xref: /freebsd/sys/dev/ioat/ioat.c (revision 3f68b24e10aeb1a1cd85f2d349da44138d52c501)
1 /*-
2  * Copyright (C) 2012 Intel Corporation
3  * All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions
7  * are met:
8  * 1. Redistributions of source code must retain the above copyright
9  *    notice, this list of conditions and the following disclaimer.
10  * 2. Redistributions in binary form must reproduce the above copyright
11  *    notice, this list of conditions and the following disclaimer in the
12  *    documentation and/or other materials provided with the distribution.
13  *
14  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
15  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
16  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
17  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
18  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
19  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
20  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
21  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
22  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
23  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
24  * SUCH DAMAGE.
25  */
26 
27 #include <sys/cdefs.h>
28 __FBSDID("$FreeBSD$");
29 
30 #include "opt_ddb.h"
31 
32 #include <sys/param.h>
33 #include <sys/systm.h>
34 #include <sys/bus.h>
35 #include <sys/conf.h>
36 #include <sys/ioccom.h>
37 #include <sys/kernel.h>
38 #include <sys/lock.h>
39 #include <sys/malloc.h>
40 #include <sys/module.h>
41 #include <sys/mutex.h>
42 #include <sys/rman.h>
43 #include <sys/sbuf.h>
44 #include <sys/sysctl.h>
45 #include <sys/taskqueue.h>
46 #include <sys/time.h>
47 #include <dev/pci/pcireg.h>
48 #include <dev/pci/pcivar.h>
49 #include <machine/bus.h>
50 #include <machine/resource.h>
51 #include <machine/stdarg.h>
52 
53 #ifdef DDB
54 #include <ddb/ddb.h>
55 #endif
56 
57 #include "ioat.h"
58 #include "ioat_hw.h"
59 #include "ioat_internal.h"
60 
61 #ifndef	BUS_SPACE_MAXADDR_40BIT
62 #define	BUS_SPACE_MAXADDR_40BIT	0xFFFFFFFFFFULL
63 #endif
64 #define	IOAT_INTR_TIMO	(hz / 10)
65 #define	IOAT_REFLK	(&ioat->submit_lock)
66 
67 static int ioat_probe(device_t device);
68 static int ioat_attach(device_t device);
69 static int ioat_detach(device_t device);
70 static int ioat_setup_intr(struct ioat_softc *ioat);
71 static int ioat_teardown_intr(struct ioat_softc *ioat);
72 static int ioat3_attach(device_t device);
73 static int ioat_start_channel(struct ioat_softc *ioat);
74 static int ioat_map_pci_bar(struct ioat_softc *ioat);
75 static void ioat_dmamap_cb(void *arg, bus_dma_segment_t *segs, int nseg,
76     int error);
77 static void ioat_interrupt_handler(void *arg);
78 static boolean_t ioat_model_resets_msix(struct ioat_softc *ioat);
79 static int chanerr_to_errno(uint32_t);
80 static void ioat_process_events(struct ioat_softc *ioat);
81 static inline uint32_t ioat_get_active(struct ioat_softc *ioat);
82 static inline uint32_t ioat_get_ring_space(struct ioat_softc *ioat);
83 static void ioat_free_ring(struct ioat_softc *, uint32_t size,
84     struct ioat_descriptor **);
85 static void ioat_free_ring_entry(struct ioat_softc *ioat,
86     struct ioat_descriptor *desc);
87 static struct ioat_descriptor *ioat_alloc_ring_entry(struct ioat_softc *,
88     int mflags);
89 static int ioat_reserve_space(struct ioat_softc *, uint32_t, int mflags);
90 static struct ioat_descriptor *ioat_get_ring_entry(struct ioat_softc *ioat,
91     uint32_t index);
92 static struct ioat_descriptor **ioat_prealloc_ring(struct ioat_softc *,
93     uint32_t size, boolean_t need_dscr, int mflags);
94 static int ring_grow(struct ioat_softc *, uint32_t oldorder,
95     struct ioat_descriptor **);
96 static int ring_shrink(struct ioat_softc *, uint32_t oldorder,
97     struct ioat_descriptor **);
98 static void ioat_halted_debug(struct ioat_softc *, uint32_t);
99 static void ioat_timer_callback(void *arg);
100 static void dump_descriptor(void *hw_desc);
101 static void ioat_submit_single(struct ioat_softc *ioat);
102 static void ioat_comp_update_map(void *arg, bus_dma_segment_t *seg, int nseg,
103     int error);
104 static int ioat_reset_hw(struct ioat_softc *ioat);
105 static void ioat_reset_hw_task(void *, int);
106 static void ioat_setup_sysctl(device_t device);
107 static int sysctl_handle_reset(SYSCTL_HANDLER_ARGS);
108 static inline struct ioat_softc *ioat_get(struct ioat_softc *,
109     enum ioat_ref_kind);
110 static inline void ioat_put(struct ioat_softc *, enum ioat_ref_kind);
111 static inline void _ioat_putn(struct ioat_softc *, uint32_t,
112     enum ioat_ref_kind, boolean_t);
113 static inline void ioat_putn(struct ioat_softc *, uint32_t,
114     enum ioat_ref_kind);
115 static inline void ioat_putn_locked(struct ioat_softc *, uint32_t,
116     enum ioat_ref_kind);
117 static void ioat_drain_locked(struct ioat_softc *);
118 
119 #define	ioat_log_message(v, ...) do {					\
120 	if ((v) <= g_ioat_debug_level) {				\
121 		device_printf(ioat->device, __VA_ARGS__);		\
122 	}								\
123 } while (0)
124 
125 MALLOC_DEFINE(M_IOAT, "ioat", "ioat driver memory allocations");
126 SYSCTL_NODE(_hw, OID_AUTO, ioat, CTLFLAG_RD, 0, "ioat node");
127 
128 static int g_force_legacy_interrupts;
129 SYSCTL_INT(_hw_ioat, OID_AUTO, force_legacy_interrupts, CTLFLAG_RDTUN,
130     &g_force_legacy_interrupts, 0, "Set to non-zero to force MSI-X disabled");
131 
132 int g_ioat_debug_level = 0;
133 SYSCTL_INT(_hw_ioat, OID_AUTO, debug_level, CTLFLAG_RWTUN, &g_ioat_debug_level,
134     0, "Set log level (0-3) for ioat(4). Higher is more verbose.");
135 
136 /*
137  * OS <-> Driver interface structures
138  */
139 static device_method_t ioat_pci_methods[] = {
140 	/* Device interface */
141 	DEVMETHOD(device_probe,     ioat_probe),
142 	DEVMETHOD(device_attach,    ioat_attach),
143 	DEVMETHOD(device_detach,    ioat_detach),
144 	DEVMETHOD_END
145 };
146 
147 static driver_t ioat_pci_driver = {
148 	"ioat",
149 	ioat_pci_methods,
150 	sizeof(struct ioat_softc),
151 };
152 
153 static devclass_t ioat_devclass;
154 DRIVER_MODULE(ioat, pci, ioat_pci_driver, ioat_devclass, 0, 0);
155 MODULE_VERSION(ioat, 1);
156 
157 /*
158  * Private data structures
159  */
160 static struct ioat_softc *ioat_channel[IOAT_MAX_CHANNELS];
161 static unsigned ioat_channel_index = 0;
162 SYSCTL_UINT(_hw_ioat, OID_AUTO, channels, CTLFLAG_RD, &ioat_channel_index, 0,
163     "Number of IOAT channels attached");
164 
165 static struct _pcsid
166 {
167 	u_int32_t   type;
168 	const char  *desc;
169 } pci_ids[] = {
170 	{ 0x34308086, "TBG IOAT Ch0" },
171 	{ 0x34318086, "TBG IOAT Ch1" },
172 	{ 0x34328086, "TBG IOAT Ch2" },
173 	{ 0x34338086, "TBG IOAT Ch3" },
174 	{ 0x34298086, "TBG IOAT Ch4" },
175 	{ 0x342a8086, "TBG IOAT Ch5" },
176 	{ 0x342b8086, "TBG IOAT Ch6" },
177 	{ 0x342c8086, "TBG IOAT Ch7" },
178 
179 	{ 0x37108086, "JSF IOAT Ch0" },
180 	{ 0x37118086, "JSF IOAT Ch1" },
181 	{ 0x37128086, "JSF IOAT Ch2" },
182 	{ 0x37138086, "JSF IOAT Ch3" },
183 	{ 0x37148086, "JSF IOAT Ch4" },
184 	{ 0x37158086, "JSF IOAT Ch5" },
185 	{ 0x37168086, "JSF IOAT Ch6" },
186 	{ 0x37178086, "JSF IOAT Ch7" },
187 	{ 0x37188086, "JSF IOAT Ch0 (RAID)" },
188 	{ 0x37198086, "JSF IOAT Ch1 (RAID)" },
189 
190 	{ 0x3c208086, "SNB IOAT Ch0" },
191 	{ 0x3c218086, "SNB IOAT Ch1" },
192 	{ 0x3c228086, "SNB IOAT Ch2" },
193 	{ 0x3c238086, "SNB IOAT Ch3" },
194 	{ 0x3c248086, "SNB IOAT Ch4" },
195 	{ 0x3c258086, "SNB IOAT Ch5" },
196 	{ 0x3c268086, "SNB IOAT Ch6" },
197 	{ 0x3c278086, "SNB IOAT Ch7" },
198 	{ 0x3c2e8086, "SNB IOAT Ch0 (RAID)" },
199 	{ 0x3c2f8086, "SNB IOAT Ch1 (RAID)" },
200 
201 	{ 0x0e208086, "IVB IOAT Ch0" },
202 	{ 0x0e218086, "IVB IOAT Ch1" },
203 	{ 0x0e228086, "IVB IOAT Ch2" },
204 	{ 0x0e238086, "IVB IOAT Ch3" },
205 	{ 0x0e248086, "IVB IOAT Ch4" },
206 	{ 0x0e258086, "IVB IOAT Ch5" },
207 	{ 0x0e268086, "IVB IOAT Ch6" },
208 	{ 0x0e278086, "IVB IOAT Ch7" },
