xref: /titanic_41/usr/src/uts/common/xen/io/xdf.c (revision c037192b037119c9fd354732fc29b38ce097d356)
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 			rc = ddi_dma_mem_alloc(vreq->v_memdmahdl,
482 			    roundup(bp->b_bcount, XB_BSIZE), &xc_acc_attr,
483 			    DDI_DMA_STREAMING, xdf_dmacallback, (caddr_t)vdp,
484 			    &aba, &bufsz, &abh);
485 			if (rc != DDI_SUCCESS) {
486 				SETDMACBON(vdp);
487 				DPRINTF(DMA_DBG, ("xdf@%s: "
488 				    "DMA mem allocation failed\n",
489 				    vdp->xdf_addr));
490 				return (DDI_FAILURE);
491 			}
492 
493 			vreq->v_abuf = aba;
494 			vreq->v_align = abh;
495 			vreq->v_status = VREQ_DMAMEM_ALLOCED;
496 
497 			ASSERT(bufsz >= bp->b_bcount);
498 			if (!IS_READ(bp))
499 				bcopy(bp->b_un.b_addr, vreq->v_abuf,
500 				    bp->b_bcount);
501 		}
502 		/*FALLTHRU*/
503 
504 	case VREQ_DMAMEM_ALLOCED:
505 		/*
506 		 * dma bind
507 		 */
508 		if (ALIGNED_XFER(bp)) {
509 			rc = ddi_dma_buf_bind_handle(vreq->v_dmahdl, bp,
510 			    dma_flags, xdf_dmacallback, (caddr_t)vdp,
511 			    &dc, &ndcs);
512 		} else {
513 			rc = ddi_dma_addr_bind_handle(vreq->v_dmahdl,
514 			    NULL, vreq->v_abuf, bp->b_bcount, dma_flags,
515 			    xdf_dmacallback, (caddr_t)vdp, &dc, &ndcs);
516 		}
517 		if (rc == DDI_DMA_MAPPED || rc == DDI_DMA_PARTIAL_MAP) {
518 			/* get num of dma windows */
519 			if (rc == DDI_DMA_PARTIAL_MAP) {
520 				rc = ddi_dma_numwin(vreq->v_dmahdl, &ndws);
521 				ASSERT(rc == DDI_SUCCESS);
522 			} else {
523 				ndws = 1;
524 			}
525 		} else {
526 			SETDMACBON(vdp);
527 			DPRINTF(DMA_DBG, ("xdf@%s: DMA bind failed\n",
528 			    vdp->xdf_addr));
529 			return (DDI_FAILURE);
530 		}
531 
532 		vreq->v_dmac = dc;
533 		vreq->v_dmaw = 0;
534 		vreq->v_ndmacs = ndcs;
535 		vreq->v_ndmaws = ndws;
536 		vreq->v_nslots = ndws;
537 		vreq->v_status = VREQ_DMABUF_BOUND;
538 		/*FALLTHRU*/
539 
540 	case VREQ_DMABUF_BOUND:
541 		/*
542 		 * get ge_slot, callback is set upon failure from gs_get(),
543 		 * if not set previously
544 		 */
545 		if ((gs = gs_get(vdp, IS_READ(bp))) == NULL) {
546 			DPRINTF(DMA_DBG, ("xdf@%s: get ge_slot failed\n",
547 			    vdp->xdf_addr));
548 			return (DDI_FAILURE);
549 		}
550 
551 		vreq->v_status = VREQ_GS_ALLOCED;
552 		gs->gs_vreq = vreq;
553 		list_insert_head(&vreq->v_gs, gs);
554 		break;
555 
556 	case VREQ_GS_ALLOCED:
557 		/* nothing need to be done */
558 		break;
559 
560 	case VREQ_DMAWIN_DONE:
561 		/*
562 		 * move to the next dma window
563 		 */
564 		ASSERT((vreq->v_dmaw + 1) < vreq->v_ndmaws);
565 
566 		/* get a ge_slot for this DMA window */
567 		if ((gs = gs_get(vdp, IS_READ(bp))) == NULL) {
568 			DPRINTF(DMA_DBG, ("xdf@%s: get ge_slot failed\n",
569 			    vdp->xdf_addr));
570 			return (DDI_FAILURE);
571 		}
572 
573 		vreq->v_dmaw++;
574 		VERIFY(ddi_dma_getwin(vreq->v_dmahdl, vreq->v_dmaw, &off, &sz,
575 		    &vreq->v_dmac, &vreq->v_ndmacs) == DDI_SUCCESS);
576 		vreq->v_status = VREQ_GS_ALLOCED;
577 		gs->gs_vreq = vreq;
578 		list_insert_head(&vreq->v_gs, gs);
579 		break;
580 
581 	default:
582 		return (DDI_FAILURE);
583 	}
584 
585 	return (DDI_SUCCESS);
586 }
587 
588 static int
589 xdf_cmlb_attach(xdf_t *vdp)
590 {
591 	dev_info_t	*dip = vdp->xdf_dip;
592 
593 	return (cmlb_attach(dip, &xdf_lb_ops,
594 	    XD_IS_CD(vdp) ? DTYPE_RODIRECT : DTYPE_DIRECT,
595 	    XD_IS_RM(vdp),
596 	    B_TRUE,
597 	    XD_IS_CD(vdp) ? DDI_NT_CD_XVMD : DDI_NT_BLOCK_XVMD,
598 #if defined(XPV_HVM_DRIVER)
599 	    (XD_IS_CD(vdp) ? 0 : CMLB_CREATE_ALTSLICE_VTOC_16_DTYPE_DIRECT) |
600 	    CMLB_INTERNAL_MINOR_NODES,
601 #else /* !XPV_HVM_DRIVER */
602 	    XD_IS_CD(vdp) ? 0 : CMLB_FAKE_LABEL_ONE_PARTITION,
603 #endif /* !XPV_HVM_DRIVER */
604 	    vdp->xdf_vd_lbl, NULL));
605 }
606 
607 static void
608 xdf_io_err(buf_t *bp, int err, size_t resid)
609 {
610 	bioerror(bp, err);
611 	if (resid == 0)
612 		bp->b_resid = bp->b_bcount;
613 	biodone(bp);
614 }
615 
616 static void
617 xdf_kstat_enter(xdf_t *vdp, buf_t *bp)
618 {
619 	v_req_t *vreq = BP_VREQ(bp);
620 
621 	ASSERT(MUTEX_HELD(&vdp->xdf_dev_lk));
622 
623 	if (vdp->xdf_xdev_iostat == NULL)
624 		return;
625 	if ((vreq != NULL) && vreq->v_runq) {
626 		kstat_runq_enter(KSTAT_IO_PTR(vdp->xdf_xdev_iostat));
627 	} else {
628 		kstat_waitq_enter(KSTAT_IO_PTR(vdp->xdf_xdev_iostat));
629 	}
630 }
631 
632 static void
633 xdf_kstat_exit(xdf_t *vdp, buf_t *bp)
634 {
635 	v_req_t *vreq = BP_VREQ(bp);
636 
637 	ASSERT(MUTEX_HELD(&vdp->xdf_dev_lk));
638 
639 	if (vdp->xdf_xdev_iostat == NULL)
640 		return;
641 	if ((vreq != NULL) && vreq->v_runq) {
642 		kstat_runq_exit(KSTAT_IO_PTR(vdp->xdf_xdev_iostat));
643 	} else {
644 		kstat_waitq_exit(KSTAT_IO_PTR(vdp->xdf_xdev_iostat));
645 	}
646 }
647 
648 static void
649 xdf_kstat_waitq_to_runq(xdf_t *vdp, buf_t *bp)
650 {
651 	v_req_t *vreq = BP_VREQ(bp);
652 
653 	ASSERT(MUTEX_HELD(&vdp->xdf_dev_lk));
654 	ASSERT(!vreq->v_runq);
655 
656 	vreq->v_runq = B_TRUE;
657 	if (vdp->xdf_xdev_iostat == NULL)
658 		return;
659 	kstat_waitq_to_runq(KSTAT_IO_PTR(vdp->xdf_xdev_iostat));
660 }
661 
662 static void
663 xdf_kstat_runq_to_waitq(xdf_t *vdp, buf_t *bp)
664 {
665 	v_req_t *vreq = BP_VREQ(bp);
666 
667 	ASSERT(MUTEX_HELD(&vdp->xdf_dev_lk));
668 	ASSERT(vreq->v_runq);
669 
670 	vreq->v_runq = B_FALSE;
671 	if (vdp->xdf_xdev_iostat == NULL)
672 		return;
673 	kstat_runq_back_to_waitq(KSTAT_IO_PTR(vdp->xdf_xdev_iostat));
674 }
675 
676 int
677 xdf_kstat_create(dev_info_t *dip, char *ks_module, int instance)
678 {
679 	xdf_t		*vdp = (xdf_t *)ddi_get_driver_private(dip);
680 	kstat_t		*kstat;
681 	buf_t		*bp;
682 
683 	if ((kstat = kstat_create(
684 	    ks_module, instance, NULL, "disk",
685 	    KSTAT_TYPE_IO, 1, KSTAT_FLAG_PERSISTENT)) == NULL)
686 		return (-1);
687 
688 	/* See comment about locking in xdf_kstat_delete(). */
689 	mutex_enter(&vdp->xdf_iostat_lk);
690 	mutex_enter(&vdp->xdf_dev_lk);
691 
692 	/* only one kstat can exist at a time */
693 	if (vdp->xdf_xdev_iostat != NULL) {
694 		mutex_exit(&vdp->xdf_dev_lk);
695 		mutex_exit(&vdp->xdf_iostat_lk);
696 		kstat_delete(kstat);
697 		return (-1);
698 	}
699 
700 	vdp->xdf_xdev_iostat = kstat;
701 	vdp->xdf_xdev_iostat->ks_lock = &vdp->xdf_dev_lk;
702 	kstat_install(vdp->xdf_xdev_iostat);
703 
704 	/*
705 	 * Now that we've created a kstat, we need to update the waitq and
706 	 * runq counts for the kstat to reflect our current state.
707 	 *
708 	 * For a buf_t structure to be on the runq, it must have a ring
709 	 * buffer slot associated with it.  To get a ring buffer slot the
710 	 * buf must first have a v_req_t and a ge_slot_t associated with it.
711 	 * Then when it is granted a ring buffer slot, v_runq will be set to
712 	 * true.
713 	 *
714 	 * For a buf_t structure to be on the waitq, it must not be on the
715 	 * runq.  So to find all the buf_t's that should be on waitq, we
716 	 * walk the active buf list and add any buf_t's which aren't on the
717 	 * runq to the waitq.
718 	 */
719 	bp = vdp->xdf_f_act;
720 	while (bp != NULL) {
721 		xdf_kstat_enter(vdp, bp);
722 		bp = bp->av_forw;
723 	}
724 	if (vdp->xdf_ready_tq_bp != NULL)
725 		xdf_kstat_enter(vdp, vdp->xdf_ready_tq_bp);
726 
727 	mutex_exit(&vdp->xdf_dev_lk);
728 	mutex_exit(&vdp->xdf_iostat_lk);
729 	return (0);
730 }
731 
732 void
733 xdf_kstat_delete(dev_info_t *dip)
734 {
735 	xdf_t		*vdp = (xdf_t *)ddi_get_driver_private(dip);
736 	kstat_t		*kstat;
737 	buf_t		*bp;
738 
739 	/*
740 	 * The locking order here is xdf_iostat_lk and then xdf_dev_lk.
741 	 * xdf_dev_lk is used to protect the xdf_xdev_iostat pointer
742 	 * and the contents of the our kstat.  xdf_iostat_lk is used
743 	 * to protect the allocation and freeing of the actual kstat.
744 	 * xdf_dev_lk can't be used for this purpose because kstat
745 	 * readers use it to access the contents of the kstat and
746 	 * hence it can't be held when calling kstat_delete().
747 	 */
748 	mutex_enter(&vdp->xdf_iostat_lk);
749 	mutex_enter(&vdp->xdf_dev_lk);
750 
751 	if (vdp->xdf_xdev_iostat == NULL) {
752 		mutex_exit(&vdp->xdf_dev_lk);
753 		mutex_exit(&vdp->xdf_iostat_lk);
754 		return;
755 	}
756 
757 	/*
758 	 * We're about to destroy the kstat structures, so it isn't really
759 	 * necessary to update the runq and waitq counts.  But, since this
760 	 * isn't a hot code path we can afford to be a little pedantic and
761 	 * go ahead and decrement the runq and waitq kstat counters to zero
762 	 * before free'ing them.  This helps us ensure that we've gotten all
763 	 * our accounting correct.
764 	 *
765 	 * For an explanation of how we determine which buffers go on the
766 	 * runq vs which go on the waitq, see the comments in
767 	 * xdf_kstat_create().
768 	 */
769 	bp = vdp->xdf_f_act;
770 	while (bp != NULL) {
771 		xdf_kstat_exit(vdp, bp);
772 		bp = bp->av_forw;
773 	}
774 	if (vdp->xdf_ready_tq_bp != NULL)
775 		xdf_kstat_exit(vdp, vdp->xdf_ready_tq_bp);
776 
777 	kstat = vdp->xdf_xdev_iostat;
778 	vdp->xdf_xdev_iostat = NULL;
779 	mutex_exit(&vdp->xdf_dev_lk);
780 	kstat_delete(kstat);
781 	mutex_exit(&vdp->xdf_iostat_lk);
782 }
783 
784 /*
785  * Add an IO requests onto the active queue.
786  *
787  * We have to detect IOs generated by xdf_ready_tq_thread.  These IOs
788  * are used to establish a connection to the backend, so they recieve
789  * priority over all other IOs.  Since xdf_ready_tq_thread only does
790  * synchronous IO, there can only be one xdf_ready_tq_thread request at any
791  * given time and we record the buf associated with that request in
792  * xdf_ready_tq_bp.
793  */
794 static void
795 xdf_bp_push(xdf_t *vdp, buf_t *bp)
796 {
797 	ASSERT(MUTEX_HELD(&vdp->xdf_dev_lk));
798 	ASSERT(bp->av_forw == NULL);
799 
800 	xdf_kstat_enter(vdp, bp);
801 
802 	if (curthread == vdp->xdf_ready_tq_thread) {
803 		/* new IO requests from the ready thread */
804 		ASSERT(vdp->xdf_ready_tq_bp == NULL);
805 		vdp->xdf_ready_tq_bp = bp;
806 		return;
807 	}
808 
809 	/* this is normal IO request */
810 	ASSERT(bp != vdp->xdf_ready_tq_bp);
811 
812 	if (vdp->xdf_f_act == NULL) {
813 		/* this is only only IO on the active queue */
814 		ASSERT(vdp->xdf_l_act == NULL);
815 		ASSERT(vdp->xdf_i_act == NULL);
816 		vdp->xdf_f_act = vdp->xdf_l_act = vdp->xdf_i_act = bp;
817 		return;
818 	}
819 
820 	/* add this IO to the tail of the active queue */
821 	vdp->xdf_l_act->av_forw = bp;
822 	vdp->xdf_l_act = bp;
823 	if (vdp->xdf_i_act == NULL)
824 		vdp->xdf_i_act = bp;
825 }
826 
827 static void
828 xdf_bp_pop(xdf_t *vdp, buf_t *bp)
829 {
830 	buf_t	*bp_iter;
831 
832 	ASSERT(MUTEX_HELD(&vdp->xdf_dev_lk));
833 	ASSERT(VREQ_DONE(BP_VREQ(bp)));
834 
835 	if (vdp->xdf_ready_tq_bp == bp) {
836 		/* we're done with a ready thread IO request */
837 		ASSERT(bp->av_forw == NULL);
838 		vdp->xdf_ready_tq_bp = NULL;
839 		return;
840 	}
841 
842 	/* we're done with a normal IO request */
843 	ASSERT((bp->av_forw != NULL) || (bp == vdp->xdf_l_act));
844 	ASSERT((bp->av_forw == NULL) || (bp != vdp->xdf_l_act));
845 	ASSERT(VREQ_DONE(BP_VREQ(vdp->xdf_f_act)));
846 	ASSERT(vdp->xdf_f_act != vdp->xdf_i_act);
847 
848 	if (bp == vdp->xdf_f_act) {
849 		/* This IO was at the head of our active queue. */
850 		vdp->xdf_f_act = bp->av_forw;
851 		if (bp == vdp->xdf_l_act)
852 			vdp->xdf_l_act = NULL;
853 	} else {
854 		/* There IO finished before some other pending IOs. */
855 		bp_iter = vdp->xdf_f_act;
856 		while (bp != bp_iter->av_forw) {
857 			bp_iter = bp_iter->av_forw;
858 			ASSERT(VREQ_DONE(BP_VREQ(bp_iter)));
859 			ASSERT(bp_iter != vdp->xdf_i_act);
860 		}
861 		bp_iter->av_forw = bp->av_forw;
862 		if (bp == vdp->xdf_l_act)
863 			vdp->xdf_l_act = bp_iter;
864 	}
865 	bp->av_forw = NULL;
866 }
867 
868 static buf_t *
869 xdf_bp_next(xdf_t *vdp)
870 {
871 	v_req_t	*vreq;
872 	buf_t	*bp;
873 
874 	if (vdp->xdf_state == XD_CONNECTED) {
875 		/*
876 		 * If we're in the XD_CONNECTED state, we only service IOs
877 		 * from the xdf_ready_tq_thread thread.
