xref: /freebsd/sys/dev/ioat/ioat.c (revision 8907f744ff167dcf9545e2cb089e2831bdd58cd6)
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 <sys/param.h>
31 #include <sys/systm.h>
32 #include <sys/bus.h>
33 #include <sys/conf.h>
34 #include <sys/ioccom.h>
35 #include <sys/kernel.h>
36 #include <sys/lock.h>
37 #include <sys/malloc.h>
38 #include <sys/module.h>
39 #include <sys/mutex.h>
40 #include <sys/rman.h>
41 #include <sys/sbuf.h>
42 #include <sys/sysctl.h>
43 #include <sys/taskqueue.h>
44 #include <sys/time.h>
45 #include <dev/pci/pcireg.h>
46 #include <dev/pci/pcivar.h>
47 #include <machine/bus.h>
48 #include <machine/resource.h>
49 #include <machine/stdarg.h>
50 
51 #include "ioat.h"
52 #include "ioat_hw.h"
53 #include "ioat_internal.h"
54 
55 #ifndef	BUS_SPACE_MAXADDR_40BIT
56 #define	BUS_SPACE_MAXADDR_40BIT	0xFFFFFFFFFFULL
57 #endif
58 #define	IOAT_INTR_TIMO	(hz / 10)
59 #define	IOAT_REFLK	(&ioat->submit_lock)
60 
61 static int ioat_probe(device_t device);
62 static int ioat_attach(device_t device);
63 static int ioat_detach(device_t device);
64 static int ioat_setup_intr(struct ioat_softc *ioat);
65 static int ioat_teardown_intr(struct ioat_softc *ioat);
66 static int ioat3_attach(device_t device);
67 static int ioat_start_channel(struct ioat_softc *ioat);
68 static int ioat_map_pci_bar(struct ioat_softc *ioat);
69 static void ioat_dmamap_cb(void *arg, bus_dma_segment_t *segs, int nseg,
70     int error);
71 static void ioat_interrupt_handler(void *arg);
72 static boolean_t ioat_model_resets_msix(struct ioat_softc *ioat);
73 static int chanerr_to_errno(uint32_t);
74 static void ioat_process_events(struct ioat_softc *ioat);
75 static inline uint32_t ioat_get_active(struct ioat_softc *ioat);
76 static inline uint32_t ioat_get_ring_space(struct ioat_softc *ioat);
77 static void ioat_free_ring(struct ioat_softc *, uint32_t size,
78     struct ioat_descriptor **);
79 static void ioat_free_ring_entry(struct ioat_softc *ioat,
80     struct ioat_descriptor *desc);
81 static struct ioat_descriptor *ioat_alloc_ring_entry(struct ioat_softc *,
82     int mflags);
83 static int ioat_reserve_space(struct ioat_softc *, uint32_t, int mflags);
84 static struct ioat_descriptor *ioat_get_ring_entry(struct ioat_softc *ioat,
85     uint32_t index);
86 static struct ioat_descriptor **ioat_prealloc_ring(struct ioat_softc *,
87     uint32_t size, boolean_t need_dscr, int mflags);
88 static int ring_grow(struct ioat_softc *, uint32_t oldorder,
89     struct ioat_descriptor **);
90 static int ring_shrink(struct ioat_softc *, uint32_t oldorder,
91     struct ioat_descriptor **);
92 static void ioat_halted_debug(struct ioat_softc *, uint32_t);
93 static void ioat_timer_callback(void *arg);
94 static void dump_descriptor(void *hw_desc);
95 static void ioat_submit_single(struct ioat_softc *ioat);
96 static void ioat_comp_update_map(void *arg, bus_dma_segment_t *seg, int nseg,
97     int error);
98 static int ioat_reset_hw(struct ioat_softc *ioat);
99 static void ioat_reset_hw_task(void *, int);
