xref: /freebsd/sys/dev/ioat/ioat.c (revision 6419bb52)
1 /*-
2  * Copyright (C) 2012 Intel Corporation
3  * All rights reserved.
4  * Copyright (C) 2018 Alexander Motin <mav@FreeBSD.org>
5  *
6  * Redistribution and use in source and binary forms, with or without
7  * modification, are permitted provided that the following conditions
8  * are met:
9  * 1. Redistributions of source code must retain the above copyright
10  *    notice, this list of conditions and the following disclaimer.
11  * 2. Redistributions in binary form must reproduce the above copyright
12  *    notice, this list of conditions and the following disclaimer in the
13  *    documentation and/or other materials provided with the distribution.
14  *
15  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
16  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
19  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
20  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
21  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
22  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
23  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
24  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
25  * SUCH DAMAGE.
26  */
27 
28 #include <sys/cdefs.h>
29 __FBSDID("$FreeBSD$");
30 
31 #include "opt_ddb.h"
32 
33 #include <sys/param.h>
34 #include <sys/systm.h>
35 #include <sys/bus.h>
36 #include <sys/conf.h>
37 #include <sys/domainset.h>
38 #include <sys/fail.h>
39 #include <sys/ioccom.h>
40 #include <sys/kernel.h>
41 #include <sys/ktr.h>
42 #include <sys/lock.h>
43 #include <sys/malloc.h>
44 #include <sys/module.h>
45 #include <sys/mutex.h>
46 #include <sys/rman.h>
47 #include <sys/sbuf.h>
48 #include <sys/smp.h>
49 #include <sys/sysctl.h>
50 #include <sys/taskqueue.h>
51 #include <sys/time.h>
52 #include <dev/pci/pcireg.h>
53 #include <dev/pci/pcivar.h>
54 #include <machine/bus.h>
55 #include <machine/resource.h>
56 #include <machine/stdarg.h>
57 
58 #ifdef DDB
59 #include <ddb/ddb.h>
60 #endif
61 
62 #include "ioat.h"
63 #include "ioat_hw.h"
64 #include "ioat_internal.h"
65 
66 #ifndef	BUS_SPACE_MAXADDR_40BIT
67 #define	BUS_SPACE_MAXADDR_40BIT	MIN(BUS_SPACE_MAXADDR, 0xFFFFFFFFFFULL)
68 #endif
69 #ifndef	BUS_SPACE_MAXADDR_46BIT
70 #define	BUS_SPACE_MAXADDR_46BIT	MIN(BUS_SPACE_MAXADDR, 0x3FFFFFFFFFFFULL)
71 #endif
72 
73 static int ioat_probe(device_t device);
74 static int ioat_attach(device_t device);
75 static int ioat_detach(device_t device);
76 static int ioat_setup_intr(struct ioat_softc *ioat);
77 static int ioat_teardown_intr(struct ioat_softc *ioat);
78 static int ioat3_attach(device_t device);
79 static int ioat_start_channel(struct ioat_softc *ioat);
80 static int ioat_map_pci_bar(struct ioat_softc *ioat);
81 static void ioat_dmamap_cb(void *arg, bus_dma_segment_t *segs, int nseg,
82     int error);
83 static void ioat_interrupt_handler(void *arg);
84 static boolean_t ioat_model_resets_msix(struct ioat_softc *ioat);
85 static int chanerr_to_errno(uint32_t);
86 static void ioat_process_events(struct ioat_softc *ioat, boolean_t intr);
87 static inline uint32_t ioat_get_active(struct ioat_softc *ioat);
88 static inline uint32_t ioat_get_ring_space(struct ioat_softc *ioat);
89 static void ioat_free_ring(struct ioat_softc *, uint32_t size,
90     struct ioat_descriptor *);
91 static int ioat_reserve_space(struct ioat_softc *, uint32_t, int mflags);
92 static union ioat_hw_descriptor *ioat_get_descriptor(struct ioat_softc *,
93     uint32_t index);
94 static struct ioat_descriptor *ioat_get_ring_entry(struct ioat_softc *,
95     uint32_t index);
96 static void ioat_halted_debug(struct ioat_softc *, uint32_t);
97 static void ioat_poll_timer_callback(void *arg);
98 static void dump_descriptor(void *hw_desc);
99 static void ioat_submit_single(struct ioat_softc *ioat);
100 static void ioat_comp_update_map(void *arg, bus_dma_segment_t *seg, int nseg,
101     int error);
102 static int ioat_reset_hw(struct ioat_softc *ioat);
103 static void ioat_reset_hw_task(void *, int);
104 static void ioat_setup_sysctl(device_t device);
105 static int sysctl_handle_reset(SYSCTL_HANDLER_ARGS);
106 static void ioat_get(struct ioat_softc *);
107 static void ioat_put(struct ioat_softc *);
108 static void ioat_drain_locked(struct ioat_softc *);
109 
110 #define	ioat_log_message(v, ...) do {					\
111 	if ((v) <= g_ioat_debug_level) {				\
112 		device_printf(ioat->device, __VA_ARGS__);		\
113 	}								\
114 } while (0)
115 
116 MALLOC_DEFINE(M_IOAT, "ioat", "ioat driver memory allocations");
117 SYSCTL_NODE(_hw, OID_AUTO, ioat, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
118     "ioat node");
119 
120 static int g_force_legacy_interrupts;
121 SYSCTL_INT(_hw_ioat, OID_AUTO, force_legacy_interrupts, CTLFLAG_RDTUN,
122     &g_force_legacy_interrupts, 0, "Set to non-zero to force MSI-X disabled");
123 
124 int g_ioat_debug_level = 0;
125 SYSCTL_INT(_hw_ioat, OID_AUTO, debug_level, CTLFLAG_RWTUN, &g_ioat_debug_level,
126     0, "Set log level (0-3) for ioat(4). Higher is more verbose.");
127 
128 unsigned g_ioat_ring_order = 13;
129 SYSCTL_UINT(_hw_ioat, OID_AUTO, ring_order, CTLFLAG_RDTUN, &g_ioat_ring_order,
130     0, "Set IOAT ring order.  (1 << this) == ring size.");
131 
132 /*
133  * OS <-> Driver interface structures
134  */
135 static device_method_t ioat_pci_methods[] = {
136 	/* Device interface */
137 	DEVMETHOD(device_probe,     ioat_probe),
138 	DEVMETHOD(device_attach,    ioat_attach),
139 	DEVMETHOD(device_detach,    ioat_detach),
140 	DEVMETHOD_END
141 };
142 
143 static driver_t ioat_pci_driver = {
144 	"ioat",
145 	ioat_pci_methods,
146 	sizeof(struct ioat_softc),
147 };
148 
149 static devclass_t ioat_devclass;
150 DRIVER_MODULE(ioat, pci, ioat_pci_driver, ioat_devclass, 0, 0);
151 MODULE_VERSION(ioat, 1);
152 
153 /*
154  * Private data structures
155  */
156 static struct ioat_softc *ioat_channel[IOAT_MAX_CHANNELS];
157 static unsigned ioat_channel_index = 0;
158 SYSCTL_UINT(_hw_ioat, OID_AUTO, channels, CTLFLAG_RD, &ioat_channel_index, 0,
159     "Number of IOAT channels attached");
160 static struct mtx ioat_list_mtx;
161 MTX_SYSINIT(ioat_list_mtx, &ioat_list_mtx, "ioat list mtx", MTX_DEF);
162 
163 static struct _pcsid
164 {
165 	u_int32_t   type;
166 	const char  *desc;
167 } pci_ids[] = {
168 	{ 0x34308086, "TBG IOAT Ch0" },
169 	{ 0x34318086, "TBG IOAT Ch1" },
170 	{ 0x34328086, "TBG IOAT Ch2" },
171 	{ 0x34338086, "TBG IOAT Ch3" },
172 	{ 0x34298086, "TBG IOAT Ch4" },
173 	{ 0x342a8086, "TBG IOAT Ch5" },
174 	{ 0x342b8086, "TBG IOAT Ch6" },
175 	{ 0x342c8086, "TBG IOAT Ch7" },
176 
177 	{ 0x37108086, "JSF IOAT Ch0" },
178 	{ 0x37118086, "JSF IOAT Ch1" },
179 	{ 0x37128086, "JSF IOAT Ch2" },
180 	{ 0x37138086, "JSF IOAT Ch3" },
181 	{ 0x37148086, "JSF IOAT Ch4" },
182 	{ 0x37158086, "JSF IOAT Ch5" },
183 	{ 0x37168086, "JSF IOAT Ch6" },
184 	{ 0x37178086, "JSF IOAT Ch7" },
185 	{ 0x37188086, "JSF IOAT Ch0 (RAID)" },
186 	{ 0x37198086, "JSF IOAT Ch1 (RAID)" },
187 
188 	{ 0x3c208086, "SNB IOAT Ch0" },
189 	{ 0x3c218086, "SNB IOAT Ch1" },
190 	{ 0x3c228086, "SNB IOAT Ch2" },
191 	{ 0x3c238086, "SNB IOAT Ch3" },
192 	{ 0x3c248086, "SNB IOAT Ch4" },
193 	{ 0x3c258086, "SNB IOAT Ch5" },
194 	{ 0x3c268086, "SNB IOAT Ch6" },
195 	{ 0x3c278086, "SNB IOAT Ch7" },
196 	{ 0x3c2e8086, "SNB IOAT Ch0 (RAID)" },
197 	{ 0x3c2f8086, "SNB IOAT Ch1 (RAID)" },
198 
199 	{ 0x0e208086, "IVB IOAT Ch0" },
200 	{ 0x0e218086, "IVB IOAT Ch1" },
201 	{ 0x0e228086, "IVB IOAT Ch2" },
202 	{ 0x0e238086, "IVB IOAT Ch3" },
203 	{ 0x0e248086, "IVB IOAT Ch4" },
204 	{ 0x0e258086, "IVB IOAT Ch5" },
205 	{ 0x0e268086, "IVB IOAT Ch6" },
206 	{ 0x0e278086, "IVB IOAT Ch7" },
207 	{ 0x0e2e8086, "IVB IOAT Ch0 (RAID)" },
208 	{ 0x0e2f8086, "IVB IOAT Ch1 (RAID)" },
209 
210 	{ 0x2f208086, "HSW IOAT Ch0" },
211 	{ 0x2f218086, "HSW IOAT Ch1" },
212 	{ 0x2f228086, "HSW IOAT Ch2" },
213 	{ 0x2f238086, "HSW IOAT Ch3" },
214 	{ 0x2f248086, "HSW IOAT Ch4" },
215 	{ 0x2f258086, "HSW IOAT Ch5" },
216 	{ 0x2f268086, "HSW IOAT Ch6" },
217 	{ 0x2f278086, "HSW IOAT Ch7" },
218 	{ 0x2f2e8086, "HSW IOAT Ch0 (RAID)" },
219 	{ 0x2f2f8086, "HSW IOAT Ch1 (RAID)" },
220 
221 	{ 0x0c508086, "BWD IOAT Ch0" },
222 	{ 0x0c518086, "BWD IOAT Ch1" },
223 	{ 0x0c528086, "BWD IOAT Ch2" },
224 	{ 0x0c538086, "BWD IOAT Ch3" },
225 
226 	{ 0x6f508086, "BDXDE IOAT Ch0" },
227 	{ 0x6f518086, "BDXDE IOAT Ch1" },
228 	{ 0x6f528086, "BDXDE IOAT Ch2" },
229 	{ 0x6f538086, "BDXDE IOAT Ch3" },
230 
231 	{ 0x6f208086, "BDX IOAT Ch0" },