878 		 */
879 		if ((bp = vdp->xdf_ready_tq_bp) == NULL)
880 			return (NULL);
881 		if (((vreq = BP_VREQ(bp)) == NULL) || (!VREQ_DONE(vreq)))
882 			return (bp);
883 		return (NULL);
884 	}
885 
886 	/* if we're not in the XD_CONNECTED or XD_READY state we can't do IO */
887 	if (vdp->xdf_state != XD_READY)
888 		return (NULL);
889 
890 	ASSERT(vdp->xdf_ready_tq_bp == NULL);
891 	for (;;) {
892 		if ((bp = vdp->xdf_i_act) == NULL)
893 			return (NULL);
894 		if (((vreq = BP_VREQ(bp)) == NULL) || (!VREQ_DONE(vreq)))
895 			return (bp);
896 
897 		/* advance the active buf index pointer */
898 		vdp->xdf_i_act = bp->av_forw;
899 	}
900 }
901 
902 static void
903 xdf_io_fini(xdf_t *vdp, uint64_t id, int bioerr)
904 {
905 	ge_slot_t	*gs = (ge_slot_t *)(uintptr_t)id;
906 	v_req_t		*vreq = gs->gs_vreq;
907 	buf_t		*bp = vreq->v_buf;
908 
909 	ASSERT(MUTEX_HELD(&vdp->xdf_dev_lk));
910 	ASSERT(BP_VREQ(bp) == vreq);
911 
912 	gs_free(gs);
913 
914 	if (bioerr != 0)
915 		bioerror(bp, bioerr);
916 	ASSERT(vreq->v_nslots > 0);
917 	if (--vreq->v_nslots > 0)
918 		return;
919 
920 	/* remove this IO from our active queue */
921 	xdf_bp_pop(vdp, bp);
922 
923 	ASSERT(vreq->v_runq);
924 	xdf_kstat_exit(vdp, bp);
925 	vreq->v_runq = B_FALSE;
926 	vreq_free(vdp, vreq);
927 
928 	if (IS_ERROR(bp)) {
929 		xdf_io_err(bp, geterror(bp), 0);
930 	} else if (bp->b_resid != 0) {
931 		/* Partial transfers are an error */
932 		xdf_io_err(bp, EIO, bp->b_resid);
933 	} else {
934 		biodone(bp);
935 	}
936 }
937 
938 /*
939  * xdf interrupt handler
940  */
941 static uint_t
942 xdf_intr_locked(xdf_t *vdp)
943 {
944 	xendev_ring_t *xbr;
945 	blkif_response_t *resp;
946 	int bioerr;
947 	uint64_t id;
948 	uint8_t op;
949 	uint16_t status;
950 	ddi_acc_handle_t acchdl;
951 
952 	ASSERT(MUTEX_HELD(&vdp->xdf_dev_lk));
953 
954 	if ((xbr = vdp->xdf_xb_ring) == NULL)
955 		return (DDI_INTR_UNCLAIMED);
956 
957 	acchdl = vdp->xdf_xb_ring_hdl;
958 
959 	/*
960 	 * complete all requests which have a response
961 	 */
962 	while (resp = xvdi_ring_get_response(xbr)) {
963 		id = ddi_get64(acchdl, &resp->id);
964 		op = ddi_get8(acchdl, &resp->operation);
965 		status = ddi_get16(acchdl, (uint16_t *)&resp->status);
966 		DPRINTF(INTR_DBG, ("resp: op %d id %"PRIu64" status %d\n",
967 		    op, id, status));
968 
969 		if (status != BLKIF_RSP_OKAY) {
970 			DPRINTF(IO_DBG, ("xdf@%s: I/O error while %s",
971 			    vdp->xdf_addr,
972 			    (op == BLKIF_OP_READ) ? "reading" : "writing"));
973 			bioerr = EIO;
974 		} else {
975 			bioerr = 0;
976 		}
977 
978 		xdf_io_fini(vdp, id, bioerr);
979 	}
980 	return (DDI_INTR_CLAIMED);
981 }
982 
983 /*
984  * xdf_intr runs at PIL 5, so no one else can grab xdf_dev_lk and
985  * block at a lower pil.
986  */
987 static uint_t
988 xdf_intr(caddr_t arg)
989 {
990 	xdf_t *vdp = (xdf_t *)arg;
991 	int rv;
992 
993 	mutex_enter(&vdp->xdf_dev_lk);
994 	rv = xdf_intr_locked(vdp);
995 	mutex_exit(&vdp->xdf_dev_lk);
996 
997 	if (!do_polled_io)
998 		xdf_io_start(vdp);
999 
1000 	return (rv);
1001 }
1002 
1003 static void
1004 xdf_ring_push(xdf_t *vdp)
1005 {
1006 	ASSERT(MUTEX_HELD(&vdp->xdf_dev_lk));
1007 
1008 	if (vdp->xdf_xb_ring == NULL)
1009 		return;
1010 
1011 	if (xvdi_ring_push_request(vdp->xdf_xb_ring)) {
1012 		DPRINTF(IO_DBG, (
1013 		    "xdf@%s: xdf_ring_push: sent request(s) to backend\n",
1014 		    vdp->xdf_addr));
1015 	}
1016 
1017 	if (xvdi_get_evtchn(vdp->xdf_dip) != INVALID_EVTCHN)
1018 		xvdi_notify_oe(vdp->xdf_dip);
1019 }
1020 
1021 static int
1022 xdf_ring_drain_locked(xdf_t *vdp)
1023 {
1024 	int		pollc, rv = 0;
1025 
1026 	ASSERT(MUTEX_HELD(&vdp->xdf_dev_lk));
1027 
1028 	if (xdf_debug & SUSRES_DBG)
1029 		xen_printf("xdf_ring_drain: start\n");
1030 
1031 	for (pollc = 0; pollc < XDF_DRAIN_RETRY_COUNT; pollc++) {
1032 		if (vdp->xdf_xb_ring == NULL)
1033 			goto out;
1034 
1035 		if (xvdi_ring_has_unconsumed_responses(vdp->xdf_xb_ring))
1036 			(void) xdf_intr_locked(vdp);
1037 		if (!xvdi_ring_has_incomp_request(vdp->xdf_xb_ring))
1038 			goto out;
1039 		xdf_ring_push(vdp);
1040 
1041 		/* file-backed devices can be slow */
1042 		mutex_exit(&vdp->xdf_dev_lk);
1043 #ifdef XPV_HVM_DRIVER
1044 		(void) HYPERVISOR_yield();
1045 #endif /* XPV_HVM_DRIVER */
1046 		delay(drv_usectohz(XDF_DRAIN_MSEC_DELAY));
1047 		mutex_enter(&vdp->xdf_dev_lk);
1048 	}
1049 	cmn_err(CE_WARN, "xdf@%s: xdf_ring_drain: timeout", vdp->xdf_addr);
1050 
1051 out:
1052 	if (vdp->xdf_xb_ring != NULL) {
1053 		if (xvdi_ring_has_incomp_request(vdp->xdf_xb_ring) ||
1054 		    xvdi_ring_has_unconsumed_responses(vdp->xdf_xb_ring))
1055 			rv = EIO;
1056 	}
1057 	if (xdf_debug & SUSRES_DBG)
1058 		xen_printf("xdf@%s: xdf_ring_drain: end, err=%d\n",
1059 		    vdp->xdf_addr, rv);
1060 	return (rv);
1061 }
1062 
1063 static int
1064 xdf_ring_drain(xdf_t *vdp)
1065 {
1066 	int rv;
1067 	mutex_enter(&vdp->xdf_dev_lk);
1068 	rv = xdf_ring_drain_locked(vdp);
1069 	mutex_exit(&vdp->xdf_dev_lk);
1070 	return (rv);
1071 }
1072 
1073 /*
1074  * Destroy all v_req_t, grant table entries, and our ring buffer.
1075  */
1076 static void
1077 xdf_ring_destroy(xdf_t *vdp)
1078 {
1079 	v_req_t		*vreq;
1080 	buf_t		*bp;
1081 	ge_slot_t	*gs;
1082 
1083 	ASSERT(MUTEX_HELD(&vdp->xdf_cb_lk));
1084 	ASSERT(MUTEX_HELD(&vdp->xdf_dev_lk));
1085 
1086 	if ((vdp->xdf_state != XD_INIT) &&
1087 	    (vdp->xdf_state != XD_CONNECTED) &&
1088 	    (vdp->xdf_state != XD_READY)) {
1089 		ASSERT(vdp->xdf_xb_ring == NULL);
1090 		ASSERT(vdp->xdf_xb_ring_hdl == NULL);
1091 		ASSERT(vdp->xdf_peer == INVALID_DOMID);
1092 		ASSERT(vdp->xdf_evtchn == INVALID_EVTCHN);
1093 		ASSERT(list_is_empty(&vdp->xdf_vreq_act));
1094 		return;
1095 	}
1096 
1097 	/*
1098 	 * We don't want to recieve async notifications from the backend
1099 	 * when it finishes processing ring entries.
1100 	 */
1101 #ifdef XPV_HVM_DRIVER
1102 	ec_unbind_evtchn(vdp->xdf_evtchn);
1103 #else /* !XPV_HVM_DRIVER */
1104 	(void) ddi_remove_intr(vdp->xdf_dip, 0, NULL);
1105 #endif /* !XPV_HVM_DRIVER */
1106 
1107 	/*
1108 	 * Drain any requests in the ring.  We need to do this before we
1109 	 * can free grant table entries, because if active ring entries
1110 	 * point to grants, then the backend could be trying to access
1111 	 * those grants.
1112 	 */
1113 	(void) xdf_ring_drain_locked(vdp);
1114 
1115 	/* We're done talking to the backend so free up our event channel */
1116 	xvdi_free_evtchn(vdp->xdf_dip);
1117 	vdp->xdf_evtchn = INVALID_EVTCHN;
1118 
1119 	while ((vreq = list_head(&vdp->xdf_vreq_act)) != NULL) {
1120 		bp = vreq->v_buf;
1121 		ASSERT(BP_VREQ(bp) == vreq);
1122 
1123 		/* Free up any grant table entries associaed with this IO */
1124 		while ((gs = list_head(&vreq->v_gs)) != NULL)
1125 			gs_free(gs);
1126 
1127 		/* If this IO was on the runq, move it back to the waitq. */
1128 		if (vreq->v_runq)
1129 			xdf_kstat_runq_to_waitq(vdp, bp);
1130 
1131 		/*
1132 		 * Reset any buf IO state since we're going to re-issue the
1133 		 * IO when we reconnect.
1134 		 */
1135 		vreq_free(vdp, vreq);
1136 		BP_VREQ_SET(bp, NULL);
1137 		bioerror(bp, 0);
1138 	}
1139 
1140 	/* reset the active queue index pointer */
1141 	vdp->xdf_i_act = vdp->xdf_f_act;
1142 
1143 	/* Destroy the ring */
1144 	xvdi_free_ring(vdp->xdf_xb_ring);
1145 	vdp->xdf_xb_ring = NULL;
1146 	vdp->xdf_xb_ring_hdl = NULL;
1147 	vdp->xdf_peer = INVALID_DOMID;
1148 }
1149 
1150 void
1151 xdfmin(struct buf *bp)
1152 {
1153 	if (bp->b_bcount > xdf_maxphys)
1154 		bp->b_bcount = xdf_maxphys;
1155 }
1156 
1157 /*
1158  * Check if we have a pending "eject" media request.
1159  */
1160 static int
1161 xdf_eject_pending(xdf_t *vdp)
1162 {
1163 	dev_info_t	*dip = vdp->xdf_dip;
1164 	char		*xsname, *str;
1165 
1166 	if (!vdp->xdf_media_req_supported)
1167 		return (B_FALSE);
1168 
1169 	if (((xsname = xvdi_get_xsname(dip)) == NULL) ||
1170 	    (xenbus_read_str(xsname, XBP_MEDIA_REQ, &str) != 0))
1171 		return (B_FALSE);
1172 
1173 	if (strcmp(str, XBV_MEDIA_REQ_EJECT) != 0) {
1174 		strfree(str);
1175 		return (B_FALSE);
1176 	}
1177 	strfree(str);
1178 	return (B_TRUE);
1179 }
1180 
1181 /*
1182  * Generate a media request.
1183  */
1184 static int
1185 xdf_media_req(xdf_t *vdp, char *req, boolean_t media_required)
1186 {
1187 	dev_info_t	*dip = vdp->xdf_dip;
1188 	char		*xsname;
1189 
1190 	/*
1191 	 * we can't be holding xdf_dev_lk because xenbus_printf() can
1192 	 * block while waiting for a PIL 1 interrupt message.  this
1193 	 * would cause a deadlock with xdf_intr() which needs to grab
1194 	 * xdf_dev_lk as well and runs at PIL 5.
1195 	 */
1196 	ASSERT(MUTEX_HELD(&vdp->xdf_cb_lk));
1197 	ASSERT(MUTEX_NOT_HELD(&vdp->xdf_dev_lk));
1198 
1199 	if ((xsname = xvdi_get_xsname(dip)) == NULL)
1200 		return (ENXIO);
1201 
1202 	/* Check if we support media requests */
1203 	if (!XD_IS_CD(vdp) || !vdp->xdf_media_req_supported)
1204 		return (ENOTTY);
1205 
1206 	/* If an eject is pending then don't allow any new requests */
1207 	if (xdf_eject_pending(vdp))
1208 		return (ENXIO);
1209 
1210 	/* Make sure that there is media present */
1211 	if (media_required && (vdp->xdf_xdev_nblocks == 0))
1212 		return (ENXIO);
1213 
1214 	/* We only allow operations when the device is ready and connected */
1215 	if (vdp->xdf_state != XD_READY)
1216 		return (EIO);
1217 
1218 	if (xenbus_printf(XBT_NULL, xsname, XBP_MEDIA_REQ, "%s", req) != 0)
1219 		return (EIO);
1220 
1221 	return (0);
1222 }
1223 
1224 /*
1225  * populate a single blkif_request_t w/ a buf
1226  */
1227 static void
1228 xdf_process_rreq(xdf_t *vdp, struct buf *bp, blkif_request_t *rreq)
1229 {
1230 	grant_ref_t	gr;
1231 	uint8_t		fsect, lsect;
1232 	size_t		bcnt;
1233 	paddr_t		dma_addr;
1234 	off_t		blk_off;
1235 	dev_info_t	*dip = vdp->xdf_dip;
1236 	blkif_vdev_t	vdev = xvdi_get_vdevnum(dip);
1237 	v_req_t		*vreq = BP_VREQ(bp);
1238 	uint64_t	blkno = vreq->v_blkno;
1239 	uint_t		ndmacs = vreq->v_ndmacs;
1240 	ddi_acc_handle_t acchdl = vdp->xdf_xb_ring_hdl;
1241 	int		seg = 0;
1242 	int		isread = IS_READ(bp);
1243 	ge_slot_t	*gs = list_head(&vreq->v_gs);
1244 
1245 	ASSERT(MUTEX_HELD(&vdp->xdf_dev_lk));
1246 	ASSERT(vreq->v_status == VREQ_GS_ALLOCED);
1247 
1248 	if (isread)
1249 		ddi_put8(acchdl, &rreq->operation, BLKIF_OP_READ);
1250 	else {
1251 		switch (vreq->v_flush_diskcache) {
1252 		case FLUSH_DISKCACHE:
1253 			ddi_put8(acchdl, &rreq->operation,
1254 			    BLKIF_OP_FLUSH_DISKCACHE);
1255 			ddi_put16(acchdl, &rreq->handle, vdev);
1256 			ddi_put64(acchdl, &rreq->id,
1257 			    (uint64_t)(uintptr_t)(gs));
1258 			ddi_put8(acchdl, &rreq->nr_segments, 0);
1259 			vreq->v_status = VREQ_DMAWIN_DONE;
1260 			return;
1261 		case WRITE_BARRIER:
1262 			ddi_put8(acchdl, &rreq->operation,
1263 			    BLKIF_OP_WRITE_BARRIER);
1264 			break;
1265 		default:
1266 			if (!vdp->xdf_wce)
1267 				ddi_put8(acchdl, &rreq->operation,
1268 				    BLKIF_OP_WRITE_BARRIER);
1269 			else
1270 				ddi_put8(acchdl, &rreq->operation,
1271 				    BLKIF_OP_WRITE);
1272 			break;
1273 		}
1274 	}
1275 
1276 	ddi_put16(acchdl, &rreq->handle, vdev);
1277 	ddi_put64(acchdl, &rreq->sector_number, blkno);
1278 	ddi_put64(acchdl, &rreq->id, (uint64_t)(uintptr_t)(gs));
1279 
1280 	/*
1281 	 * loop until all segments are populated or no more dma cookie in buf
1282 	 */
1283 	for (;;) {
1284 		/*
1285 		 * Each segment of a blkif request can transfer up to
1286 		 * one 4K page of data.