100 static void ioat_setup_sysctl(device_t device);
101 static int sysctl_handle_reset(SYSCTL_HANDLER_ARGS);
102 static inline struct ioat_softc *ioat_get(struct ioat_softc *,
103     enum ioat_ref_kind);
104 static inline void ioat_put(struct ioat_softc *, enum ioat_ref_kind);
105 static inline void _ioat_putn(struct ioat_softc *, uint32_t,
106     enum ioat_ref_kind, boolean_t);
107 static inline void ioat_putn(struct ioat_softc *, uint32_t,
108     enum ioat_ref_kind);
109 static inline void ioat_putn_locked(struct ioat_softc *, uint32_t,
110     enum ioat_ref_kind);
111 static void ioat_drain_locked(struct ioat_softc *);
112 
113 #define	ioat_log_message(v, ...) do {					\
114 	if ((v) <= g_ioat_debug_level) {				\
115 		device_printf(ioat->device, __VA_ARGS__);		\
116 	}								\
117 } while (0)
118 
119 MALLOC_DEFINE(M_IOAT, "ioat", "ioat driver memory allocations");
120 SYSCTL_NODE(_hw, OID_AUTO, ioat, CTLFLAG_RD, 0, "ioat node");
121 
122 static int g_force_legacy_interrupts;
123 SYSCTL_INT(_hw_ioat, OID_AUTO, force_legacy_interrupts, CTLFLAG_RDTUN,
124     &g_force_legacy_interrupts, 0, "Set to non-zero to force MSI-X disabled");
125 
126 int g_ioat_debug_level = 0;
127 SYSCTL_INT(_hw_ioat, OID_AUTO, debug_level, CTLFLAG_RWTUN, &g_ioat_debug_level,
128     0, "Set log level (0-3) for ioat(4). Higher is more verbose.");
129 
130 /*
131  * OS <-> Driver interface structures
132  */
133 static device_method_t ioat_pci_methods[] = {
134 	/* Device interface */
135 	DEVMETHOD(device_probe,     ioat_probe),
136 	DEVMETHOD(device_attach,    ioat_attach),
137 	DEVMETHOD(device_detach,    ioat_detach),
138 	DEVMETHOD_END
139 };
140 
141 static driver_t ioat_pci_driver = {
142 	"ioat",
143 	ioat_pci_methods,
144 	sizeof(struct ioat_softc),
145 };
146 
147 static devclass_t ioat_devclass;
148 DRIVER_MODULE(ioat, pci, ioat_pci_driver, ioat_devclass, 0, 0);
149 MODULE_VERSION(ioat, 1);
150 
151 /*
152  * Private data structures
153  */
154 static struct ioat_softc *ioat_channel[IOAT_MAX_CHANNELS];
155 static int ioat_channel_index = 0;
156 SYSCTL_INT(_hw_ioat, OID_AUTO, channels, CTLFLAG_RD, &ioat_channel_index, 0,
157     "Number of IOAT channels attached");
158 
159 static struct _pcsid
160 {
161 	u_int32_t   type;
162 	const char  *desc;
163 } pci_ids[] = {
164 	{ 0x34308086, "TBG IOAT Ch0" },
165 	{ 0x34318086, "TBG IOAT Ch1" },
166 	{ 0x34328086, "TBG IOAT Ch2" },
167 	{ 0x34338086, "TBG IOAT Ch3" },
168 	{ 0x34298086, "TBG IOAT Ch4" },
169 	{ 0x342a8086, "TBG IOAT Ch5" },
170 	{ 0x342b8086, "TBG IOAT Ch6" },
171 	{ 0x342c8086, "TBG IOAT Ch7" },
172 
173 	{ 0x37108086, "JSF IOAT Ch0" },
174 	{ 0x37118086, "JSF IOAT Ch1" },
175 	{ 0x37128086, "JSF IOAT Ch2" },
176 	{ 0x37138086, "JSF IOAT Ch3" },
177 	{ 0x37148086, "JSF IOAT Ch4" },
178 	{ 0x37158086, "JSF IOAT Ch5" },
179 	{ 0x37168086, "JSF IOAT Ch6" },
180 	{ 0x37178086, "JSF IOAT Ch7" },
181 	{ 0x37188086, "JSF IOAT Ch0 (RAID)" },
182 	{ 0x37198086, "JSF IOAT Ch1 (RAID)" },
183 
184 	{ 0x3c208086, "SNB IOAT Ch0" },