209 	{ 0x0e2e8086, "IVB IOAT Ch0 (RAID)" },
210 	{ 0x0e2f8086, "IVB IOAT Ch1 (RAID)" },
211 
212 	{ 0x2f208086, "HSW IOAT Ch0" },
213 	{ 0x2f218086, "HSW IOAT Ch1" },
214 	{ 0x2f228086, "HSW IOAT Ch2" },
215 	{ 0x2f238086, "HSW IOAT Ch3" },
216 	{ 0x2f248086, "HSW IOAT Ch4" },
217 	{ 0x2f258086, "HSW IOAT Ch5" },
218 	{ 0x2f268086, "HSW IOAT Ch6" },
219 	{ 0x2f278086, "HSW IOAT Ch7" },
220 	{ 0x2f2e8086, "HSW IOAT Ch0 (RAID)" },
221 	{ 0x2f2f8086, "HSW IOAT Ch1 (RAID)" },
222 
223 	{ 0x0c508086, "BWD IOAT Ch0" },
224 	{ 0x0c518086, "BWD IOAT Ch1" },
225 	{ 0x0c528086, "BWD IOAT Ch2" },
226 	{ 0x0c538086, "BWD IOAT Ch3" },
227 
228 	{ 0x6f508086, "BDXDE IOAT Ch0" },
229 	{ 0x6f518086, "BDXDE IOAT Ch1" },
230 	{ 0x6f528086, "BDXDE IOAT Ch2" },
231 	{ 0x6f538086, "BDXDE IOAT Ch3" },
232 
233 	{ 0x6f208086, "BDX IOAT Ch0" },
234 	{ 0x6f218086, "BDX IOAT Ch1" },
235 	{ 0x6f228086, "BDX IOAT Ch2" },
236 	{ 0x6f238086, "BDX IOAT Ch3" },
237 	{ 0x6f248086, "BDX IOAT Ch4" },
238 	{ 0x6f258086, "BDX IOAT Ch5" },
239 	{ 0x6f268086, "BDX IOAT Ch6" },
240 	{ 0x6f278086, "BDX IOAT Ch7" },
241 	{ 0x6f2e8086, "BDX IOAT Ch0 (RAID)" },
242 	{ 0x6f2f8086, "BDX IOAT Ch1 (RAID)" },
243 
244 	{ 0x00000000, NULL           }
245 };
246 
247 /*
248  * OS <-> Driver linkage functions
249  */
250 static int
251 ioat_probe(device_t device)
252 {
253 	struct _pcsid *ep;
254 	u_int32_t type;
255 
256 	type = pci_get_devid(device);
257 	for (ep = pci_ids; ep->type; ep++) {
258 		if (ep->type == type) {
259 			device_set_desc(device, ep->desc);
260 			return (0);
261 		}
262 	}
263 	return (ENXIO);
264 }
265 
266 static int
267 ioat_attach(device_t device)
268 {
269 	struct ioat_softc *ioat;
270 	int error;
271 
272 	ioat = DEVICE2SOFTC(device);
273 	ioat->device = device;
274 
275 	error = ioat_map_pci_bar(ioat);
276 	if (error != 0)
277 		goto err;
278 
279 	ioat->version = ioat_read_cbver(ioat);
280 	if (ioat->version < IOAT_VER_3_0) {
281 		error = ENODEV;
282 		goto err;
283 	}
284 
285 	error = ioat3_attach(device);
286 	if (error != 0)
287 		goto err;
288 
289 	error = pci_enable_busmaster(device);
290 	if (error != 0)
291 		goto err;
292 
293 	error = ioat_setup_intr(ioat);
294 	if (error != 0)
295 		goto err;
296 
297 	error = ioat_reset_hw(ioat);
298 	if (error != 0)
299 		goto err;
300 
301 	ioat_process_events(ioat);
302 	ioat_setup_sysctl(device);
303 
304 	ioat->chan_idx = ioat_channel_index;
305 	ioat_channel[ioat_channel_index++] = ioat;
306 	ioat_test_attach();
307 
308 err:
309 	if (error != 0)
310 		ioat_detach(device);
311 	return (error);
312 }
313 
314 static int
315 ioat_detach(device_t device)
316 {
317 	struct ioat_softc *ioat;
318 
319 	ioat = DEVICE2SOFTC(device);
320 
321 	ioat_test_detach();
322 	taskqueue_drain(taskqueue_thread, &ioat->reset_task);
323 
324 	mtx_lock(IOAT_REFLK);
325 	ioat->quiescing = TRUE;
326 	ioat->destroying = TRUE;
327 	wakeup(&ioat->quiescing);
328 
329 	ioat_channel[ioat->chan_idx] = NULL;
330 
331 	ioat_drain_locked(ioat);
332 	mtx_unlock(IOAT_REFLK);
333 
334 	ioat_teardown_intr(ioat);
335 	callout_drain(&ioat->timer);
336 
337 	pci_disable_busmaster(device);
338 
339 	if (ioat->pci_resource != NULL)
340 		bus_release_resource(device, SYS_RES_MEMORY,
341 		    ioat->pci_resource_id, ioat->pci_resource);
342 
343 	if (ioat->ring != NULL)
344 		ioat_free_ring(ioat, 1 << ioat->ring_size_order, ioat->ring);
345 
346 	if (ioat->comp_update != NULL) {
347 		bus_dmamap_unload(ioat->comp_update_tag, ioat->comp_update_map);
348 		bus_dmamem_free(ioat->comp_update_tag, ioat->comp_update,
349 		    ioat->comp_update_map);
350 		bus_dma_tag_destroy(ioat->comp_update_tag);
351 	}
352 
353 	bus_dma_tag_destroy(ioat->hw_desc_tag);
354 
355 	return (0);
356 }
357 
358 static int
359 ioat_teardown_intr(struct ioat_softc *ioat)
360 {
361 
362 	if (ioat->tag != NULL)
363 		bus_teardown_intr(ioat->device, ioat->res, ioat->tag);
364 
365 	if (ioat->res != NULL)
366 		bus_release_resource(ioat->device, SYS_RES_IRQ,
367 		    rman_get_rid(ioat->res), ioat->res);
368 
369 	pci_release_msi(ioat->device);
370 	return (0);
371 }
372 
373 static int
374 ioat_start_channel(struct ioat_softc *ioat)
375 {
376 	uint64_t status;
377 	uint32_t chanerr;
378 	int i;
379 
380 	ioat_acquire(&ioat->dmaengine);
381 	ioat_null(&ioat->dmaengine, NULL, NULL, 0);
382 	ioat_release(&ioat->dmaengine);
383 
384 	for (i = 0; i < 100; i++) {
385 		DELAY(1);
386 		status = ioat_get_chansts(ioat);
387 		if (is_ioat_idle(status))
388 			return (0);
389 	}
390 
391 	chanerr = ioat_read_4(ioat, IOAT_CHANERR_OFFSET);
392 	ioat_log_message(0, "could not start channel: "
393 	    "status = %#jx error = %b\n", (uintmax_t)status, (int)chanerr,
394 	    IOAT_CHANERR_STR);
395 	return (ENXIO);
396 }
397 
398 /*
399  * Initialize Hardware
400  */
401 static int
402 ioat3_attach(device_t device)
403 {
404 	struct ioat_softc *ioat;
405 	struct ioat_descriptor **ring;
406 	struct ioat_descriptor *next;
407 	struct ioat_dma_hw_descriptor *dma_hw_desc;
408 	int i, num_descriptors;
409 	int error;
410 	uint8_t xfercap;
411 
412 	error = 0;
413 	ioat = DEVICE2SOFTC(device);
414 	ioat->capabilities = ioat_read_dmacapability(ioat);
415 
416 	ioat_log_message(0, "Capabilities: %b\n", (int)ioat->capabilities,
417 	    IOAT_DMACAP_STR);
418 
419 	xfercap = ioat_read_xfercap(ioat);
420 	ioat->max_xfer_size = 1 << xfercap;
421 
422 	ioat->intrdelay_supported = (ioat_read_2(ioat, IOAT_INTRDELAY_OFFSET) &
423 	    IOAT_INTRDELAY_SUPPORTED) != 0;
424 	if (ioat->intrdelay_supported)
425 		ioat->intrdelay_max = IOAT_INTRDELAY_US_MASK;
426 
427 	/* TODO: need to check DCA here if we ever do XOR/PQ */
428 
429 	mtx_init(&ioat->submit_lock, "ioat_submit", NULL, MTX_DEF);
430 	mtx_init(&ioat->cleanup_lock, "ioat_cleanup", NULL, MTX_DEF);
431 	callout_init(&ioat->timer, 1);
432 	TASK_INIT(&ioat->reset_task, 0, ioat_reset_hw_task, ioat);
433 
434 	/* Establish lock order for Witness */
435 	mtx_lock(&ioat->submit_lock);
436 	mtx_lock(&ioat->cleanup_lock);
437 	mtx_unlock(&ioat->cleanup_lock);
438 	mtx_unlock(&ioat->submit_lock);
439 
440 	ioat->is_resize_pending = FALSE;
441 	ioat->is_completion_pending = FALSE;
442 	ioat->is_reset_pending = FALSE;
443 	ioat->is_channel_running = FALSE;
444 
445 	bus_dma_tag_create(bus_get_dma_tag(ioat->device), sizeof(uint64_t), 0x0,
446 	    BUS_SPACE_MAXADDR, BUS_SPACE_MAXADDR, NULL, NULL,
447 	    sizeof(uint64_t), 1, sizeof(uint64_t), 0, NULL, NULL,
448 	    &ioat->comp_update_tag);
449 
450 	error = bus_dmamem_alloc(ioat->comp_update_tag,
451 	    (void **)&ioat->comp_update, BUS_DMA_ZERO, &ioat->comp_update_map);
452 	if (ioat->comp_update == NULL)
453 		return (ENOMEM);
454 
455 	error = bus_dmamap_load(ioat->comp_update_tag, ioat->comp_update_map,
456 	    ioat->comp_update, sizeof(uint64_t), ioat_comp_update_map, ioat,
457 	    0);
458 	if (error != 0)
459 		return (error);
460 
461 	ioat->ring_size_order = IOAT_MIN_ORDER;
462 
463 	num_descriptors = 1 << ioat->ring_size_order;
464 
465 	bus_dma_tag_create(bus_get_dma_tag(ioat->device), 0x40, 0x0,
466 	    BUS_SPACE_MAXADDR_40BIT, BUS_SPACE_MAXADDR, NULL, NULL,
467 	    sizeof(struct ioat_dma_hw_descriptor), 1,
468 	    sizeof(struct ioat_dma_hw_descriptor), 0, NULL, NULL,
469 	    &ioat->hw_desc_tag);
470 
471 	ioat->ring = malloc(num_descriptors * sizeof(*ring), M_IOAT,
472 	    M_ZERO | M_WAITOK);
473 
474 	ring = ioat->ring;
475 	for (i = 0; i < num_descriptors; i++) {
476 		ring[i] = ioat_alloc_ring_entry(ioat, M_WAITOK);
477 		if (ring[i] == NULL)
478 			return (ENOMEM);
479 
480 		ring[i]->id = i;
481 	}
482 
483 	for (i = 0; i < num_descriptors - 1; i++) {
484 		next = ring[i + 1];
485 		dma_hw_desc = ring[i]->u.dma;
486 
487 		dma_hw_desc->next = next->hw_desc_bus_addr;
488 	}
489 
490 	ring[i]->u.dma->next = ring[0]->hw_desc_bus_addr;
491 
492 	ioat->head = ioat->hw_head = 0;
493 	ioat->tail = 0;
494 	ioat->last_seen = 0;
495 	return (0);