1287 		 */
1288 		bcnt = vreq->v_dmac.dmac_size;
1289 		dma_addr = vreq->v_dmac.dmac_laddress;
1290 		blk_off = (uint_t)((paddr_t)XB_SEGOFFSET & dma_addr);
1291 		fsect = blk_off >> XB_BSHIFT;
1292 		lsect = fsect + (bcnt >> XB_BSHIFT) - 1;
1293 
1294 		ASSERT(bcnt <= PAGESIZE);
1295 		ASSERT((bcnt % XB_BSIZE) == 0);
1296 		ASSERT((blk_off & XB_BMASK) == 0);
1297 		ASSERT(fsect < XB_MAX_SEGLEN / XB_BSIZE &&
1298 		    lsect < XB_MAX_SEGLEN / XB_BSIZE);
1299 
1300 		gr = gs_grant(gs, PATOMA(dma_addr) >> PAGESHIFT);
1301 		ddi_put32(acchdl, &rreq->seg[seg].gref, gr);
1302 		ddi_put8(acchdl, &rreq->seg[seg].first_sect, fsect);
1303 		ddi_put8(acchdl, &rreq->seg[seg].last_sect, lsect);
1304 
1305 		DPRINTF(IO_DBG, (
1306 		    "xdf@%s: seg%d: dmacS %lu blk_off %ld\n",
1307 		    vdp->xdf_addr, seg, vreq->v_dmac.dmac_size, blk_off));
1308 		DPRINTF(IO_DBG, (
1309 		    "xdf@%s: seg%d: fs %d ls %d gr %d dma 0x%"PRIx64"\n",
1310 		    vdp->xdf_addr, seg, fsect, lsect, gr, dma_addr));
1311 
1312 		blkno += (bcnt >> XB_BSHIFT);
1313 		seg++;
1314 		ASSERT(seg <= BLKIF_MAX_SEGMENTS_PER_REQUEST);
1315 		if (--ndmacs) {
1316 			ddi_dma_nextcookie(vreq->v_dmahdl, &vreq->v_dmac);
1317 			continue;
1318 		}
1319 
1320 		vreq->v_status = VREQ_DMAWIN_DONE;
1321 		vreq->v_blkno = blkno;
1322 		break;
1323 	}
1324 	ddi_put8(acchdl,  &rreq->nr_segments, seg);
1325 	DPRINTF(IO_DBG, (
1326 	    "xdf@%s: xdf_process_rreq: request id=%"PRIx64" ready\n",
1327 	    vdp->xdf_addr, rreq->id));
1328 }
1329 
1330 static void
1331 xdf_io_start(xdf_t *vdp)
1332 {
1333 	struct buf	*bp;
1334 	v_req_t		*vreq;
1335 	blkif_request_t	*rreq;
1336 	boolean_t	rreqready = B_FALSE;
1337 
1338 	mutex_enter(&vdp->xdf_dev_lk);
1339 
1340 	/*
1341 	 * Populate the ring request(s).  Loop until there is no buf to
1342 	 * transfer or no free slot available in I/O ring.
1343 	 */
1344 	for (;;) {
1345 		/* don't start any new IO if we're suspending */
1346 		if (vdp->xdf_suspending)
1347 			break;
1348 		if ((bp = xdf_bp_next(vdp)) == NULL)
1349 			break;
1350 
1351 		/* if the buf doesn't already have a vreq, allocate one */
1352 		if (((vreq = BP_VREQ(bp)) == NULL) &&
1353 		    ((vreq = vreq_get(vdp, bp)) == NULL))
1354 			break;
1355 
1356 		/* alloc DMA/GTE resources */
1357 		if (vreq_setup(vdp, vreq) != DDI_SUCCESS)
1358 			break;
1359 
1360 		/* get next blkif_request in the ring */
1361 		if ((rreq = xvdi_ring_get_request(vdp->xdf_xb_ring)) == NULL)
1362 			break;
1363 		bzero(rreq, sizeof (blkif_request_t));
1364 		rreqready = B_TRUE;
1365 
1366 		/* populate blkif_request with this buf */
1367 		xdf_process_rreq(vdp, bp, rreq);
1368 
1369 		/*
1370 		 * This buffer/vreq pair is has been allocated a ring buffer
1371 		 * resources, so if it isn't already in our runq, add it.
1372 		 */
1373 		if (!vreq->v_runq)
1374 			xdf_kstat_waitq_to_runq(vdp, bp);
1375 	}
1376 
1377 	/* Send the request(s) to the backend */
1378 	if (rreqready)
1379 		xdf_ring_push(vdp);
1380 
1381 	mutex_exit(&vdp->xdf_dev_lk);
1382 }
1383 
1384 
1385 /* check if partition is open, -1 - check all partitions on the disk */
1386 static boolean_t
1387 xdf_isopen(xdf_t *vdp, int partition)
1388 {
1389 	int i;
1390 	ulong_t parbit;
1391 	boolean_t rval = B_FALSE;
1392 
1393 	ASSERT((partition == -1) ||
1394 	    ((partition >= 0) || (partition < XDF_PEXT)));
1395 
1396 	if (partition == -1)
1397 		parbit = (ulong_t)-1;
1398 	else
1399 		parbit = 1 << partition;
1400 
1401 	for (i = 0; i < OTYPCNT; i++) {
1402 		if (vdp->xdf_vd_open[i] & parbit)
1403 			rval = B_TRUE;
1404 	}
1405 
1406 	return (rval);
1407 }
1408 
1409 /*
1410  * The connection should never be closed as long as someone is holding
1411  * us open, there is pending IO, or someone is waiting waiting for a
1412  * connection.
1413  */
1414 static boolean_t
1415 xdf_busy(xdf_t *vdp)
1416 {
1417 	ASSERT(MUTEX_HELD(&vdp->xdf_dev_lk));
1418 
1419 	if ((vdp->xdf_xb_ring != NULL) &&
1420 	    xvdi_ring_has_unconsumed_responses(vdp->xdf_xb_ring)) {
1421 		ASSERT(vdp->xdf_state != XD_CLOSED);
1422 		return (B_TRUE);
1423 	}
1424 
1425 	if (!list_is_empty(&vdp->xdf_vreq_act) || (vdp->xdf_f_act != NULL)) {
1426 		ASSERT(vdp->xdf_state != XD_CLOSED);
1427 		return (B_TRUE);
1428 	}
1429 
1430 	if (xdf_isopen(vdp, -1)) {
1431 		ASSERT(vdp->xdf_state != XD_CLOSED);
1432 		return (B_TRUE);
1433 	}
1434 
1435 	if (vdp->xdf_connect_req > 0) {
1436 		ASSERT(vdp->xdf_state != XD_CLOSED);
1437 		return (B_TRUE);
1438 	}
1439 
1440 	return (B_FALSE);
1441 }
1442 
1443 static void
1444 xdf_set_state(xdf_t *vdp, xdf_state_t new_state)
1445 {
1446 	ASSERT(MUTEX_HELD(&vdp->xdf_cb_lk));
1447 	ASSERT(MUTEX_HELD(&vdp->xdf_dev_lk));
1448 	DPRINTF(DDI_DBG, ("xdf@%s: state change %d -> %d\n",
1449 	    vdp->xdf_addr, vdp->xdf_state, new_state));
1450 	vdp->xdf_state = new_state;
1451 	cv_broadcast(&vdp->xdf_dev_cv);
1452 }
1453 
1454 static void
1455 xdf_disconnect(xdf_t *vdp, xdf_state_t new_state, boolean_t quiet)
1456 {
1457 	dev_info_t	*dip = vdp->xdf_dip;
1458 	boolean_t	busy;
1459 
1460 	ASSERT(MUTEX_HELD(&vdp->xdf_cb_lk));
1461 	ASSERT(MUTEX_NOT_HELD(&vdp->xdf_dev_lk));
1462 	ASSERT((new_state == XD_UNKNOWN) || (new_state == XD_CLOSED));
1463 
1464 	/* Check if we're already there. */
1465 	if (vdp->xdf_state == new_state)
1466 		return;
1467 
1468 	mutex_enter(&vdp->xdf_dev_lk);
1469 	busy = xdf_busy(vdp);
1470 
1471 	/* If we're already closed then there's nothing todo. */
1472 	if (vdp->xdf_state == XD_CLOSED) {
1473 		ASSERT(!busy);
1474 		xdf_set_state(vdp, new_state);
1475 		mutex_exit(&vdp->xdf_dev_lk);
1476 		return;
1477 	}
1478 
1479 #ifdef DEBUG
1480 	/* UhOh.  Warn the user that something bad has happened. */
1481 	if (!quiet && busy && (vdp->xdf_state == XD_READY) &&
1482 	    (vdp->xdf_xdev_nblocks != 0)) {
1483 		cmn_err(CE_WARN, "xdf@%s: disconnected while in use",
1484 		    vdp->xdf_addr);
1485 	}
1486 #endif /* DEBUG */
1487 
1488 	xdf_ring_destroy(vdp);
1489 
1490 	/* If we're busy then we can only go into the unknown state */
1491 	xdf_set_state(vdp, (busy) ? XD_UNKNOWN : new_state);
1492 	mutex_exit(&vdp->xdf_dev_lk);
1493 
1494 	/* if we're closed now, let the other end know */
1495 	if (vdp->xdf_state == XD_CLOSED)
1496 		(void) xvdi_switch_state(dip, XBT_NULL, XenbusStateClosed);
1497 }
1498 
1499 
1500 /*
1501  * Kick-off connect process
1502  * Status should be XD_UNKNOWN or XD_CLOSED
1503  * On success, status will be changed to XD_INIT
1504  * On error, it will be changed to XD_UNKNOWN
1505  */
1506 static int
1507 xdf_setstate_init(xdf_t *vdp)
1508 {
1509 	dev_info_t		*dip = vdp->xdf_dip;
1510 	xenbus_transaction_t	xbt;
1511 	grant_ref_t		gref;
1512 	char			*xsname, *str;
1513 	int 			rv;
1514 
1515 	ASSERT(MUTEX_HELD(&vdp->xdf_cb_lk));
1516 	ASSERT(MUTEX_NOT_HELD(&vdp->xdf_dev_lk));
1517 	ASSERT((vdp->xdf_state == XD_UNKNOWN) ||
1518 	    (vdp->xdf_state == XD_CLOSED));
1519 
1520 	DPRINTF(DDI_DBG,
1521 	    ("xdf@%s: starting connection process\n", vdp->xdf_addr));
1522 
1523 	/*
1524 	 * If an eject is pending then don't allow a new connection.
1525 	 * (Only the backend can clear media request eject request.)
1526 	 */
1527 	if (xdf_eject_pending(vdp))
1528 		return (DDI_FAILURE);
1529 
1530 	if ((xsname = xvdi_get_xsname(dip)) == NULL)
1531 		goto errout;
1532 
1533 	if ((vdp->xdf_peer = xvdi_get_oeid(dip)) == INVALID_DOMID)
1534 		goto errout;
1535 
1536 	(void) xvdi_switch_state(dip, XBT_NULL, XenbusStateInitialising);
1537 
1538 	/*
1539 	 * Sanity check for the existance of the xenbus device-type property.
1540 	 * This property might not exist if we our xenbus device nodes was
1541 	 * force destroyed while we were still connected to the backend.
1542 	 */
1543 	if (xenbus_read_str(xsname, XBP_DEV_TYPE, &str) != 0)
1544 		goto errout;
1545 	strfree(str);
1546 
1547 	if (xvdi_alloc_evtchn(dip) != DDI_SUCCESS)
1548 		goto errout;
1549 
1550 	vdp->xdf_evtchn = xvdi_get_evtchn(dip);
1551 #ifdef XPV_HVM_DRIVER
1552 	ec_bind_evtchn_to_handler(vdp->xdf_evtchn, IPL_VBD, xdf_intr, vdp);
1553 #else /* !XPV_HVM_DRIVER */
1554 	if (ddi_add_intr(dip, 0, NULL, NULL, xdf_intr, (caddr_t)vdp) !=
1555 	    DDI_SUCCESS) {
1556 		cmn_err(CE_WARN, "xdf@%s: xdf_setstate_init: "
1557 		    "failed to add intr handler", vdp->xdf_addr);
1558 		goto errout1;
1559 	}
1560 #endif /* !XPV_HVM_DRIVER */
1561 
1562 	if (xvdi_alloc_ring(dip, BLKIF_RING_SIZE,
1563 	    sizeof (union blkif_sring_entry), &gref, &vdp->xdf_xb_ring) !=
1564 	    DDI_SUCCESS) {
1565 		cmn_err(CE_WARN, "xdf@%s: failed to alloc comm ring",
1566 		    vdp->xdf_addr);
1567 		goto errout2;
1568 	}
1569 	vdp->xdf_xb_ring_hdl = vdp->xdf_xb_ring->xr_acc_hdl; /* ugly!! */
1570 
1571 	/*
1572 	 * Write into xenstore the info needed by backend
1573 	 */
1574 trans_retry:
1575 	if (xenbus_transaction_start(&xbt)) {
1576 		cmn_err(CE_WARN, "xdf@%s: failed to start transaction",
1577 		    vdp->xdf_addr);
1578 		xvdi_fatal_error(dip, EIO, "connect transaction init");
1579 		goto fail_trans;
1580 	}
1581 
1582 	/*
1583 	 * XBP_PROTOCOL is written by the domain builder in the case of PV
1584 	 * domains. However, it is not written for HVM domains, so let's
1585 	 * write it here.
1586 	 */
1587 	if (((rv = xenbus_printf(xbt, xsname,
1588 	    XBP_MEDIA_REQ, "%s", XBV_MEDIA_REQ_NONE)) != 0) ||
1589 	    ((rv = xenbus_printf(xbt, xsname,
1590 	    XBP_RING_REF, "%u", gref)) != 0) ||
1591 	    ((rv = xenbus_printf(xbt, xsname,
1592 	    XBP_EVENT_CHAN, "%u", vdp->xdf_evtchn)) != 0) ||
1593 	    ((rv = xenbus_printf(xbt, xsname,
1594 	    XBP_PROTOCOL, "%s", XEN_IO_PROTO_ABI_NATIVE)) != 0) ||
1595 	    ((rv = xvdi_switch_state(dip, xbt, XenbusStateInitialised)) > 0)) {
1596 		(void) xenbus_transaction_end(xbt, 1);
1597 		xvdi_fatal_error(dip, rv, "connect transaction setup");
1598 		goto fail_trans;
1599 	}
1600 
1601 	/* kick-off connect process */
1602 	if (rv = xenbus_transaction_end(xbt, 0)) {
1603 		if (rv == EAGAIN)
1604 			goto trans_retry;
1605 		xvdi_fatal_error(dip, rv, "connect transaction commit");
1606 		goto fail_trans;
1607 	}
1608 
1609 	ASSERT(MUTEX_HELD(&vdp->xdf_cb_lk));
1610 	mutex_enter(&vdp->xdf_dev_lk);
1611 	xdf_set_state(vdp, XD_INIT);
1612 	mutex_exit(&vdp->xdf_dev_lk);
1613 
1614 	return (DDI_SUCCESS);
1615 
1616 fail_trans:
1617 	xvdi_free_ring(vdp->xdf_xb_ring);
1618 errout2:
1619 #ifdef XPV_HVM_DRIVER
1620 	ec_unbind_evtchn(vdp->xdf_evtchn);
1621 #else /* !XPV_HVM_DRIVER */
1622 	(void) ddi_remove_intr(vdp->xdf_dip, 0, NULL);
1623 #endif /* !XPV_HVM_DRIVER */
1624 errout1:
1625 	xvdi_free_evtchn(dip);
1626 	vdp->xdf_evtchn = INVALID_EVTCHN;
1627 errout:
1628 	xdf_disconnect(vdp, XD_UNKNOWN, B_FALSE);
1629 	cmn_err(CE_WARN, "xdf@%s: failed to start connection to backend",
1630 	    vdp->xdf_addr);
1631 	return (DDI_FAILURE);
1632 }
1633 
1634 int
1635 xdf_get_flush_block(xdf_t *vdp)
1636 {
1637 	/*
1638 	 * Get a DEV_BSIZE aligned bufer
1639 	 */
1640 	vdp->xdf_flush_mem = kmem_alloc(vdp->xdf_xdev_secsize * 2, KM_SLEEP);
1641 	vdp->xdf_cache_flush_block =
1642 	    (char *)P2ROUNDUP((uintptr_t)(vdp->xdf_flush_mem),
1643 	    (int)vdp->xdf_xdev_secsize);
1644 
1645 	if (xdf_lb_rdwr(vdp->xdf_dip, TG_READ, vdp->xdf_cache_flush_block,
1646 	    xdf_flush_block, vdp->xdf_xdev_secsize, NULL) != 0)
1647 		return (DDI_FAILURE);
1648 	return (DDI_SUCCESS);
1649 }
1650 
1651 static void
1652 xdf_setstate_ready(void *arg)
1653 {
1654 	xdf_t	*vdp = (xdf_t *)arg;
1655 
1656 	vdp->xdf_ready_tq_thread = curthread;
1657 
1658 	/*
1659 	 * We've created all the minor nodes via cmlb_attach() using default
1660 	 * value in xdf_attach() to make it possible to block in xdf_open(),
1661 	 * in case there's anyone (say, booting thread) ever trying to open
1662 	 * it before connected to backend. We will refresh all those minor
1663 	 * nodes w/ latest info we've got now when we are almost connected.
