCESS)) {
3141 		mutex_exit(&vdp->xdf_cb_lk);
3142 		goto err;
3143 	}
3144 
3145 	mutex_enter(&vdp->xdf_dev_lk);
3146 	ASSERT(vdp->xdf_state != XD_READY);
3147 	xdf_set_state(vdp, XD_UNKNOWN);
3148 	mutex_exit(&vdp->xdf_dev_lk);
3149 
3150 	if (xdf_setstate_init(vdp) != DDI_SUCCESS) {
3151 		mutex_exit(&vdp->xdf_cb_lk);
3152 		goto err;
3153 	}
3154 
3155 	mutex_exit(&vdp->xdf_cb_lk);
3156 
3157 	if (xdf_debug & SUSRES_DBG)
3158 		xen_printf("xdf@%s: xdf_resume: done\n", vdp->xdf_addr);
3159 	return (DDI_SUCCESS);
3160 err:
3161 	if (xdf_debug & SUSRES_DBG)
3162 		xen_printf("xdf@%s: xdf_resume: fail\n", vdp->xdf_addr);
3163 	return (DDI_FAILURE);
3164 }
3165 
3166 static int
3167 xdf_attach(dev_info_t *dip, ddi_attach_cmd_t cmd)
3168 {
3169 	int			n, instance = ddi_get_instance(dip);
3170 	ddi_iblock_cookie_t	ibc, softibc;
3171 	boolean_t		dev_iscd = B_FALSE;
3172 	xdf_t			*vdp;
3173 	char			*oename, *xsname, *str;
3174 
3175 	if ((n = ddi_prop_get_int(DDI_DEV_T_ANY, dip, DDI_PROP_NOTPROM,
3176 	    "xdf_debug", 0)) != 0)
3177 		xdf_debug = n;
3178 
3179 	switch (cmd) {
3180 	case DDI_RESUME:
3181 		return (xdf_resume(dip));
3182 	case DDI_ATTACH:
3183 		break;
3184 	default:
3185 		return (DDI_FAILURE);
3186 	}
3187 	/* DDI_ATTACH */
3188 
3189 	if (((xsname = xvdi_get_xsname(dip)) == NULL) ||
3190 	    ((oename = xvdi_get_oename(dip)) == NULL))
3191 		return (DDI_FAILURE);
3192 
3193 	/*
3194 	 * Disable auto-detach.  This is necessary so that we don't get
3195 	 * detached while we're disconnected from the back end.
3196 	 */
3197 	if ((ddi_prop_update_int(DDI_DEV_T_NONE, dip,
3198 	    DDI_NO_AUTODETACH, 1) != DDI_PROP_SUCCESS))
3199 		return (DDI_FAILURE);
3200 
3201 	/* driver handles kernel-issued IOCTLs */
3202 	if (ddi_prop_create(DDI_DEV_T_NONE, dip,
3203 	    DDI_PROP_CANSLEEP, DDI_KERNEL_IOCTL, NULL, 0) != DDI_PROP_SUCCESS)
3204 		return (DDI_FAILURE);
3205 
3206 	if (ddi_get_iblock_cookie(dip, 0, &ibc) != DDI_SUCCESS)
3207 		return (DDI_FAILURE);
3208 
3209 	if (ddi_get_soft_iblock_cookie(dip,
3210 	    DDI_SOFTINT_LOW, &softibc) != DDI_SUCCESS)
3211 		return (DDI_FAILURE);
3212 
3213 	if (xenbus_read_str(xsname, XBP_DEV_TYPE, &str) != 0) {
3214 		cmn_err(CE_WARN, "xdf@%s: cannot read device-type",