496 }
497 
498 static int
499 ioat_map_pci_bar(struct ioat_softc *ioat)
500 {
501 
502 	ioat->pci_resource_id = PCIR_BAR(0);
503 	ioat->pci_resource = bus_alloc_resource_any(ioat->device,
504 	    SYS_RES_MEMORY, &ioat->pci_resource_id, RF_ACTIVE);
505 
506 	if (ioat->pci_resource == NULL) {
507 		ioat_log_message(0, "unable to allocate pci resource\n");
508 		return (ENODEV);
509 	}
510 
511 	ioat->pci_bus_tag = rman_get_bustag(ioat->pci_resource);
512 	ioat->pci_bus_handle = rman_get_bushandle(ioat->pci_resource);
513 	return (0);
514 }
515 
516 static void
517 ioat_comp_update_map(void *arg, bus_dma_segment_t *seg, int nseg, int error)
518 {
519 	struct ioat_softc *ioat = arg;
520 
521 	KASSERT(error == 0, ("%s: error:%d", __func__, error));
522 	ioat->comp_update_bus_addr = seg[0].ds_addr;
523 }
524 
525 static void
526 ioat_dmamap_cb(void *arg, bus_dma_segment_t *segs, int nseg, int error)
527 {
528 	bus_addr_t *baddr;
529 
530 	KASSERT(error == 0, ("%s: error:%d", __func__, error));
531 	baddr = arg;
532 	*baddr = segs->ds_addr;
533 }
534 
535 /*
536  * Interrupt setup and handlers
537  */
538 static int
539 ioat_setup_intr(struct ioat_softc *ioat)
540 {
541 	uint32_t num_vectors;
542 	int error;
543 	boolean_t use_msix;
544 	boolean_t force_legacy_interrupts;
545 
546 	use_msix = FALSE;
547 	force_legacy_interrupts = FALSE;
548 
549 	if (!g_force_legacy_interrupts && pci_msix_count(ioat->device) >= 1) {
550 		num_vectors = 1;
551 		pci_alloc_msix(ioat->device, &num_vectors);
552 		if (num_vectors == 1)
553 			use_msix = TRUE;
554 	}
555 
556 	if (use_msix) {
557 		ioat->rid = 1;
558 		ioat->res = bus_alloc_resource_any(ioat->device, SYS_RES_IRQ,
559 		    &ioat->rid, RF_ACTIVE);
560 	} else {
561 		ioat->rid = 0;
562 		ioat->res = bus_alloc_resource_any(ioat->device, SYS_RES_IRQ,
563 		    &ioat->rid, RF_SHAREABLE | RF_ACTIVE);
564 	}
565 	if (ioat->res == NULL) {
566 		ioat_log_message(0, "bus_alloc_resource failed\n");
567 		return (ENOMEM);
568 	}
569 
570 	ioat->tag = NULL;
571 	error = bus_setup_intr(ioat->device, ioat->res, INTR_MPSAFE |
572 	    INTR_TYPE_MISC, NULL, ioat_interrupt_handler, ioat, &ioat->tag);
573 	if (error != 0) {
574 		ioat_log_message(0, "bus_setup_intr failed\n");
575 		return (error);
576 	}
577 
578 	ioat_write_intrctrl(ioat, IOAT_INTRCTRL_MASTER_INT_EN);
579 	return (0);
580 }
581 
582 static boolean_t
583 ioat_model_resets_msix(struct ioat_softc *ioat)
584 {
585 	u_int32_t pciid;
586 
587 	pciid = pci_get_devid(ioat->device);
588 	switch (pciid) {
589 		/* BWD: */
590 	case 0x0c508086:
591 	case 0x0c518086:
592 	case 0x0c528086:
593 	case 0x0c538086:
594 		/* BDXDE: */
595 	case 0x6f508086:
596 	case 0x6f518086:
597 	case 0x6f528086:
598 	case 0x6f538086:
599 		return (TRUE);
600 	}
601 
602 	return (FALSE);
603 }
604 
605 static void
606 ioat_interrupt_handler(void *arg)
607 {
608 	struct ioat_softc *ioat = arg;
609 
610 	ioat->stats.interrupts++;
611 	ioat_process_events(ioat);
612 }
613 
614 static int
615 chanerr_to_errno(uint32_t chanerr)
616 {
617 
618 	if (chanerr == 0)
619 		return (0);
620 	if ((chanerr & (IOAT_CHANERR_XSADDERR | IOAT_CHANERR_XDADDERR)) != 0)
621 		return (EFAULT);
622 	if ((chanerr & (IOAT_CHANERR_RDERR | IOAT_CHANERR_WDERR)) != 0)
623 		return (EIO);
624 	/* This one is probably our fault: */
625 	if ((chanerr & IOAT_CHANERR_NDADDERR) != 0)
626 		return (EIO);
627 	return (EIO);
628 }
629 
630 static void
631 ioat_process_events(struct ioat_softc *ioat)
632 {
633 	struct ioat_descriptor *desc;
634 	struct bus_dmadesc *dmadesc;
635 	uint64_t comp_update, status;
636 	uint32_t completed, chanerr;
637 	int error;
638 
639 	mtx_lock(&ioat->cleanup_lock);
640 
641 	completed = 0;
642 	comp_update = *ioat->comp_update;
643 	status = comp_update & IOAT_CHANSTS_COMPLETED_DESCRIPTOR_MASK;
644 
645 	CTR0(KTR_IOAT, __func__);
646 
647 	if (status == ioat->last_seen) {
648 		/*
649 		 * If we landed in process_events and nothing has been
650 		 * completed, check for a timeout due to channel halt.
651 		 */
652 		comp_update = ioat_get_chansts(ioat);
653 		goto out;
654 	}
655 
656 	while (1) {
657 		desc = ioat_get_ring_entry(ioat, ioat->tail);
658 		dmadesc = &desc->bus_dmadesc;
659 		CTR1(KTR_IOAT, "completing desc %d", ioat->tail);
660 
661 		if (dmadesc->callback_fn != NULL)
662 			dmadesc->callback_fn(dmadesc->callback_arg, 0);
663 
664 		completed++;
665 		ioat->tail++;
666 		if (desc->hw_desc_bus_addr == status)
667 			break;
668 	}
669 
670 	ioat->last_seen = desc->hw_desc_bus_addr;
671 
672 	if (ioat->head == ioat->tail) {
673 		ioat->is_completion_pending = FALSE;
674 		callout_reset(&ioat->timer, IOAT_INTR_TIMO,
675 		    ioat_timer_callback, ioat);
676 	}
677 
678 	ioat->stats.descriptors_processed += completed;
679 
680 out:
681 	ioat_write_chanctrl(ioat, IOAT_CHANCTRL_RUN);
682 	mtx_unlock(&ioat->cleanup_lock);
683 
684 	if (completed != 0) {
685 		ioat_putn(ioat, completed, IOAT_ACTIVE_DESCR_REF);
686 		wakeup(&ioat->tail);
687 	}
688 
689 	if (!is_ioat_halted(comp_update) && !is_ioat_suspended(comp_update))
690 		return;
691 
692 	ioat->stats.channel_halts++;
693 
694 	/*
695 	 * Fatal programming error on this DMA channel.  Flush any outstanding
696 	 * work with error status and restart the engine.
697 	 */
698 	ioat_log_message(0, "Channel halted due to fatal programming error\n");
699 	mtx_lock(&ioat->submit_lock);
700 	mtx_lock(&ioat->cleanup_lock);
701 	ioat->quiescing = TRUE;
702 
703 	chanerr = ioat_read_4(ioat, IOAT_CHANERR_OFFSET);
704 	ioat_halted_debug(ioat, chanerr);
705 	ioat->stats.last_halt_chanerr = chanerr;
706 
707 	while (ioat_get_active(ioat) > 0) {
708 		desc = ioat_get_ring_entry(ioat, ioat->tail);
709 		dmadesc = &desc->bus_dmadesc;
710 		CTR1(KTR_IOAT, "completing err desc %d", ioat->tail);
711 
712 		if (dmadesc->callback_fn != NULL)
713 			dmadesc->callback_fn(dmadesc->callback_arg,
714 			    chanerr_to_errno(chanerr));
715 
716 		ioat_putn_locked(ioat, 1, IOAT_ACTIVE_DESCR_REF);
717 		ioat->tail++;
718 		ioat->stats.descriptors_processed++;
719 		ioat->stats.descriptors_error++;
720 	}
721 
722 	/* Clear error status */
723 	ioat_write_4(ioat, IOAT_CHANERR_OFFSET, chanerr);
724 
725 	mtx_unlock(&ioat->cleanup_lock);
726 	mtx_unlock(&ioat->submit_lock);
727 
728 	ioat_log_message(0, "Resetting channel to recover from error\n");
729 	error = taskqueue_enqueue(taskqueue_thread, &ioat->reset_task);
730 	KASSERT(error == 0,
731 	    ("%s: taskqueue_enqueue failed: %d", __func__, error));
732 }
733 
734 static void
735 ioat_reset_hw_task(void *ctx, int pending __unused)
736 {
737 	struct ioat_softc *ioat;
738 	int error;
739 
740 	ioat = ctx;
741 	ioat_log_message(1, "%s: Resetting channel\n", __func__);
742 
743 	error = ioat_reset_hw(ioat);
744 	KASSERT(error == 0, ("%s: reset failed: %d", __func__, error));
745 	(void)error;
746 }
747 
748 /*
749  * User API functions
750  */
751 unsigned
752 ioat_get_nchannels(void)
753 {
754 
755 	return (ioat_channel_index);
756 }
757 
758 bus_dmaengine_t
759 ioat_get_dmaengine(uint32_t index, int flags)
760 {
761 	struct ioat_softc *ioat;
762 
763 	KASSERT((flags & ~(M_NOWAIT | M_WAITOK)) == 0,
764 	    ("invalid flags: 0x%08x", flags));
765 	KASSERT((flags & (M_NOWAIT | M_WAITOK)) != (M_NOWAIT | M_WAITOK),
766 	    ("invalid wait | nowait"));
767 
768 	if (index >= ioat_channel_index)
769 		return (NULL);
770 
771 	ioat = ioat_channel[index];
772 	if (ioat == NULL || ioat->destroying)
773 		return (NULL);
774 
775 	if (ioat->quiescing) {
776 		if ((flags & M_NOWAIT) != 0)
777 			return (NULL);
778 
779 		mtx_lock(IOAT_REFLK);
780 		while (ioat->quiescing && !ioat->destroying)
781 			msleep(&ioat->quiescing, IOAT_REFLK, 0, "getdma", 0);
782 		mtx_unlock(IOAT_REFLK);
783 
784 		if (ioat->destroying)
785 			return (NULL);
786 	}
787 
788 	/*
789 	 * There's a race here between the quiescing check and HW reset or
790 	 * module destroy.