1664 	 */
1665 	mutex_enter(&vdp->xdf_dev_lk);
1666 	if (vdp->xdf_cmbl_reattach) {
1667 		vdp->xdf_cmbl_reattach = B_FALSE;
1668 
1669 		mutex_exit(&vdp->xdf_dev_lk);
1670 		if (xdf_cmlb_attach(vdp) != 0) {
1671 			xdf_disconnect(vdp, XD_UNKNOWN, B_FALSE);
1672 			return;
1673 		}
1674 		mutex_enter(&vdp->xdf_dev_lk);
1675 	}
1676 
1677 	/* If we're not still trying to get to the ready state, then bail. */
1678 	if (vdp->xdf_state != XD_CONNECTED) {
1679 		mutex_exit(&vdp->xdf_dev_lk);
1680 		return;
1681 	}
1682 	mutex_exit(&vdp->xdf_dev_lk);
1683 
1684 	/*
1685 	 * If backend has feature-barrier, see if it supports disk
1686 	 * cache flush op.
1687 	 */
1688 	vdp->xdf_flush_supported = B_FALSE;
1689 	if (vdp->xdf_feature_barrier) {
1690 		/*
1691 		 * Pretend we already know flush is supported so probe
1692 		 * will attempt the correct op.
1693 		 */
1694 		vdp->xdf_flush_supported = B_TRUE;
1695 		if (xdf_lb_rdwr(vdp->xdf_dip, TG_WRITE, NULL, 0, 0, 0) == 0) {
1696 			vdp->xdf_flush_supported = B_TRUE;
1697 		} else {
1698 			vdp->xdf_flush_supported = B_FALSE;
1699 			/*
1700 			 * If the other end does not support the cache flush op
1701 			 * then we must use a barrier-write to force disk
1702 			 * cache flushing.  Barrier writes require that a data
1703 			 * block actually be written.
1704 			 * Cache a block to barrier-write when we are
1705 			 * asked to perform a flush.
1706 			 * XXX - would it be better to just copy 1 block
1707 			 * (512 bytes) from whatever write we did last
1708 			 * and rewrite that block?
1709 			 */
1710 			if (xdf_get_flush_block(vdp) != DDI_SUCCESS) {
1711 				xdf_disconnect(vdp, XD_UNKNOWN, B_FALSE);
1712 				return;
1713 			}
1714 		}
1715 	}
1716 
1717 	mutex_enter(&vdp->xdf_cb_lk);
1718 	mutex_enter(&vdp->xdf_dev_lk);
1719 	if (vdp->xdf_state == XD_CONNECTED)
1720 		xdf_set_state(vdp, XD_READY);
1721 	mutex_exit(&vdp->xdf_dev_lk);
1722 
1723 	/* Restart any currently queued up io */
1724 	xdf_io_start(vdp);
1725 
1726 	mutex_exit(&vdp->xdf_cb_lk);
1727 }
1728 
1729 /*
1730  * synthetic geometry
1731  */
1732 #define	XDF_NSECTS	256
1733 #define	XDF_NHEADS	16
1734 
1735 static void
1736 xdf_synthetic_pgeom(dev_info_t *dip, cmlb_geom_t *geomp)
1737 {
1738 	xdf_t *vdp;
1739 	uint_t ncyl;
1740 
1741 	vdp = ddi_get_soft_state(xdf_ssp, ddi_get_instance(dip));
1742 
1743 	ncyl = vdp->xdf_xdev_nblocks / (XDF_NHEADS * XDF_NSECTS);
1744 
1745 	bzero(geomp, sizeof (*geomp));
1746 	geomp->g_ncyl = ncyl == 0 ? 1 : ncyl;
1747 	geomp->g_acyl = 0;
1748 	geomp->g_nhead = XDF_NHEADS;
1749 	geomp->g_nsect = XDF_NSECTS;
1750 	geomp->g_secsize = vdp->xdf_xdev_secsize;
1751 	geomp->g_capacity = vdp->xdf_xdev_nblocks;
1752 	geomp->g_intrlv = 0;
1753 	geomp->g_rpm = 7200;
1754 }
1755 
1756 /*
1757  * Finish other initialization after we've connected to backend
1758  * Status should be XD_INIT before calling this routine
1759  * On success, status should be changed to XD_CONNECTED.
1760  * On error, status should stay XD_INIT
1761  */
1762 static int
1763 xdf_setstate_connected(xdf_t *vdp)
1764 {
1765 	dev_info_t	*dip = vdp->xdf_dip;
1766 	cmlb_geom_t	pgeom;
1767 	diskaddr_t	nblocks = 0;
1768 	uint_t		secsize = 0;
1769 	char		*oename, *xsname, *str;
1770 	uint_t		dinfo;
1771 
1772 	ASSERT(MUTEX_HELD(&vdp->xdf_cb_lk));
1773 	ASSERT(MUTEX_NOT_HELD(&vdp->xdf_dev_lk));
1774 	ASSERT(vdp->xdf_state == XD_INIT);
1775 
1776 	if (((xsname = xvdi_get_xsname(dip)) == NULL) ||
1777 	    ((oename = xvdi_get_oename(dip)) == NULL))
1778 		return (DDI_FAILURE);
1779 
1780 	/* Make sure the other end is XenbusStateConnected */
1781 	if (xenbus_read_driver_state(oename) != XenbusStateConnected)
1782 		return (DDI_FAILURE);
1783 
1784 	/* Determine if feature barrier is supported by backend */
1785 	if (!(vdp->xdf_feature_barrier = xenbus_exists(oename, XBP_FB)))
1786 		cmn_err(CE_NOTE, "!xdf@%s: feature-barrier not supported",
1787 		    vdp->xdf_addr);
1788 
1789 	/*
1790 	 * Probe backend.  Read the device size into xdf_xdev_nblocks
1791 	 * and set the VDISK_READONLY, VDISK_CDROM, and VDISK_REMOVABLE
1792 	 * flags in xdf_dinfo.  If the emulated device type is "cdrom",
1793 	 * we always set VDISK_CDROM, regardless of if it's present in
1794 	 * the xenbus info parameter.
1795 	 */
1796 	if (xenbus_gather(XBT_NULL, oename,
1797 	    XBP_SECTORS, "%"SCNu64, &nblocks,
1798 	    XBP_SECTOR_SIZE, "%u", &secsize,
1799 	    XBP_INFO, "%u", &dinfo,
1800 	    NULL) != 0) {
1801 		cmn_err(CE_WARN, "xdf@%s: xdf_setstate_connected: "
1802 		    "cannot read backend info", vdp->xdf_addr);
1803 		return (DDI_FAILURE);
1804 	}
1805 	if (xenbus_read_str(xsname, XBP_DEV_TYPE, &str) != 0) {
1806 		cmn_err(CE_WARN, "xdf@%s: cannot read device-type",
1807 		    vdp->xdf_addr);
1808 		return (DDI_FAILURE);
1809 	}
1810 	if (strcmp(str, XBV_DEV_TYPE_CD) == 0)
1811 		dinfo |= VDISK_CDROM;
1812 	strfree(str);
1813 
1814 	if (secsize == 0 || !(ISP2(secsize / DEV_BSIZE)))
1815 		secsize = DEV_BSIZE;
1816 	vdp->xdf_xdev_nblocks = nblocks;
1817 	vdp->xdf_xdev_secsize = secsize;
1818 #ifdef _ILP32
1819 	if (vdp->xdf_xdev_nblocks > DK_MAX_BLOCKS) {
1820 		cmn_err(CE_WARN, "xdf@%s: xdf_setstate_connected: "
1821 		    "backend disk device too large with %llu blocks for"
1822 		    " 32-bit kernel", vdp->xdf_addr, vdp->xdf_xdev_nblocks);
1823 		xvdi_fatal_error(dip, EFBIG, "reading backend info");
1824 		return (DDI_FAILURE);
1825 	}
1826 #endif
1827 
1828 	/*
1829 	 * If the physical geometry for a fixed disk has been explicity
1830 	 * set then make sure that the specified physical geometry isn't
1831 	 * larger than the device we connected to.
1832 	 */
1833 	if (vdp->xdf_pgeom_fixed &&
1834 	    (vdp->xdf_pgeom.g_capacity > vdp->xdf_xdev_nblocks)) {
1835 		cmn_err(CE_WARN,
1836 		    "xdf@%s: connect failed, fixed geometry too large",
1837 		    vdp->xdf_addr);
1838 		return (DDI_FAILURE);
1839 	}
1840 
1841 	vdp->xdf_media_req_supported = xenbus_exists(oename, XBP_MEDIA_REQ_SUP);
1842 
1843 	/* mark vbd is ready for I/O */
1844 	mutex_enter(&vdp->xdf_dev_lk);
1845 	xdf_set_state(vdp, XD_CONNECTED);
1846 
1847 	/* check if the cmlb label should be updated */
1848 	xdf_synthetic_pgeom(dip, &pgeom);
1849 	if ((vdp->xdf_dinfo != dinfo) ||
1850 	    (!vdp->xdf_pgeom_fixed &&
1851 	    (memcmp(&vdp->xdf_pgeom, &pgeom, sizeof (pgeom)) != 0))) {
1852 		vdp->xdf_cmbl_reattach = B_TRUE;
1853 
1854 		vdp->xdf_dinfo = dinfo;
1855 		if (!vdp->xdf_pgeom_fixed)
1856 			vdp->xdf_pgeom = pgeom;
1857 	}
1858 
1859 	if (XD_IS_CD(vdp) || XD_IS_RM(vdp)) {
1860 		if (vdp->xdf_xdev_nblocks == 0) {
1861 			vdp->xdf_mstate = DKIO_EJECTED;
1862 			cv_broadcast(&vdp->xdf_mstate_cv);
1863 		} else {
1864 			vdp->xdf_mstate = DKIO_INSERTED;
1865 			cv_broadcast(&vdp->xdf_mstate_cv);
1866 		}
1867 	} else {
1868 		if (vdp->xdf_mstate != DKIO_NONE) {
1869 			vdp->xdf_mstate = DKIO_NONE;
1870 			cv_broadcast(&vdp->xdf_mstate_cv);
1871 		}
1872 	}
1873 
1874 	mutex_exit(&vdp->xdf_dev_lk);
1875 
1876 	cmn_err(CE_CONT, "?xdf@%s: %"PRIu64" blocks", vdp->xdf_addr,
1877 	    (uint64_t)vdp->xdf_xdev_nblocks);
1878 
1879 	/* Restart any currently queued up io */
1880 	xdf_io_start(vdp);
1881 
1882 	/*
1883 	 * To get to the ready state we have to do IO to the backend device,
1884 	 * but we can't initiate IO from the other end change callback thread
1885 	 * (which is the current context we're executing in.)  This is because
1886 	 * if the other end disconnects while we're doing IO from the callback
1887 	 * thread, then we can't recieve that disconnect event and we hang
1888 	 * waiting for an IO that can never complete.
1889 	 */
1890 	(void) ddi_taskq_dispatch(vdp->xdf_ready_tq, xdf_setstate_ready, vdp,
1891 	    DDI_SLEEP);
1892 
1893 	(void) xvdi_switch_state(dip, XBT_NULL, XenbusStateConnected);
1894 	return (DDI_SUCCESS);
1895 }
1896 
1897 /*ARGSUSED*/
1898 static void
1899 xdf_oe_change(dev_info_t *dip, ddi_eventcookie_t id, void *arg, void *impl_data)
1900 {
1901 	XenbusState new_state = *(XenbusState *)impl_data;
1902 	xdf_t *vdp = (xdf_t *)ddi_get_driver_private(dip);
1903 
1904 	DPRINTF(DDI_DBG, ("xdf@%s: otherend state change to %d!\n",
1905 	    vdp->xdf_addr, new_state));
1906 
1907 	mutex_enter(&vdp->xdf_cb_lk);
1908 
1909 	/* We assume that this callback is single threaded */
1910 	ASSERT(vdp->xdf_oe_change_thread == NULL);
1911 	DEBUG_EVAL(vdp->xdf_oe_change_thread = curthread);
1912 
1913 	/* ignore any backend state changes if we're suspending/suspended */
1914 	if (vdp->xdf_suspending || (vdp->xdf_state == XD_SUSPEND)) {
1915 		DEBUG_EVAL(vdp->xdf_oe_change_thread = NULL);
1916 		mutex_exit(&vdp->xdf_cb_lk);
1917 		return;
1918 	}
1919 
1920 	switch (new_state) {
1921 	case XenbusStateUnknown:
1922 	case XenbusStateInitialising:
1923 	case XenbusStateInitWait:
1924 	case XenbusStateInitialised:
1925 		if (vdp->xdf_state == XD_INIT)
1926 			break;
1927 
1928 		xdf_disconnect(vdp, XD_UNKNOWN, B_FALSE);
1929 		if (xdf_setstate_init(vdp) != DDI_SUCCESS)
1930 			break;
1931 		ASSERT(vdp->xdf_state == XD_INIT);
1932 		break;
1933 
1934 	case XenbusStateConnected:
1935 		if ((vdp->xdf_state == XD_CONNECTED) ||
1936 		    (vdp->xdf_state == XD_READY))
1937 			break;
1938 
1939 		if (vdp->xdf_state != XD_INIT) {
1940 			xdf_disconnect(vdp, XD_UNKNOWN, B_FALSE);
1941 			if (xdf_setstate_init(vdp) != DDI_SUCCESS)
1942 				break;
1943 			ASSERT(vdp->xdf_state == XD_INIT);
1944 		}
1945 
1946 		if (xdf_setstate_connected(vdp) != DDI_SUCCESS) {
1947 			xdf_disconnect(vdp, XD_UNKNOWN, B_FALSE);
1948 			break;
1949 		}
1950 		ASSERT(vdp->xdf_state == XD_CONNECTED);
1951 		break;
1952 
1953 	case XenbusStateClosing:
1954 		if (xdf_isopen(vdp, -1)) {
1955 			cmn_err(CE_NOTE,
1956 			    "xdf@%s: hot-unplug failed, still in use",
1957 			    vdp->xdf_addr);
1958 			break;
1959 		}
1960 		/*FALLTHROUGH*/
1961 	case XenbusStateClosed:
1962 		xdf_disconnect(vdp, XD_CLOSED, B_FALSE);
1963 		break;
1964 	}
1965 
1966 	/* notify anybody waiting for oe state change */
1967 	cv_broadcast(&vdp->xdf_dev_cv);
1968 	DEBUG_EVAL(vdp->xdf_oe_change_thread = NULL);
1969 	mutex_exit(&vdp->xdf_cb_lk);
1970 }
1971 
1972 static int
1973 xdf_connect_locked(xdf_t *vdp, boolean_t wait)
1974 {
1975 	int	rv, timeouts = 0, reset = 20;
1976 
1977 	ASSERT(MUTEX_HELD(&vdp->xdf_cb_lk));
1978 	ASSERT(MUTEX_HELD(&vdp->xdf_dev_lk));
1979 
1980 	/* we can't connect once we're in the closed state */
1981 	if (vdp->xdf_state == XD_CLOSED)
1982 		return (XD_CLOSED);
1983 
1984 	vdp->xdf_connect_req++;
1985 	while (vdp->xdf_state != XD_READY) {
1986 		mutex_exit(&vdp->xdf_dev_lk);
1987 
1988 		/* only one thread at a time can be the connection thread */
1989 		if (vdp->xdf_connect_thread == NULL)
1990 			vdp->xdf_connect_thread = curthread;
1991 
1992 		if (vdp->xdf_connect_thread == curthread) {
1993 			if ((timeouts > 0) && ((timeouts % reset) == 0)) {
1994 				/*
1995 				 * If we haven't establised a connection
1996 				 * within the reset time, then disconnect
1997 				 * so we can try again, and double the reset
1998 				 * time.  The reset time starts at 2 sec.
1999 				 */
2000 				(void) xdf_disconnect(vdp, XD_UNKNOWN, B_TRUE);
2001 				reset *= 2;
2002 			}
2003 			if (vdp->xdf_state == XD_UNKNOWN)
2004 				(void) xdf_setstate_init(vdp);
2005 			if (vdp->xdf_state == XD_INIT)
2006 				(void) xdf_setstate_connected(vdp);
2007 		}
2008 
2009 		mutex_enter(&vdp->xdf_dev_lk);
2010 		if (!wait || (vdp->xdf_state == XD_READY))
2011 			goto out;
2012 
2013 		mutex_exit((&vdp->xdf_cb_lk));
2014 		if (vdp->xdf_connect_thread != curthread) {
2015 			rv = cv_wait_sig(&vdp->xdf_dev_cv, &vdp->xdf_dev_lk);
2016 		} else {
2017 			/* delay for 0.1 sec */
2018 			rv = cv_reltimedwait_sig(&vdp->xdf_dev_cv,
2019 			    &vdp->xdf_dev_lk, drv_usectohz(100*1000),
2020 			    TR_CLOCK_TICK);
2021 			if (rv == -1)
2022 				timeouts++;
2023 		}
2024 		mutex_exit((&vdp->xdf_dev_lk));
2025 		mutex_enter((&vdp->xdf_cb_lk));
2026 		mutex_enter((&vdp->xdf_dev_lk));
2027 		if (rv == 0)
2028 			goto out;
2029 	}
2030 
2031 out:
2032 	ASSERT(MUTEX_HELD(&vdp->xdf_cb_lk));
2033 	ASSERT(MUTEX_HELD(&vdp->xdf_dev_lk));
2034 
2035 	if (vdp->xdf_connect_thread == curthread) {
2036 		/*
2037 		 * wake up someone else so they can become the connection
2038 		 * thread.