232 	{ 0x6f218086, "BDX IOAT Ch1" },
233 	{ 0x6f228086, "BDX IOAT Ch2" },
234 	{ 0x6f238086, "BDX IOAT Ch3" },
235 	{ 0x6f248086, "BDX IOAT Ch4" },
236 	{ 0x6f258086, "BDX IOAT Ch5" },
237 	{ 0x6f268086, "BDX IOAT Ch6" },
238 	{ 0x6f278086, "BDX IOAT Ch7" },
239 	{ 0x6f2e8086, "BDX IOAT Ch0 (RAID)" },
240 	{ 0x6f2f8086, "BDX IOAT Ch1 (RAID)" },
241 
242 	{ 0x20218086, "SKX IOAT" },
243 };
244 
245 MODULE_PNP_INFO("W32:vendor/device;D:#", pci, ioat, pci_ids,
246     nitems(pci_ids));
247 
248 /*
249  * OS <-> Driver linkage functions
250  */
251 static int
252 ioat_probe(device_t device)
253 {
254 	struct _pcsid *ep;
255 	u_int32_t type;
256 
257 	type = pci_get_devid(device);
258 	for (ep = pci_ids; ep < &pci_ids[nitems(pci_ids)]; ep++) {
259 		if (ep->type == type) {
260 			device_set_desc(device, ep->desc);
261 			return (0);
262 		}
263 	}
264 	return (ENXIO);
265 }
266 
267 static int
268 ioat_attach(device_t device)
269 {
270 	struct ioat_softc *ioat;
271 	int error, i;
272 
273 	ioat = DEVICE2SOFTC(device);
274 	ioat->device = device;
275 	if (bus_get_domain(device, &ioat->domain) != 0)
276 		ioat->domain = 0;
277 	ioat->cpu = CPU_FFS(&cpuset_domain[ioat->domain]) - 1;
278 	if (ioat->cpu < 0)
279 		ioat->cpu = CPU_FIRST();
280 
281 	error = ioat_map_pci_bar(ioat);
282 	if (error != 0)
283 		goto err;
284 
285 	ioat->version = ioat_read_cbver(ioat);
286 	if (ioat->version < IOAT_VER_3_0) {
287 		error = ENODEV;
288 		goto err;
289 	}
290 
291 	error = ioat3_attach(device);
292 	if (error != 0)
293 		goto err;
294 
295 	error = pci_enable_busmaster(device);
296 	if (error != 0)
297 		goto err;
298 
299 	error = ioat_setup_intr(ioat);
300 	if (error != 0)
301 		goto err;
302 
303 	error = ioat_reset_hw(ioat);
304 	if (error != 0)
305 		goto err;
306 
307 	ioat_process_events(ioat, FALSE);
308 	ioat_setup_sysctl(device);
309 
310 	mtx_lock(&ioat_list_mtx);
311 	for (i = 0; i < IOAT_MAX_CHANNELS; i++) {
312 		if (ioat_channel[i] == NULL)
313 			break;
314 	}
315 	if (i >= IOAT_MAX_CHANNELS) {
316 		mtx_unlock(&ioat_list_mtx);
317 		device_printf(device, "Too many I/OAT devices in system\n");
318 		error = ENXIO;
319 		goto err;
320 	}
321 	ioat->chan_idx = i;
322 	ioat_channel[i] = ioat;
323 	if (i >= ioat_channel_index)
324 		ioat_channel_index = i + 1;
325 	mtx_unlock(&ioat_list_mtx);
326 
327 	ioat_test_attach();
328 
329 err:
330 	if (error != 0)
331 		ioat_detach(device);
332 	return (error);
333 }
334 
335 static inline int
336 ioat_bus_dmamap_destroy(struct ioat_softc *ioat, const char *func,
337     bus_dma_tag_t dmat, bus_dmamap_t map)
338 {
339 	int error;
340 
341 	error = bus_dmamap_destroy(dmat, map);
342 	if (error != 0) {
343 		ioat_log_message(0,
344 		    "%s: bus_dmamap_destroy failed %d\n", func, error);
345 	}
346 
347 	return (error);
348 }
349 
350 static int
351 ioat_detach(device_t device)
352 {
353 	struct ioat_softc *ioat;
354 	int i, error;
355 
356 	ioat = DEVICE2SOFTC(device);
357 
358 	mtx_lock(&ioat_list_mtx);
359 	ioat_channel[ioat->chan_idx] = NULL;
360 	while (ioat_channel_index > 0 &&
361 	    ioat_channel[ioat_channel_index - 1] == NULL)
362 		ioat_channel_index--;
363 	mtx_unlock(&ioat_list_mtx);
364 
365 	ioat_test_detach();
366 	taskqueue_drain(taskqueue_thread, &ioat->reset_task);
367 
368 	mtx_lock(&ioat->submit_lock);
369 	ioat->quiescing = TRUE;
370 	ioat->destroying = TRUE;
371 	wakeup(&ioat->quiescing);
372 	wakeup(&ioat->resetting);
373 
374 	ioat_drain_locked(ioat);
375 	mtx_unlock(&ioat->submit_lock);
376 	mtx_lock(&ioat->cleanup_lock);
377 	while (ioat_get_active(ioat) > 0)
378 		msleep(&ioat->tail, &ioat->cleanup_lock, 0, "ioat_drain", 1);
379 	mtx_unlock(&ioat->cleanup_lock);
380 
381 	ioat_teardown_intr(ioat);
382 	callout_drain(&ioat->poll_timer);
383 
384 	pci_disable_busmaster(device);
385 
386 	if (ioat->pci_resource != NULL)
387 		bus_release_resource(device, SYS_RES_MEMORY,
388 		    ioat->pci_resource_id, ioat->pci_resource);
389 
390 	if (ioat->data_tag != NULL) {
391 		for (i = 0; i < 1 << ioat->ring_size_order; i++) {
392 			error = ioat_bus_dmamap_destroy(ioat, __func__,
393 			    ioat->data_tag, ioat->ring[i].src_dmamap);
394 			if (error != 0)
395 				return (error);
396 		}
397 		for (i = 0; i < 1 << ioat->ring_size_order; i++) {
398 			error = ioat_bus_dmamap_destroy(ioat, __func__,
399 			    ioat->data_tag, ioat->ring[i].dst_dmamap);
400 			if (error != 0)
401 				return (error);
402 		}
403 
404 		for (i = 0; i < 1 << ioat->ring_size_order; i++) {
405 			error = ioat_bus_dmamap_destroy(ioat, __func__,
406 			    ioat->data_tag, ioat->ring[i].src2_dmamap);
407 			if (error != 0)
408 				return (error);
409 		}
410 		for (i = 0; i < 1 << ioat->ring_size_order; i++) {
411 			error = ioat_bus_dmamap_destroy(ioat, __func__,
412 			    ioat->data_tag, ioat->ring[i].dst2_dmamap);
413 			if (error != 0)
414 				return (error);
415 		}
416 
417 		bus_dma_tag_destroy(ioat->data_tag);
418 	}
419 
420 	if (ioat->ring != NULL)
421 		ioat_free_ring(ioat, 1 << ioat->ring_size_order, ioat->ring);
422 
423 	if (ioat->comp_update != NULL) {
424 		bus_dmamap_unload(ioat->comp_update_tag, ioat->comp_update_map);
425 		bus_dmamem_free(ioat->comp_update_tag, ioat->comp_update,
426 		    ioat->comp_update_map);
427 		bus_dma_tag_destroy(ioat->comp_update_tag);
428 	}
429 
430 	if (ioat->hw_desc_ring != NULL) {
431 		bus_dmamap_unload(ioat->hw_desc_tag, ioat->hw_desc_map);
432 		bus_dmamem_free(ioat->hw_desc_tag, ioat->hw_desc_ring,
433 		    ioat->hw_desc_map);
434 		bus_dma_tag_destroy(ioat->hw_desc_tag);
435 	}
436 
437 	return (0);
438 }
439 
440 static int
441 ioat_teardown_intr(struct ioat_softc *ioat)
442 {
443 
444 	if (ioat->tag != NULL)
445 		bus_teardown_intr(ioat->device, ioat->res, ioat->tag);
446 
447 	if (ioat->res != NULL)
448 		bus_release_resource(ioat->device, SYS_RES_IRQ,
449 		    rman_get_rid(ioat->res), ioat->res);
450 
451 	pci_release_msi(ioat->device);
452 	return (0);
453 }
454 
455 static int
456 ioat_start_channel(struct ioat_softc *ioat)
457 {
458 	struct ioat_dma_hw_descriptor *hw_desc;
459 	struct ioat_descriptor *desc;
460 	struct bus_dmadesc *dmadesc;
461 	uint64_t status;
462 	uint32_t chanerr;
463 	int i;
464 
465 	ioat_acquire(&ioat->dmaengine);
466 
467 	/* Submit 'NULL' operation manually to avoid quiescing flag */
468 	desc = ioat_get_ring_entry(ioat, ioat->head);
469 	hw_desc = &ioat_get_descriptor(ioat, ioat->head)->dma;
470 	dmadesc = &desc->bus_dmadesc;
471 
472 	dmadesc->callback_fn = NULL;
473 	dmadesc->callback_arg = NULL;
474 
475 	hw_desc->u.control_raw = 0;
476 	hw_desc->u.control_generic.op = IOAT_OP_COPY;
477 	hw_desc->u.control_generic.completion_update = 1;
478 	hw_desc->size = 8;
479 	hw_desc->src_addr = 0;
480 	hw_desc->dest_addr = 0;
481 	hw_desc->u.control.null = 1;
482 
483 	ioat_submit_single(ioat);
484 	ioat_release(&ioat->dmaengine);
485 
486 	for (i = 0; i < 100; i++) {
487 		DELAY(1);
488 		status = ioat_get_chansts(ioat);
489 		if (is_ioat_idle(status))
490 			return (0);
491 	}
492 
493 	chanerr = ioat_read_4(ioat, IOAT_CHANERR_OFFSET);
494 	ioat_log_message(0, "could not start channel: "
495 	    "status = %#jx error = %b\n", (uintmax_t)status, (int)chanerr,
496 	    IOAT_CHANERR_STR);
497 	return (ENXIO);
498 }
499 
500 /*
501  * Initialize Hardware
502  */
503 static int
504 ioat3_attach(device_t device)
505 {
506 	struct ioat_softc *ioat;
507 	struct ioat_descriptor *ring;
508 	struct ioat_dma_hw_descriptor *dma_hw_desc;
509 	void *hw_desc;
510 	bus_addr_t lowaddr;
511 	size_t ringsz;
512 	int i, num_descriptors;
513 	int error;
514 	uint8_t xfercap;
515 
516 	error = 0;
517 	ioat = DEVICE2SOFTC(device);
518 	ioat->capabilities = ioat_read_dmacapability(ioat);
519 
520 	ioat_log_message(0, "Capabilities: %b\n", (int)ioat->capabilities,
521 	    IOAT_DMACAP_STR);
522 
523 	xfercap = ioat_read_xfercap(ioat);
524 	ioat->max_xfer_size = 1 << xfercap;
525 
526 	ioat->intrdelay_supported = (ioat_read_2(ioat, IOAT_INTRDELAY_OFFSET) &
527 	    IOAT_INTRDELAY_SUPPORTED) != 0;
528 	if (ioat->intrdelay_supported)
529 		ioat->intrdelay_max = IOAT_INTRDELAY_US_MASK;
530 
531 	/* TODO: need to check DCA here if we ever do XOR/PQ */
532 
533 	mtx_init(&ioat->submit_lock, "ioat_submit", NULL, MTX_DEF);
534 	mtx_init(&ioat->cleanup_lock, "ioat_cleanup", NULL, MTX_DEF);
535 	callout_init(&ioat->poll_timer, 1);
536 	TASK_INIT(&ioat->reset_task, 0, ioat_reset_hw_task, ioat);
537 
538 	/* Establish lock order for Witness */
539 	mtx_lock(&ioat->cleanup_lock);
540 	mtx_lock(&ioat->submit_lock);