185 	{ 0x3c218086, "SNB IOAT Ch1" },
186 	{ 0x3c228086, "SNB IOAT Ch2" },
187 	{ 0x3c238086, "SNB IOAT Ch3" },
188 	{ 0x3c248086, "SNB IOAT Ch4" },
189 	{ 0x3c258086, "SNB IOAT Ch5" },
190 	{ 0x3c268086, "SNB IOAT Ch6" },
191 	{ 0x3c278086, "SNB IOAT Ch7" },
192 	{ 0x3c2e8086, "SNB IOAT Ch0 (RAID)" },
193 	{ 0x3c2f8086, "SNB IOAT Ch1 (RAID)" },
194 
195 	{ 0x0e208086, "IVB IOAT Ch0" },
196 	{ 0x0e218086, "IVB IOAT Ch1" },
197 	{ 0x0e228086, "IVB IOAT Ch2" },
198 	{ 0x0e238086, "IVB IOAT Ch3" },
199 	{ 0x0e248086, "IVB IOAT Ch4" },
200 	{ 0x0e258086, "IVB IOAT Ch5" },
201 	{ 0x0e268086, "IVB IOAT Ch6" },
202 	{ 0x0e278086, "IVB IOAT Ch7" },
203 	{ 0x0e2e8086, "IVB IOAT Ch0 (RAID)" },
204 	{ 0x0e2f8086, "IVB IOAT Ch1 (RAID)" },
205 
206 	{ 0x2f208086, "HSW IOAT Ch0" },
207 	{ 0x2f218086, "HSW IOAT Ch1" },
208 	{ 0x2f228086, "HSW IOAT Ch2" },
209 	{ 0x2f238086, "HSW IOAT Ch3" },
210 	{ 0x2f248086, "HSW IOAT Ch4" },
211 	{ 0x2f258086, "HSW IOAT Ch5" },
212 	{ 0x2f268086, "HSW IOAT Ch6" },
213 	{ 0x2f278086, "HSW IOAT Ch7" },
214 	{ 0x2f2e8086, "HSW IOAT Ch0 (RAID)" },
215 	{ 0x2f2f8086, "HSW IOAT Ch1 (RAID)" },
216 
217 	{ 0x0c508086, "BWD IOAT Ch0" },
218 	{ 0x0c518086, "BWD IOAT Ch1" },
219 	{ 0x0c528086, "BWD IOAT Ch2" },
220 	{ 0x0c538086, "BWD IOAT Ch3" },
221 
222 	{ 0x6f508086, "BDXDE IOAT Ch0" },
223 	{ 0x6f518086, "BDXDE IOAT Ch1" },
224 	{ 0x6f528086, "BDXDE IOAT Ch2" },
225 	{ 0x6f538086, "BDXDE IOAT Ch3" },
226 
227 	{ 0x6f208086, "BDX IOAT Ch0" },
228 	{ 0x6f218086, "BDX IOAT Ch1" },
229 	{ 0x6f228086, "BDX IOAT Ch2" },
230 	{ 0x6f238086, "BDX IOAT Ch3" },
231 	{ 0x6f248086, "BDX IOAT Ch4" },
232 	{ 0x6f258086, "BDX IOAT Ch5" },
233 	{ 0x6f268086, "BDX IOAT Ch6" },
234 	{ 0x6f278086, "BDX IOAT Ch7" },
235 	{ 0x6f2e8086, "BDX IOAT Ch0 (RAID)" },
236 	{ 0x6f2f8086, "BDX IOAT Ch1 (RAID)" },
237 
238 	{ 0x00000000, NULL           }
239 };
240 
241 /*
242  * OS <-> Driver linkage functions
243  */
244 static int
245 ioat_probe(device_t device)
246 {
247 	struct _pcsid *ep;
248 	u_int32_t type;
249 
250 	type = pci_get_devid(device);
251 	for (ep = pci_ids; ep->type; ep++) {
252 		if (ep->type == type) {
253 			device_set_desc(device, ep->desc);
254 			return (0);
255 		}
256 	}
257 	return (ENXIO);
258 }
259 
260 static int
261 ioat_attach(device_t device)
262 {
263 	struct ioat_softc *ioat;
264 	int error;
265 
266 	ioat = DEVICE2SOFTC(device);
267 	ioat->device = device;
268 
269 	error = ioat_map_pci_bar(ioat);
270 	if (error != 0)
271 		goto err;
272 
273 	ioat->version = ioat_read_cbver(ioat);
274 	if (ioat->version < IOAT_VER_3_0) {
275 		error = ENODEV;
276 		goto err;
277 	}
278 
279 	error = ioat3_attach(device);
280 	if (error != 0)
281 		goto err;
282 
283 	error = pci_enable_busmaster(device);
284 	if (error != 0)
285 		goto err;
286 
287 	error = ioat_setup_intr(ioat);
288 	if (error != 0)
289 		goto err;
290 
291 	error = ioat_reset_hw(ioat);
292 	if (error != 0)
293 		goto err;
294 
295 	ioat_process_events(ioat);
296 	ioat_setup_sysctl(device);
297 
298 	ioat->chan_idx = ioat_channel_index;