2039 		 */
2040 		cv_signal(&vdp->xdf_dev_cv);
2041 		vdp->xdf_connect_thread = NULL;
2042 	}
2043 
2044 	/* Try to lock the media */
2045 	mutex_exit((&vdp->xdf_dev_lk));
2046 	(void) xdf_media_req(vdp, XBV_MEDIA_REQ_LOCK, B_TRUE);
2047 	mutex_enter((&vdp->xdf_dev_lk));
2048 
2049 	vdp->xdf_connect_req--;
2050 	return (vdp->xdf_state);
2051 }
2052 
2053 static uint_t
2054 xdf_iorestart(caddr_t arg)
2055 {
2056 	xdf_t *vdp = (xdf_t *)arg;
2057 
2058 	ASSERT(vdp != NULL);
2059 
2060 	mutex_enter(&vdp->xdf_dev_lk);
2061 	ASSERT(ISDMACBON(vdp));
2062 	SETDMACBOFF(vdp);
2063 	mutex_exit(&vdp->xdf_dev_lk);
2064 
2065 	xdf_io_start(vdp);
2066 
2067 	return (DDI_INTR_CLAIMED);
2068 }
2069 
2070 #if defined(XPV_HVM_DRIVER)
2071 
2072 typedef struct xdf_hvm_entry {
2073 	list_node_t	xdf_he_list;
2074 	char		*xdf_he_path;
2075 	dev_info_t	*xdf_he_dip;
2076 } xdf_hvm_entry_t;
2077 
2078 static list_t xdf_hvm_list;
2079 static kmutex_t xdf_hvm_list_lock;
2080 
2081 static xdf_hvm_entry_t *
2082 i_xdf_hvm_find(const char *path, dev_info_t *dip)
2083 {
2084 	xdf_hvm_entry_t	*i;
2085 
2086 	ASSERT((path != NULL) || (dip != NULL));
2087 	ASSERT(MUTEX_HELD(&xdf_hvm_list_lock));
2088 
2089 	i = list_head(&xdf_hvm_list);
2090 	while (i != NULL) {
2091 		if ((path != NULL) && strcmp(i->xdf_he_path, path) != 0) {
2092 			i = list_next(&xdf_hvm_list, i);
2093 			continue;
2094 		}
2095 		if ((dip != NULL) && (i->xdf_he_dip != dip)) {
2096 			i = list_next(&xdf_hvm_list, i);
2097 			continue;
2098 		}
2099 		break;
2100 	}
2101 	return (i);
2102 }
2103 
2104 dev_info_t *
2105 xdf_hvm_hold(const char *path)
2106 {
2107 	xdf_hvm_entry_t	*i;
2108 	dev_info_t	*dip;
2109 
2110 	mutex_enter(&xdf_hvm_list_lock);
2111 	i = i_xdf_hvm_find(path, NULL);
2112 	if (i == NULL) {
2113 		mutex_exit(&xdf_hvm_list_lock);
2114 		return (B_FALSE);
2115 	}
2116 	ndi_hold_devi(dip = i->xdf_he_dip);
2117 	mutex_exit(&xdf_hvm_list_lock);
2118 	return (dip);
2119 }
2120 
2121 static void
2122 xdf_hvm_add(dev_info_t *dip)
2123 {
2124 	xdf_hvm_entry_t	*i;
2125 	char		*path;
2126 
2127 	/* figure out the path for the dip */
2128 	path = kmem_zalloc(MAXPATHLEN, KM_SLEEP);
2129 	(void) ddi_pathname(dip, path);
2130 
2131 	i = kmem_alloc(sizeof (*i), KM_SLEEP);
2132 	i->xdf_he_dip = dip;
2133 	i->xdf_he_path = i_ddi_strdup(path, KM_SLEEP);
2134 
2135 	mutex_enter(&xdf_hvm_list_lock);
2136 	ASSERT(i_xdf_hvm_find(path, NULL) == NULL);
2137 	ASSERT(i_xdf_hvm_find(NULL, dip) == NULL);
2138 	list_insert_head(&xdf_hvm_list, i);
2139 	mutex_exit(&xdf_hvm_list_lock);
2140 
2141 	kmem_free(path, MAXPATHLEN);
2142 }
2143 
2144 static void
2145 xdf_hvm_rm(dev_info_t *dip)
2146 {
2147 	xdf_hvm_entry_t	*i;
2148 
2149 	mutex_enter(&xdf_hvm_list_lock);
2150 	VERIFY((i = i_xdf_hvm_find(NULL, dip)) != NULL);
2151 	list_remove(&xdf_hvm_list, i);
2152 	mutex_exit(&xdf_hvm_list_lock);
2153 
2154 	kmem_free(i->xdf_he_path, strlen(i->xdf_he_path) + 1);
2155 	kmem_free(i, sizeof (*i));
2156 }
2157 
2158 static void
2159 xdf_hvm_init(void)
2160 {
2161 	list_create(&xdf_hvm_list, sizeof (xdf_hvm_entry_t),
2162 	    offsetof(xdf_hvm_entry_t, xdf_he_list));
2163 	mutex_init(&xdf_hvm_list_lock, NULL, MUTEX_DEFAULT, NULL);
2164 }
2165 
2166 static void
2167 xdf_hvm_fini(void)
2168 {
2169 	ASSERT(list_head(&xdf_hvm_list) == NULL);
2170 	list_destroy(&xdf_hvm_list);
2171 	mutex_destroy(&xdf_hvm_list_lock);
2172 }
2173 
2174 boolean_t
2175 xdf_hvm_connect(dev_info_t *dip)
2176 {
2177 	xdf_t	*vdp = (xdf_t *)ddi_get_driver_private(dip);
2178 	char	*oename, *str;
2179 	int	rv;
2180 
2181 	mutex_enter(&vdp->xdf_cb_lk);
2182 
2183 	/*
2184 	 * Before try to establish a connection we need to wait for the
2185 	 * backend hotplug scripts to have run.  Once they are run the
2186 	 * "<oename>/hotplug-status" property will be set to "connected".
2187 	 */
2188 	for (;;) {
2189 		ASSERT(MUTEX_HELD(&vdp->xdf_cb_lk));
2190 
2191 		/*
2192 		 * Get the xenbus path to the backend device.  Note that
2193 		 * we can't cache this path (and we look it up on each pass
2194 		 * through this loop) because it could change during
2195 		 * suspend, resume, and migration operations.
2196 		 */
2197 		if ((oename = xvdi_get_oename(dip)) == NULL) {
2198 			mutex_exit(&vdp->xdf_cb_lk);
2199 			return (B_FALSE);
2200 		}
2201 
2202 		str = NULL;
2203 		if ((xenbus_read_str(oename, XBP_HP_STATUS, &str) == 0) &&
2204 		    (strcmp(str, XBV_HP_STATUS_CONN) == 0))
2205 			break;
2206 
2207 		if (str != NULL)
2208 			strfree(str);
2209 
2210 		/* wait for an update to "<oename>/hotplug-status" */
2211 		if (cv_wait_sig(&vdp->xdf_hp_status_cv, &vdp->xdf_cb_lk) == 0) {
2212 			/* we got interrupted by a signal */
2213 			mutex_exit(&vdp->xdf_cb_lk);
2214 			return (B_FALSE);
2215 		}
2216 	}
2217 
2218 	/* Good news.  The backend hotplug scripts have been run. */
2219 	ASSERT(MUTEX_HELD(&vdp->xdf_cb_lk));
2220 	ASSERT(strcmp(str, XBV_HP_STATUS_CONN) == 0);
2221 	strfree(str);
2222 
2223 	/*
2224 	 * If we're emulating a cd device and if the backend doesn't support
2225 	 * media request opreations, then we're not going to bother trying
2226 	 * to establish a connection for a couple reasons.  First off, media
2227 	 * requests support is required to support operations like eject and
2228 	 * media locking.  Second, other backend platforms like Linux don't
2229 	 * support hvm pv cdrom access.  They don't even have a backend pv
2230 	 * driver for cdrom device nodes, so we don't want to block forever
2231 	 * waiting for a connection to a backend driver that doesn't exist.
2232 	 */
2233 	if (XD_IS_CD(vdp) && !xenbus_exists(oename, XBP_MEDIA_REQ_SUP)) {
2234 		mutex_exit(&vdp->xdf_cb_lk);
2235 		return (B_FALSE);
2236 	}
2237 
2238 	mutex_enter(&vdp->xdf_dev_lk);
2239 	rv = xdf_connect_locked(vdp, B_TRUE);
2240 	mutex_exit(&vdp->xdf_dev_lk);
2241 	mutex_exit(&vdp->xdf_cb_lk);
2242 
2243 	return ((rv == XD_READY) ? B_TRUE : B_FALSE);
2244 }
2245 
2246 int
2247 xdf_hvm_setpgeom(dev_info_t *dip, cmlb_geom_t *geomp)
2248 {
2249 	xdf_t	*vdp = (xdf_t *)ddi_get_driver_private(dip);
2250 
2251 	/* sanity check the requested physical geometry */
2252 	mutex_enter(&vdp->xdf_dev_lk);
2253 	if ((geomp->g_secsize != XB_BSIZE) ||
2254 	    (geomp->g_capacity == 0)) {
2255 		mutex_exit(&vdp->xdf_dev_lk);
2256 		return (EINVAL);
2257 	}
2258 
2259 	/*
2260 	 * If we've already connected to the backend device then make sure
2261 	 * we're not defining a physical geometry larger than our backend
2262 	 * device.
2263 	 */
2264 	if ((vdp->xdf_xdev_nblocks != 0) &&
2265 	    (geomp->g_capacity > vdp->xdf_xdev_nblocks)) {
2266 		mutex_exit(&vdp->xdf_dev_lk);
2267 		return (EINVAL);
2268 	}
2269 
2270 	bzero(&vdp->xdf_pgeom, sizeof (vdp->xdf_pgeom));
2271 	vdp->xdf_pgeom.g_ncyl = geomp->g_ncyl;
2272 	vdp->xdf_pgeom.g_acyl = geomp->g_acyl;
2273 	vdp->xdf_pgeom.g_nhead = geomp->g_nhead;
2274 	vdp->xdf_pgeom.g_nsect = geomp->g_nsect;
2275 	vdp->xdf_pgeom.g_secsize = geomp->g_secsize;
2276 	vdp->xdf_pgeom.g_capacity = geomp->g_capacity;
2277 	vdp->xdf_pgeom.g_intrlv = geomp->g_intrlv;
2278 	vdp->xdf_pgeom.g_rpm = geomp->g_rpm;
2279 
2280 	vdp->xdf_pgeom_fixed = B_TRUE;
2281 	mutex_exit(&vdp->xdf_dev_lk);
2282 
2283 	/* force a re-validation */
2284 	cmlb_invalidate(vdp->xdf_vd_lbl, NULL);
2285 
2286 	return (0);
2287 }
2288 
2289 boolean_t
2290 xdf_is_cd(dev_info_t *dip)
2291 {
2292 	xdf_t		*vdp = (xdf_t *)ddi_get_driver_private(dip);
2293 	boolean_t	rv;
2294 
2295 	mutex_enter(&vdp->xdf_cb_lk);
2296 	rv = XD_IS_CD(vdp);
2297 	mutex_exit(&vdp->xdf_cb_lk);
2298 	return (rv);
2299 }
2300 
2301 boolean_t
2302 xdf_is_rm(dev_info_t *dip)
2303 {
2304 	xdf_t		*vdp = (xdf_t *)ddi_get_driver_private(dip);
2305 	boolean_t	rv;
2306 
2307 	mutex_enter(&vdp->xdf_cb_lk);
2308 	rv = XD_IS_RM(vdp);
2309 	mutex_exit(&vdp->xdf_cb_lk);
2310 	return (rv);
2311 }
2312 
2313 boolean_t
2314 xdf_media_req_supported(dev_info_t *dip)
2315 {
2316 	xdf_t		*vdp = (xdf_t *)ddi_get_driver_private(dip);
2317 	boolean_t	rv;
2318 
2319 	mutex_enter(&vdp->xdf_cb_lk);
2320 	rv = vdp->xdf_media_req_supported;
2321 	mutex_exit(&vdp->xdf_cb_lk);
2322 	return (rv);
2323 }
2324 
2325 #endif /* XPV_HVM_DRIVER */
2326 
2327 static int
2328 xdf_lb_getcap(dev_info_t *dip, diskaddr_t *capp)
2329 {
2330 	xdf_t *vdp;
2331 	vdp = ddi_get_soft_state(xdf_ssp, ddi_get_instance(dip));
2332 
2333 	if (vdp == NULL)
2334 		return (ENXIO);
2335 
2336 	mutex_enter(&vdp->xdf_dev_lk);
2337 	*capp = vdp->xdf_pgeom.g_capacity;
2338 	DPRINTF(LBL_DBG, ("xdf@%s:capacity %llu\n", vdp->xdf_addr, *capp));
2339 	mutex_exit(&vdp->xdf_dev_lk);
2340 	return (0);
2341 }
2342 
2343 static int
2344 xdf_lb_getpgeom(dev_info_t *dip, cmlb_geom_t *geomp)
2345 {
2346 	xdf_t *vdp;
2347 
2348 	if ((vdp = ddi_get_soft_state(xdf_ssp, ddi_get_instance(dip))) == NULL)
2349 		return (ENXIO);
2350 	*geomp = vdp->xdf_pgeom;
2351 	return (0);
2352 }
2353 
2354 /*
2355  * No real HBA, no geometry available from it
2356  */
2357 /*ARGSUSED*/
2358 static int
2359 xdf_lb_getvgeom(dev_info_t *dip, cmlb_geom_t *geomp)
2360 {
2361 	return (EINVAL);
2362 }
2363 
2364 static int
2365 xdf_lb_getattribute(dev_info_t *dip, tg_attribute_t *tgattributep)
2366 {
2367 	xdf_t *vdp;
2368 
2369 	if (!(vdp = ddi_get_soft_state(xdf_ssp, ddi_get_instance(dip))))
2370 		return (ENXIO);
2371 
2372 	if (XD_IS_RO(vdp))
2373 		tgattributep->media_is_writable = 0;
2374 	else
2375 		tgattributep->media_is_writable = 1;
2376 	return (0);
2377 }
2378 
2379 /* ARGSUSED3 */
2380 int
2381 xdf_lb_getinfo(dev_info_t *dip, int cmd, void *arg, void *tg_cookie)
2382 {
2383 	int instance;
2384 	xdf_t   *vdp;
2385 
2386 	instance = ddi_get_instance(dip);
2387 
2388 	if ((vdp = ddi_get_soft_state(xdf_ssp, instance)) == NULL)
2389 		return (ENXIO);
2390 
2391 	switch (cmd) {
2392 	case TG_GETPHYGEOM:
2393 		return (xdf_lb_getpgeom(dip, (cmlb_geom_t *)arg));
2394 	case TG_GETVIRTGEOM:
2395 		return (xdf_lb_getvgeom(dip, (cmlb_geom_t *)arg));
2396 	case TG_GETCAPACITY:
2397 		return (xdf_lb_getcap(dip, (diskaddr_t *)arg));
2398 	case TG_GETBLOCKSIZE:
2399 		mutex_enter(&vdp->xdf_cb_lk);
2400 		*(uint32_t *)arg = vdp->xdf_xdev_secsize;
2401 		mutex_exit(&vdp->xdf_cb_lk);
2402 		return (0);
2403 	case TG_GETATTR:
2404 		return (xdf_lb_getattribute(dip, (tg_attribute_t *)arg));
2405 	default:
2406 		return (ENOTTY);
2407 	}
2408 }
2409 
2410 /* ARGSUSED5 */
2411 int
2412 xdf_lb_rdwr(dev_info_t *dip, uchar_t cmd, void *bufp,
2413     diskaddr_t start, size_t reqlen, void *tg_cookie)
2414 {
2415 	xdf_t *vdp;
2416 	struct buf *bp;
2417 	int err = 0;
2418 
2419 	vdp = ddi_get_soft_state(xdf_ssp, ddi_get_instance(dip));
2420 
2421 	/* We don't allow IO from the oe_change callback thread */
2422 	ASSERT(curthread != vdp->xdf_oe_change_thread);
2423 
2424 	if ((start + ((reqlen / (vdp->xdf_xdev_secsize / DEV_BSIZE))
2425 	    >> DEV_BSHIFT)) > vdp->xdf_pgeom.g_capacity)
2426 		return (EINVAL);
2427 
2428 	bp = getrbuf(KM_SLEEP);
2429 	if (cmd == TG_READ)
2430 		bp->b_flags = B_BUSY | B_READ;
2431 	else
2432 		bp->b_flags = B_BUSY | B_WRITE;
2433 
2434 	bp->b_un.b_addr = bufp;
2435 	bp->b_bcount = reqlen;
2436 	bp->b_blkno = start * (vdp->xdf_xdev_secsize / DEV_BSIZE);
2437 	bp->b_edev = DDI_DEV_T_NONE; /* don't have dev_t */
2438 
2439 	mutex_enter(&vdp->xdf_dev_lk);
2440 	xdf_bp_push(vdp, bp);
2441 	mutex_exit(&vdp->xdf_dev_lk);
2442 	xdf_io_start(vdp);
2443 	if (curthread == vdp->xdf_ready_tq_thread)
2444 		(void) xdf_ring_drain(vdp);
2445 	err = biowait(bp);
2446 	ASSERT(bp->b_flags & B_DONE);
2447 	freerbuf(bp);
2448 	return (err);
2449 }
2450 
2451 /*
2452  * Lock the current media.  Set the media state to "lock".