541 	mtx_unlock(&ioat->submit_lock);
542 	mtx_unlock(&ioat->cleanup_lock);
543 
544 	ioat->is_submitter_processing = FALSE;
545 
546 	if (ioat->version >= IOAT_VER_3_3)
547 		lowaddr = BUS_SPACE_MAXADDR_48BIT;
548 	else if (ioat->version >= IOAT_VER_3_2)
549 		lowaddr = BUS_SPACE_MAXADDR_46BIT;
550 	else
551 		lowaddr = BUS_SPACE_MAXADDR_40BIT;
552 
553 	error = bus_dma_tag_create(bus_get_dma_tag(ioat->device),
554 	    sizeof(uint64_t), 0x0, lowaddr, BUS_SPACE_MAXADDR, NULL, NULL,
555 	    sizeof(uint64_t), 1, sizeof(uint64_t), 0, NULL, NULL,
556 	    &ioat->comp_update_tag);
557 	if (error != 0)
558 		return (error);
559 
560 	error = bus_dmamem_alloc(ioat->comp_update_tag,
561 	    (void **)&ioat->comp_update, BUS_DMA_ZERO | BUS_DMA_WAITOK,
562 	    &ioat->comp_update_map);
563 	if (error != 0)
564 		return (error);
565 
566 	error = bus_dmamap_load(ioat->comp_update_tag, ioat->comp_update_map,
567 	    ioat->comp_update, sizeof(uint64_t), ioat_comp_update_map, ioat,
568 	    BUS_DMA_NOWAIT);
569 	if (error != 0)
570 		return (error);
571 
572 	ioat->ring_size_order = g_ioat_ring_order;
573 	num_descriptors = 1 << ioat->ring_size_order;
574 	ringsz = sizeof(struct ioat_dma_hw_descriptor) * num_descriptors;
575 
576 	error = bus_dma_tag_create(bus_get_dma_tag(ioat->device),
577 	    2 * 1024 * 1024, 0x0, lowaddr, BUS_SPACE_MAXADDR, NULL, NULL,
578 	    ringsz, 1, ringsz, 0, NULL, NULL, &ioat->hw_desc_tag);
579 	if (error != 0)
580 		return (error);
581 
582 	error = bus_dmamem_alloc(ioat->hw_desc_tag, &hw_desc,
583 	    BUS_DMA_ZERO | BUS_DMA_WAITOK, &ioat->hw_desc_map);
584 	if (error != 0)
585 		return (error);
586 
587 	error = bus_dmamap_load(ioat->hw_desc_tag, ioat->hw_desc_map, hw_desc,
588 	    ringsz, ioat_dmamap_cb, &ioat->hw_desc_bus_addr, BUS_DMA_NOWAIT);
589 	if (error)
590 		return (error);
591 
592 	ioat->hw_desc_ring = hw_desc;
593 
594 	error = bus_dma_tag_create(bus_get_dma_tag(ioat->device),
595 	    1, 0, lowaddr, BUS_SPACE_MAXADDR, NULL, NULL,
596 	    ioat->max_xfer_size, 1, ioat->max_xfer_size, 0, NULL, NULL,
597 	    &ioat->data_tag);
598 	if (error != 0)
599 		return (error);
600 	ioat->ring = malloc_domainset(num_descriptors * sizeof(*ring), M_IOAT,
601 	    DOMAINSET_PREF(ioat->domain), M_ZERO | M_WAITOK);
602 
603 	ring = ioat->ring;
604 	for (i = 0; i < num_descriptors; i++) {
605 		memset(&ring[i].bus_dmadesc, 0, sizeof(ring[i].bus_dmadesc));
606 		ring[i].id = i;
607 		error = bus_dmamap_create(ioat->data_tag, 0,
608                     &ring[i].src_dmamap);
609 		if (error != 0) {
610 			ioat_log_message(0,
611 			    "%s: bus_dmamap_create failed %d\n", __func__,
612 			    error);
613 			return (error);
614 		}
615 		error = bus_dmamap_create(ioat->data_tag, 0,
616                     &ring[i].dst_dmamap);
617 		if (error != 0) {
618 			ioat_log_message(0,
619 			    "%s: bus_dmamap_create failed %d\n", __func__,
620 			    error);
621 			return (error);
622 		}
623 		error = bus_dmamap_create(ioat->data_tag, 0,
624                     &ring[i].src2_dmamap);
625 		if (error != 0) {
626 			ioat_log_message(0,
627 			    "%s: bus_dmamap_create failed %d\n", __func__,
628 			    error);
629 			return (error);
630 		}
631 		error = bus_dmamap_create(ioat->data_tag, 0,
632                     &ring[i].dst2_dmamap);
633 		if (error != 0) {
634 			ioat_log_message(0,
635 			    "%s: bus_dmamap_create failed %d\n", __func__,
636 			    error);
637 			return (error);
638 		}
639 	}
640 
641 	for (i = 0; i < num_descriptors; i++) {
642 		dma_hw_desc = &ioat->hw_desc_ring[i].dma;
643 		dma_hw_desc->next = RING_PHYS_ADDR(ioat, i + 1);
644 	}
645 
646 	ioat->tail = ioat->head = 0;
647 	*ioat->comp_update = ioat->last_seen =
648 	    RING_PHYS_ADDR(ioat, ioat->tail - 1);
649 	return (0);
650 }
651 
652 static int
653 ioat_map_pci_bar(struct ioat_softc *ioat)
654 {
655 
656 	ioat->pci_resource_id = PCIR_BAR(0);
657 	ioat->pci_resource = bus_alloc_resource_any(ioat->device,
658 	    SYS_RES_MEMORY, &ioat->pci_resource_id, RF_ACTIVE);
659 
660 	if (ioat->pci_resource == NULL) {
661 		ioat_log_message(0, "unable to allocate pci resource\n");
662 		return (ENODEV);
663 	}
664 
665 	ioat->pci_bus_tag = rman_get_bustag(ioat->pci_resource);
666 	ioat->pci_bus_handle = rman_get_bushandle(ioat->pci_resource);
667 	return (0);
668 }
669 
670 static void
671 ioat_comp_update_map(void *arg, bus_dma_segment_t *seg, int nseg, int error)
672 {
673 	struct ioat_softc *ioat = arg;
674 
675 	KASSERT(error == 0, ("%s: error:%d", __func__, error));
676 	ioat->comp_update_bus_addr = seg[0].ds_addr;
677 }
678 
679 static void
680 ioat_dmamap_cb(void *arg, bus_dma_segment_t *segs, int nseg, int error)
681 {
682 	bus_addr_t *baddr;
683 
684 	KASSERT(error == 0, ("%s: error:%d", __func__, error));
685 	baddr = arg;
686 	*baddr = segs->ds_addr;
687 }
688 
689 /*
690  * Interrupt setup and handlers
691  */
692 static int
693 ioat_setup_intr(struct ioat_softc *ioat)
694 {
695 	uint32_t num_vectors;
696 	int error;
697 	boolean_t use_msix;
698 	boolean_t force_legacy_interrupts;
699 
700 	use_msix = FALSE;
701 	force_legacy_interrupts = FALSE;
702 
703 	if (!g_force_legacy_interrupts && pci_msix_count(ioat->device) >= 1) {
704 		num_vectors = 1;
705 		pci_alloc_msix(ioat->device, &num_vectors);
706 		if (num_vectors == 1)
707 			use_msix = TRUE;
708 	}
709 
710 	if (use_msix) {
711 		ioat->rid = 1;
712 		ioat->res = bus_alloc_resource_any(ioat->device, SYS_RES_IRQ,
713 		    &ioat->rid, RF_ACTIVE);
714 	} else {
715 		ioat->rid = 0;
716 		ioat->res = bus_alloc_resource_any(ioat->device, SYS_RES_IRQ,
717 		    &ioat->rid, RF_SHAREABLE | RF_ACTIVE);
718 	}
719 	if (ioat->res == NULL) {
720 		ioat_log_message(0, "bus_alloc_resource failed\n");
721 		return (ENOMEM);
722 	}
723 
724 	ioat->tag = NULL;
725 	error = bus_setup_intr(ioat->device, ioat->res, INTR_MPSAFE |
726 	    INTR_TYPE_MISC, NULL, ioat_interrupt_handler, ioat, &ioat->tag);
727 	if (error != 0) {
728 		ioat_log_message(0, "bus_setup_intr failed\n");
729 		return (error);
730 	}
731 
732 	ioat_write_intrctrl(ioat, IOAT_INTRCTRL_MASTER_INT_EN);
733 	return (0);
734 }
735 
736 static boolean_t
737 ioat_model_resets_msix(struct ioat_softc *ioat)
738 {
739 	u_int32_t pciid;
740 
741 	pciid = pci_get_devid(ioat->device);
742 	switch (pciid) {
743 		/* BWD: */
744 	case 0x0c508086:
745 	case 0x0c518086:
746 	case 0x0c528086:
747 	case 0x0c538086:
748 		/* BDXDE: */
749 	case 0x6f508086:
750 	case 0x6f518086:
751 	case 0x6f528086:
752 	case 0x6f538086:
753 		return (TRUE);
754 	}
755 
756 	return (FALSE);
757 }
758 
759 static void
760 ioat_interrupt_handler(void *arg)
761 {
762 	struct ioat_softc *ioat = arg;
763 
764 	ioat->stats.interrupts++;
765 	ioat_process_events(ioat, TRUE);
766 }
767 
768 static int
769 chanerr_to_errno(uint32_t chanerr)
770 {
771 
772 	if (chanerr == 0)
773 		return (0);
774 	if ((chanerr & (IOAT_CHANERR_XSADDERR | IOAT_CHANERR_XDADDERR)) != 0)
775 		return (EFAULT);
776 	if ((chanerr & (IOAT_CHANERR_RDERR | IOAT_CHANERR_WDERR)) != 0)
777 		return (EIO);
778 	/* This one is probably our fault: */
779 	if ((chanerr & IOAT_CHANERR_NDADDERR) != 0)
780 		return (EIO);
781 	return (EIO);
782 }
783 
784 static void
785 ioat_process_events(struct ioat_softc *ioat, boolean_t intr)
786 {
787 	struct ioat_descriptor *desc;
788 	struct bus_dmadesc *dmadesc;
789 	uint64_t comp_update, status;
790 	uint32_t completed, chanerr;
791 	int error;
792 
793 	if (intr) {
794 		mtx_lock(&ioat->cleanup_lock);
795 	} else {
796 		if (!mtx_trylock(&ioat->cleanup_lock))
797 			return;
798 	}
799 
800 	/*
801 	 * Don't run while the hardware is being reset.  Reset is responsible
802 	 * for blocking new work and draining & completing existing work, so
803 	 * there is nothing to do until new work is queued after reset anyway.
804 	 */
805 	if (ioat->resetting_cleanup) {
806 		mtx_unlock(&ioat->cleanup_lock);
807 		return;
808 	}
809 
810 	completed = 0;
811 	comp_update = *ioat->comp_update;
812 	status = comp_update & IOAT_CHANSTS_COMPLETED_DESCRIPTOR_MASK;
813 
814 	if (status < ioat->hw_desc_bus_addr ||
815 	    status >= ioat->hw_desc_bus_addr + (1 << ioat->ring_size_order) *
816 	    sizeof(struct ioat_generic_hw_descriptor))
817 		panic("Bogus completion address %jx (channel %u)",
818 		    (uintmax_t)status, ioat->chan_idx);
819 
820 	if (status == ioat->last_seen) {
821 		/*
822 		 * If we landed in process_events and nothing has been
823 		 * completed, check for a timeout due to channel halt.