3215 		    ddi_get_name_addr(dip));
3216 		return (DDI_FAILURE);
3217 	}
3218 	if (strcmp(str, XBV_DEV_TYPE_CD) == 0)
3219 		dev_iscd = B_TRUE;
3220 	strfree(str);
3221 
3222 	if (ddi_soft_state_zalloc(xdf_ssp, instance) != DDI_SUCCESS)
3223 		return (DDI_FAILURE);
3224 
3225 	DPRINTF(DDI_DBG, ("xdf@%s: attaching\n", ddi_get_name_addr(dip)));
3226 	vdp = ddi_get_soft_state(xdf_ssp, instance);
3227 	ddi_set_driver_private(dip, vdp);
3228 	vdp->xdf_dip = dip;
3229 	vdp->xdf_addr = ddi_get_name_addr(dip);
3230 	vdp->xdf_suspending = B_FALSE;
3231 	vdp->xdf_media_req_supported = B_FALSE;
3232 	vdp->xdf_peer = INVALID_DOMID;
3233 	vdp->xdf_evtchn = INVALID_EVTCHN;
3234 	list_create(&vdp->xdf_vreq_act, sizeof (v_req_t),
3235 	    offsetof(v_req_t, v_link));
3236 	cv_init(&vdp->xdf_dev_cv, NULL, CV_DEFAULT, NULL);
3237 	cv_init(&vdp->xdf_hp_status_cv, NULL, CV_DEFAULT, NULL);
3238 	cv_init(&vdp->xdf_mstate_cv, NULL, CV_DEFAULT, NULL);
3239 	mutex_init(&vdp->xdf_dev_lk, NULL, MUTEX_DRIVER, (void *)ibc);
3240 	mutex_init(&vdp->xdf_cb_lk, NULL, MUTEX_DRIVER, (void *)ibc);
3241 	mutex_init(&vdp->xdf_iostat_lk, NULL, MUTEX_DRIVER, (void *)ibc);
3242 	vdp->xdf_cmbl_reattach = B_TRUE;
3243 	if (dev_iscd) {
3244 		vdp->xdf_dinfo |= VDISK_CDROM;
3245 		vdp->xdf_mstate = DKIO_EJECTED;
3246 	} else {
3247 		vdp->xdf_mstate = DKIO_NONE;
3248 	}
3249 
3250 	if ((vdp->xdf_ready_tq = ddi_taskq_create(dip, "xdf_ready_tq",
3251 	    1, TASKQ_DEFAULTPRI, 0)) == NULL)
3252 		goto errout0;
3253 
3254 	if (xvdi_add_xb_watch_handler(dip, oename, XBP_HP_STATUS,
3255 	    xdf_watch_hp_status_cb, NULL) != DDI_SUCCESS)
3256 		goto errout0;
3257 
3258 	if (ddi_add_softintr(dip, DDI_SOFTINT_LOW, &vdp->xdf_softintr_id,
3259 	    &softibc, NULL, xdf_iorestart, (caddr_t)vdp) != DDI_SUCCESS) {
3260 		cmn_err(CE_WARN, "xdf@%s: failed to add softintr",
3261 		    ddi_get_name_addr(dip));
3262 		goto errout0;
3263 	}
3264 
3265 	/*
3266 	 * Initialize the physical geometry stucture.  Note that currently
3267 	 * we don't know the size of the backend device so the number
3268 	 * of blocks on the device will be initialized to zero.  Once
3269 	 * we connect to the backend device we'll update the physical
3270 	 * geometry to reflect the real size of the device.