2453  * (Media locks are only respected by the backend driver.)
2454  */
2455 static int
2456 xdf_ioctl_mlock(xdf_t *vdp)
2457 {
2458 	int rv;
2459 	mutex_enter(&vdp->xdf_cb_lk);
2460 	rv = xdf_media_req(vdp, XBV_MEDIA_REQ_LOCK, B_TRUE);
2461 	mutex_exit(&vdp->xdf_cb_lk);
2462 	return (rv);
2463 }
2464 
2465 /*
2466  * Release a media lock.  Set the media state to "none".
2467  */
2468 static int
2469 xdf_ioctl_munlock(xdf_t *vdp)
2470 {
2471 	int rv;
2472 	mutex_enter(&vdp->xdf_cb_lk);
2473 	rv = xdf_media_req(vdp, XBV_MEDIA_REQ_NONE, B_TRUE);
2474 	mutex_exit(&vdp->xdf_cb_lk);
2475 	return (rv);
2476 }
2477 
2478 /*
2479  * Eject the current media.  Ignores any media locks.  (Media locks
2480  * are only for benifit of the the backend.)
2481  */
2482 static int
2483 xdf_ioctl_eject(xdf_t *vdp)
2484 {
2485 	int rv;
2486 
2487 	mutex_enter(&vdp->xdf_cb_lk);
2488 	if ((rv = xdf_media_req(vdp, XBV_MEDIA_REQ_EJECT, B_FALSE)) != 0) {
2489 		mutex_exit(&vdp->xdf_cb_lk);
2490 		return (rv);
2491 	}
2492 
2493 	/*
2494 	 * We've set the media requests xenbus parameter to eject, so now
2495 	 * disconnect from the backend, wait for the backend to clear
2496 	 * the media requets xenbus paramter, and then we can reconnect
2497 	 * to the backend.
2498 	 */
2499 	(void) xdf_disconnect(vdp, XD_UNKNOWN, B_TRUE);
2500 	mutex_enter(&vdp->xdf_dev_lk);
2501 	if (xdf_connect_locked(vdp, B_TRUE) != XD_READY) {
2502 		mutex_exit(&vdp->xdf_dev_lk);
2503 		mutex_exit(&vdp->xdf_cb_lk);
2504 		return (EIO);
2505 	}
2506 	mutex_exit(&vdp->xdf_dev_lk);
2507 	mutex_exit(&vdp->xdf_cb_lk);
2508 	return (0);
2509 }
2510 
2511 /*
2512  * Watch for media state changes.  This can be an insertion of a device
2513  * (triggered by a 'xm block-configure' request in another domain) or
2514  * the ejection of a device (triggered by a local "eject" operation).
2515  * For a full description of the DKIOCSTATE ioctl behavior see dkio(7I).
2516  */
2517 static int
2518 xdf_dkstate(xdf_t *vdp, enum dkio_state mstate)
2519 {
2520 	enum dkio_state		prev_state;
2521 
2522 	mutex_enter(&vdp->xdf_cb_lk);
2523 	prev_state = vdp->xdf_mstate;
2524 
2525 	if (vdp->xdf_mstate == mstate) {
2526 		while (vdp->xdf_mstate == prev_state) {
2527 			if (cv_wait_sig(&vdp->xdf_mstate_cv,
2528 			    &vdp->xdf_cb_lk) == 0) {
2529 				mutex_exit(&vdp->xdf_cb_lk);
2530 				return (EINTR);
2531 			}
2532 		}
2533 	}
2534 
2535 	if ((prev_state != DKIO_INSERTED) &&
2536 	    (vdp->xdf_mstate == DKIO_INSERTED)) {
2537 		(void) xdf_media_req(vdp, XBV_MEDIA_REQ_LOCK, B_TRUE);
2538 		mutex_exit(&vdp->xdf_cb_lk);
2539 		return (0);
2540 	}
2541 
2542 	mutex_exit(&vdp->xdf_cb_lk);
2543 	return (0);
2544 }
2545 
2546 /*ARGSUSED*/
2547 static int
2548 xdf_ioctl(dev_t dev, int cmd, intptr_t arg, int mode, cred_t *credp,
2549     int *rvalp)
2550 {
2551 	minor_t		minor = getminor(dev);
2552 	int		part = XDF_PART(minor);
2553 	xdf_t		*vdp;
2554 	int		rv;
2555 
2556 	if (((vdp = ddi_get_soft_state(xdf_ssp, XDF_INST(minor))) == NULL) ||
2557 	    (!xdf_isopen(vdp, part)))
2558 		return (ENXIO);
2559 
2560 	DPRINTF(IOCTL_DBG, ("xdf@%s:ioctl: cmd %d (0x%x)\n",
2561 	    vdp->xdf_addr, cmd, cmd));
2562 
2563 	switch (cmd) {
2564 	default:
2565 		return (ENOTTY);
2566 	case DKIOCG_PHYGEOM:
2567 	case DKIOCG_VIRTGEOM:
2568 	case DKIOCGGEOM:
2569 	case DKIOCSGEOM:
2570 	case DKIOCGAPART:
2571 	case DKIOCSAPART:
2572 	case DKIOCGVTOC:
2573 	case DKIOCSVTOC:
2574 	case DKIOCPARTINFO:
2575 	case DKIOCGEXTVTOC:
2576 	case DKIOCSEXTVTOC:
2577 	case DKIOCEXTPARTINFO:
2578 	case DKIOCGMBOOT:
2579 	case DKIOCSMBOOT:
2580 	case DKIOCGETEFI:
2581 	case DKIOCSETEFI:
2582 	case DKIOCSETEXTPART:
2583 	case DKIOCPARTITION:
2584 		return (cmlb_ioctl(vdp->xdf_vd_lbl, dev, cmd, arg, mode, credp,
2585 		    rvalp, NULL));
2586 	case FDEJECT:
2587 	case DKIOCEJECT:
2588 	case CDROMEJECT:
2589 		return (xdf_ioctl_eject(vdp));
2590 	case DKIOCLOCK:
2591 		return (xdf_ioctl_mlock(vdp));
2592 	case DKIOCUNLOCK:
2593 		return (xdf_ioctl_munlock(vdp));
2594 	case CDROMREADOFFSET: {
2595 		int offset = 0;
2596 		if (!XD_IS_CD(vdp))
2597 			return (ENOTTY);
2598 		if (ddi_copyout(&offset, (void *)arg, sizeof (int), mode))
2599 			return (EFAULT);
2600 		return (0);
2601 	}
2602 	case DKIOCGMEDIAINFO: {
2603 		struct dk_minfo media_info;
2604 
2605 		media_info.dki_lbsize = vdp->xdf_xdev_secsize;
2606 		media_info.dki_capacity = vdp->xdf_pgeom.g_capacity;
2607 		if (XD_IS_CD(vdp))
2608 			media_info.dki_media_type = DK_CDROM;
2609 		else
2610 			media_info.dki_media_type = DK_FIXED_DISK;
2611 
2612 		if (ddi_copyout(&media_info, (void *)arg,
2613 		    sizeof (struct dk_minfo), mode))
2614 			return (EFAULT);
2615 		return (0);
2616 	}
2617 	case DKIOCINFO: {
2618 		struct dk_cinfo info;
2619 
2620 		/* controller information */
2621 		if (XD_IS_CD(vdp))
2622 			info.dki_ctype = DKC_CDROM;
2623 		else
2624 			info.dki_ctype = DKC_VBD;
2625 
2626 		info.dki_cnum = 0;
2627 		(void) strncpy((char *)(&info.dki_cname), "xdf", 8);
2628 
2629 		/* unit information */
2630 		info.dki_unit = ddi_get_instance(vdp->xdf_dip);
2631 		(void) strncpy((char *)(&info.dki_dname), "xdf", 8);
2632 		info.dki_flags = DKI_FMTVOL;
2633 		info.dki_partition = part;
2634 		info.dki_maxtransfer = maxphys / DEV_BSIZE;
2635 		info.dki_addr = 0;
2636 		info.dki_space = 0;
2637 		info.dki_prio = 0;
2638 		info.dki_vec = 0;
2639 
2640 		if (ddi_copyout(&info, (void *)arg, sizeof (info), mode))
2641 			return (EFAULT);
2642 		return (0);
2643 	}
2644 	case DKIOCSTATE: {
2645 		enum dkio_state mstate;
2646 
2647 		if (ddi_copyin((void *)arg, &mstate,
2648 		    sizeof (mstate), mode) != 0)
2649 			return (EFAULT);
2650 		if ((rv = xdf_dkstate(vdp, mstate)) != 0)
2651 			return (rv);
2652 		mstate = vdp->xdf_mstate;
2653 		if (ddi_copyout(&mstate, (void *)arg,
2654 		    sizeof (mstate), mode) != 0)
2655 			return (EFAULT);
2656 		return (0);
2657 	}
2658 	case DKIOCREMOVABLE: {
2659 		int i = BOOLEAN2VOID(XD_IS_RM(vdp));
2660 		if (ddi_copyout(&i, (caddr_t)arg, sizeof (i), mode))
2661 			return (EFAULT);
2662 		return (0);
2663 	}
2664 	case DKIOCGETWCE: {
2665 		int i = BOOLEAN2VOID(XD_IS_RM(vdp));
2666 		if (ddi_copyout(&i, (void *)arg, sizeof (i), mode))
2667 			return (EFAULT);
2668 		return (0);
2669 	}
2670 	case DKIOCSETWCE: {
2671 		int i;
2672 		if (ddi_copyin((void *)arg, &i, sizeof (i), mode))
2673 			return (EFAULT);
2674 		vdp->xdf_wce = VOID2BOOLEAN(i);
2675 		return (0);
2676 	}
2677 	case DKIOCFLUSHWRITECACHE: {
2678 		struct dk_callback *dkc = (struct dk_callback *)arg;
2679 
2680 		if (vdp->xdf_flush_supported) {
2681 			rv = xdf_lb_rdwr(vdp->xdf_dip, TG_WRITE,
2682 			    NULL, 0, 0, (void *)dev);
2683 		} else if (vdp->xdf_feature_barrier &&
2684 		    !xdf_barrier_flush_disable) {
2685 			rv = xdf_lb_rdwr(vdp->xdf_dip, TG_WRITE,
2686 			    vdp->xdf_cache_flush_block, xdf_flush_block,
2687 			    vdp->xdf_xdev_secsize, (void *)dev);
2688 		} else {
2689 			return (ENOTTY);
2690 		}
2691 		if ((mode & FKIOCTL) && (dkc != NULL) &&
2692 		    (dkc->dkc_callback != NULL)) {
2693 			(*dkc->dkc_callback)(dkc->dkc_cookie, rv);
2694 			/* need to return 0 after calling callback */
2695 			rv = 0;
2696 		}
2697 		return (rv);
2698 	}
2699 	}
2700 	/*NOTREACHED*/
2701 }
2702 
2703 static int
2704 xdf_strategy(struct buf *bp)
2705 {
2706 	xdf_t	*vdp;
2707 	minor_t minor;
2708 	diskaddr_t p_blkct, p_blkst;
2709 	daddr_t blkno;
2710 	ulong_t nblks;
2711 	int part;
2712 
2713 	minor = getminor(bp->b_edev);
2714 	part = XDF_PART(minor);
2715 	vdp = ddi_get_soft_state(xdf_ssp, XDF_INST(minor));
2716 
2717 	mutex_enter(&vdp->xdf_dev_lk);
2718 	if (!xdf_isopen(vdp, part)) {
2719 		mutex_exit(&vdp->xdf_dev_lk);
2720 		xdf_io_err(bp, ENXIO, 0);
2721 		return (0);
2722 	}
2723 
2724 	/* We don't allow IO from the oe_change callback thread */
2725 	ASSERT(curthread != vdp->xdf_oe_change_thread);
2726 
2727 	/* Check for writes to a read only device */
2728 	if (!IS_READ(bp) && XD_IS_RO(vdp)) {
2729 		mutex_exit(&vdp->xdf_dev_lk);
2730 		xdf_io_err(bp, EROFS, 0);
2731 		return (0);
2732 	}
2733 
2734 	/* Check if this I/O is accessing a partition or the entire disk */
2735 	if ((long)bp->b_private == XB_SLICE_NONE) {
2736 		/* This I/O is using an absolute offset */
2737 		p_blkct = vdp->xdf_xdev_nblocks;
2738 		p_blkst = 0;
2739 	} else {
2740 		/* This I/O is using a partition relative offset */
2741 		mutex_exit(&vdp->xdf_dev_lk);
2742 		if (cmlb_partinfo(vdp->xdf_vd_lbl, part, &p_blkct,
2743 		    &p_blkst, NULL, NULL, NULL)) {
2744 			xdf_io_err(bp, ENXIO, 0);
2745 			return (0);
2746 		}
2747 		mutex_enter(&vdp->xdf_dev_lk);
2748 	}
2749 
2750 	/*
2751 	 * Adjust the real blkno and bcount according to the underline
2752 	 * physical sector size.