299 	ioat_channel[ioat_channel_index++] = ioat;
300 	ioat_test_attach();
301 
302 err:
303 	if (error != 0)
304 		ioat_detach(device);
305 	return (error);
306 }
307 
308 static int
309 ioat_detach(device_t device)
310 {
311 	struct ioat_softc *ioat;
312 
313 	ioat = DEVICE2SOFTC(device);
314 
315 	ioat_test_detach();
316 	taskqueue_drain(taskqueue_thread, &ioat->reset_task);
317 
318 	mtx_lock(IOAT_REFLK);
319 	ioat->quiescing = TRUE;
320 	ioat->destroying = TRUE;
321 	wakeup(&ioat->quiescing);
322 
323 	ioat_channel[ioat->chan_idx] = NULL;
324 
325 	ioat_drain_locked(ioat);
326 	mtx_unlock(IOAT_REFLK);
327 
328 	ioat_teardown_intr(ioat);
329 	callout_drain(&ioat->timer);
330 
331 	pci_disable_busmaster(device);
332 
333 	if (ioat->pci_resource != NULL)
334 		bus_release_resource(device, SYS_RES_MEMORY,
335 		    ioat->pci_resource_id, ioat->pci_resource);
336 
337 	if (ioat->ring != NULL)
338 		ioat_free_ring(ioat, 1 << ioat->ring_size_order, ioat->ring);
339 
340 	if (ioat->comp_update != NULL) {
341 		bus_dmamap_unload(ioat->comp_update_tag, ioat->comp_update_map);
342 		bus_dmamem_free(ioat->comp_update_tag, ioat->comp_update,
343 		    ioat->comp_update_map);
344 		bus_dma_tag_destroy(ioat->comp_update_tag);
345 	}
346 
347 	bus_dma_tag_destroy(ioat->hw_desc_tag);
348 
349 	return (0);
350 }
351 
352 static int
353 ioat_teardown_intr(struct ioat_softc *ioat)
354 {
355 
356 	if (ioat->tag != NULL)
357 		bus_teardown_intr(ioat->device, ioat->res, ioat->tag);
358 
359 	if (ioat->res != NULL)
360 		bus_release_resource(ioat->device, SYS_RES_IRQ,
361 		    rman_get_rid(ioat->res), ioat->res);
362 
363 	pci_release_msi(ioat->device);
364 	return (0);
365 }
366 
367 static int
368 ioat_start_channel(struct ioat_softc *ioat)
369 {
370 	uint64_t status;
371 	uint32_t chanerr;
372 	int i;
373 
374 	ioat_acquire(&ioat->dmaengine);
375 	ioat_null(&ioat->dmaengine, NULL, NULL, 0);
376 	ioat_release(&ioat->dmaengine);
377 
378 	for (i = 0; i < 100; i++) {
379 		DELAY(1);
380 		status = ioat_get_chansts(ioat);
381 		if (is_ioat_idle(status))
382 			return (0);
383 	}
384 
385 	chanerr = ioat_read_4(ioat, IOAT_CHANERR_OFFSET);
386 	ioat_log_message(0, "could not start channel: "
387 	    "status = %#jx error = %b\n", (uintmax_t)status, (int)chanerr,
388 	    IOAT_CHANERR_STR);
389 	return (ENXIO);
390 }
391 
392 /*
393  * Initialize Hardware
394  */
395 static int
396 ioat3_attach(device_t device)
397 {
398 	struct ioat_softc *ioat;
399 	struct ioat_descriptor **ring;
400 	struct ioat_descriptor *next;
401 	struct ioat_dma_hw_descriptor *dma_hw_desc;
402 	int i, num_descriptors;
403 	int error;
404 	uint8_t xfercap;
405 
406 	error = 0;
407 	ioat = DEVICE2SOFTC(device);
408 	ioat->capabilities = ioat_read_dmacapability(ioat);
409 
410 	ioat_log_message(1, "Capabilities: %b\n", (int)ioat->capabilities,
411 	    IOAT_DMACAP_STR);
412 
413 	xfercap = ioat_read_xfercap(ioat);
414 	ioat->max_xfer_size = 1 << xfercap;
415 
416 	ioat->intrdelay_supported = (ioat_read_2(ioat, IOAT_INTRDELAY_OFFSET) &