824 		 */
825 		goto out;
826 	}
827 	CTR4(KTR_IOAT, "%s channel=%u hw_status=0x%lx last_seen=0x%lx",
828 	    __func__, ioat->chan_idx, comp_update, ioat->last_seen);
829 
830 	while (RING_PHYS_ADDR(ioat, ioat->tail - 1) != status) {
831 		desc = ioat_get_ring_entry(ioat, ioat->tail);
832 		dmadesc = &desc->bus_dmadesc;
833 		CTR5(KTR_IOAT, "channel=%u completing desc idx %u (%p) ok  cb %p(%p)",
834 		    ioat->chan_idx, ioat->tail, dmadesc, dmadesc->callback_fn,
835 		    dmadesc->callback_arg);
836 
837 		bus_dmamap_unload(ioat->data_tag, desc->src_dmamap);
838 		bus_dmamap_unload(ioat->data_tag, desc->dst_dmamap);
839 		bus_dmamap_unload(ioat->data_tag, desc->src2_dmamap);
840 		bus_dmamap_unload(ioat->data_tag, desc->dst2_dmamap);
841 
842 		if (dmadesc->callback_fn != NULL)
843 			dmadesc->callback_fn(dmadesc->callback_arg, 0);
844 
845 		completed++;
846 		ioat->tail++;
847 	}
848 	CTR5(KTR_IOAT, "%s channel=%u head=%u tail=%u active=%u", __func__,
849 	    ioat->chan_idx, ioat->head, ioat->tail, ioat_get_active(ioat));
850 
851 	if (completed != 0) {
852 		ioat->last_seen = RING_PHYS_ADDR(ioat, ioat->tail - 1);
853 		ioat->stats.descriptors_processed += completed;
854 		wakeup(&ioat->tail);
855 	}
856 
857 out:
858 	ioat_write_chanctrl(ioat, IOAT_CHANCTRL_RUN);
859 	mtx_unlock(&ioat->cleanup_lock);
860 
861 	/*
862 	 * The device doesn't seem to reliably push suspend/halt statuses to
863 	 * the channel completion memory address, so poll the device register
864 	 * here.  For performance reasons skip it on interrupts, do it only
865 	 * on much more rare polling events.
866 	 */
867 	if (!intr)
868 		comp_update = ioat_get_chansts(ioat) & IOAT_CHANSTS_STATUS;
869 	if (!is_ioat_halted(comp_update) && !is_ioat_suspended(comp_update))
870 		return;
871 
872 	ioat->stats.channel_halts++;
873 
874 	/*
875 	 * Fatal programming error on this DMA channel.  Flush any outstanding
876 	 * work with error status and restart the engine.
877 	 */
878 	mtx_lock(&ioat->submit_lock);
879 	ioat->quiescing = TRUE;
880 	mtx_unlock(&ioat->submit_lock);
881 
882 	/*
883 	 * This is safe to do here because the submit queue is quiesced.  We
884 	 * know that we will drain all outstanding events, so ioat_reset_hw
885 	 * can't deadlock. It is necessary to protect other ioat_process_event
886 	 * threads from racing ioat_reset_hw, reading an indeterminate hw
887 	 * state, and attempting to continue issuing completions.
888 	 */
889 	mtx_lock(&ioat->cleanup_lock);
890 	ioat->resetting_cleanup = TRUE;
891 
892 	chanerr = ioat_read_4(ioat, IOAT_CHANERR_OFFSET);
893 	if (1 <= g_ioat_debug_level)
894 		ioat_halted_debug(ioat, chanerr);
895 	ioat->stats.last_halt_chanerr = chanerr;
896 
897 	while (ioat_get_active(ioat) > 0) {
898 		desc = ioat_get_ring_entry(ioat, ioat->tail);
899 		dmadesc = &desc->bus_dmadesc;
900 		CTR5(KTR_IOAT, "channel=%u completing desc idx %u (%p) err cb %p(%p)",
901 		    ioat->chan_idx, ioat->tail, dmadesc, dmadesc->callback_fn,
902 		    dmadesc->callback_arg);
903 
904 		if (dmadesc->callback_fn != NULL)
905 			dmadesc->callback_fn(dmadesc->callback_arg,
906 			    chanerr_to_errno(chanerr));
907 
908 		ioat->tail++;
909 		ioat->stats.descriptors_processed++;
910 		ioat->stats.descriptors_error++;
911 	}
912 	CTR5(KTR_IOAT, "%s channel=%u head=%u tail=%u active=%u", __func__,
913 	    ioat->chan_idx, ioat->head, ioat->tail, ioat_get_active(ioat));
914 
915 	/* Clear error status */
916 	ioat_write_4(ioat, IOAT_CHANERR_OFFSET, chanerr);
917 
918 	mtx_unlock(&ioat->cleanup_lock);
919 
920 	ioat_log_message(0, "Resetting channel to recover from error\n");
921 	error = taskqueue_enqueue(taskqueue_thread, &ioat->reset_task);
922 	KASSERT(error == 0,
923 	    ("%s: taskqueue_enqueue failed: %d", __func__, error));
924 }
925 
926 static void
927 ioat_reset_hw_task(void *ctx, int pending __unused)
928 {
929 	struct ioat_softc *ioat;
930 	int error;
931 
932 	ioat = ctx;
933 	ioat_log_message(1, "%s: Resetting channel\n", __func__);
934 
935 	error = ioat_reset_hw(ioat);
936 	KASSERT(error == 0, ("%s: reset failed: %d", __func__, error));
937 	(void)error;
938 }
939 
940 /*
941  * User API functions
942  */
943 unsigned
944 ioat_get_nchannels(void)
945 {
946 
947 	return (ioat_channel_index);
948 }
949 
950 bus_dmaengine_t
951 ioat_get_dmaengine(uint32_t index, int flags)
952 {
953 	struct ioat_softc *ioat;
954 
955 	KASSERT((flags & ~(M_NOWAIT | M_WAITOK)) == 0,
956 	    ("invalid flags: 0x%08x", flags));
957 	KASSERT((flags & (M_NOWAIT | M_WAITOK)) != (M_NOWAIT | M_WAITOK),
958 	    ("invalid wait | nowait"));
959 
960 	mtx_lock(&ioat_list_mtx);
961 	if (index >= ioat_channel_index ||
962 	    (ioat = ioat_channel[index]) == NULL) {
963 		mtx_unlock(&ioat_list_mtx);
964 		return (NULL);
965 	}
966 	mtx_lock(&ioat->submit_lock);
967 	mtx_unlock(&ioat_list_mtx);
968 
969 	if (ioat->destroying) {
970 		mtx_unlock(&ioat->submit_lock);
971 		return (NULL);
972 	}
973 
974 	ioat_get(ioat);
975 	if (ioat->quiescing) {
976 		if ((flags & M_NOWAIT) != 0) {
977 			ioat_put(ioat);
978 			mtx_unlock(&ioat->submit_lock);
979 			return (NULL);
980 		}
981 
982 		while (ioat->quiescing && !ioat->destroying)
983 			msleep(&ioat->quiescing, &ioat->submit_lock, 0, "getdma", 0);
984 
985 		if (ioat->destroying) {
986 			ioat_put(ioat);
987 			mtx_unlock(&ioat->submit_lock);
988 			return (NULL);
989 		}
990 	}
991 	mtx_unlock(&ioat->submit_lock);
992 	return (&ioat->dmaengine);
993 }
994 
995 void
996 ioat_put_dmaengine(bus_dmaengine_t dmaengine)
997 {
998 	struct ioat_softc *ioat;
999 
1000 	ioat = to_ioat_softc(dmaengine);
1001 	mtx_lock(&ioat->submit_lock);
1002 	ioat_put(ioat);
1003 	mtx_unlock(&ioat->submit_lock);
1004 }
1005 
1006 int
1007 ioat_get_hwversion(bus_dmaengine_t dmaengine)
1008 {
1009 	struct ioat_softc *ioat;
1010 
1011 	ioat = to_ioat_softc(dmaengine);
1012 	return (ioat->version);
1013 }
1014 
1015 size_t
1016 ioat_get_max_io_size(bus_dmaengine_t dmaengine)
1017 {
1018 	struct ioat_softc *ioat;
1019 
1020 	ioat = to_ioat_softc(dmaengine);
1021 	return (ioat->max_xfer_size);
1022 }
1023 
1024 uint32_t
1025 ioat_get_capabilities(bus_dmaengine_t dmaengine)
1026 {
1027 	struct ioat_softc *ioat;
1028 
1029 	ioat = to_ioat_softc(dmaengine);
1030 	return (ioat->capabilities);
1031 }
1032 
1033 int
1034 ioat_get_domain(bus_dmaengine_t dmaengine, int *domain)
1035 {
1036 	struct ioat_softc *ioat;
1037 
1038 	ioat = to_ioat_softc(dmaengine);
1039 	return (bus_get_domain(ioat->device, domain));
1040 }
1041 
1042 int
1043 ioat_set_interrupt_coalesce(bus_dmaengine_t dmaengine, uint16_t delay)
1044 {
1045 	struct ioat_softc *ioat;
1046 
1047 	ioat = to_ioat_softc(dmaengine);
1048 	if (!ioat->intrdelay_supported)
1049 		return (ENODEV);
1050 	if (delay > ioat->intrdelay_max)
1051 		return (ERANGE);
1052 
1053 	ioat_write_2(ioat, IOAT_INTRDELAY_OFFSET, delay);
1054 	ioat->cached_intrdelay =
1055 	    ioat_read_2(ioat, IOAT_INTRDELAY_OFFSET) & IOAT_INTRDELAY_US_MASK;
1056 	return (0);
1057 }
1058 
1059 uint16_t
1060 ioat_get_max_coalesce_period(bus_dmaengine_t dmaengine)
1061 {
1062 	struct ioat_softc *ioat;
1063 
1064 	ioat = to_ioat_softc(dmaengine);
1065 	return (ioat->intrdelay_max);
1066 }
1067 
1068 void
1069 ioat_acquire(bus_dmaengine_t dmaengine)
1070 {
1071 	struct ioat_softc *ioat;
1072 
1073 	ioat = to_ioat_softc(dmaengine);
1074 	mtx_lock(&ioat->submit_lock);
1075 	CTR2(KTR_IOAT, "%s channel=%u", __func__, ioat->chan_idx);
1076 	ioat->acq_head = ioat->head;
1077 }
1078 
1079 int
1080 ioat_acquire_reserve(bus_dmaengine_t dmaengine, unsigned n, int mflags)
1081 {
1082 	struct ioat_softc *ioat;
1083 	int error;
1084 
1085 	ioat = to_ioat_softc(dmaengine);
1086 	ioat_acquire(dmaengine);
1087 
1088 	error = ioat_reserve_space(ioat, n, mflags);
1089 	if (error != 0)
1090 		ioat_release(dmaengine);
1091 	return (error);
1092 }
1093 
1094 void
1095 ioat_release(bus_dmaengine_t dmaengine)
1096 {
1097 	struct ioat_softc *ioat;
1098 
1099 	ioat = to_ioat_softc(dmaengine);
1100 	CTR3(KTR_IOAT, "%s channel=%u dispatch1 head=%u", __func__,
1101 	    ioat->chan_idx, ioat->head);
1102 	KFAIL_POINT_CODE(DEBUG_FP, ioat_release, /* do nothing */);
1103 	CTR3(KTR_IOAT, "%s channel=%u dispatch2 head=%u", __func__,
1104 	    ioat->chan_idx, ioat->head);
1105 
1106 	if (ioat->acq_head != ioat->head) {
1107 		ioat_write_2(ioat, IOAT_DMACOUNT_OFFSET,
1108 		    (uint16_t)ioat->head);
1109 
1110 		if (!callout_pending(&ioat->poll_timer)) {
1111 			callout_reset_on(&ioat->poll_timer, 1,
1112 			    ioat_poll_timer_callback, ioat, ioat->cpu);
1113 		}
1114 	}
1115 	mtx_unlock(&ioat->submit_lock);
1116 }
1117 
1118 static struct ioat_descriptor *
1119 ioat_op_generic(struct ioat_softc *ioat, uint8_t op,