3271 	 */
3272 	xdf_synthetic_pgeom(dip, &vdp->xdf_pgeom);
3273 	vdp->xdf_pgeom_fixed = B_FALSE;
3274 
3275 	/*
3276 	 * create default device minor nodes: non-removable disk
3277 	 * we will adjust minor nodes after we are connected w/ backend
3278 	 */
3279 	cmlb_alloc_handle(&vdp->xdf_vd_lbl);
3280 	if (xdf_cmlb_attach(vdp) != 0) {
3281 		cmn_err(CE_WARN,
3282 		    "xdf@%s: attach failed, cmlb attach failed",
3283 		    ddi_get_name_addr(dip));
3284 		goto errout0;
3285 	}
3286 
3287 	/*
3288 	 * We ship with cache-enabled disks
3289 	 */
3290 	vdp->xdf_wce = B_TRUE;
3291 
3292 	mutex_enter(&vdp->xdf_cb_lk);
3293 	/* Watch backend XenbusState change */
3294 	if (xvdi_add_event_handler(dip,
3295 	    XS_OE_STATE, xdf_oe_change, NULL) != DDI_SUCCESS) {
3296 		mutex_exit(&vdp->xdf_cb_lk);
3297 		goto errout0;
3298 	}
3299 
3300 	if (xdf_setstate_init(vdp) != DDI_SUCCESS) {
3301 		cmn_err(CE_WARN, "xdf@%s: start connection failed",
3302 		    ddi_get_name_addr(dip));
3303 		mutex_exit(&vdp->xdf_cb_lk);
3304 		goto errout1;
3305 	}
3306 	mutex_exit(&vdp->xdf_cb_lk);
3307 
3308 #if defined(XPV_HVM_DRIVER)
3309 
3310 	xdf_hvm_add(dip);
3311 
3312 	/* Report our version to dom0.  */
3313 	if (xenbus_printf(XBT_NULL, "hvmpv/xdf", "version", "%d",
3314 	    HVMPV_XDF_VERS))
3315 		cmn_err(CE_WARN, "xdf: couldn't write version\n");
3316 
3317 #else /* !XPV_HVM_DRIVER */
3318 
3319 	/* create kstat for iostat(1M) */
3320 	if (xdf_kstat_create(dip, "xdf", instance) != 0) {
3321 		cmn_err(CE_WARN, "xdf@%s: failed to create kstat",
3322 		    ddi_get_name_addr(dip));
3323 		goto errout1;
3324 	}
3325 
3326 #endif /* !XPV_HVM_DRIVER */
3327 
3328 	ddi_report_dev(dip);
3329 	DPRINTF(DDI_DBG, ("xdf@%s: attached\n", vdp->xdf_addr));
3330 	return (DDI_SUCCESS);
3331 
3332 errout1:
3333 	(void) xvdi_switch_state(vdp->xdf_dip, XBT_NULL, XenbusStateClosed);
3334 	xvdi_remove_event_handler(dip, XS_OE_STATE);
3335 errout0:
3336 	if (vdp->xdf_vd_lbl != NULL) {
3337 		cmlb_detach(vdp->xdf_vd_lbl, NULL);
3338 		cmlb_free_handle(&vdp->xdf_vd_lbl);
3339 		vdp->xdf_vd_lbl = NULL;
3340 	}
3341 	if (vdp->xdf_softintr_id != NULL)
3342 		ddi_remove_softintr(vdp->xdf_softintr_id);
3343 	xvdi_remove_xb_watch_handlers(dip);
3344 	if (vdp->xdf_ready_tq != NULL)
3345 		ddi_taskq_destroy(vdp->xdf_ready_tq);
3346 	mutex_destroy(&vdp->xdf_cb_lk);
3347 	mutex_destroy(&vdp->xdf_dev_lk);
3348 	cv_destroy(&vdp->xdf_dev_cv);
3349 	cv_destroy(&vdp->xdf_hp_status_cv);
3350 	ddi_soft_state_free(xdf_ssp, instance);
3351 	ddi_set_driver_private(dip, NULL);
3352 	ddi_prop_remove_all(dip);
3353 	cmn_err(CE_WARN, "xdf@%s: attach failed", ddi_get_name_addr(dip));
3354 	return (DDI_FAILURE);
3355 }
3356 
3357 static int
3358 xdf_suspend(dev_info_t *dip)