791 	 */
792 	return (&ioat_get(ioat, IOAT_DMAENGINE_REF)->dmaengine);
793 }
794 
795 void
796 ioat_put_dmaengine(bus_dmaengine_t dmaengine)
797 {
798 	struct ioat_softc *ioat;
799 
800 	ioat = to_ioat_softc(dmaengine);
801 	ioat_put(ioat, IOAT_DMAENGINE_REF);
802 }
803 
804 int
805 ioat_get_hwversion(bus_dmaengine_t dmaengine)
806 {
807 	struct ioat_softc *ioat;
808 
809 	ioat = to_ioat_softc(dmaengine);
810 	return (ioat->version);
811 }
812 
813 size_t
814 ioat_get_max_io_size(bus_dmaengine_t dmaengine)
815 {
816 	struct ioat_softc *ioat;
817 
818 	ioat = to_ioat_softc(dmaengine);
819 	return (ioat->max_xfer_size);
820 }
821 
822 int
823 ioat_set_interrupt_coalesce(bus_dmaengine_t dmaengine, uint16_t delay)
824 {
825 	struct ioat_softc *ioat;
826 
827 	ioat = to_ioat_softc(dmaengine);
828 	if (!ioat->intrdelay_supported)
829 		return (ENODEV);
830 	if (delay > ioat->intrdelay_max)
831 		return (ERANGE);
832 
833 	ioat_write_2(ioat, IOAT_INTRDELAY_OFFSET, delay);
834 	ioat->cached_intrdelay =
835 	    ioat_read_2(ioat, IOAT_INTRDELAY_OFFSET) & IOAT_INTRDELAY_US_MASK;
836 	return (0);
837 }
838 
839 uint16_t
840 ioat_get_max_coalesce_period(bus_dmaengine_t dmaengine)
841 {
842 	struct ioat_softc *ioat;
843 
844 	ioat = to_ioat_softc(dmaengine);
845 	return (ioat->intrdelay_max);
846 }
847 
848 void
849 ioat_acquire(bus_dmaengine_t dmaengine)
850 {
851 	struct ioat_softc *ioat;
852 
853 	ioat = to_ioat_softc(dmaengine);
854 	mtx_lock(&ioat->submit_lock);
855 	CTR0(KTR_IOAT, __func__);
856 }
857 
858 int
859 ioat_acquire_reserve(bus_dmaengine_t dmaengine, unsigned n, int mflags)
860 {
861 	struct ioat_softc *ioat;
862 	int error;
863 
864 	ioat = to_ioat_softc(dmaengine);
865 	ioat_acquire(dmaengine);
866 
867 	error = ioat_reserve_space(ioat, n, mflags);
868 	if (error != 0)
869 		ioat_release(dmaengine);
870 	return (error);
871 }
872 
873 void
874 ioat_release(bus_dmaengine_t dmaengine)
875 {
876 	struct ioat_softc *ioat;
877 
878 	ioat = to_ioat_softc(dmaengine);
879 	CTR0(KTR_IOAT, __func__);
880 	ioat_write_2(ioat, IOAT_DMACOUNT_OFFSET, (uint16_t)ioat->hw_head);
881 	mtx_unlock(&ioat->submit_lock);
882 }
883 
884 static struct ioat_descriptor *
885 ioat_op_generic(struct ioat_softc *ioat, uint8_t op,
886     uint32_t size, uint64_t src, uint64_t dst,
887     bus_dmaengine_callback_t callback_fn, void *callback_arg,
888     uint32_t flags)
889 {
890 	struct ioat_generic_hw_descriptor *hw_desc;
891 	struct ioat_descriptor *desc;
892 	int mflags;
893 
894 	mtx_assert(&ioat->submit_lock, MA_OWNED);
895 
896 	KASSERT((flags & ~_DMA_GENERIC_FLAGS) == 0,
897 	    ("Unrecognized flag(s): %#x", flags & ~_DMA_GENERIC_FLAGS));
898 	if ((flags & DMA_NO_WAIT) != 0)
899 		mflags = M_NOWAIT;
900 	else
901 		mflags = M_WAITOK;
902 
903 	if (size > ioat->max_xfer_size) {
904 		ioat_log_message(0, "%s: max_xfer_size = %d, requested = %u\n",
905 		    __func__, ioat->max_xfer_size, (unsigned)size);
906 		return (NULL);
907 	}
908 
909 	if (ioat_reserve_space(ioat, 1, mflags) != 0)
910 		return (NULL);
911 
912 	desc = ioat_get_ring_entry(ioat, ioat->head);
913 	hw_desc = desc->u.generic;
914 
915 	hw_desc->u.control_raw = 0;
916 	hw_desc->u.control_generic.op = op;
917 	hw_desc->u.control_generic.completion_update = 1;
918 
919 	if ((flags & DMA_INT_EN) != 0)
920 		hw_desc->u.control_generic.int_enable = 1;
921 	if ((flags & DMA_FENCE) != 0)
922 		hw_desc->u.control_generic.fence = 1;
923 
924 	hw_desc->size = size;
925 	hw_desc->src_addr = src;
926 	hw_desc->dest_addr = dst;
927 
928 	desc->bus_dmadesc.callback_fn = callback_fn;
929 	desc->bus_dmadesc.callback_arg = callback_arg;
930 	return (desc);
931 }
932 
933 struct bus_dmadesc *
934 ioat_null(bus_dmaengine_t dmaengine, bus_dmaengine_callback_t callback_fn,
935     void *callback_arg, uint32_t flags)
936 {
937 	struct ioat_dma_hw_descriptor *hw_desc;
938 	struct ioat_descriptor *desc;
939 	struct ioat_softc *ioat;
940 
941 	CTR0(KTR_IOAT, __func__);
942 	ioat = to_ioat_softc(dmaengine);
943 
944 	desc = ioat_op_generic(ioat, IOAT_OP_COPY, 8, 0, 0, callback_fn,
945 	    callback_arg, flags);
946 	if (desc == NULL)
947 		return (NULL);
948 
949 	hw_desc = desc->u.dma;
950 	hw_desc->u.control.null = 1;
951 	ioat_submit_single(ioat);
952 	return (&desc->bus_dmadesc);
953 }
954 
955 struct bus_dmadesc *
956 ioat_copy(bus_dmaengine_t dmaengine, bus_addr_t dst,
957     bus_addr_t src, bus_size_t len, bus_dmaengine_callback_t callback_fn,
958     void *callback_arg, uint32_t flags)
959 {
960 	struct ioat_dma_hw_descriptor *hw_desc;
961 	struct ioat_descriptor *desc;
962 	struct ioat_softc *ioat;
963 
964 	CTR0(KTR_IOAT, __func__);
965 	ioat = to_ioat_softc(dmaengine);
966 
967 	if (((src | dst) & (0xffffull << 48)) != 0) {
968 		ioat_log_message(0, "%s: High 16 bits of src/dst invalid\n",
969 		    __func__);
970 		return (NULL);
971 	}
972 
973 	desc = ioat_op_generic(ioat, IOAT_OP_COPY, len, src, dst, callback_fn,
974 	    callback_arg, flags);
975 	if (desc == NULL)
976 		return (NULL);
977 
978 	hw_desc = desc->u.dma;
979 	if (g_ioat_debug_level >= 3)
980 		dump_descriptor(hw_desc);
981 
982 	ioat_submit_single(ioat);
983 	return (&desc->bus_dmadesc);
984 }
985 
986 struct bus_dmadesc *
987 ioat_copy_8k_aligned(bus_dmaengine_t dmaengine, bus_addr_t dst1,
988     bus_addr_t dst2, bus_addr_t src1, bus_addr_t src2,
989     bus_dmaengine_callback_t callback_fn, void *callback_arg, uint32_t flags)
990 {
991 	struct ioat_dma_hw_descriptor *hw_desc;
992 	struct ioat_descriptor *desc;
993 	struct ioat_softc *ioat;
994 
995 	CTR0(KTR_IOAT, __func__);
996 	ioat = to_ioat_softc(dmaengine);
997 
998 	if (((src1 | src2 | dst1 | dst2) & (0xffffull << 48)) != 0) {
999 		ioat_log_message(0, "%s: High 16 bits of src/dst invalid\n",
1000 		    __func__);
1001 		return (NULL);
1002 	}
1003 	if (((src1 | src2 | dst1 | dst2) & PAGE_MASK) != 0) {
1004 		ioat_log_message(0, "%s: Addresses must be page-aligned\n",
1005 		    __func__);
1006 		return (NULL);
1007 	}
1008 
1009 	desc = ioat_op_generic(ioat, IOAT_OP_COPY, 2 * PAGE_SIZE, src1, dst1,
1010 	    callback_fn, callback_arg, flags);
1011 	if (desc == NULL)
1012 		return (NULL);
1013 
1014 	hw_desc = desc->u.dma;
1015 	if (src2 != src1 + PAGE_SIZE) {
1016 		hw_desc->u.control.src_page_break = 1;
1017 		hw_desc->next_src_addr = src2;
1018 	}
1019 	if (dst2 != dst1 + PAGE_SIZE) {
1020 		hw_desc->u.control.dest_page_break = 1;
1021 		hw_desc->next_dest_addr = dst2;
1022 	}
1023 
1024 	if (g_ioat_debug_level >= 3)
1025 		dump_descriptor(hw_desc);
1026 
1027 	ioat_submit_single(ioat);
1028 	return (&desc->bus_dmadesc);
1029 }
1030 
1031 struct bus_dmadesc *
1032 ioat_copy_crc(bus_dmaengine_t dmaengine, bus_addr_t dst, bus_addr_t src,
1033     bus_size_t len, uint32_t *initialseed, bus_addr_t crcptr,
1034     bus_dmaengine_callback_t callback_fn, void *callback_arg, uint32_t flags)
1035 {
1036 	struct ioat_crc32_hw_descriptor *hw_desc;
1037 	struct ioat_descriptor *desc;
1038 	struct ioat_softc *ioat;
1039 	uint32_t teststore;
1040 	uint8_t op;
1041 
1042 	CTR0(KTR_IOAT, __func__);
1043 	ioat = to_ioat_softc(dmaengine);
1044 
1045 	if ((ioat->capabilities & IOAT_DMACAP_MOVECRC) == 0) {
1046 		ioat_log_message(0, "%s: Device lacks MOVECRC capability\n",
1047 		    __func__);
1048 		return (NULL);
1049 	}
1050 	if (((src | dst) & (0xffffffull << 40)) != 0) {
1051 		ioat_log_message(0, "%s: High 24 bits of src/dst invalid\n",
1052 		    __func__);
1053 		return (NULL);
1054 	}
1055 	teststore = (flags & _DMA_CRC_TESTSTORE);
1056 	if (teststore == _DMA_CRC_TESTSTORE) {
1057 		ioat_log_message(0, "%s: TEST and STORE invalid\n", __func__);
1058 		return (NULL);
1059 	}
1060 	if (teststore == 0 && (flags & DMA_CRC_INLINE) != 0) {
1061 		ioat_log_message(0, "%s: INLINE invalid without TEST or STORE\n",
1062 		    __func__);
1063 		return (NULL);
1064 	}
1065 
1066 	switch (teststore) {
1067 	case DMA_CRC_STORE:
1068 		op = IOAT_OP_MOVECRC_STORE;
1069 		break;
1070 	case DMA_CRC_TEST:
1071 		op = IOAT_OP_MOVECRC_TEST;
1072 		break;
1073 	default:
1074 		KASSERT(teststore == 0, ("bogus"));
1075 		op = IOAT_OP_MOVECRC;
1076 		break;
1077 	}
1078 
1079 	if ((flags & DMA_CRC_INLINE) == 0 &&
1080 	    (crcptr & (0xffffffull << 40)) != 0) {
1081 		ioat_log_message(0,
1082 		    "%s: High 24 bits of crcptr invalid\n", __func__);
1083 		return (NULL);
1084 	}
1085 
1086 	desc = ioat_op_generic(ioat, op, len, src, dst, callback_fn,
1087 	    callback_arg, flags & ~_DMA_CRC_FLAGS);
1088 	if (desc == NULL)
1089 		return (NULL);
1090 
1091 	hw_desc = desc->u.crc32;
1092 
1093 	if ((flags & DMA_CRC_INLINE) == 0)
1094 		hw_desc->crc_address = crcptr;
1095 	else
1096 		hw_desc->u.control.crc_location = 1;
1097 
1098 	if (initialseed != NULL) {
1099 		hw_desc->u.control.use_seed = 1;
1100 		hw_desc->seed = *initialseed;
1101 	}
1102 
1103 	if (g_ioat_debug_level >= 3)
1104 		dump_descriptor(hw_desc);
1105 
1106 	ioat_submit_single(ioat);
1107 	return (&desc->bus_dmadesc);
1108 }
1109 
1110 struct bus_dmadesc *
1111 ioat_crc(bus_dmaengine_t dmaengine, bus_addr_t src, bus_size_t len,
1112     uint32_t *initialseed, bus_addr_t crcptr,
1113     bus_dmaengine_callback_t callback_fn, void *callback_arg, uint32_t flags)
1114 {
1115 	struct ioat_crc32_hw_descriptor *hw_desc;
1116 	struct ioat_descriptor *desc;
1117 	struct ioat_softc *ioat;
1118 	uint32_t teststore;
1119 	uint8_t op;
1120 
1121 	CTR0(KTR_IOAT, __func__);
1122 	ioat = to_ioat_softc(dmaengine);
1123 