1120     uint32_t size, uint64_t src, uint64_t dst,
1121     bus_dmaengine_callback_t callback_fn, void *callback_arg,
1122     uint32_t flags)
1123 {
1124 	struct ioat_generic_hw_descriptor *hw_desc;
1125 	struct ioat_descriptor *desc;
1126 	bus_dma_segment_t seg;
1127 	int mflags, nseg, error;
1128 
1129 	mtx_assert(&ioat->submit_lock, MA_OWNED);
1130 
1131 	KASSERT((flags & ~_DMA_GENERIC_FLAGS) == 0,
1132 	    ("Unrecognized flag(s): %#x", flags & ~_DMA_GENERIC_FLAGS));
1133 	KASSERT(size <= ioat->max_xfer_size, ("%s: size too big (%u > %u)",
1134 	    __func__, (unsigned)size, ioat->max_xfer_size));
1135 
1136 	if ((flags & DMA_NO_WAIT) != 0)
1137 		mflags = M_NOWAIT;
1138 	else
1139 		mflags = M_WAITOK;
1140 
1141 	if (ioat_reserve_space(ioat, 1, mflags) != 0)
1142 		return (NULL);
1143 
1144 	desc = ioat_get_ring_entry(ioat, ioat->head);
1145 	hw_desc = &ioat_get_descriptor(ioat, ioat->head)->generic;
1146 
1147 	hw_desc->u.control_raw = 0;
1148 	hw_desc->u.control_generic.op = op;
1149 	hw_desc->u.control_generic.completion_update = 1;
1150 
1151 	if ((flags & DMA_INT_EN) != 0)
1152 		hw_desc->u.control_generic.int_enable = 1;
1153 	if ((flags & DMA_FENCE) != 0)
1154 		hw_desc->u.control_generic.fence = 1;
1155 
1156 	hw_desc->size = size;
1157 
1158 	if (src != 0) {
1159 		nseg = -1;
1160 		error = _bus_dmamap_load_phys(ioat->data_tag, desc->src_dmamap,
1161 		    src, size, 0, &seg, &nseg);
1162 		if (error != 0) {
1163 			ioat_log_message(0, "%s: _bus_dmamap_load_phys"
1164 			    " failed %d\n", __func__, error);
1165 			return (NULL);
1166 		}
1167 		hw_desc->src_addr = seg.ds_addr;
1168 	}
1169 
1170 	if (dst != 0) {
1171 		nseg = -1;
1172 		error = _bus_dmamap_load_phys(ioat->data_tag, desc->dst_dmamap,
1173 		    dst, size, 0, &seg, &nseg);
1174 		if (error != 0) {
1175 			ioat_log_message(0, "%s: _bus_dmamap_load_phys"
1176 			    " failed %d\n", __func__, error);
1177 			return (NULL);
1178 		}
1179 		hw_desc->dest_addr = seg.ds_addr;
1180 	}
1181 
1182 	desc->bus_dmadesc.callback_fn = callback_fn;
1183 	desc->bus_dmadesc.callback_arg = callback_arg;
1184 	return (desc);
1185 }
1186 
1187 struct bus_dmadesc *
1188 ioat_null(bus_dmaengine_t dmaengine, bus_dmaengine_callback_t callback_fn,
1189     void *callback_arg, uint32_t flags)
1190 {
1191 	struct ioat_dma_hw_descriptor *hw_desc;
1192 	struct ioat_descriptor *desc;
1193 	struct ioat_softc *ioat;
1194 
1195 	ioat = to_ioat_softc(dmaengine);
1196 	CTR2(KTR_IOAT, "%s channel=%u", __func__, ioat->chan_idx);
1197 
1198 	desc = ioat_op_generic(ioat, IOAT_OP_COPY, 8, 0, 0, callback_fn,
1199 	    callback_arg, flags);
1200 	if (desc == NULL)
1201 		return (NULL);
1202 
1203 	hw_desc = &ioat_get_descriptor(ioat, desc->id)->dma;
1204 	hw_desc->u.control.null = 1;
1205 	ioat_submit_single(ioat);
1206 	return (&desc->bus_dmadesc);
1207 }
1208 
1209 struct bus_dmadesc *
1210 ioat_copy(bus_dmaengine_t dmaengine, bus_addr_t dst,
1211     bus_addr_t src, bus_size_t len, bus_dmaengine_callback_t callback_fn,
1212     void *callback_arg, uint32_t flags)
1213 {
1214 	struct ioat_dma_hw_descriptor *hw_desc;
1215 	struct ioat_descriptor *desc;
1216 	struct ioat_softc *ioat;
1217 
1218 	ioat = to_ioat_softc(dmaengine);
1219 	desc = ioat_op_generic(ioat, IOAT_OP_COPY, len, src, dst, callback_fn,
1220 	    callback_arg, flags);
1221 	if (desc == NULL)
1222 		return (NULL);
1223 
1224 	hw_desc = &ioat_get_descriptor(ioat, desc->id)->dma;
1225 	if (g_ioat_debug_level >= 3)
1226 		dump_descriptor(hw_desc);
1227 
1228 	ioat_submit_single(ioat);
1229 	CTR6(KTR_IOAT, "%s channel=%u desc=%p dest=%lx src=%lx len=%lx",
1230 	    __func__, ioat->chan_idx, &desc->bus_dmadesc, dst, src, len);
1231 	return (&desc->bus_dmadesc);
1232 }
1233 
1234 struct bus_dmadesc *
1235 ioat_copy_8k_aligned(bus_dmaengine_t dmaengine, bus_addr_t dst1,
1236     bus_addr_t dst2, bus_addr_t src1, bus_addr_t src2,
1237     bus_dmaengine_callback_t callback_fn, void *callback_arg, uint32_t flags)
1238 {
1239 	struct ioat_dma_hw_descriptor *hw_desc;
1240 	struct ioat_descriptor *desc;
1241 	struct ioat_softc *ioat;
1242 	bus_size_t src1_len, dst1_len;
1243 	bus_dma_segment_t seg;
1244 	int nseg, error;
1245 
1246 	ioat = to_ioat_softc(dmaengine);
1247 	CTR2(KTR_IOAT, "%s channel=%u", __func__, ioat->chan_idx);
1248 
1249 	KASSERT(((src1 | src2 | dst1 | dst2) & PAGE_MASK) == 0,
1250 	    ("%s: addresses are not page-aligned", __func__));
1251 
1252 	desc = ioat_op_generic(ioat, IOAT_OP_COPY, 2 * PAGE_SIZE, 0, 0,
1253 	    callback_fn, callback_arg, flags);
1254 	if (desc == NULL)
1255 		return (NULL);
1256 
1257 	hw_desc = &ioat_get_descriptor(ioat, desc->id)->dma;
1258 
1259 	src1_len = (src2 != src1 + PAGE_SIZE) ? PAGE_SIZE : 2 * PAGE_SIZE;
1260 	nseg = -1;
1261 	error = _bus_dmamap_load_phys(ioat->data_tag,
1262 	    desc->src_dmamap, src1, src1_len, 0, &seg, &nseg);
1263 	if (error != 0) {
1264 		ioat_log_message(0, "%s: _bus_dmamap_load_phys"
1265 		    " failed %d\n", __func__, error);
1266 		return (NULL);
1267 	}
1268 	hw_desc->src_addr = seg.ds_addr;
1269 	if (src1_len != 2 * PAGE_SIZE) {
1270 		hw_desc->u.control.src_page_break = 1;
1271 		nseg = -1;
1272 		error = _bus_dmamap_load_phys(ioat->data_tag,
1273 		    desc->src2_dmamap, src2, PAGE_SIZE, 0, &seg, &nseg);
1274 		if (error != 0) {
1275 			ioat_log_message(0, "%s: _bus_dmamap_load_phys"
1276 			    " failed %d\n", __func__, error);
1277 			return (NULL);
1278 		}
1279 		hw_desc->next_src_addr = seg.ds_addr;
1280 	}
1281 
1282 	dst1_len = (dst2 != dst1 + PAGE_SIZE) ? PAGE_SIZE : 2 * PAGE_SIZE;
1283 	nseg = -1;
1284 	error = _bus_dmamap_load_phys(ioat->data_tag,
1285 	    desc->dst_dmamap, dst1, dst1_len, 0, &seg, &nseg);
1286 	if (error != 0) {
1287 		ioat_log_message(0, "%s: _bus_dmamap_load_phys"
1288 		    " failed %d\n", __func__, error);
1289 		return (NULL);
1290 	}
1291 	hw_desc->dest_addr = seg.ds_addr;
1292 	if (dst1_len != 2 * PAGE_SIZE) {
1293 		hw_desc->u.control.dest_page_break = 1;
1294 		nseg = -1;
1295 		error = _bus_dmamap_load_phys(ioat->data_tag,
1296 		    desc->dst2_dmamap, dst2, PAGE_SIZE, 0, &seg, &nseg);
1297 		if (error != 0) {
1298 			ioat_log_message(0, "%s: _bus_dmamap_load_phys"
1299 			    " failed %d\n", __func__, error);
1300 			return (NULL);
1301 		}
1302 		hw_desc->next_dest_addr = seg.ds_addr;
1303 	}
1304 
1305 	if (g_ioat_debug_level >= 3)
1306 		dump_descriptor(hw_desc);
1307 
1308 	ioat_submit_single(ioat);
1309 	return (&desc->bus_dmadesc);
1310 }
1311 
1312 struct bus_dmadesc *
1313 ioat_copy_crc(bus_dmaengine_t dmaengine, bus_addr_t dst, bus_addr_t src,
1314     bus_size_t len, uint32_t *initialseed, bus_addr_t crcptr,
1315     bus_dmaengine_callback_t callback_fn, void *callback_arg, uint32_t flags)
1316 {
1317 	struct ioat_crc32_hw_descriptor *hw_desc;
1318 	struct ioat_descriptor *desc;
1319 	struct ioat_softc *ioat;
1320 	uint32_t teststore;
1321 	uint8_t op;
1322 	bus_dma_segment_t seg;
1323 	int nseg, error;
1324 
1325 	ioat = to_ioat_softc(dmaengine);
1326 	CTR2(KTR_IOAT, "%s channel=%u", __func__, ioat->chan_idx);
1327 
1328 	KASSERT((ioat->capabilities & IOAT_DMACAP_MOVECRC) != 0,
1329 	    ("%s: device lacks MOVECRC capability", __func__));
1330 	teststore = (flags & _DMA_CRC_TESTSTORE);
1331 	KASSERT(teststore != _DMA_CRC_TESTSTORE,
1332 	    ("%s: TEST and STORE invalid", __func__));
1333 	KASSERT(teststore != 0 || (flags & DMA_CRC_INLINE) == 0,
1334 	    ("%s: INLINE invalid without TEST or STORE", __func__));
1335 
1336 	switch (teststore) {
1337 	case DMA_CRC_STORE:
1338 		op = IOAT_OP_MOVECRC_STORE;
1339 		break;
1340 	case DMA_CRC_TEST:
1341 		op = IOAT_OP_MOVECRC_TEST;
1342 		break;
1343 	default:
1344 		KASSERT(teststore == 0, ("bogus"));
1345 		op = IOAT_OP_MOVECRC;
1346 		break;
1347 	}
1348 
1349 	desc = ioat_op_generic(ioat, op, len, src, dst, callback_fn,
1350 	    callback_arg, flags & ~_DMA_CRC_FLAGS);
1351 	if (desc == NULL)
1352 		return (NULL);
1353 
1354 	hw_desc = &ioat_get_descriptor(ioat, desc->id)->crc32;
1355 
1356 	if ((flags & DMA_CRC_INLINE) == 0) {
1357 		nseg = -1;
1358 		error = _bus_dmamap_load_phys(ioat->data_tag,
1359 		    desc->dst2_dmamap, crcptr, sizeof(uint32_t), 0,
1360 		    &seg, &nseg);
1361 		if (error != 0) {
1362 			ioat_log_message(0, "%s: _bus_dmamap_load_phys"
1363 			    " failed %d\n", __func__, error);
1364 			return (NULL);
1365 		}
1366 		hw_desc->crc_address = seg.ds_addr;
1367 	} else
1368 		hw_desc->u.control.crc_location = 1;
1369 
1370 	if (initialseed != NULL) {
1371 		hw_desc->u.control.use_seed = 1;
1372 		hw_desc->seed = *initialseed;
1373 	}
1374 
1375 	if (g_ioat_debug_level >= 3)
1376 		dump_descriptor(hw_desc);
1377 
1378 	ioat_submit_single(ioat);
1379 	return (&desc->bus_dmadesc);