2753 	 */
2754 	blkno = bp->b_blkno / (vdp->xdf_xdev_secsize / XB_BSIZE);
2755 
2756 	/* check for a starting block beyond the disk or partition limit */
2757 	if (blkno > p_blkct) {
2758 		DPRINTF(IO_DBG, ("xdf@%s: block %lld exceeds VBD size %"PRIu64,
2759 		    vdp->xdf_addr, (longlong_t)blkno, (uint64_t)p_blkct));
2760 		mutex_exit(&vdp->xdf_dev_lk);
2761 		xdf_io_err(bp, EINVAL, 0);
2762 		return (0);
2763 	}
2764 
2765 	/* Legacy: don't set error flag at this case */
2766 	if (blkno == p_blkct) {
2767 		mutex_exit(&vdp->xdf_dev_lk);
2768 		bp->b_resid = bp->b_bcount;
2769 		biodone(bp);
2770 		return (0);
2771 	}
2772 
2773 	/* sanitize the input buf */
2774 	bioerror(bp, 0);
2775 	bp->b_resid = 0;
2776 	bp->av_back = bp->av_forw = NULL;
2777 
2778 	/* Adjust for partial transfer, this will result in an error later */
2779 	if (vdp->xdf_xdev_secsize != 0 &&
2780 	    vdp->xdf_xdev_secsize != XB_BSIZE) {
2781 		nblks = bp->b_bcount / vdp->xdf_xdev_secsize;
2782 	} else {
2783 		nblks = bp->b_bcount >> XB_BSHIFT;
2784 	}
2785 
2786 	if ((blkno + nblks) > p_blkct) {
2787 		if (vdp->xdf_xdev_secsize != 0 &&
2788 		    vdp->xdf_xdev_secsize != XB_BSIZE) {
2789 			bp->b_resid =
2790 			    ((blkno + nblks) - p_blkct) *
2791 			    vdp->xdf_xdev_secsize;
2792 		} else {
2793 			bp->b_resid =
2794 			    ((blkno + nblks) - p_blkct) <<
2795 			    XB_BSHIFT;
2796 		}
2797 		bp->b_bcount -= bp->b_resid;
2798 	}
2799 
2800 	DPRINTF(IO_DBG, ("xdf@%s: strategy blk %lld len %lu\n",
2801 	    vdp->xdf_addr, (longlong_t)blkno, (ulong_t)bp->b_bcount));
2802 
2803 	/* Fix up the buf struct */
2804 	bp->b_flags |= B_BUSY;
2805 	bp->b_private = (void *)(uintptr_t)p_blkst;
2806 
2807 	xdf_bp_push(vdp, bp);
2808 	mutex_exit(&vdp->xdf_dev_lk);
2809 	xdf_io_start(vdp);
2810 	if (do_polled_io)
2811 		(void) xdf_ring_drain(vdp);
2812 	return (0);
2813 }
2814 
2815 /*ARGSUSED*/
2816 static int
2817 xdf_read(dev_t dev, struct uio *uiop, cred_t *credp)
2818 {
2819 	xdf_t	*vdp;
2820 	minor_t minor;
2821 	diskaddr_t p_blkcnt;
2822 	int part;
2823 
2824 	minor = getminor(dev);
2825 	if ((vdp = ddi_get_soft_state(xdf_ssp, XDF_INST(minor))) == NULL)
2826 		return (ENXIO);
2827 
2828 	DPRINTF(IO_DBG, ("xdf@%s: read offset 0x%"PRIx64"\n",
2829 	    vdp->xdf_addr, (int64_t)uiop->uio_offset));
2830 
2831 	part = XDF_PART(minor);
2832 	if (!xdf_isopen(vdp, part))
2833 		return (ENXIO);
2834 
2835 	if (cmlb_partinfo(vdp->xdf_vd_lbl, part, &p_blkcnt,
2836 	    NULL, NULL, NULL, NULL))
2837 		return (ENXIO);
2838 
2839 	if (uiop->uio_loffset >= XB_DTOB(p_blkcnt, vdp))
2840 		return (ENOSPC);
2841 
2842 	if (U_INVAL(uiop))
2843 		return (EINVAL);
2844 
2845 	return (physio(xdf_strategy, NULL, dev, B_READ, xdfmin, uiop));
2846 }
2847 
2848 /*ARGSUSED*/
2849 static int
2850 xdf_write(dev_t dev, struct uio *uiop, cred_t *credp)
2851 {
2852 	xdf_t *vdp;
2853 	minor_t minor;
2854 	diskaddr_t p_blkcnt;
2855 	int part;
2856 
2857 	minor = getminor(dev);
2858 	if ((vdp = ddi_get_soft_state(xdf_ssp, XDF_INST(minor))) == NULL)
2859 		return (ENXIO);
2860 
2861 	DPRINTF(IO_DBG, ("xdf@%s: write offset 0x%"PRIx64"\n",
2862 	    vdp->xdf_addr, (int64_t)uiop->uio_offset));
2863 
2864 	part = XDF_PART(minor);
2865 	if (!xdf_isopen(vdp, part))
2866 		return (ENXIO);
2867 
2868 	if (cmlb_partinfo(vdp->xdf_vd_lbl, part, &p_blkcnt,
2869 	    NULL, NULL, NULL, NULL))
2870 		return (ENXIO);
2871 
2872 	if (uiop->uio_loffset >= XB_DTOB(p_blkcnt, vdp))
2873 		return (ENOSPC);
2874 
2875 	if (U_INVAL(uiop))
2876 		return (EINVAL);
2877 
2878 	return (physio(xdf_strategy, NULL, dev, B_WRITE, xdfmin, uiop));
2879 }
2880 
2881 /*ARGSUSED*/
2882 static int
2883 xdf_aread(dev_t dev, struct aio_req *aiop, cred_t *credp)
2884 {
2885 	xdf_t	*vdp;
2886 	minor_t minor;
2887 	struct uio *uiop = aiop->aio_uio;
2888 	diskaddr_t p_blkcnt;
2889 	int part;
2890 
2891 	minor = getminor(dev);
2892 	if ((vdp = ddi_get_soft_state(xdf_ssp, XDF_INST(minor))) == NULL)
2893 		return (ENXIO);
2894 
2895 	part = XDF_PART(minor);
2896 	if (!xdf_isopen(vdp, part))
2897 		return (ENXIO);
2898 
2899 	if (cmlb_partinfo(vdp->xdf_vd_lbl, part, &p_blkcnt,
2900 	    NULL, NULL, NULL, NULL))
2901 		return (ENXIO);
2902 
2903 	if (uiop->uio_loffset >= XB_DTOB(p_blkcnt, vdp))
2904 		return (ENOSPC);
2905 
2906 	if (U_INVAL(uiop))
2907 		return (EINVAL);
2908 
2909 	return (aphysio(xdf_strategy, anocancel, dev, B_READ, xdfmin, aiop));
2910 }
2911 
2912 /*ARGSUSED*/
2913 static int
2914 xdf_awrite(dev_t dev, struct aio_req *aiop, cred_t *credp)
2915 {
2916 	xdf_t *vdp;
2917 	minor_t minor;
2918 	struct uio *uiop = aiop->aio_uio;
2919 	diskaddr_t p_blkcnt;
2920 	int part;
2921 
2922 	minor = getminor(dev);
2923 	if ((vdp = ddi_get_soft_state(xdf_ssp, XDF_INST(minor))) == NULL)
2924 		return (ENXIO);
2925 
2926 	part = XDF_PART(minor);
2927 	if (!xdf_isopen(vdp, part))
2928 		return (ENXIO);
2929 
2930 	if (cmlb_partinfo(vdp->xdf_vd_lbl, part, &p_blkcnt,
2931 	    NULL, NULL, NULL, NULL))
2932 		return (ENXIO);
2933 
2934 	if (uiop->uio_loffset >= XB_DTOB(p_blkcnt, vdp))
2935 		return (ENOSPC);
2936 
2937 	if (U_INVAL(uiop))
2938 		return (EINVAL);
2939 
2940 	return (aphysio(xdf_strategy, anocancel, dev, B_WRITE, xdfmin, aiop));
2941 }
2942 
2943 static int
2944 xdf_dump(dev_t dev, caddr_t addr, daddr_t blkno, int nblk)
2945 {
2946 	struct buf dumpbuf, *dbp = &dumpbuf;
2947 	xdf_t	*vdp;
2948 	minor_t minor;
2949 	int err = 0;
2950 	int part;
2951 	diskaddr_t p_blkcnt, p_blkst;
2952 
2953 	minor = getminor(dev);
2954 	if ((vdp = ddi_get_soft_state(xdf_ssp, XDF_INST(minor))) == NULL)
2955 		return (ENXIO);
2956 
2957 	DPRINTF(IO_DBG, ("xdf@%s: dump addr (0x%p) blk (%ld) nblks (%d)\n",
2958 	    vdp->xdf_addr, (void *)addr, blkno, nblk));
2959 
2960 	/* We don't allow IO from the oe_change callback thread */
2961 	ASSERT(curthread != vdp->xdf_oe_change_thread);
2962 
2963 	part = XDF_PART(minor);
2964 	if (!xdf_isopen(vdp, part))
2965 		return (ENXIO);
2966 
2967 	if (cmlb_partinfo(vdp->xdf_vd_lbl, part, &p_blkcnt, &p_blkst,
2968 	    NULL, NULL, NULL))
2969 		return (ENXIO);
2970 
2971 	if ((blkno + nblk) >
2972 	    (p_blkcnt * (vdp->xdf_xdev_secsize / XB_BSIZE))) {
2973 		cmn_err(CE_WARN, "xdf@%s: block %ld exceeds VBD size %"PRIu64,
2974 		    vdp->xdf_addr, (daddr_t)((blkno + nblk) /
2975 		    (vdp->xdf_xdev_secsize / XB_BSIZE)), (uint64_t)p_blkcnt);
2976 		return (EINVAL);
2977 	}
2978 
2979 	bioinit(dbp);
2980 	dbp->b_flags = B_BUSY;
2981 	dbp->b_un.b_addr = addr;
2982 	dbp->b_bcount = nblk << DEV_BSHIFT;
2983 	dbp->b_blkno = blkno;
2984 	dbp->b_edev = dev;
2985 	dbp->b_private = (void *)(uintptr_t)p_blkst;
2986 
2987 	mutex_enter(&vdp->xdf_dev_lk);
2988 	xdf_bp_push(vdp, dbp);
2989 	mutex_exit(&vdp->xdf_dev_lk);
2990 	xdf_io_start(vdp);
2991 	err = xdf_ring_drain(vdp);
2992 	biofini(dbp);
2993 	return (err);
2994 }
2995 
2996 /*ARGSUSED*/
2997 static int
2998 xdf_close(dev_t dev, int flag, int otyp, struct cred *credp)
2999 {
3000 	minor_t	minor;
3001 	xdf_t	*vdp;
3002 	int part;
3003 	ulong_t parbit;
3004 
3005 	minor = getminor(dev);
3006 	if ((vdp = ddi_get_soft_state(xdf_ssp, XDF_INST(minor))) == NULL)
3007 		return (ENXIO);
3008 
3009 	mutex_enter(&vdp->xdf_dev_lk);
3010 	part = XDF_PART(minor);
3011 	if (!xdf_isopen(vdp, part)) {
3012 		mutex_exit(&vdp->xdf_dev_lk);
3013 		return (ENXIO);
3014 	}
3015 	parbit = 1 << part;
3016 
3017 	ASSERT((vdp->xdf_vd_open[otyp] & parbit) != 0);
3018 	if (otyp == OTYP_LYR) {
3019 		ASSERT(vdp->xdf_vd_lyropen[part] > 0);
3020 		if (--vdp->xdf_vd_lyropen[part] == 0)
3021 			vdp->xdf_vd_open[otyp] &= ~parbit;
3022 	} else {
3023 		vdp->xdf_vd_open[otyp] &= ~parbit;
3024 	}
3025 	vdp->xdf_vd_exclopen &= ~parbit;
3026 
3027 	mutex_exit(&vdp->xdf_dev_lk);
3028 	return (0);
3029 }
3030 
3031 static int
3032 xdf_open(dev_t *devp, int flag, int otyp, cred_t *credp)
3033 {
3034 	minor_t	minor;
3035 	xdf_t	*vdp;
3036 	int part;
3037 	ulong_t parbit;
3038 	diskaddr_t p_blkct = 0;
3039 	boolean_t firstopen;
3040 	boolean_t nodelay;
3041 
3042 	minor = getminor(*devp);
3043 	if ((vdp = ddi_get_soft_state(xdf_ssp, XDF_INST(minor))) == NULL)
3044 		return (ENXIO);
3045 
3046 	nodelay = (flag & (FNDELAY | FNONBLOCK));
3047 
3048 	DPRINTF(DDI_DBG, ("xdf@%s: opening\n", vdp->xdf_addr));
3049 
3050 	/* do cv_wait until connected or failed */
3051 	mutex_enter(&vdp->xdf_cb_lk);
3052 	mutex_enter(&vdp->xdf_dev_lk);
3053 	if (!nodelay && (xdf_connect_locked(vdp, B_TRUE) != XD_READY)) {
3054 		mutex_exit(&vdp->xdf_dev_lk);
3055 		mutex_exit(&vdp->xdf_cb_lk);
3056 		return (ENXIO);
3057 	}
3058 	mutex_exit(&vdp->xdf_cb_lk);
3059 
3060 	if ((flag & FWRITE) && XD_IS_RO(vdp)) {
3061 		mutex_exit(&vdp->xdf_dev_lk);
3062 		return (EROFS);
3063 	}
3064 
3065 	part = XDF_PART(minor);
3066 	parbit = 1 << part;
3067 	if ((vdp->xdf_vd_exclopen & parbit) ||
3068 	    ((flag & FEXCL) && xdf_isopen(vdp, part))) {
3069 		mutex_exit(&vdp->xdf_dev_lk);
3070 		return (EBUSY);
3071 	}
3072 
3073 	/* are we the first one to open this node? */
3074 	firstopen = !xdf_isopen(vdp, -1);
3075 
3076 	if (otyp == OTYP_LYR)
3077 		vdp->xdf_vd_lyropen[part]++;
3078 
3079 	vdp->xdf_vd_open[otyp] |= parbit;
3080 
3081 	if (flag & FEXCL)
3082 		vdp->xdf_vd_exclopen |= parbit;
3083 
3084 	mutex_exit(&vdp->xdf_dev_lk);
3085 
3086 	/* force a re-validation */
3087 	if (firstopen)
3088 		cmlb_invalidate(vdp->xdf_vd_lbl, NULL);
3089 
3090 	/* If this is a non-blocking open then we're done */
3091 	if (nodelay)
3092 		return (0);
3093 
3094 	/*
3095 	 * This is a blocking open, so we require:
3096 	 * - that the disk have a valid label on it
3097 	 * - that the size of the partition that we're opening is non-zero
3098 	 */
3099 	if ((cmlb_partinfo(vdp->xdf_vd_lbl, part, &p_blkct,
3100 	    NULL, NULL, NULL, NULL) != 0) || (p_blkct == 0)) {
3101 		(void) xdf_close(*devp, flag, otyp, credp);
3102 		return (ENXIO);
3103 	}
3104 
3105 	return (0);
3106 }
3107 
3108 /*ARGSUSED*/
3109 static void
3110 xdf_watch_hp_status_cb(dev_info_t *dip, const char *path, void *arg)
3111 {
3112 	xdf_t *vdp = (xdf_t *)ddi_get_driver_private(dip);
3113 	cv_broadcast(&vdp->xdf_hp_status_cv);
3114 }
3115 
3116 static int
3117 xdf_prop_op(dev_t dev, dev_info_t *dip, ddi_prop_op_t prop_op, int flags,
3118 	char *name, caddr_t valuep, int *lengthp)
3119 {
3120 	xdf_t	*vdp = ddi_get_soft_state(xdf_ssp, ddi_get_instance(dip));
3121 
3122 	/*
3123 	 * Sanity check that if a dev_t or dip were specified that they
3124 	 * correspond to this device driver.  On debug kernels we'll
3125 	 * panic and on non-debug kernels we'll return failure.
3126 	 */
3127 	ASSERT(ddi_driver_major(dip) == xdf_major);
3128 	ASSERT((dev == DDI_DEV_T_ANY) || (getmajor(dev) == xdf_major));
3129 	if ((ddi_driver_major(dip) != xdf_major) ||
3130 	    ((dev != DDI_DEV_T_ANY) && (getmajor(dev) != xdf_major)))
3131 		return (DDI_PROP_NOT_FOUND);
3132 
3133 	if (vdp == NULL)
3134 		return (ddi_prop_op(dev, dip, prop_op, flags,
3135 		    name, valuep, lengthp));
3136 
3137 	return (cmlb_prop_op(vdp->xdf_vd_lbl,
3138 	    dev, dip, prop_op, flags, name, valuep, lengthp,
3139 	    XDF_PART(getminor(dev)), NULL));
3140 }
3141 
3142 /*ARGSUSED*/
3143 static int
3144 xdf_getinfo(dev_info_t *dip, ddi_info_cmd_t cmd, void *arg, void **rp)
3145 {
3146 	int	instance = XDF_INST(getminor((dev_t)arg));
3147 	xdf_t	*vbdp;
3148 
3149 	switch (cmd) {
3150 	case DDI_INFO_DEVT2DEVINFO:
3151 		if ((vbdp = ddi_get_soft_state(xdf_ssp, instance)) == NULL) {
3152 			*rp = NULL;
3153 			return (DDI_FAILURE);
3154 		}
3155 		*rp = vbdp->xdf_dip;
3156 		return (DDI_SUCCESS);
3157 
3158 	case DDI_INFO_DEVT2INSTANCE:
3159 		*rp = (void *)(uintptr_t)instance;
3160 		return (DDI_SUCCESS);
3161 
3162 	default:
3163 		return (DDI_FAILURE);
3164 	}
3165 }
3166 
3167 /*ARGSUSED*/
3168 static int
3169 xdf_resume(dev_info_t *dip)
3170 {
3171 	xdf_t	*vdp;
3172 	char	*oename;
3173 
3174 	if ((vdp = ddi_get_soft_state(xdf_ssp, ddi_get_instance(dip))) == NULL)
3175 		goto err;
3176 
3177 	if (xdf_debug & SUSRES_DBG)
3178 		xen_printf("xdf@%s: xdf_resume\n", vdp->xdf_addr);
3179 
3180 	mutex_enter(&vdp->xdf_cb_lk);
3181 
3182 	if (xvdi_resume(dip) != DDI_SUCCESS) {
3183 		mutex_exit(&vdp->xdf_cb_lk);
3184 		goto err;
3185 	}
3186 
3187 	if (((oename = xvdi_get_oename(dip)) == NULL) ||
3188 	    (xvdi_add_xb_watch_handler(dip, oename, XBP_HP_STATUS,
3189 	    xdf_watch_hp_status_cb, NULL) != DDI_SUCCESS)) {
3190 		mutex_exit(&vdp->xdf_cb_lk);
3191 		goto err;
3192 	}
3193 
3194 	mutex_enter(&vdp->xdf_dev_lk);
3195 	ASSERT(vdp->xdf_state != XD_READY);
3196 	xdf_set_state(vdp, XD_UNKNOWN);
3197 	mutex_exit(&vdp->xdf_dev_lk);
3198 
3199 	if (xdf_setstate_init(vdp) != DDI_SUCCESS) {
3200 		mutex_exit(&vdp->xdf_cb_lk);
3201 		goto err;
3202 	}
3203 
3204 	mutex_exit(&vdp->xdf_cb_lk);
3205 
3206 	if (xdf_debug & SUSRES_DBG)
3207 		xen_printf("xdf@%s: xdf_resume: done\n", vdp->xdf_addr);
3208 	return (DDI_SUCCESS);
3209 err:
3210 	if (xdf_debug & SUSRES_DBG)
3211 		xen_printf("xdf@%s: xdf_resume: fail\n", vdp->xdf_addr);
3212 	return (DDI_FAILURE);
3213 }
3214 
3215 static int
3216 xdf_attach(dev_info_t *dip, ddi_attach_cmd_t cmd)
3217 {
3218 	int			n, instance = ddi_get_instance(dip);
3219 	ddi_iblock_cookie_t	ibc, softibc;
3220 	boolean_t		dev_iscd = B_FALSE;
3221 	xdf_t			*vdp;
3222 	char			*oename, *xsname, *str;
3223 
3224 	if ((n = ddi_prop_get_int(DDI_DEV_T_ANY, dip, DDI_PROP_NOTPROM,
3225 	    "xdf_debug", 0)) != 0)
3226 		xdf_debug = n;
3227 
3228 	switch (cmd) {
3229 	case DDI_RESUME:
3230 		return (xdf_resume(dip));
3231 	case DDI_ATTACH:
3232 		break;
3233 	default:
3234 		return (DDI_FAILURE);
3235 	}
3236 	/* DDI_ATTACH */
3237 
3238 	if (((xsname = xvdi_get_xsname(dip)) == NULL) ||
3239 	    ((oename = xvdi_get_oename(dip)) == NULL))
3240 		return (DDI_FAILURE);
3241 
3242 	/*
3243 	 * Disable auto-detach.  This is necessary so that we don't get
3244 	 * detached while we're disconnected from the back end.