1124 	if ((ioat->capabilities & IOAT_DMACAP_CRC) == 0) {
1125 		ioat_log_message(0, "%s: Device lacks CRC capability\n",
1126 		    __func__);
1127 		return (NULL);
1128 	}
1129 	if ((src & (0xffffffull << 40)) != 0) {
1130 		ioat_log_message(0, "%s: High 24 bits of src invalid\n",
1131 		    __func__);
1132 		return (NULL);
1133 	}
1134 	teststore = (flags & _DMA_CRC_TESTSTORE);
1135 	if (teststore == _DMA_CRC_TESTSTORE) {
1136 		ioat_log_message(0, "%s: TEST and STORE invalid\n", __func__);
1137 		return (NULL);
1138 	}
1139 	if (teststore == 0 && (flags & DMA_CRC_INLINE) != 0) {
1140 		ioat_log_message(0, "%s: INLINE invalid without TEST or STORE\n",
1141 		    __func__);
1142 		return (NULL);
1143 	}
1144 
1145 	switch (teststore) {
1146 	case DMA_CRC_STORE:
1147 		op = IOAT_OP_CRC_STORE;
1148 		break;
1149 	case DMA_CRC_TEST:
1150 		op = IOAT_OP_CRC_TEST;
1151 		break;
1152 	default:
1153 		KASSERT(teststore == 0, ("bogus"));
1154 		op = IOAT_OP_CRC;
1155 		break;
1156 	}
1157 
1158 	if ((flags & DMA_CRC_INLINE) == 0 &&
1159 	    (crcptr & (0xffffffull << 40)) != 0) {
1160 		ioat_log_message(0,
1161 		    "%s: High 24 bits of crcptr invalid\n", __func__);
1162 		return (NULL);
1163 	}
1164 
1165 	desc = ioat_op_generic(ioat, op, len, src, 0, callback_fn,
1166 	    callback_arg, flags & ~_DMA_CRC_FLAGS);
1167 	if (desc == NULL)
1168 		return (NULL);
1169 
1170 	hw_desc = desc->u.crc32;
1171 
1172 	if ((flags & DMA_CRC_INLINE) == 0)
1173 		hw_desc->crc_address = crcptr;
1174 	else
1175 		hw_desc->u.control.crc_location = 1;
1176 
1177 	if (initialseed != NULL) {
1178 		hw_desc->u.control.use_seed = 1;
1179 		hw_desc->seed = *initialseed;
1180 	}
1181 
1182 	if (g_ioat_debug_level >= 3)
1183 		dump_descriptor(hw_desc);
1184 
1185 	ioat_submit_single(ioat);
1186 	return (&desc->bus_dmadesc);
1187 }
1188 
1189 struct bus_dmadesc *
1190 ioat_blockfill(bus_dmaengine_t dmaengine, bus_addr_t dst, uint64_t fillpattern,
1191     bus_size_t len, bus_dmaengine_callback_t callback_fn, void *callback_arg,
1192     uint32_t flags)
1193 {
1194 	struct ioat_fill_hw_descriptor *hw_desc;
1195 	struct ioat_descriptor *desc;
1196 	struct ioat_softc *ioat;
1197 
1198 	CTR0(KTR_IOAT, __func__);
1199 	ioat = to_ioat_softc(dmaengine);
1200 
1201 	if ((ioat->capabilities & IOAT_DMACAP_BFILL) == 0) {
1202 		ioat_log_message(0, "%s: Device lacks BFILL capability\n",
1203 		    __func__);
1204 		return (NULL);
1205 	}
1206 
1207 	if ((dst & (0xffffull << 48)) != 0) {
1208 		ioat_log_message(0, "%s: High 16 bits of dst invalid\n",
1209 		    __func__);
1210 		return (NULL);
1211 	}
1212 
1213 	desc = ioat_op_generic(ioat, IOAT_OP_FILL, len, fillpattern, dst,
1214 	    callback_fn, callback_arg, flags);
1215 	if (desc == NULL)
1216 		return (NULL);
1217 
1218 	hw_desc = desc->u.fill;
1219 	if (g_ioat_debug_level >= 3)
1220 		dump_descriptor(hw_desc);
1221 
1222 	ioat_submit_single(ioat);
1223 	return (&desc->bus_dmadesc);
1224 }
1225 
1226 /*
1227  * Ring Management
1228  */
1229 static inline uint32_t
1230 ioat_get_active(struct ioat_softc *ioat)
1231 {
1232 
1233 	return ((ioat->head - ioat->tail) & ((1 << ioat->ring_size_order) - 1));
1234 }
1235 
1236 static inline uint32_t
1237 ioat_get_ring_space(struct ioat_softc *ioat)
1238 {
1239 
1240 	return ((1 << ioat->ring_size_order) - ioat_get_active(ioat) - 1);
1241 }
1242 
1243 static struct ioat_descriptor *
1244 ioat_alloc_ring_entry(struct ioat_softc *ioat, int mflags)
1245 {
1246 	struct ioat_generic_hw_descriptor *hw_desc;
1247 	struct ioat_descriptor *desc;
1248 	int error, busdmaflag;
1249 
1250 	error = ENOMEM;
1251 	hw_desc = NULL;
1252 
1253 	if ((mflags & M_WAITOK) != 0)
1254 		busdmaflag = BUS_DMA_WAITOK;
1255 	else
1256 		busdmaflag = BUS_DMA_NOWAIT;
1257 
1258 	desc = malloc(sizeof(*desc), M_IOAT, mflags);
1259 	if (desc == NULL)
1260 		goto out;
1261 
1262 	bus_dmamem_alloc(ioat->hw_desc_tag, (void **)&hw_desc,
1263 	    BUS_DMA_ZERO | busdmaflag, &ioat->hw_desc_map);
1264 	if (hw_desc == NULL)
1265 		goto out;
1266 
1267 	memset(&desc->bus_dmadesc, 0, sizeof(desc->bus_dmadesc));
1268 	desc->u.generic = hw_desc;
1269 
1270 	error = bus_dmamap_load(ioat->hw_desc_tag, ioat->hw_desc_map, hw_desc,
1271 	    sizeof(*hw_desc), ioat_dmamap_cb, &desc->hw_desc_bus_addr,
1272 	    busdmaflag);
1273 	if (error)
1274 		goto out;
1275 
1276 out:
1277 	if (error) {
1278 		ioat_free_ring_entry(ioat, desc);
1279 		return (NULL);
1280 	}
1281 	return (desc);
1282 }
1283 
1284 static void
1285 ioat_free_ring_entry(struct ioat_softc *ioat, struct ioat_descriptor *desc)
1286 {
1287 
1288 	if (desc == NULL)
1289 		return;
1290 
1291 	if (desc->u.generic)
1292 		bus_dmamem_free(ioat->hw_desc_tag, desc->u.generic,
1293 		    ioat->hw_desc_map);
1294 	free(desc, M_IOAT);
1295 }
1296 
1297 /*
1298  * Reserves space in this IOAT descriptor ring by ensuring enough slots remain
1299  * for 'num_descs'.
1300  *
1301  * If mflags contains M_WAITOK, blocks until enough space is available.
1302  *
1303  * Returns zero on success, or an errno on error.  If num_descs is beyond the
1304  * maximum ring size, returns EINVAl; if allocation would block and mflags
1305  * contains M_NOWAIT, returns EAGAIN.
1306  *
1307  * Must be called with the submit_lock held; returns with the lock held.  The
1308  * lock may be dropped to allocate the ring.
1309  *
1310  * (The submit_lock is needed to add any entries to the ring, so callers are
1311  * assured enough room is available.)
1312  */
1313 static int
1314 ioat_reserve_space(struct ioat_softc *ioat, uint32_t num_descs, int mflags)
1315 {
1316 	struct ioat_descriptor **new_ring;
1317 	uint32_t order;
1318 	int error;
1319 
1320 	mtx_assert(&ioat->submit_lock, MA_OWNED);
1321 	error = 0;
1322 
1323 	if (num_descs < 1 || num_descs > (1 << IOAT_MAX_ORDER)) {
1324 		error = EINVAL;
1325 		goto out;
1326 	}
1327 	if (ioat->quiescing) {
1328 		error = ENXIO;
1329 		goto out;
1330 	}
1331 
1332 	for (;;) {
1333 		if (ioat_get_ring_space(ioat) >= num_descs)
1334 			goto out;
1335 
1336 		order = ioat->ring_size_order;
1337 		if (ioat->is_resize_pending || order == IOAT_MAX_ORDER) {
1338 			if ((mflags & M_WAITOK) != 0) {
1339 				msleep(&ioat->tail, &ioat->submit_lock, 0,
1340 				    "ioat_rsz", 0);
1341 				continue;
1342 			}
1343 
1344 			error = EAGAIN;
1345 			break;
1346 		}
1347 
1348 		ioat->is_resize_pending = TRUE;
1349 		for (;;) {
1350 			mtx_unlock(&ioat->submit_lock);
1351 
1352 			new_ring = ioat_prealloc_ring(ioat, 1 << (order + 1),
1353 			    TRUE, mflags);
1354 
1355 			mtx_lock(&ioat->submit_lock);
1356 			KASSERT(ioat->ring_size_order == order,
1357 			    ("is_resize_pending should protect order"));
1358 
1359 			if (new_ring == NULL) {
1360 				KASSERT((mflags & M_WAITOK) == 0,
1361 				    ("allocation failed"));
1362 				error = EAGAIN;
1363 				break;
1364 			}
1365 
1366 			error = ring_grow(ioat, order, new_ring);
1367 			if (error == 0)
1368 				break;
1369 		}
1370 		ioat->is_resize_pending = FALSE;
1371 		wakeup(&ioat->tail);
1372 		if (error)
1373 			break;
1374 	}
1375 
1376 out:
1377 	mtx_assert(&ioat->submit_lock, MA_OWNED);
1378 	return (error);
1379 }
1380 
1381 static struct ioat_descriptor **
1382 ioat_prealloc_ring(struct ioat_softc *ioat, uint32_t size, boolean_t need_dscr,
1383     int mflags)
1384 {
1385 	struct ioat_descriptor **ring;
1386 	uint32_t i;
1387 	int error;
1388 
1389 	KASSERT(size > 0 && powerof2(size), ("bogus size"));
1390 
1391 	ring = malloc(size * sizeof(*ring), M_IOAT, M_ZERO | mflags);
1392 	if (ring == NULL)
1393 		return (NULL);
1394 
1395 	if (need_dscr) {
1396 		error = ENOMEM;
1397 		for (i = size / 2; i < size; i++) {
1398 			ring[i] = ioat_alloc_ring_entry(ioat, mflags);
1399 			if (ring[i] == NULL)
1400 				goto out;
1401 			ring[i]->id = i;
1402 		}
1403 	}
1404 	error = 0;
1405 
1406 out:
1407 	if (error != 0 && ring != NULL) {
1408 		ioat_free_ring(ioat, size, ring);
1409 		ring = NULL;
1410 	}
1411 	return (ring);
1412 }
1413 
1414 static void
1415 ioat_free_ring(struct ioat_softc *ioat, uint32_t size,
1416     struct ioat_descriptor **ring)
1417 {
1418 	uint32_t i;
1419 
1420 	for (i = 0; i < size; i++) {
1421 		if (ring[i] != NULL)
1422 			ioat_free_ring_entry(ioat, ring[i]);
1423 	}
1424 	free(ring, M_IOAT);
1425 }
1426 
1427 static struct ioat_descriptor *
1428 ioat_get_ring_entry(struct ioat_softc *ioat, uint32_t index)
1429 {
1430 
1431 	return (ioat->ring[index % (1 << ioat->ring_size_order)]);
1432 }
1433 
1434 static int
1435 ring_grow(struct ioat_softc *ioat, uint32_t oldorder,
1436     struct ioat_descriptor **newring)
1437 {
1438 	struct ioat_descriptor *tmp, *next;
1439 	struct ioat_dma_hw_descriptor *hw;
1440 	uint32_t oldsize, newsize, head, tail, i, end;
1441 	int error;
1442 
1443 	CTR0(KTR_IOAT, __func__);
1444 
1445 	mtx_assert(&ioat->submit_lock, MA_OWNED);
1446 
1447 	if (oldorder != ioat->ring_size_order || oldorder >= IOAT_MAX_ORDER) {
1448 		error = EINVAL;
1449 		goto out;
1450 	}
1451 
1452 	oldsize = (1 << oldorder);
1453 	newsize = (1 << (oldorder + 1));
1454 
1455 	mtx_lock(&ioat->cleanup_lock);
1456 
1457 	head = ioat->head & (oldsize - 1);
1458 	tail = ioat->tail & (oldsize - 1);
1459 
1460 	/* Copy old descriptors to new ring */
1461 	for (i = 0; i < oldsize; i++)
1462 		newring[i] = ioat->ring[i];
1463 
1464 	/*
1465 	 * If head has wrapped but tail hasn't, we must swap some descriptors
1466 	 * around so that tail can increment directly to head.