3359 {
3360 	int		instance = ddi_get_instance(dip);
3361 	xdf_t		*vdp;
3362 
3363 	if ((vdp = ddi_get_soft_state(xdf_ssp, instance)) == NULL)
3364 		return (DDI_FAILURE);
3365 
3366 	if (xdf_debug & SUSRES_DBG)
3367 		xen_printf("xdf@%s: xdf_suspend\n", vdp->xdf_addr);
3368 
3369 	xvdi_suspend(dip);
3370 
3371 	mutex_enter(&vdp->xdf_cb_lk);
3372 	mutex_enter(&vdp->xdf_dev_lk);
3373 
3374 	vdp->xdf_suspending = B_TRUE;
3375 	xdf_ring_destroy(vdp);
3376 	xdf_set_state(vdp, XD_SUSPEND);
3377 	vdp->xdf_suspending = B_FALSE;
3378 
3379 	mutex_exit(&vdp->xdf_dev_lk);
3380 	mutex_exit(&vdp->xdf_cb_lk);
3381 
3382 	if (xdf_debug & SUSRES_DBG)
3383 		xen_printf("xdf@%s: xdf_suspend: done\n", vdp->xdf_addr);
3384 
3385 	return (DDI_SUCCESS);
3386 }
3387 
3388 static int
3389 xdf_detach(dev_info_t *dip, ddi_detach_cmd_t cmd)
3390 {
3391 	xdf_t *vdp;
3392 	int instance;
3393 
3394 	switch (cmd) {
3395 
3396 	case DDI_PM_SUSPEND:
3397 		break;
3398 
3399 	case DDI_SUSPEND:
3400 		return (xdf_suspend(dip));
3401 
3402 	case DDI_DETACH:
3403 		break;
3404 
3405 	default:
3406 		return (DDI_FAILURE);
3407 	}
3408 
3409 	instance = ddi_get_instance(dip);
3410 	DPRINTF(DDI_DBG, ("xdf@%s: detaching\n", ddi_get_name_addr(dip)));
3411 	vdp = ddi_get_soft_state(xdf_ssp, instance);
3412 
3413 	if (vdp == NULL)
3414 		return (DDI_FAILURE);
3415 
3416 	mutex_enter(&vdp->xdf_cb_lk);
3417 	xdf_disconnect(vdp, XD_CLOSED, B_FALSE);
3418 	if (vdp->xdf_state != XD_CLOSED) {
3419 		mutex_exit(&vdp->xdf_cb_lk);
3420 		return (DDI_FAILURE);
3421 	}
3422 	mutex_exit(&vdp->xdf_cb_lk);
3423 
3424 	ASSERT(!ISDMACBON(vdp));
3425 
3426 #if defined(XPV_HVM_DRIVER)
3427 	xdf_hvm_rm(dip);
3428 #endif /* XPV_HVM_DRIVER */
3429 
3430 	if (vdp->xdf_timeout_id != 0)
3431 		(void) untimeout(vdp->xdf_timeout_id);
3432 
3433 	xvdi_remove_event_handler(dip, XS_OE_STATE);
3434 	ddi_taskq_destroy(vdp->xdf_ready_tq);
3435 
3436 	cmlb_detach(vdp->xdf_vd_lbl, NULL);
3437 	cmlb_free_handle(&vdp->xdf_vd_lbl);
3438 
3439 	/* we'll support backend running in domU later */
3440 #ifdef	DOMU_BACKEND
3441 	(void) xvdi_post_event(dip, XEN_HP_REMOVE);
3442 #endif
3443 
3444 	list_destroy(&vdp->xdf_vreq_act);
3445 	ddi_prop_remove_all(dip);
3446 	xdf_kstat_delete(dip);
3447 	ddi_remove_softintr(vdp->xdf_softintr_id);
3448 	xvdi_remove_xb_watch_handlers(dip);
3449 	ddi_set_driver_private(dip, NULL);
3450 	cv_destroy(&vdp->xdf_dev_cv);
3451 	mutex_destroy(&vdp->xdf_cb_lk);
3452 	mutex_destroy(&vdp->xdf_dev_lk);
3453 	if (vdp->xdf_cache_flush_block != NULL)
3454 		kmem_free(vdp->xdf_flush_mem, 2 * DEV_BSIZE);
3455 	ddi_soft_state_free(xdf_ssp, instance);
3456 	return (DDI_SUCCESS);
3457 }
3458 
3459 /*
3460  * Driver linkage structures.