3245 	 */
3246 	if ((ddi_prop_update_int(DDI_DEV_T_NONE, dip,
3247 	    DDI_NO_AUTODETACH, 1) != DDI_PROP_SUCCESS))
3248 		return (DDI_FAILURE);
3249 
3250 	/* driver handles kernel-issued IOCTLs */
3251 	if (ddi_prop_create(DDI_DEV_T_NONE, dip,
3252 	    DDI_PROP_CANSLEEP, DDI_KERNEL_IOCTL, NULL, 0) != DDI_PROP_SUCCESS)
3253 		return (DDI_FAILURE);
3254 
3255 	if (ddi_get_iblock_cookie(dip, 0, &ibc) != DDI_SUCCESS)
3256 		return (DDI_FAILURE);
3257 
3258 	if (ddi_get_soft_iblock_cookie(dip,
3259 	    DDI_SOFTINT_LOW, &softibc) != DDI_SUCCESS)
3260 		return (DDI_FAILURE);
3261 
3262 	if (xenbus_read_str(xsname, XBP_DEV_TYPE, &str) != 0) {
3263 		cmn_err(CE_WARN, "xdf@%s: cannot read device-type",
3264 		    ddi_get_name_addr(dip));
3265 		return (DDI_FAILURE);
3266 	}
3267 	if (strcmp(str, XBV_DEV_TYPE_CD) == 0)
3268 		dev_iscd = B_TRUE;
3269 	strfree(str);
3270 
3271 	if (ddi_soft_state_zalloc(xdf_ssp, instance) != DDI_SUCCESS)
3272 		return (DDI_FAILURE);
3273 
3274 	DPRINTF(DDI_DBG, ("xdf@%s: attaching\n", ddi_get_name_addr(dip)));
3275 	vdp = ddi_get_soft_state(xdf_ssp, instance);
3276 	ddi_set_driver_private(dip, vdp);
3277 	vdp->xdf_dip = dip;
3278 	vdp->xdf_addr = ddi_get_name_addr(dip);
3279 	vdp->xdf_suspending = B_FALSE;
3280 	vdp->xdf_media_req_supported = B_FALSE;
3281 	vdp->xdf_peer = INVALID_DOMID;
3282 	vdp->xdf_evtchn = INVALID_EVTCHN;
3283 	list_create(&vdp->xdf_vreq_act, sizeof (v_req_t),
3284 	    offsetof(v_req_t, v_link));
3285 	cv_init(&vdp->xdf_dev_cv, NULL, CV_DEFAULT, NULL);
3286 	cv_init(&vdp->xdf_hp_status_cv, NULL, CV_DEFAULT, NULL);
3287 	cv_init(&vdp->xdf_mstate_cv, NULL, CV_DEFAULT, NULL);
3288 	mutex_init(&vdp->xdf_dev_lk, NULL, MUTEX_DRIVER, (void *)ibc);
3289 	mutex_init(&vdp->xdf_cb_lk, NULL, MUTEX_DRIVER, (void *)ibc);
3290 	mutex_init(&vdp->xdf_iostat_lk, NULL, MUTEX_DRIVER, (void *)ibc);
3291 	vdp->xdf_cmbl_reattach = B_TRUE;
3292 	if (dev_iscd) {
3293 		vdp->xdf_dinfo |= VDISK_CDROM;
3294 		vdp->xdf_mstate = DKIO_EJECTED;
3295 	} else {
3296 		vdp->xdf_mstate = DKIO_NONE;
3297 	}
3298 
3299 	if ((vdp->xdf_ready_tq = ddi_taskq_create(dip, "xdf_ready_tq",
3300 	    1, TASKQ_DEFAULTPRI, 0)) == NULL)
3301 		goto errout0;
3302 
3303 	if (xvdi_add_xb_watch_handler(dip, oename, XBP_HP_STATUS,
3304 	    xdf_watch_hp_status_cb, NULL) != DDI_SUCCESS)
3305 		goto errout0;
3306 
3307 	if (ddi_add_softintr(dip, DDI_SOFTINT_LOW, &vdp->xdf_softintr_id,
3308 	    &softibc, NULL, xdf_iorestart, (caddr_t)vdp) != DDI_SUCCESS) {
3309 		cmn_err(CE_WARN, "xdf@%s: failed to add softintr",
3310 		    ddi_get_name_addr(dip));
3311 		goto errout0;
3312 	}
3313 
3314 	/*
3315 	 * Initialize the physical geometry stucture.  Note that currently
3316 	 * we don't know the size of the backend device so the number
3317 	 * of blocks on the device will be initialized to zero.  Once
3318 	 * we connect to the backend device we'll update the physical
3319 	 * geometry to reflect the real size of the device.
3320 	 */
3321 	xdf_synthetic_pgeom(dip, &vdp->xdf_pgeom);
3322 	vdp->xdf_pgeom_fixed = B_FALSE;
3323 
3324 	/*
3325 	 * create default device minor nodes: non-removable disk
3326 	 * we will adjust minor nodes after we are connected w/ backend
3327 	 */
3328 	cmlb_alloc_handle(&vdp->xdf_vd_lbl);
3329 	if (xdf_cmlb_attach(vdp) != 0) {
3330 		cmn_err(CE_WARN,
3331 		    "xdf@%s: attach failed, cmlb attach failed",
3332 		    ddi_get_name_addr(dip));
3333 		goto errout0;
3334 	}
3335 
3336 	/*
3337 	 * We ship with cache-enabled disks
3338 	 */
3339 	vdp->xdf_wce = B_TRUE;
3340 
3341 	mutex_enter(&vdp->xdf_cb_lk);
3342 	/* Watch backend XenbusState change */
3343 	if (xvdi_add_event_handler(dip,
3344 	    XS_OE_STATE, xdf_oe_change, NULL) != DDI_SUCCESS) {
3345 		mutex_exit(&vdp->xdf_cb_lk);
3346 		goto errout0;
3347 	}
3348 
3349 	if (xdf_setstate_init(vdp) != DDI_SUCCESS) {
3350 		cmn_err(CE_WARN, "xdf@%s: start connection failed",
3351 		    ddi_get_name_addr(dip));
3352 		mutex_exit(&vdp->xdf_cb_lk);
3353 		goto errout1;
3354 	}
3355 	mutex_exit(&vdp->xdf_cb_lk);
3356 
3357 #if defined(XPV_HVM_DRIVER)
3358 
3359 	xdf_hvm_add(dip);
3360 
3361 	/* Report our version to dom0.  */
3362 	if (xenbus_printf(XBT_NULL, "guest/xdf", "version", "%d",
3363 	    HVMPV_XDF_VERS))
3364 		cmn_err(CE_WARN, "xdf: couldn't write version\n");
3365 
3366 #else /* !XPV_HVM_DRIVER */
3367 
3368 	/* create kstat for iostat(1M) */
3369 	if (xdf_kstat_create(dip, "xdf", instance) != 0) {
3370 		cmn_err(CE_WARN, "xdf@%s: failed to create kstat",
3371 		    ddi_get_name_addr(dip));
3372 		goto errout1;
3373 	}
3374 
3375 #endif /* !XPV_HVM_DRIVER */
3376 
3377 	ddi_report_dev(dip);
3378 	DPRINTF(DDI_DBG, ("xdf@%s: attached\n", vdp->xdf_addr));
3379 	return (DDI_SUCCESS);
3380 
3381 errout1:
3382 	(void) xvdi_switch_state(vdp->xdf_dip, XBT_NULL, XenbusStateClosed);
3383 	xvdi_remove_event_handler(dip, XS_OE_STATE);
3384 errout0:
3385 	if (vdp->xdf_vd_lbl != NULL) {
3386 		cmlb_detach(vdp->xdf_vd_lbl, NULL);
3387 		cmlb_free_handle(&vdp->xdf_vd_lbl);
3388 		vdp->xdf_vd_lbl = NULL;
3389 	}
3390 	if (vdp->xdf_softintr_id != NULL)
3391 		ddi_remove_softintr(vdp->xdf_softintr_id);
3392 	xvdi_remove_xb_watch_handlers(dip);
3393 	if (vdp->xdf_ready_tq != NULL)
3394 		ddi_taskq_destroy(vdp->xdf_ready_tq);
3395 	mutex_destroy(&vdp->xdf_cb_lk);
3396 	mutex_destroy(&vdp->xdf_dev_lk);
3397 	cv_destroy(&vdp->xdf_dev_cv);
3398 	cv_destroy(&vdp->xdf_hp_status_cv);
3399 	ddi_soft_state_free(xdf_ssp, instance);
3400 	ddi_set_driver_private(dip, NULL);
3401 	ddi_prop_remove_all(dip);
3402 	cmn_err(CE_WARN, "xdf@%s: attach failed", ddi_get_name_addr(dip));
3403 	return (DDI_FAILURE);
3404 }
3405 
3406 static int
3407 xdf_suspend(dev_info_t *dip)
3408 {
3409 	int		instance = ddi_get_instance(dip);
3410 	xdf_t		*vdp;
3411 
3412 	if ((vdp = ddi_get_soft_state(xdf_ssp, instance)) == NULL)
3413 		return (DDI_FAILURE);
3414 
3415 	if (xdf_debug & SUSRES_DBG)
3416 		xen_printf("xdf@%s: xdf_suspend\n", vdp->xdf_addr);
3417 
3418 	xvdi_suspend(dip);
3419 
3420 	mutex_enter(&vdp->xdf_cb_lk);
3421 	mutex_enter(&vdp->xdf_dev_lk);
3422 
3423 	vdp->xdf_suspending = B_TRUE;
3424 	xdf_ring_destroy(vdp);
3425 	xdf_set_state(vdp, XD_SUSPEND);
3426 	vdp->xdf_suspending = B_FALSE;
3427 
3428 	mutex_exit(&vdp->xdf_dev_lk);
3429 	mutex_exit(&vdp->xdf_cb_lk);
3430 
3431 	if (xdf_debug & SUSRES_DBG)
3432 		xen_printf("xdf@%s: xdf_suspend: done\n", vdp->xdf_addr);
3433 
3434 	return (DDI_SUCCESS);
3435 }
3436 
3437 static int
3438 xdf_detach(dev_info_t *dip, ddi_detach_cmd_t cmd)
3439 {
3440 	xdf_t *vdp;
3441 	int instance;
3442 
3443 	switch (cmd) {
3444 
3445 	case DDI_PM_SUSPEND:
3446 		break;
3447 
3448 	case DDI_SUSPEND:
3449 		return (xdf_suspend(dip));
3450 
3451 	case DDI_DETACH:
3452 		break;
3453 
3454 	default:
3455 		return (DDI_FAILURE);
3456 	}
3457 
3458 	instance = ddi_get_instance(dip);
3459 	DPRINTF(DDI_DBG, ("xdf@%s: detaching\n", ddi_get_name_addr(dip)));
3460 	vdp = ddi_get_soft_state(xdf_ssp, instance);
3461 
3462 	if (vdp == NULL)
3463 		return (DDI_FAILURE);
3464 
3465 	mutex_enter(&vdp->xdf_cb_lk);
3466 	xdf_disconnect(vdp, XD_CLOSED, B_FALSE);
3467 	if (vdp->xdf_state != XD_CLOSED) {
3468 		mutex_exit(&vdp->xdf_cb_lk);
3469 		return (DDI_FAILURE);
3470 	}
3471 	mutex_exit(&vdp->xdf_cb_lk);
3472 
3473 	ASSERT(!ISDMACBON(vdp));
3474 
3475 #if defined(XPV_HVM_DRIVER)
3476 	xdf_hvm_rm(dip);
3477 #endif /* XPV_HVM_DRIVER */
3478 
3479 	if (vdp->xdf_timeout_id != 0)
3480 		(void) untimeout(vdp->xdf_timeout_id);
3481 
3482 	xvdi_remove_event_handler(dip, XS_OE_STATE);
3483 	ddi_taskq_destroy(vdp->xdf_ready_tq);
3484 
3485 	cmlb_detach(vdp->xdf_vd_lbl, NULL);
3486 	cmlb_free_handle(&vdp->xdf_vd_lbl);
3487 
3488 	/* we'll support backend running in domU later */
3489 #ifdef	DOMU_BACKEND
3490 	(void) xvdi_post_event(dip, XEN_HP_REMOVE);
3491 #endif
3492 
3493 	list_destroy(&vdp->xdf_vreq_act);
3494 	ddi_prop_remove_all(dip);
3495 	xdf_kstat_delete(dip);
3496 	ddi_remove_softintr(vdp->xdf_softintr_id);
3497 	xvdi_remove_xb_watch_handlers(dip);
3498 	ddi_set_driver_private(dip, NULL);
3499 	cv_destroy(&vdp->xdf_dev_cv);
3500 	mutex_destroy(&vdp->xdf_cb_lk);
3501 	mutex_destroy(&vdp->xdf_dev_lk);
3502 	if (vdp->xdf_cache_flush_block != NULL)
3503 		kmem_free(vdp->xdf_flush_mem, 2 * vdp->xdf_xdev_secsize);
3504 	ddi_soft_state_free(xdf_ssp, instance);
3505 	return (DDI_SUCCESS);
3506 }
3507 
3508 /*
3509  * Driver linkage structures.
3510  */
3511 static struct cb_ops xdf_cbops = {
3512 	xdf_open,
3513 	xdf_close,
3514 	xdf_strategy,
3515 	nodev,
3516 	xdf_dump,
3517 	xdf_read,
3518 	xdf_write,
3519 	xdf_ioctl,
3520 	nodev,
3521 	nodev,
3522 	nodev,
3523 	nochpoll,
3524 	xdf_prop_op,
3525 	NULL,
3526 	D_MP | D_NEW | D_64BIT,
3527 	CB_REV,
3528 	xdf_aread,
3529 	xdf_awrite
3530 };
3531 
3532 struct dev_ops xdf_devops = {
3533 	DEVO_REV,		/* devo_rev */
3534 	0,			/* devo_refcnt */
3535 	xdf_getinfo,		/* devo_getinfo */
3536 	nulldev,		/* devo_identify */
3537 	nulldev,		/* devo_probe */
3538 	xdf_attach,		/* devo_attach */
3539 	xdf_detach,		/* devo_detach */
3540 	nodev,			/* devo_reset */
3541 	&xdf_cbops,		/* devo_cb_ops */
3542 	NULL,			/* devo_bus_ops */
3543 	NULL,			/* devo_power */
3544 	ddi_quiesce_not_supported, /* devo_quiesce */
3545 };
3546 
3547 /*
3548  * Module linkage structures.
3549  */
3550 static struct modldrv modldrv = {
3551 	&mod_driverops,		/* Type of module.  This one is a driver */
3552 	"virtual block driver",	/* short description */
3553 	&xdf_devops		/* driver specific ops */
3554 };
3555 
3556 static struct modlinkage xdf_modlinkage = {
3557 	MODREV_1, (void *)&modldrv, NULL
3558 };
3559 
3560 /*
3561  * standard module entry points
3562  */
3563 int
3564 _init(void)
3565 {
3566 	int rc;
3567 
3568 	xdf_major = ddi_name_to_major("xdf");
3569 	if (xdf_major == (major_t)-1)
3570 		return (EINVAL);
3571 
3572 	if ((rc = ddi_soft_state_init(&xdf_ssp, sizeof (xdf_t), 0)) != 0)
3573 		return (rc);
3574 
3575 	xdf_vreq_cache = kmem_cache_create("xdf_vreq_cache",
3576 	    sizeof (v_req_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
3577 	xdf_gs_cache = kmem_cache_create("xdf_gs_cache",
3578 	    sizeof (ge_slot_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
3579 
3580 #if defined(XPV_HVM_DRIVER)
3581 	xdf_hvm_init();
3582 #endif /* XPV_HVM_DRIVER */
3583 
3584 	if ((rc = mod_install(&xdf_modlinkage)) != 0) {
3585 #if defined(XPV_HVM_DRIVER)
3586 		xdf_hvm_fini();
3587 #endif /* XPV_HVM_DRIVER */
3588 		kmem_cache_destroy(xdf_vreq_cache);
3589 		kmem_cache_destroy(xdf_gs_cache);
3590 		ddi_soft_state_fini(&xdf_ssp);
3591 		return (rc);
3592 	}
3593 
3594 	return (rc);
3595 }
3596 
3597 int
3598 _fini(void)
3599 {
3600 	int err;
3601 	if ((err = mod_remove(&xdf_modlinkage)) != 0)
3602 		return (err);
3603 
3604 #if defined(XPV_HVM_DRIVER)
3605 	xdf_hvm_fini();
3606 #endif /* XPV_HVM_DRIVER */
3607 
3608 	kmem_cache_destroy(xdf_vreq_cache);
3609 	kmem_cache_destroy(xdf_gs_cache);
3610 	ddi_soft_state_fini(&xdf_ssp);
3611 
3612 	return (0);
3613 }
3614 
3615 int
3616 _info(struct modinfo *modinfop)
3617 {
3618 	return (mod_info(&xdf_modlinkage, modinfop));
3619 }
3620