1467 	 */
1468 	if (head < tail) {
1469 		for (i = 0; i <= head; i++) {
1470 			tmp = newring[oldsize + i];
1471 
1472 			newring[oldsize + i] = newring[i];
1473 			newring[oldsize + i]->id = oldsize + i;
1474 
1475 			newring[i] = tmp;
1476 			newring[i]->id = i;
1477 		}
1478 		head += oldsize;
1479 	}
1480 
1481 	KASSERT(head >= tail, ("invariants"));
1482 
1483 	/* Head didn't wrap; we only need to link in oldsize..newsize */
1484 	if (head < oldsize) {
1485 		i = oldsize - 1;
1486 		end = newsize;
1487 	} else {
1488 		/* Head did wrap; link newhead..newsize and 0..oldhead */
1489 		i = head;
1490 		end = newsize + (head - oldsize) + 1;
1491 	}
1492 
1493 	/*
1494 	 * Fix up hardware ring, being careful not to trample the active
1495 	 * section (tail -> head).
1496 	 */
1497 	for (; i < end; i++) {
1498 		KASSERT((i & (newsize - 1)) < tail ||
1499 		    (i & (newsize - 1)) >= head, ("trampling snake"));
1500 
1501 		next = newring[(i + 1) & (newsize - 1)];
1502 		hw = newring[i & (newsize - 1)]->u.dma;
1503 		hw->next = next->hw_desc_bus_addr;
1504 	}
1505 
1506 	free(ioat->ring, M_IOAT);
1507 	ioat->ring = newring;
1508 	ioat->ring_size_order = oldorder + 1;
1509 	ioat->tail = tail;
1510 	ioat->head = head;
1511 	error = 0;
1512 
1513 	mtx_unlock(&ioat->cleanup_lock);
1514 out:
1515 	if (error)
1516 		ioat_free_ring(ioat, (1 << (oldorder + 1)), newring);
1517 	return (error);
1518 }
1519 
1520 static int
1521 ring_shrink(struct ioat_softc *ioat, uint32_t oldorder,
1522     struct ioat_descriptor **newring)
1523 {
1524 	struct ioat_dma_hw_descriptor *hw;
1525 	struct ioat_descriptor *ent, *next;
1526 	uint32_t oldsize, newsize, current_idx, new_idx, i;
1527 	int error;
1528 
1529 	CTR0(KTR_IOAT, __func__);
1530 
1531 	mtx_assert(&ioat->submit_lock, MA_OWNED);
1532 
1533 	if (oldorder != ioat->ring_size_order || oldorder <= IOAT_MIN_ORDER) {
1534 		error = EINVAL;
1535 		goto out_unlocked;
1536 	}
1537 
1538 	oldsize = (1 << oldorder);
1539 	newsize = (1 << (oldorder - 1));
1540 
1541 	mtx_lock(&ioat->cleanup_lock);
1542 
1543 	/* Can't shrink below current active set! */
1544 	if (ioat_get_active(ioat) >= newsize) {
1545 		error = ENOMEM;
1546 		goto out;
1547 	}
1548 
1549 	/*
1550 	 * Copy current descriptors to the new ring, dropping the removed
1551 	 * descriptors.
1552 	 */
1553 	for (i = 0; i < newsize; i++) {
1554 		current_idx = (ioat->tail + i) & (oldsize - 1);
1555 		new_idx = (ioat->tail + i) & (newsize - 1);
1556 
1557 		newring[new_idx] = ioat->ring[current_idx];
1558 		newring[new_idx]->id = new_idx;
1559 	}
1560 
1561 	/* Free deleted descriptors */
1562 	for (i = newsize; i < oldsize; i++) {
1563 		ent = ioat_get_ring_entry(ioat, ioat->tail + i);
1564 		ioat_free_ring_entry(ioat, ent);
1565 	}
1566 
1567 	/* Fix up hardware ring. */
1568 	hw = newring[(ioat->tail + newsize - 1) & (newsize - 1)]->u.dma;
1569 	next = newring[(ioat->tail + newsize) & (newsize - 1)];
1570 	hw->next = next->hw_desc_bus_addr;
1571 
1572 	free(ioat->ring, M_IOAT);
1573 	ioat->ring = newring;
1574 	ioat->ring_size_order = oldorder - 1;
1575 	error = 0;
1576 
1577 out:
1578 	mtx_unlock(&ioat->cleanup_lock);
1579 out_unlocked:
1580 	if (error)
1581 		ioat_free_ring(ioat, (1 << (oldorder - 1)), newring);
1582 	return (error);
1583 }
1584 
1585 static void
1586 ioat_halted_debug(struct ioat_softc *ioat, uint32_t chanerr)
1587 {
1588 	struct ioat_descriptor *desc;
1589 
1590 	ioat_log_message(0, "Channel halted (%b)\n", (int)chanerr,
1591 	    IOAT_CHANERR_STR);
1592 	if (chanerr == 0)
1593 		return;
1594 
1595 	mtx_assert(&ioat->cleanup_lock, MA_OWNED);
1596 
1597 	desc = ioat_get_ring_entry(ioat, ioat->tail + 0);
1598 	dump_descriptor(desc->u.raw);
1599 
1600 	desc = ioat_get_ring_entry(ioat, ioat->tail + 1);
1601 	dump_descriptor(desc->u.raw);
1602 }
1603 
1604 static void
1605 ioat_timer_callback(void *arg)
1606 {
1607 	struct ioat_descriptor **newring;
1608 	struct ioat_softc *ioat;
1609 	uint32_t order;
1610 
1611 	ioat = arg;
1612 	ioat_log_message(1, "%s\n", __func__);
1613 
1614 	if (ioat->is_completion_pending) {
1615 		ioat_process_events(ioat);
1616 		return;
1617 	}
1618 
1619 	/* Slowly scale the ring down if idle. */
1620 	mtx_lock(&ioat->submit_lock);
1621 	order = ioat->ring_size_order;
1622 	if (ioat->is_resize_pending || order == IOAT_MIN_ORDER) {
1623 		mtx_unlock(&ioat->submit_lock);
1624 		goto out;
1625 	}
1626 	ioat->is_resize_pending = TRUE;
1627 	mtx_unlock(&ioat->submit_lock);
1628 
1629 	newring = ioat_prealloc_ring(ioat, 1 << (order - 1), FALSE,
1630 	    M_NOWAIT);
1631 
1632 	mtx_lock(&ioat->submit_lock);
1633 	KASSERT(ioat->ring_size_order == order,
1634 	    ("resize_pending protects order"));
1635 
1636 	if (newring != NULL)
1637 		ring_shrink(ioat, order, newring);
1638 
1639 	ioat->is_resize_pending = FALSE;
1640 	mtx_unlock(&ioat->submit_lock);
1641 
1642 out:
1643 	if (ioat->ring_size_order > IOAT_MIN_ORDER)
1644 		callout_reset(&ioat->timer, 10 * hz,
1645 		    ioat_timer_callback, ioat);
1646 }
1647 
1648 /*
1649  * Support Functions
1650  */
1651 static void
1652 ioat_submit_single(struct ioat_softc *ioat)
1653 {
1654 
1655 	ioat_get(ioat, IOAT_ACTIVE_DESCR_REF);
1656 	atomic_add_rel_int(&ioat->head, 1);
1657 	atomic_add_rel_int(&ioat->hw_head, 1);
1658 
1659 	if (!ioat->is_completion_pending) {
1660 		ioat->is_completion_pending = TRUE;
1661 		callout_reset(&ioat->timer, IOAT_INTR_TIMO,
1662 		    ioat_timer_callback, ioat);
1663 	}
1664 
1665 	ioat->stats.descriptors_submitted++;
1666 }
1667 
1668 static int
1669 ioat_reset_hw(struct ioat_softc *ioat)
1670 {
1671 	uint64_t status;
1672 	uint32_t chanerr;
1673 	unsigned timeout;
1674 	int error;
1675 
1676 	mtx_lock(IOAT_REFLK);
1677 	ioat->quiescing = TRUE;
1678 	ioat_drain_locked(ioat);
1679 	mtx_unlock(IOAT_REFLK);
1680 
1681 	status = ioat_get_chansts(ioat);
1682 	if (is_ioat_active(status) || is_ioat_idle(status))
1683 		ioat_suspend(ioat);
1684 
1685 	/* Wait at most 20 ms */
1686 	for (timeout = 0; (is_ioat_active(status) || is_ioat_idle(status)) &&
1687 	    timeout < 20; timeout++) {
1688 		DELAY(1000);
1689 		status = ioat_get_chansts(ioat);
1690 	}
1691 	if (timeout == 20) {
1692 		error = ETIMEDOUT;
1693 		goto out;
1694 	}
1695 
1696 	KASSERT(ioat_get_active(ioat) == 0, ("active after quiesce"));
1697 
1698 	chanerr = ioat_read_4(ioat, IOAT_CHANERR_OFFSET);
1699 	ioat_write_4(ioat, IOAT_CHANERR_OFFSET, chanerr);
1700 
1701 	/*
1702 	 * IOAT v3 workaround - CHANERRMSK_INT with 3E07h to masks out errors
1703 	 *  that can cause stability issues for IOAT v3.
1704 	 */
1705 	pci_write_config(ioat->device, IOAT_CFG_CHANERRMASK_INT_OFFSET, 0x3e07,
1706 	    4);
1707 	chanerr = pci_read_config(ioat->device, IOAT_CFG_CHANERR_INT_OFFSET, 4);
1708 	pci_write_config(ioat->device, IOAT_CFG_CHANERR_INT_OFFSET, chanerr, 4);
1709 
1710 	/*
1711 	 * BDXDE and BWD models reset MSI-X registers on device reset.
1712 	 * Save/restore their contents manually.