1380 }
1381 
1382 struct bus_dmadesc *
1383 ioat_crc(bus_dmaengine_t dmaengine, bus_addr_t src, bus_size_t len,
1384     uint32_t *initialseed, bus_addr_t crcptr,
1385     bus_dmaengine_callback_t callback_fn, void *callback_arg, uint32_t flags)
1386 {
1387 	struct ioat_crc32_hw_descriptor *hw_desc;
1388 	struct ioat_descriptor *desc;
1389 	struct ioat_softc *ioat;
1390 	uint32_t teststore;
1391 	uint8_t op;
1392 	bus_dma_segment_t seg;
1393 	int nseg, error;
1394 
1395 	ioat = to_ioat_softc(dmaengine);
1396 	CTR2(KTR_IOAT, "%s channel=%u", __func__, ioat->chan_idx);
1397 
1398 	KASSERT((ioat->capabilities & IOAT_DMACAP_CRC) != 0,
1399 	    ("%s: device lacks CRC capability", __func__));
1400 	teststore = (flags & _DMA_CRC_TESTSTORE);
1401 	KASSERT(teststore != _DMA_CRC_TESTSTORE,
1402 	    ("%s: TEST and STORE invalid", __func__));
1403 	KASSERT(teststore != 0 || (flags & DMA_CRC_INLINE) == 0,
1404 	    ("%s: INLINE invalid without TEST or STORE", __func__));
1405 
1406 	switch (teststore) {
1407 	case DMA_CRC_STORE:
1408 		op = IOAT_OP_CRC_STORE;
1409 		break;
1410 	case DMA_CRC_TEST:
1411 		op = IOAT_OP_CRC_TEST;
1412 		break;
1413 	default:
1414 		KASSERT(teststore == 0, ("bogus"));
1415 		op = IOAT_OP_CRC;
1416 		break;
1417 	}
1418 
1419 	desc = ioat_op_generic(ioat, op, len, src, 0, callback_fn,
1420 	    callback_arg, flags & ~_DMA_CRC_FLAGS);
1421 	if (desc == NULL)
1422 		return (NULL);
1423 
1424 	hw_desc = &ioat_get_descriptor(ioat, desc->id)->crc32;
1425 
1426 	if ((flags & DMA_CRC_INLINE) == 0) {
1427 		nseg = -1;
1428 		error = _bus_dmamap_load_phys(ioat->data_tag,
1429 		    desc->dst2_dmamap, crcptr, sizeof(uint32_t), 0,
1430 		    &seg, &nseg);
1431 		if (error != 0) {
1432 			ioat_log_message(0, "%s: _bus_dmamap_load_phys"
1433 			    " failed %d\n", __func__, error);
1434 			return (NULL);
1435 		}
1436 		hw_desc->crc_address = seg.ds_addr;
1437 	} else
1438 		hw_desc->u.control.crc_location = 1;
1439 
1440 	if (initialseed != NULL) {
1441 		hw_desc->u.control.use_seed = 1;
1442 		hw_desc->seed = *initialseed;
1443 	}
1444 
1445 	if (g_ioat_debug_level >= 3)
1446 		dump_descriptor(hw_desc);
1447 
1448 	ioat_submit_single(ioat);
1449 	return (&desc->bus_dmadesc);
1450 }
1451 
1452 struct bus_dmadesc *
1453 ioat_blockfill(bus_dmaengine_t dmaengine, bus_addr_t dst, uint64_t fillpattern,
1454     bus_size_t len, bus_dmaengine_callback_t callback_fn, void *callback_arg,
1455     uint32_t flags)
1456 {
1457 	struct ioat_fill_hw_descriptor *hw_desc;
1458 	struct ioat_descriptor *desc;
1459 	struct ioat_softc *ioat;
1460 
1461 	ioat = to_ioat_softc(dmaengine);
1462 	CTR2(KTR_IOAT, "%s channel=%u", __func__, ioat->chan_idx);
1463 
1464 	KASSERT((ioat->capabilities & IOAT_DMACAP_BFILL) != 0,
1465 	    ("%s: device lacks BFILL capability", __func__));
1466 
1467 	desc = ioat_op_generic(ioat, IOAT_OP_FILL, len, 0, dst,
1468 	    callback_fn, callback_arg, flags);
1469 	if (desc == NULL)
1470 		return (NULL);
1471 
1472 	hw_desc = &ioat_get_descriptor(ioat, desc->id)->fill;
1473 	hw_desc->src_data = fillpattern;
1474 	if (g_ioat_debug_level >= 3)
1475 		dump_descriptor(hw_desc);
1476 
1477 	ioat_submit_single(ioat);
1478 	return (&desc->bus_dmadesc);
1479 }
1480 
1481 /*
1482  * Ring Management
1483  */
1484 static inline uint32_t
1485 ioat_get_active(struct ioat_softc *ioat)
1486 {
1487 
1488 	return ((ioat->head - ioat->tail) & ((1 << ioat->ring_size_order) - 1));
1489 }
1490 
1491 static inline uint32_t
1492 ioat_get_ring_space(struct ioat_softc *ioat)
1493 {
1494 
1495 	return ((1 << ioat->ring_size_order) - ioat_get_active(ioat) - 1);
1496 }
1497 
1498 /*
1499  * Reserves space in this IOAT descriptor ring by ensuring enough slots remain
1500  * for 'num_descs'.
1501  *
1502  * If mflags contains M_WAITOK, blocks until enough space is available.
1503  *
1504  * Returns zero on success, or an errno on error.  If num_descs is beyond the
1505  * maximum ring size, returns EINVAl; if allocation would block and mflags
1506  * contains M_NOWAIT, returns EAGAIN.
1507  *
1508  * Must be called with the submit_lock held; returns with the lock held.  The
1509  * lock may be dropped to allocate the ring.
1510  *
1511  * (The submit_lock is needed to add any entries to the ring, so callers are
1512  * assured enough room is available.)
1513  */
1514 static int
1515 ioat_reserve_space(struct ioat_softc *ioat, uint32_t num_descs, int mflags)
1516 {
1517 	boolean_t dug;
1518 	int error;
1519 
1520 	mtx_assert(&ioat->submit_lock, MA_OWNED);
1521 	error = 0;
1522 	dug = FALSE;
1523 
1524 	if (num_descs < 1 || num_descs >= (1 << ioat->ring_size_order)) {
1525 		error = EINVAL;
1526 		goto out;
1527 	}
1528 
1529 	for (;;) {
1530 		if (ioat->quiescing) {
1531 			error = ENXIO;
1532 			goto out;
1533 		}
1534 
1535 		if (ioat_get_ring_space(ioat) >= num_descs)
1536 			goto out;
1537 
1538 		CTR3(KTR_IOAT, "%s channel=%u starved (%u)", __func__,
1539 		    ioat->chan_idx, num_descs);
1540 
1541 		if (!dug && !ioat->is_submitter_processing) {
1542 			ioat->is_submitter_processing = TRUE;
1543 			mtx_unlock(&ioat->submit_lock);
1544 
1545 			CTR2(KTR_IOAT, "%s channel=%u attempting to process events",
1546 			    __func__, ioat->chan_idx);
1547 			ioat_process_events(ioat, FALSE);
1548 
1549 			mtx_lock(&ioat->submit_lock);
1550 			dug = TRUE;
1551 			KASSERT(ioat->is_submitter_processing == TRUE,
1552 			    ("is_submitter_processing"));
1553 			ioat->is_submitter_processing = FALSE;
1554 			wakeup(&ioat->tail);
1555 			continue;
1556 		}
1557 
1558 		if ((mflags & M_WAITOK) == 0) {
1559 			error = EAGAIN;
1560 			break;
1561 		}
1562 		CTR2(KTR_IOAT, "%s channel=%u blocking on completions",
1563 		    __func__, ioat->chan_idx);
1564 		msleep(&ioat->tail, &ioat->submit_lock, 0,
1565 		    "ioat_full", 0);
1566 		continue;
1567 	}
1568 
1569 out:
1570 	mtx_assert(&ioat->submit_lock, MA_OWNED);
1571 	KASSERT(!ioat->quiescing || error == ENXIO,
1572 	    ("reserved during quiesce"));
1573 	return (error);
1574 }
1575 
1576 static void
1577 ioat_free_ring(struct ioat_softc *ioat, uint32_t size,
1578     struct ioat_descriptor *ring)
1579 {
1580 
1581 	free_domain(ring, M_IOAT);
1582 }
1583 
1584 static struct ioat_descriptor *
1585 ioat_get_ring_entry(struct ioat_softc *ioat, uint32_t index)
1586 {
1587 
1588 	return (&ioat->ring[index % (1 << ioat->ring_size_order)]);
1589 }
1590 
1591 static union ioat_hw_descriptor *
1592 ioat_get_descriptor(struct ioat_softc *ioat, uint32_t index)
1593 {
1594 
1595 	return (&ioat->hw_desc_ring[index % (1 << ioat->ring_size_order)]);
1596 }
1597 
1598 static void
1599 ioat_halted_debug(struct ioat_softc *ioat, uint32_t chanerr)
1600 {
1601 	union ioat_hw_descriptor *desc;
1602 
1603 	ioat_log_message(0, "Channel halted (%b)\n", (int)chanerr,
1604 	    IOAT_CHANERR_STR);
1605 	if (chanerr == 0)
1606 		return;
1607 
1608 	mtx_assert(&ioat->cleanup_lock, MA_OWNED);
1609 
1610 	desc = ioat_get_descriptor(ioat, ioat->tail + 0);
1611 	dump_descriptor(desc);
1612 
1613 	desc = ioat_get_descriptor(ioat, ioat->tail + 1);
1614 	dump_descriptor(desc);
1615 }
1616 
1617 static void
1618 ioat_poll_timer_callback(void *arg)
1619 {
1620 	struct ioat_softc *ioat;
1621 
1622 	ioat = arg;
1623 	CTR1(KTR_IOAT, "%s", __func__);
1624 
1625 	ioat_process_events(ioat, FALSE);
1626 
1627 	mtx_lock(&ioat->submit_lock);
1628 	if (ioat_get_active(ioat) > 0)
1629 		callout_schedule(&ioat->poll_timer, 1);
1630 	mtx_unlock(&ioat->submit_lock);
1631 }
1632 
1633 /*
1634  * Support Functions
1635  */
1636 static void
1637 ioat_submit_single(struct ioat_softc *ioat)
1638 {
1639 
1640 	mtx_assert(&ioat->submit_lock, MA_OWNED);
1641 
1642 	ioat->head++;
1643 	CTR4(KTR_IOAT, "%s channel=%u head=%u tail=%u", __func__,
1644 	    ioat->chan_idx, ioat->head, ioat->tail);
1645 
1646 	ioat->stats.descriptors_submitted++;
1647 }
1648 
1649 static int
1650 ioat_reset_hw(struct ioat_softc *ioat)
1651 {
1652 	uint64_t status;
1653 	uint32_t chanerr;
1654 	unsigned timeout;
1655 	int error;
1656 
1657 	CTR2(KTR_IOAT, "%s channel=%u", __func__, ioat->chan_idx);
1658 
1659 	mtx_lock(&ioat->submit_lock);
1660 	while (ioat->resetting && !ioat->destroying)
1661 		msleep(&ioat->resetting, &ioat->submit_lock, 0, "IRH_drain", 0);
1662 	if (ioat->destroying) {
1663 		mtx_unlock(&ioat->submit_lock);
1664 		return (ENXIO);
1665 	}
1666 	ioat->resetting = TRUE;
1667 	ioat->quiescing = TRUE;
1668 	mtx_unlock(&ioat->submit_lock);
1669 	mtx_lock(&ioat->cleanup_lock);
1670 	while (ioat_get_active(ioat) > 0)
1671 		msleep(&ioat->tail, &ioat->cleanup_lock, 0, "ioat_drain", 1);
1672 
1673 	/*
1674 	 * Suspend ioat_process_events while the hardware and softc are in an
1675 	 * indeterminate state.