3461  */
3462 static struct cb_ops xdf_cbops = {
3463 	xdf_open,
3464 	xdf_close,
3465 	xdf_strategy,
3466 	nodev,
3467 	xdf_dump,
3468 	xdf_read,
3469 	xdf_write,
3470 	xdf_ioctl,
3471 	nodev,
3472 	nodev,
3473 	nodev,
3474 	nochpoll,
3475 	xdf_prop_op,
3476 	NULL,
3477 	D_MP | D_NEW | D_64BIT,
3478 	CB_REV,
3479 	xdf_aread,
3480 	xdf_awrite
3481 };
3482 
3483 struct dev_ops xdf_devops = {
3484 	DEVO_REV,		/* devo_rev */
3485 	0,			/* devo_refcnt */
3486 	xdf_getinfo,		/* devo_getinfo */
3487 	nulldev,		/* devo_identify */
3488 	nulldev,		/* devo_probe */
3489 	xdf_attach,		/* devo_attach */
3490 	xdf_detach,		/* devo_detach */
3491 	nodev,			/* devo_reset */
3492 	&xdf_cbops,		/* devo_cb_ops */
3493 	NULL,			/* devo_bus_ops */
3494 	NULL,			/* devo_power */
3495 	ddi_quiesce_not_supported, /* devo_quiesce */
3496 };
3497 
3498 /*
3499  * Module linkage structures.
3500  */
3501 static struct modldrv modldrv = {
3502 	&mod_driverops,		/* Type of module.  This one is a driver */
3503 	"virtual block driver",	/* short description */
3504 	&xdf_devops		/* driver specific ops */
3505 };
3506 
3507 static struct modlinkage xdf_modlinkage = {
3508 	MODREV_1, (void *)&modldrv, NULL
3509 };
3510 
3511 /*
3512  * standard module entry points
3513  */
3514 int
3515 _init(void)
3516 {
3517 	int rc;
3518 
3519 	xdf_major = ddi_name_to_major("xdf");
3520 	if (xdf_major == (major_t)-1)
3521 		return (EINVAL);
3522 
3523 	if ((rc = ddi_soft_state_init(&xdf_ssp, sizeof (xdf_t), 0)) != 0)
3524 		return (rc);
3525 
3526 	xdf_vreq_cache = kmem_cache_create("xdf_vreq_cache",
3527 	    sizeof (v_req_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
3528 	xdf_gs_cache = kmem_cache_create("xdf_gs_cache",
3529 	    sizeof (ge_slot_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
3530 
3531 #if defined(XPV_HVM_DRIVER)
3532 	xdf_hvm_init();
3533 #endif /* XPV_HVM_DRIVER */
3534 
3535 	if ((rc = mod_install(&xdf_modlinkage)) != 0) {
3536 #if defined(XPV_HVM_DRIVER)
3537 		xdf_hvm_fini();
3538 #endif /* XPV_HVM_DRIVER */
3539 		kmem_cache_destroy(xdf_vreq_cache);
3540 		kmem_cache_destroy(xdf_gs_cache);
3541 		ddi_soft_state_fini(&xdf_ssp);
3542 		return (rc);
3543 	}
3544 
3545 	return (rc);
3546 }
3547 
3548 int
3549 _fini(void)
3550 {
3551 	int err;
3552 	if ((err = mod_remove(&xdf_modlinkage)) != 0)
3553 		return (err);
3554 
3555 #if defined(XPV_HVM_DRIVER)
3556 	xdf_hvm_fini();
3557 #endif /* XPV_HVM_DRIVER */
3558 
3559 	kmem_cache_destroy(xdf_vreq_cache);
3560 	kmem_cache_destroy(xdf_gs_cache);
3561 	ddi_soft_state_fini(&xdf_ssp);
3562 
3563 	return (0);
3564 }
3565 
3566 int
3567 _info(struct modinfo *modinfop)
3568 {
3569 	return (mod_info(&xdf_modlinkage, modinfop));
3570 }
3571