1713 	 */
1714 	if (ioat_model_resets_msix(ioat)) {
1715 		ioat_log_message(1, "device resets MSI-X registers; saving\n");
1716 		pci_save_state(ioat->device);
1717 	}
1718 
1719 	ioat_reset(ioat);
1720 
1721 	/* Wait at most 20 ms */
1722 	for (timeout = 0; ioat_reset_pending(ioat) && timeout < 20; timeout++)
1723 		DELAY(1000);
1724 	if (timeout == 20) {
1725 		error = ETIMEDOUT;
1726 		goto out;
1727 	}
1728 
1729 	if (ioat_model_resets_msix(ioat)) {
1730 		ioat_log_message(1, "device resets registers; restored\n");
1731 		pci_restore_state(ioat->device);
1732 	}
1733 
1734 	/* Reset attempts to return the hardware to "halted." */
1735 	status = ioat_get_chansts(ioat);
1736 	if (is_ioat_active(status) || is_ioat_idle(status)) {
1737 		/* So this really shouldn't happen... */
1738 		ioat_log_message(0, "Device is active after a reset?\n");
1739 		ioat_write_chanctrl(ioat, IOAT_CHANCTRL_RUN);
1740 		error = 0;
1741 		goto out;
1742 	}
1743 
1744 	chanerr = ioat_read_4(ioat, IOAT_CHANERR_OFFSET);
1745 	if (chanerr != 0) {
1746 		mtx_lock(&ioat->cleanup_lock);
1747 		ioat_halted_debug(ioat, chanerr);
1748 		mtx_unlock(&ioat->cleanup_lock);
1749 		error = EIO;
1750 		goto out;
1751 	}
1752 
1753 	/*
1754 	 * Bring device back online after reset.  Writing CHAINADDR brings the
1755 	 * device back to active.
1756 	 *
1757 	 * The internal ring counter resets to zero, so we have to start over
1758 	 * at zero as well.
1759 	 */
1760 	ioat->tail = ioat->head = ioat->hw_head = 0;
1761 	ioat->last_seen = 0;
1762 
1763 	ioat_write_chanctrl(ioat, IOAT_CHANCTRL_RUN);
1764 	ioat_write_chancmp(ioat, ioat->comp_update_bus_addr);
1765 	ioat_write_chainaddr(ioat, ioat->ring[0]->hw_desc_bus_addr);
1766 	error = 0;
1767 
1768 out:
1769 	mtx_lock(IOAT_REFLK);
1770 	ioat->quiescing = FALSE;
1771 	wakeup(&ioat->quiescing);
1772 	mtx_unlock(IOAT_REFLK);
1773 
1774 	if (error == 0)
1775 		error = ioat_start_channel(ioat);
1776 
1777 	return (error);
1778 }
1779 
1780 static int
1781 sysctl_handle_chansts(SYSCTL_HANDLER_ARGS)
1782 {
1783 	struct ioat_softc *ioat;
1784 	struct sbuf sb;
1785 	uint64_t status;
1786 	int error;
1787 
1788 	ioat = arg1;
1789 
1790 	status = ioat_get_chansts(ioat) & IOAT_CHANSTS_STATUS;
1791 
1792 	sbuf_new_for_sysctl(&sb, NULL, 256, req);
1793 	switch (status) {
1794 	case IOAT_CHANSTS_ACTIVE:
1795 		sbuf_printf(&sb, "ACTIVE");
1796 		break;
1797 	case IOAT_CHANSTS_IDLE:
1798 		sbuf_printf(&sb, "IDLE");
1799 		break;
1800 	case IOAT_CHANSTS_SUSPENDED:
1801 		sbuf_printf(&sb, "SUSPENDED");
1802 		break;
1803 	case IOAT_CHANSTS_HALTED:
1804 		sbuf_printf(&sb, "HALTED");
1805 		break;
1806 	case IOAT_CHANSTS_ARMED:
1807 		sbuf_printf(&sb, "ARMED");
1808 		break;
1809 	default:
1810 		sbuf_printf(&sb, "UNKNOWN");
1811 		break;
1812 	}
1813 	error = sbuf_finish(&sb);
1814 	sbuf_delete(&sb);
1815 
1816 	if (error != 0 || req->newptr == NULL)
1817 		return (error);
1818 	return (EINVAL);
1819 }
1820 
1821 static int
1822 sysctl_handle_dpi(SYSCTL_HANDLER_ARGS)
1823 {
1824 	struct ioat_softc *ioat;
1825 	struct sbuf sb;
1826 #define	PRECISION	"1"
1827 	const uintmax_t factor = 10;
1828 	uintmax_t rate;
1829 	int error;
1830 
1831 	ioat = arg1;
1832 	sbuf_new_for_sysctl(&sb, NULL, 16, req);
1833 
1834 	if (ioat->stats.interrupts == 0) {
1835 		sbuf_printf(&sb, "NaN");
1836 		goto out;
1837 	}
1838 	rate = ioat->stats.descriptors_processed * factor /
1839 	    ioat->stats.interrupts;
1840 	sbuf_printf(&sb, "%ju.%." PRECISION "ju", rate / factor,
1841 	    rate % factor);
1842 #undef	PRECISION
1843 out:
1844 	error = sbuf_finish(&sb);
1845 	sbuf_delete(&sb);
1846 	if (error != 0 || req->newptr == NULL)
1847 		return (error);
1848 	return (EINVAL);
1849 }
1850 
1851 static int
1852 sysctl_handle_error(SYSCTL_HANDLER_ARGS)
1853 {
1854 	struct ioat_descriptor *desc;
1855 	struct ioat_softc *ioat;
1856 	int error, arg;
1857 
1858 	ioat = arg1;
1859 
1860 	arg = 0;
1861 	error = SYSCTL_OUT(req, &arg, sizeof(arg));
1862 	if (error != 0 || req->newptr == NULL)
1863 		return (error);
1864 
1865 	error = SYSCTL_IN(req, &arg, sizeof(arg));
1866 	if (error != 0)
1867 		return (error);
1868 
1869 	if (arg != 0) {
1870 		ioat_acquire(&ioat->dmaengine);
1871 		desc = ioat_op_generic(ioat, IOAT_OP_COPY, 1,
1872 		    0xffff000000000000ull, 0xffff000000000000ull, NULL, NULL,
1873 		    0);
1874 		if (desc == NULL)
1875 			error = ENOMEM;
1876 		else
1877 			ioat_submit_single(ioat);
1878 		ioat_release(&ioat->dmaengine);
1879 	}
1880 	return (error);
1881 }
1882 
1883 static int
1884 sysctl_handle_reset(SYSCTL_HANDLER_ARGS)
1885 {
1886 	struct ioat_softc *ioat;
1887 	int error, arg;
1888 
1889 	ioat = arg1;
1890 
1891 	arg = 0;
1892 	error = SYSCTL_OUT(req, &arg, sizeof(arg));
1893 	if (error != 0 || req->newptr == NULL)
1894 		return (error);
1895 
1896 	error = SYSCTL_IN(req, &arg, sizeof(arg));
1897 	if (error != 0)
1898 		return (error);
1899 
1900 	if (arg != 0)
1901 		error = ioat_reset_hw(ioat);
1902 
1903 	return (error);
1904 }
1905 
1906 static void
1907 dump_descriptor(void *hw_desc)
1908 {
1909 	int i, j;
1910 
1911 	for (i = 0; i < 2; i++) {
1912 		for (j = 0; j < 8; j++)
1913 			printf("%08x ", ((uint32_t *)hw_desc)[i * 8 + j]);
1914 		printf("\n");
1915 	}
1916 }
1917 
1918 static void
1919 ioat_setup_sysctl(device_t device)
1920 {
1921 	struct sysctl_oid_list *par, *statpar, *state, *hammer;
1922 	struct sysctl_ctx_list *ctx;
1923 	struct sysctl_oid *tree, *tmp;
1924 	struct ioat_softc *ioat;
1925 
1926 	ioat = DEVICE2SOFTC(device);
1927 	ctx = device_get_sysctl_ctx(device);
1928 	tree = device_get_sysctl_tree(device);
1929 	par = SYSCTL_CHILDREN(tree);
1930 
1931 	SYSCTL_ADD_INT(ctx, par, OID_AUTO, "version", CTLFLAG_RD,
1932 	    &ioat->version, 0, "HW version (0xMM form)");
1933 	SYSCTL_ADD_UINT(ctx, par, OID_AUTO, "max_xfer_size", CTLFLAG_RD,
1934 	    &ioat->max_xfer_size, 0, "HW maximum transfer size");
1935 	SYSCTL_ADD_INT(ctx, par, OID_AUTO, "intrdelay_supported", CTLFLAG_RD,
1936 	    &ioat->intrdelay_supported, 0, "Is INTRDELAY supported");
1937 	SYSCTL_ADD_U16(ctx, par, OID_AUTO, "intrdelay_max", CTLFLAG_RD,
1938 	    &ioat->intrdelay_max, 0,
1939 	    "Maximum configurable INTRDELAY on this channel (microseconds)");
1940 
1941 	tmp = SYSCTL_ADD_NODE(ctx, par, OID_AUTO, "state", CTLFLAG_RD, NULL,
1942 	    "IOAT channel internal state");
1943 	state = SYSCTL_CHILDREN(tmp);
1944 
1945 	SYSCTL_ADD_UINT(ctx, state, OID_AUTO, "ring_size_order", CTLFLAG_RD,
1946 	    &ioat->ring_size_order, 0, "SW descriptor ring size order");
1947 	SYSCTL_ADD_UINT(ctx, state, OID_AUTO, "head", CTLFLAG_RD, &ioat->head,
1948 	    0, "SW descriptor head pointer index");
1949 	SYSCTL_ADD_UINT(ctx, state, OID_AUTO, "tail", CTLFLAG_RD, &ioat->tail,
1950 	    0, "SW descriptor tail pointer index");
1951 	SYSCTL_ADD_UINT(ctx, state, OID_AUTO, "hw_head", CTLFLAG_RD,
1952 	    &ioat->hw_head, 0, "HW DMACOUNT");
1953 
1954 	SYSCTL_ADD_UQUAD(ctx, state, OID_AUTO, "last_completion", CTLFLAG_RD,
1955 	    ioat->comp_update, "HW addr of last completion");
1956 
1957 	SYSCTL_ADD_INT(ctx, state, OID_AUTO, "is_resize_pending", CTLFLAG_RD,
1958 	    &ioat->is_resize_pending, 0, "resize pending");
1959 	SYSCTL_ADD_INT(ctx, state, OID_AUTO, "is_completion_pending",
1960 	    CTLFLAG_RD, &ioat->is_completion_pending, 0, "completion pending");
1961 	SYSCTL_ADD_INT(ctx, state, OID_AUTO, "is_reset_pending", CTLFLAG_RD,
1962 	    &ioat->is_reset_pending, 0, "reset pending");
1963 	SYSCTL_ADD_INT(ctx, state, OID_AUTO, "is_channel_running", CTLFLAG_RD,
1964 	    &ioat->is_channel_running, 0, "channel running");
1965 
1966 	SYSCTL_ADD_PROC(ctx, state, OID_AUTO, "chansts",
1967 	    CTLTYPE_STRING | CTLFLAG_RD, ioat, 0, sysctl_handle_chansts, "A",
1968 	    "String of the channel status");
1969 
1970 	SYSCTL_ADD_U16(ctx, state, OID_AUTO, "intrdelay", CTLFLAG_RD,
1971 	    &ioat->cached_intrdelay, 0,