1676 	 */
1677 	ioat->resetting_cleanup = TRUE;
1678 	mtx_unlock(&ioat->cleanup_lock);
1679 
1680 	CTR2(KTR_IOAT, "%s channel=%u quiesced and drained", __func__,
1681 	    ioat->chan_idx);
1682 
1683 	status = ioat_get_chansts(ioat);
1684 	if (is_ioat_active(status) || is_ioat_idle(status))
1685 		ioat_suspend(ioat);
1686 
1687 	/* Wait at most 20 ms */
1688 	for (timeout = 0; (is_ioat_active(status) || is_ioat_idle(status)) &&
1689 	    timeout < 20; timeout++) {
1690 		DELAY(1000);
1691 		status = ioat_get_chansts(ioat);
1692 	}
1693 	if (timeout == 20) {
1694 		error = ETIMEDOUT;
1695 		goto out;
1696 	}
1697 
1698 	KASSERT(ioat_get_active(ioat) == 0, ("active after quiesce"));
1699 
1700 	chanerr = ioat_read_4(ioat, IOAT_CHANERR_OFFSET);
1701 	ioat_write_4(ioat, IOAT_CHANERR_OFFSET, chanerr);
1702 
1703 	CTR2(KTR_IOAT, "%s channel=%u hardware suspended", __func__,
1704 	    ioat->chan_idx);
1705 
1706 	/*
1707 	 * IOAT v3 workaround - CHANERRMSK_INT with 3E07h to masks out errors
1708 	 *  that can cause stability issues for IOAT v3.
1709 	 */
1710 	pci_write_config(ioat->device, IOAT_CFG_CHANERRMASK_INT_OFFSET, 0x3e07,
1711 	    4);
1712 	chanerr = pci_read_config(ioat->device, IOAT_CFG_CHANERR_INT_OFFSET, 4);
1713 	pci_write_config(ioat->device, IOAT_CFG_CHANERR_INT_OFFSET, chanerr, 4);
1714 
1715 	/*
1716 	 * BDXDE and BWD models reset MSI-X registers on device reset.
1717 	 * Save/restore their contents manually.
1718 	 */
1719 	if (ioat_model_resets_msix(ioat)) {
1720 		ioat_log_message(1, "device resets MSI-X registers; saving\n");
1721 		pci_save_state(ioat->device);
1722 	}
1723 
1724 	ioat_reset(ioat);
1725 	CTR2(KTR_IOAT, "%s channel=%u hardware reset", __func__,
1726 	    ioat->chan_idx);
1727 
1728 	/* Wait at most 20 ms */
1729 	for (timeout = 0; ioat_reset_pending(ioat) && timeout < 20; timeout++)
1730 		DELAY(1000);
1731 	if (timeout == 20) {
1732 		error = ETIMEDOUT;
1733 		goto out;
1734 	}
1735 
1736 	if (ioat_model_resets_msix(ioat)) {
1737 		ioat_log_message(1, "device resets registers; restored\n");
1738 		pci_restore_state(ioat->device);
1739 	}
1740 
1741 	/* Reset attempts to return the hardware to "halted." */
1742 	status = ioat_get_chansts(ioat);
1743 	if (is_ioat_active(status) || is_ioat_idle(status)) {
1744 		/* So this really shouldn't happen... */
1745 		ioat_log_message(0, "Device is active after a reset?\n");
1746 		ioat_write_chanctrl(ioat, IOAT_CHANCTRL_RUN);
1747 		error = 0;
1748 		goto out;
1749 	}
1750 
1751 	chanerr = ioat_read_4(ioat, IOAT_CHANERR_OFFSET);
1752 	if (chanerr != 0) {
1753 		mtx_lock(&ioat->cleanup_lock);
1754 		ioat_halted_debug(ioat, chanerr);
1755 		mtx_unlock(&ioat->cleanup_lock);
1756 		error = EIO;
1757 		goto out;
1758 	}
1759 
1760 	/*
1761 	 * Bring device back online after reset.  Writing CHAINADDR brings the
1762 	 * device back to active.
1763 	 *
1764 	 * The internal ring counter resets to zero, so we have to start over
1765 	 * at zero as well.
1766 	 */
1767 	ioat->tail = ioat->head = 0;
1768 	*ioat->comp_update = ioat->last_seen =
1769 	    RING_PHYS_ADDR(ioat, ioat->tail - 1);
1770 
1771 	ioat_write_chanctrl(ioat, IOAT_CHANCTRL_RUN);
1772 	ioat_write_chancmp(ioat, ioat->comp_update_bus_addr);
1773 	ioat_write_chainaddr(ioat, RING_PHYS_ADDR(ioat, 0));
1774 	error = 0;
1775 	CTR2(KTR_IOAT, "%s channel=%u configured channel", __func__,
1776 	    ioat->chan_idx);
1777 
1778 out:
1779 	/* Enqueues a null operation and ensures it completes. */
1780 	if (error == 0) {
1781 		error = ioat_start_channel(ioat);
1782 		CTR2(KTR_IOAT, "%s channel=%u started channel", __func__,
1783 		    ioat->chan_idx);
1784 	}
1785 
1786 	/*
1787 	 * Resume completions now that ring state is consistent.
1788 	 */
1789 	mtx_lock(&ioat->cleanup_lock);
1790 	ioat->resetting_cleanup = FALSE;
1791 	mtx_unlock(&ioat->cleanup_lock);
1792 
1793 	/* Unblock submission of new work */
1794 	mtx_lock(&ioat->submit_lock);
1795 	ioat->quiescing = FALSE;
1796 	wakeup(&ioat->quiescing);
1797 
1798 	ioat->resetting = FALSE;
1799 	wakeup(&ioat->resetting);
1800 
1801 	CTR2(KTR_IOAT, "%s channel=%u reset done", __func__, ioat->chan_idx);
1802 	mtx_unlock(&ioat->submit_lock);
1803 
1804 	return (error);
1805 }
1806 
1807 static int
1808 sysctl_handle_chansts(SYSCTL_HANDLER_ARGS)
1809 {
1810 	struct ioat_softc *ioat;
1811 	struct sbuf sb;
1812 	uint64_t status;
1813 	int error;
1814 
1815 	ioat = arg1;
1816 
1817 	status = ioat_get_chansts(ioat) & IOAT_CHANSTS_STATUS;
1818 
1819 	sbuf_new_for_sysctl(&sb, NULL, 256, req);
1820 	switch (status) {
1821 	case IOAT_CHANSTS_ACTIVE:
1822 		sbuf_printf(&sb, "ACTIVE");
1823 		break;
1824 	case IOAT_CHANSTS_IDLE:
1825 		sbuf_printf(&sb, "IDLE");
1826 		break;
1827 	case IOAT_CHANSTS_SUSPENDED:
1828 		sbuf_printf(&sb, "SUSPENDED");
1829 		break;
1830 	case IOAT_CHANSTS_HALTED:
1831 		sbuf_printf(&sb, "HALTED");
1832 		break;
1833 	case IOAT_CHANSTS_ARMED:
1834 		sbuf_printf(&sb, "ARMED");
1835 		break;
1836 	default:
1837 		sbuf_printf(&sb, "UNKNOWN");
1838 		break;
1839 	}
1840 	error = sbuf_finish(&sb);
1841 	sbuf_delete(&sb);
1842 
1843 	if (error != 0 || req->newptr == NULL)
1844 		return (error);
1845 	return (EINVAL);
1846 }
1847 
1848 static int
1849 sysctl_handle_dpi(SYSCTL_HANDLER_ARGS)
1850 {
1851 	struct ioat_softc *ioat;
1852 	struct sbuf sb;
1853 #define	PRECISION	"1"
1854 	const uintmax_t factor = 10;
1855 	uintmax_t rate;
1856 	int error;
1857 
1858 	ioat = arg1;
1859 	sbuf_new_for_sysctl(&sb, NULL, 16, req);
1860 
1861 	if (ioat->stats.interrupts == 0) {
1862 		sbuf_printf(&sb, "NaN");
1863 		goto out;
1864 	}
1865 	rate = ioat->stats.descriptors_processed * factor /
1866 	    ioat->stats.interrupts;
1867 	sbuf_printf(&sb, "%ju.%." PRECISION "ju", rate / factor,
1868 	    rate % factor);
1869 #undef	PRECISION
1870 out:
1871 	error = sbuf_finish(&sb);
1872 	sbuf_delete(&sb);
1873 	if (error != 0 || req->newptr == NULL)
1874 		return (error);
1875 	return (EINVAL);
1876 }
1877 
1878 static int
1879 sysctl_handle_reset(SYSCTL_HANDLER_ARGS)
1880 {
1881 	struct ioat_softc *ioat;
1882 	int error, arg;
1883 
1884 	ioat = arg1;
1885 
1886 	arg = 0;
1887 	error = SYSCTL_OUT(req, &arg, sizeof(arg));
1888 	if (error != 0 || req->newptr == NULL)
1889 		return (error);
1890 
1891 	error = SYSCTL_IN(req, &arg, sizeof(arg));
1892 	if (error != 0)
1893 		return (error);
1894 
1895 	if (arg != 0)
1896 		error = ioat_reset_hw(ioat);
1897 
1898 	return (error);
1899 }
1900 
1901 static void
1902 dump_descriptor(void *hw_desc)
1903 {
1904 	int i, j;
1905 
1906 	for (i = 0; i < 2; i++) {
1907 		for (j = 0; j < 8; j++)
1908 			printf("%08x ", ((uint32_t *)hw_desc)[i * 8 + j]);
1909 		printf("\n");
1910 	}
1911 }
1912 
1913 static void
1914 ioat_setup_sysctl(device_t device)
1915 {
1916 	struct sysctl_oid_list *par, *statpar, *state, *hammer;
1917 	struct sysctl_ctx_list *ctx;
1918 	struct sysctl_oid *tree, *tmp;
1919 	struct ioat_softc *ioat;
1920 
1921 	ioat = DEVICE2SOFTC(device);
1922 	ctx = device_get_sysctl_ctx(device);
1923 	tree = device_get_sysctl_tree(device);
1924 	par = SYSCTL_CHILDREN(tree);
1925 
1926 	SYSCTL_ADD_INT(ctx, par, OID_AUTO, "version", CTLFLAG_RD,
1927 	    &ioat->version, 0, "HW version (0xMM form)");
1928 	SYSCTL_ADD_UINT(ctx, par, OID_AUTO, "max_xfer_size", CTLFLAG_RD,
1929 	    &ioat->max_xfer_size, 0, "HW maximum transfer size");
1930 	SYSCTL_ADD_INT(ctx, par, OID_AUTO, "intrdelay_supported", CTLFLAG_RD,
1931 	    &ioat->intrdelay_supported, 0, "Is INTRDELAY supported");
1932 	SYSCTL_ADD_U16(ctx, par, OID_AUTO, "intrdelay_max", CTLFLAG_RD,
1933 	    &ioat->intrdelay_max, 0,
1934 	    "Maximum configurable INTRDELAY on this channel (microseconds)");
1935 
1936 	tmp = SYSCTL_ADD_NODE(ctx, par, OID_AUTO, "state",
1937 	    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "IOAT channel internal state");
1938 	state = SYSCTL_CHILDREN(tmp);
1939 
1940 	SYSCTL_ADD_UINT(ctx, state, OID_AUTO, "ring_size_order", CTLFLAG_RD,
1941 	    &ioat->ring_size_order, 0, "SW descriptor ring size order");
1942 	SYSCTL_ADD_UINT(ctx, state, OID_AUTO, "head", CTLFLAG_RD, &ioat->head,
1943 	    0, "SW descriptor head pointer index");
1944 	SYSCTL_ADD_UINT(ctx, state, OID_AUTO, "tail", CTLFLAG_RD, &ioat->tail,
1945 	    0, "SW descriptor tail pointer index");
1946 
1947 	SYSCTL_ADD_UQUAD(ctx, state, OID_AUTO, "last_completion", CTLFLAG_RD,
1948 	    ioat->comp_update, "HW addr of last completion");
1949 
1950 	SYSCTL_ADD_INT(ctx, state, OID_AUTO, "is_submitter_processing",
1951 	    CTLFLAG_RD, &ioat->is_submitter_processing, 0,
1952 	    "submitter processing");
1953 
1954 	SYSCTL_ADD_PROC(ctx, state, OID_AUTO, "chansts",
1955 	    CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_NEEDGIANT, ioat, 0,
1956 	    sysctl_handle_chansts, "A", "String of the channel status");
1957 
1958 	SYSCTL_ADD_U16(ctx, state, OID_AUTO, "intrdelay", CTLFLAG_RD,
1959 	    &ioat->cached_intrdelay, 0,
1960 	    "Current INTRDELAY on this channel (cached, microseconds)");
1961 
1962 	tmp = SYSCTL_ADD_NODE(ctx, par, OID_AUTO, "hammer",
1963 	    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL,
1964 	    "Big hammers (mostly for testing)");
1965 	hammer = SYSCTL_CHILDREN(tmp);
1966 
1967 	SYSCTL_ADD_PROC(ctx, hammer, OID_AUTO, "force_hw_reset",
1968 	    CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT, ioat, 0,
1969 	    sysctl_handle_reset, "I", "Set to non-zero to reset the hardware");
1970 
1971 	tmp = SYSCTL_ADD_NODE(ctx, par, OID_AUTO, "stats",
1972 	    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "IOAT channel statistics");
1973 	statpar = SYSCTL_CHILDREN(tmp);
1974 
1975 	SYSCTL_ADD_UQUAD(ctx, statpar, OID_AUTO, "interrupts",
1976 	    CTLFLAG_RW | CTLFLAG_STATS, &ioat->stats.interrupts,
1977 	    "Number of interrupts processed on this channel");
1978 	SYSCTL_ADD_UQUAD(ctx, statpar, OID_AUTO, "descriptors",
1979 	    CTLFLAG_RW | CTLFLAG_STATS, &ioat->stats.descriptors_processed,
1980 	    "Number of descriptors processed on this channel");
1981 	SYSCTL_ADD_UQUAD(ctx, statpar, OID_AUTO, "submitted",
1982 	    CTLFLAG_RW | CTLFLAG_STATS, &ioat->stats.descriptors_submitted,
1983 	    "Number of descriptors submitted to this channel");
1984 	SYSCTL_ADD_UQUAD(ctx, statpar, OID_AUTO, "errored",
1985 	    CTLFLAG_RW | CTLFLAG_STATS, &ioat->stats.descriptors_error,
1986 	    "Number of descriptors failed by channel errors");
1987 	SYSCTL_ADD_U32(ctx, statpar, OID_AUTO, "halts",
1988 	    CTLFLAG_RW | CTLFLAG_STATS, &ioat->stats.channel_halts, 0,
1989 	    "Number of times the channel has halted");
1990 	SYSCTL_ADD_U32(ctx, statpar, OID_AUTO, "last_halt_chanerr",
1991 	    CTLFLAG_RW | CTLFLAG_STATS, &ioat->stats.last_halt_chanerr, 0,
1992 	    "The raw CHANERR when the channel was last halted");
1993 
1994 	SYSCTL_ADD_PROC(ctx, statpar, OID_AUTO, "desc_per_interrupt",
1995 	    CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_NEEDGIANT, ioat, 0,
1996 	    sysctl_handle_dpi, "A", "Descriptors per interrupt");
1997 }
1998 
1999 static void
2000 ioat_get(struct ioat_softc *ioat)
2001 {
2002 
2003 	mtx_assert(&ioat->submit_lock, MA_OWNED);
2004 	KASSERT(ioat->refcnt < UINT32_MAX, ("refcnt overflow"));
2005 
2006 	ioat->refcnt++;
2007 }
2008 
2009 static void
2010 ioat_put(struct ioat_softc *ioat)
2011 {
2012 
2013 	mtx_assert(&ioat->submit_lock, MA_OWNED);
2014 	KASSERT(ioat->refcnt >= 1, ("refcnt error"));
2015 
2016 	if (--ioat->refcnt == 0)
2017 		wakeup(&ioat->refcnt);
2018 }
2019 
2020 static void
2021 ioat_drain_locked(struct ioat_softc *ioat)
2022 {
2023 
2024 	mtx_assert(&ioat->submit_lock, MA_OWNED);
2025 
2026 	while (ioat->refcnt > 0)
2027 		msleep(&ioat->refcnt, &ioat->submit_lock, 0, "ioat_drain", 0);
2028 }
2029 
2030 #ifdef DDB
2031 #define	_db_show_lock(lo)	LOCK_CLASS(lo)->lc_ddb_show(lo)
2032 #define	db_show_lock(lk)	_db_show_lock(&(lk)->lock_object)
2033 DB_SHOW_COMMAND(ioat, db_show_ioat)
2034 {
2035 	struct ioat_softc *sc;
2036 	unsigned idx;
2037 
2038 	if (!have_addr)
2039 		goto usage;
2040 	idx = (unsigned)addr;
2041 	if (idx >= ioat_channel_index)
2042 		goto usage;
2043 
2044 	sc = ioat_channel[idx];
2045 	db_printf("ioat softc at %p\n", sc);
2046 	if (sc == NULL)
2047 		return;
2048 
2049 	db_printf(" version: %d\n", sc->version);
2050 	db_printf(" chan_idx: %u\n", sc->chan_idx);
2051 	db_printf(" submit_lock: ");
2052 	db_show_lock(&sc->submit_lock);
2053 
2054 	db_printf(" capabilities: %b\n", (int)sc->capabilities,
2055 	    IOAT_DMACAP_STR);
2056 	db_printf(" cached_intrdelay: %u\n", sc->cached_intrdelay);
2057 	db_printf(" *comp_update: 0x%jx\n", (uintmax_t)*sc->comp_update);
2058 
2059 	db_printf(" poll_timer:\n");
2060 	db_printf("  c_time: %ju\n", (uintmax_t)sc->poll_timer.c_time);
2061 	db_printf("  c_arg: %p\n", sc->poll_timer.c_arg);
2062 	db_printf("  c_func: %p\n", sc->poll_timer.c_func);
2063 	db_printf("  c_lock: %p\n", sc->poll_timer.c_lock);
2064 	db_printf("  c_flags: 0x%x\n", (unsigned)sc->poll_timer.c_flags);
2065 
2066 	db_printf(" quiescing: %d\n", (int)sc->quiescing);
2067 	db_printf(" destroying: %d\n", (int)sc->destroying);
2068 	db_printf(" is_submitter_processing: %d\n",
2069 	    (int)sc->is_submitter_processing);
2070 	db_printf(" intrdelay_supported: %d\n", (int)sc->intrdelay_supported);
2071 	db_printf(" resetting: %d\n", (int)sc->resetting);
2072 
2073 	db_printf(" head: %u\n", sc->head);
2074 	db_printf(" tail: %u\n", sc->tail);
2075 	db_printf(" ring_size_order: %u\n", sc->ring_size_order);
2076 	db_printf(" last_seen: 0x%lx\n", sc->last_seen);
2077 	db_printf(" ring: %p\n", sc->ring);
2078 	db_printf(" descriptors: %p\n", sc->hw_desc_ring);
2079 	db_printf(" descriptors (phys): 0x%jx\n",
2080 	    (uintmax_t)sc->hw_desc_bus_addr);
2081 
2082 	db_printf("  ring[%u] (tail):\n", sc->tail %
2083 	    (1 << sc->ring_size_order));
2084 	db_printf("   id: %u\n", ioat_get_ring_entry(sc, sc->tail)->id);
2085 	db_printf("   addr: 0x%lx\n",
2086 	    RING_PHYS_ADDR(sc, sc->tail));
2087 	db_printf("   next: 0x%lx\n",
2088 	     ioat_get_descriptor(sc, sc->tail)->generic.next);
2089 
2090 	db_printf("  ring[%u] (head - 1):\n", (sc->head - 1) %
2091 	    (1 << sc->ring_size_order));
2092 	db_printf("   id: %u\n", ioat_get_ring_entry(sc, sc->head - 1)->id);
2093 	db_printf("   addr: 0x%lx\n",
2094 	    RING_PHYS_ADDR(sc, sc->head - 1));
2095 	db_printf("   next: 0x%lx\n",
2096 	     ioat_get_descriptor(sc, sc->head - 1)->generic.next);
2097 
2098 	db_printf("  ring[%u] (head):\n", (sc->head) %
2099 	    (1 << sc->ring_size_order));
2100 	db_printf("   id: %u\n", ioat_get_ring_entry(sc, sc->head)->id);
2101 	db_printf("   addr: 0x%lx\n",
2102 	    RING_PHYS_ADDR(sc, sc->head));
2103 	db_printf("   next: 0x%lx\n",
2104 	     ioat_get_descriptor(sc, sc->head)->generic.next);
2105 
2106 	for (idx = 0; idx < (1 << sc->ring_size_order); idx++)
2107 		if ((*sc->comp_update & IOAT_CHANSTS_COMPLETED_DESCRIPTOR_MASK)
2108 		    == RING_PHYS_ADDR(sc, idx))
2109 			db_printf("  ring[%u] == hardware tail\n", idx);
2110 
2111 	db_printf(" cleanup_lock: ");
2112 	db_show_lock(&sc->cleanup_lock);
2113 
2114 	db_printf(" refcnt: %u\n", sc->refcnt);
2115 	db_printf(" stats:\n");
2116 	db_printf("  interrupts: %lu\n", sc->stats.interrupts);
2117 	db_printf("  descriptors_processed: %lu\n", sc->stats.descriptors_processed);
2118 	db_printf("  descriptors_error: %lu\n", sc->stats.descriptors_error);
2119 	db_printf("  descriptors_submitted: %lu\n", sc->stats.descriptors_submitted);
2120 
2121 	db_printf("  channel_halts: %u\n", sc->stats.channel_halts);
2122 	db_printf("  last_halt_chanerr: %u\n", sc->stats.last_halt_chanerr);
2123 
2124 	if (db_pager_quit)
2125 		return;
2126 
2127 	db_printf(" hw status:\n");
2128 	db_printf("  status: 0x%lx\n", ioat_get_chansts(sc));
2129 	db_printf("  chanctrl: 0x%x\n",
2130 	    (unsigned)ioat_read_2(sc, IOAT_CHANCTRL_OFFSET));
2131 	db_printf("  chancmd: 0x%x\n",
2132 	    (unsigned)ioat_read_1(sc, IOAT_CHANCMD_OFFSET));
2133 	db_printf("  dmacount: 0x%x\n",
2134 	    (unsigned)ioat_read_2(sc, IOAT_DMACOUNT_OFFSET));
2135 	db_printf("  chainaddr: 0x%lx\n",
2136 	    ioat_read_double_4(sc, IOAT_CHAINADDR_OFFSET_LOW));
2137 	db_printf("  chancmp: 0x%lx\n",
2138 	    ioat_read_double_4(sc, IOAT_CHANCMP_OFFSET_LOW));
2139 	db_printf("  chanerr: %b\n",
2140 	    (int)ioat_read_4(sc, IOAT_CHANERR_OFFSET), IOAT_CHANERR_STR);
2141 	return;
2142 usage:
2143 	db_printf("usage: show ioat <0-%u>\n", ioat_channel_index);
2144 	return;
2145 }
2146 #endif /* DDB */
2147