1972 	    "Current INTRDELAY on this channel (cached, microseconds)");
1973 
1974 	tmp = SYSCTL_ADD_NODE(ctx, par, OID_AUTO, "hammer", CTLFLAG_RD, NULL,
1975 	    "Big hammers (mostly for testing)");
1976 	hammer = SYSCTL_CHILDREN(tmp);
1977 
1978 	SYSCTL_ADD_PROC(ctx, hammer, OID_AUTO, "force_hw_reset",
1979 	    CTLTYPE_INT | CTLFLAG_RW, ioat, 0, sysctl_handle_reset, "I",
1980 	    "Set to non-zero to reset the hardware");
1981 	SYSCTL_ADD_PROC(ctx, hammer, OID_AUTO, "force_hw_error",
1982 	    CTLTYPE_INT | CTLFLAG_RW, ioat, 0, sysctl_handle_error, "I",
1983 	    "Set to non-zero to inject a recoverable hardware error");
1984 
1985 	tmp = SYSCTL_ADD_NODE(ctx, par, OID_AUTO, "stats", CTLFLAG_RD, NULL,
1986 	    "IOAT channel statistics");
1987 	statpar = SYSCTL_CHILDREN(tmp);
1988 
1989 	SYSCTL_ADD_UQUAD(ctx, statpar, OID_AUTO, "interrupts", CTLFLAG_RW,
1990 	    &ioat->stats.interrupts,
1991 	    "Number of interrupts processed on this channel");
1992 	SYSCTL_ADD_UQUAD(ctx, statpar, OID_AUTO, "descriptors", CTLFLAG_RW,
1993 	    &ioat->stats.descriptors_processed,
1994 	    "Number of descriptors processed on this channel");
1995 	SYSCTL_ADD_UQUAD(ctx, statpar, OID_AUTO, "submitted", CTLFLAG_RW,
1996 	    &ioat->stats.descriptors_submitted,
1997 	    "Number of descriptors submitted to this channel");
1998 	SYSCTL_ADD_UQUAD(ctx, statpar, OID_AUTO, "errored", CTLFLAG_RW,
1999 	    &ioat->stats.descriptors_error,
2000 	    "Number of descriptors failed by channel errors");
2001 	SYSCTL_ADD_U32(ctx, statpar, OID_AUTO, "halts", CTLFLAG_RW,
2002 	    &ioat->stats.channel_halts, 0,
2003 	    "Number of times the channel has halted");
2004 	SYSCTL_ADD_U32(ctx, statpar, OID_AUTO, "last_halt_chanerr", CTLFLAG_RW,
2005 	    &ioat->stats.last_halt_chanerr, 0,
2006 	    "The raw CHANERR when the channel was last halted");
2007 
2008 	SYSCTL_ADD_PROC(ctx, statpar, OID_AUTO, "desc_per_interrupt",
2009 	    CTLTYPE_STRING | CTLFLAG_RD, ioat, 0, sysctl_handle_dpi, "A",
2010 	    "Descriptors per interrupt");
2011 }
2012 
2013 static inline struct ioat_softc *
2014 ioat_get(struct ioat_softc *ioat, enum ioat_ref_kind kind)
2015 {
2016 	uint32_t old;
2017 
2018 	KASSERT(kind < IOAT_NUM_REF_KINDS, ("bogus"));
2019 
2020 	old = atomic_fetchadd_32(&ioat->refcnt, 1);
2021 	KASSERT(old < UINT32_MAX, ("refcnt overflow"));
2022 
2023 #ifdef INVARIANTS
2024 	old = atomic_fetchadd_32(&ioat->refkinds[kind], 1);
2025 	KASSERT(old < UINT32_MAX, ("refcnt kind overflow"));
2026 #endif
2027 
2028 	return (ioat);
2029 }
2030 
2031 static inline void
2032 ioat_putn(struct ioat_softc *ioat, uint32_t n, enum ioat_ref_kind kind)
2033 {
2034 
2035 	_ioat_putn(ioat, n, kind, FALSE);
2036 }
2037 
2038 static inline void
2039 ioat_putn_locked(struct ioat_softc *ioat, uint32_t n, enum ioat_ref_kind kind)
2040 {
2041 
2042 	_ioat_putn(ioat, n, kind, TRUE);
2043 }
2044 
2045 static inline void
2046 _ioat_putn(struct ioat_softc *ioat, uint32_t n, enum ioat_ref_kind kind,
2047     boolean_t locked)
2048 {
2049 	uint32_t old;
2050 
2051 	KASSERT(kind < IOAT_NUM_REF_KINDS, ("bogus"));
2052 
2053 	if (n == 0)
2054 		return;
2055 
2056 #ifdef INVARIANTS
2057 	old = atomic_fetchadd_32(&ioat->refkinds[kind], -n);
2058 	KASSERT(old >= n, ("refcnt kind underflow"));
2059 #endif
2060 
2061 	/* Skip acquiring the lock if resulting refcnt > 0. */
2062 	for (;;) {
2063 		old = ioat->refcnt;
2064 		if (old <= n)
2065 			break;
2066 		if (atomic_cmpset_32(&ioat->refcnt, old, old - n))
2067 			return;
2068 	}
2069 
2070 	if (locked)
2071 		mtx_assert(IOAT_REFLK, MA_OWNED);
2072 	else
2073 		mtx_lock(IOAT_REFLK);
2074 
2075 	old = atomic_fetchadd_32(&ioat->refcnt, -n);
2076 	KASSERT(old >= n, ("refcnt error"));
2077 
2078 	if (old == n)
2079 		wakeup(IOAT_REFLK);
2080 	if (!locked)
2081 		mtx_unlock(IOAT_REFLK);
2082 }
2083 
2084 static inline void
2085 ioat_put(struct ioat_softc *ioat, enum ioat_ref_kind kind)
2086 {
2087 
2088 	ioat_putn(ioat, 1, kind);
2089 }
2090 
2091 static void
2092 ioat_drain_locked(struct ioat_softc *ioat)
2093 {
2094 
2095 	mtx_assert(IOAT_REFLK, MA_OWNED);
2096 	while (ioat->refcnt > 0)
2097 		msleep(IOAT_REFLK, IOAT_REFLK, 0, "ioat_drain", 0);
2098 }
2099 
2100 #ifdef DDB
2101 #define	_db_show_lock(lo)	LOCK_CLASS(lo)->lc_ddb_show(lo)
2102 #define	db_show_lock(lk)	_db_show_lock(&(lk)->lock_object)
2103 DB_SHOW_COMMAND(ioat, db_show_ioat)
2104 {
2105 	struct ioat_softc *sc;
2106 	unsigned idx;
2107 
2108 	if (!have_addr)
2109 		goto usage;
2110 	idx = (unsigned)addr;
2111 	if (addr >= ioat_channel_index)
2112 		goto usage;
2113 
2114 	sc = ioat_channel[idx];
2115 	db_printf("ioat softc at %p\n", sc);
2116 	if (sc == NULL)
2117 		return;
2118 
2119 	db_printf(" version: %d\n", sc->version);
2120 	db_printf(" chan_idx: %u\n", sc->chan_idx);
2121 	db_printf(" submit_lock: ");
2122 	db_show_lock(&sc->submit_lock);
2123 
2124 	db_printf(" capabilities: %b\n", (int)sc->capabilities,
2125 	    IOAT_DMACAP_STR);
2126 	db_printf(" cached_intrdelay: %u\n", sc->cached_intrdelay);
2127 	db_printf(" *comp_update: 0x%jx\n", (uintmax_t)*sc->comp_update);
2128 
2129 	db_printf(" timer:\n");
2130 	db_printf("  c_time: %ju\n", (uintmax_t)sc->timer.c_time);
2131 	db_printf("  c_arg: %p\n", sc->timer.c_arg);
2132 	db_printf("  c_func: %p\n", sc->timer.c_func);
2133 	db_printf("  c_lock: %p\n", sc->timer.c_lock);
2134 	db_printf("  c_flags: 0x%x\n", (unsigned)sc->timer.c_flags);
2135 
2136 	db_printf(" quiescing: %d\n", (int)sc->quiescing);
2137 	db_printf(" destroying: %d\n", (int)sc->destroying);
2138 	db_printf(" is_resize_pending: %d\n", (int)sc->is_resize_pending);
2139 	db_printf(" is_completion_pending: %d\n", (int)sc->is_completion_pending);
2140 	db_printf(" is_reset_pending: %d\n", (int)sc->is_reset_pending);
2141 	db_printf(" is_channel_running: %d\n", (int)sc->is_channel_running);
2142 	db_printf(" intrdelay_supported: %d\n", (int)sc->intrdelay_supported);
2143 
2144 	db_printf(" head: %u\n", sc->head);
2145 	db_printf(" tail: %u\n", sc->tail);
2146 	db_printf(" hw_head: %u\n", sc->hw_head);
2147 	db_printf(" ring_size_order: %u\n", sc->ring_size_order);
2148 	db_printf(" last_seen: 0x%lx\n", sc->last_seen);
2149 	db_printf(" ring: %p\n", sc->ring);
2150 
2151 	db_printf(" cleanup_lock: ");
2152 	db_show_lock(&sc->cleanup_lock);
2153 
2154 	db_printf(" refcnt: %u\n", sc->refcnt);
2155 #ifdef INVARIANTS
2156 	CTASSERT(IOAT_NUM_REF_KINDS == 2);
2157 	db_printf(" refkinds: [ENG=%u, DESCR=%u]\n", sc->refkinds[0],
2158 	    sc->refkinds[1]);
2159 #endif
2160 	db_printf(" stats:\n");
2161 	db_printf("  interrupts: %lu\n", sc->stats.interrupts);
2162 	db_printf("  descriptors_processed: %lu\n", sc->stats.descriptors_processed);
2163 	db_printf("  descriptors_error: %lu\n", sc->stats.descriptors_error);
2164 	db_printf("  descriptors_submitted: %lu\n", sc->stats.descriptors_submitted);
2165 
2166 	db_printf("  channel_halts: %u\n", sc->stats.channel_halts);
2167 	db_printf("  last_halt_chanerr: %u\n", sc->stats.last_halt_chanerr);
2168 
2169 	if (db_pager_quit)
2170 		return;
2171 
2172 	db_printf(" hw status:\n");
2173 	db_printf("  status: 0x%lx\n", ioat_get_chansts(sc));
2174 	db_printf("  chanctrl: 0x%x\n",
2175 	    (unsigned)ioat_read_2(sc, IOAT_CHANCTRL_OFFSET));
2176 	db_printf("  chancmd: 0x%x\n",
2177 	    (unsigned)ioat_read_1(sc, IOAT_CHANCMD_OFFSET));
2178 	db_printf("  dmacount: 0x%x\n",
2179 	    (unsigned)ioat_read_2(sc, IOAT_DMACOUNT_OFFSET));
2180 	db_printf("  chainaddr: 0x%lx\n",
2181 	    ioat_read_double_4(sc, IOAT_CHAINADDR_OFFSET_LOW));
2182 	db_printf("  chancmp: 0x%lx\n",
2183 	    ioat_read_double_4(sc, IOAT_CHANCMP_OFFSET_LOW));
2184 	db_printf("  chanerr: %b\n",
2185 	    (int)ioat_read_4(sc, IOAT_CHANERR_OFFSET), IOAT_CHANERR_STR);
2186 	return;
2187 usage:
2188 	db_printf("usage: show ioat <0-%u>\n", ioat_channel_index);
2189 	return;
2190 }
2191 #endif /* DDB */
2192