417 	    IOAT_INTRDELAY_SUPPORTED) != 0;
418 	if (ioat->intrdelay_supported)
419 		ioat->intrdelay_max = IOAT_INTRDELAY_US_MASK;
420 
421 	/* TODO: need to check DCA here if we ever do XOR/PQ */
422 
423 	mtx_init(&ioat->submit_lock, "ioat_submit", NULL, MTX_DEF);
424 	mtx_init(&ioat->cleanup_lock, "ioat_cleanup", NULL, MTX_DEF);
425 	callout_init(&ioat->timer, 1);
426 	TASK_INIT(&ioat->reset_task, 0, ioat_reset_hw_task, ioat);
427 
428 	/* Establish lock order for Witness */
429 	mtx_lock(&ioat->submit_lock);
430 	mtx_lock(&ioat->cleanup_lock);
431 	mtx_unlock(&ioat->cleanup_lock);
432 	mtx_unlock(&ioat->submit_lock);
433 
434 	ioat->is_resize_pending = FALSE;
435 	ioat->is_completion_pending = FALSE;
436 	ioat->is_reset_pending = FALSE;
437 	ioat->is_channel_running = FALSE;
438 
439 	bus_dma_tag_create(bus_get_dma_tag(ioat->device), sizeof(uint64_t), 0x0,
440 	    BUS_SPACE_MAXADDR, BUS_SPACE_MAXADDR, NULL, NULL,
441 	    sizeof(uint64_t), 1, sizeof(uint64_t), 0, NULL, NULL,
442 	    &ioat->comp_update_tag);
443 
444 	error = bus_dmamem_alloc(ioat->comp_update_tag,
445 	    (void **)&ioat->comp_update, BUS_DMA_ZERO, &ioat->comp_update_map);
446 	if (ioat->comp_update == NULL)
447 		return (ENOMEM);
448 
449 	error = bus_dmamap_load(ioat->comp_update_tag, ioat->comp_update_map,
450 	    ioat->comp_update, sizeof(uint64_t), ioat_comp_update_map, ioat,
451 	    0);
452 	if (error != 0)
453 		return (error);
454 
455 	ioat->ring_size_order = IOAT_MIN_ORDER;
456 
457 	num_descriptors = 1 << ioat->ring_size_order;
458 
459 	bus_dma_tag_create(bus_get_dma_tag(ioat->device), 0x40, 0x0,
460 	    BUS_SPACE_MAXADDR_40BIT, BUS_SPACE_MAXADDR, NULL, NULL,
461 	    sizeof(struct ioat_dma_hw_descriptor), 1,
462 	    sizeof(struct ioat_dma_hw_descriptor), 0, NULL, NULL,
463 	    &ioat->hw_desc_tag);
464 
465 	ioat->ring = malloc(num_descriptors * sizeof(*ring), M_IOAT,
466 	    M_ZERO | M_WAITOK);
467 	if (ioat->ring == NULL)
468 		return (ENOMEM);
469 
470 	ring = ioat->ring;
471 	for (i = 0; i < num_descriptors; i++) {
472 		ring[i] = ioat_alloc_ring_entry(ioat, M_WAITOK);
473 		if (ring[i] == NULL)
474 			return (ENOMEM);
475 
476 		ring[i]->id = i;
477 	}
478 
479 	for (i = 0; i < num_descriptors - 1; i++) {
480 		next = ring[i + 1];
481 		dma_hw_desc = ring[i]->u.dma;
482 
483 		dma_hw_desc->next = next->hw_desc_bus_addr;
484 	}
485 
486 	ring[i]->u.dma->next = ring[0]->hw_desc_bus_addr;
487 
488 	ioat->head = ioat->hw_head = 0;
489 	ioat->tail = 0;
490 	ioat->last_seen = 0;
491 	return (0);
492 }
493 
494 static int
495 ioat_map_pci_bar(struct ioat_softc *ioat)
496 {
497 
498 	ioat->pci_resource_id = PCIR_BAR(0);
499 	ioat->pci_resource = bus_alloc_resource_any(ioat->device,
500 	    SYS_RES_MEMORY, &ioat->pci_resource_id, RF_ACTIVE);
501 
502 	if (ioat->pci_resource == NULL) {
503 		ioat_log_message(0, "unable to allocate pci resource\n");
504 		return (ENODEV);
505 	}
506 
507 	ioat->pci_bus_tag = rman_get_bustag(ioat->pci_resource);
508 	ioat->pci_bus_handle = rman_get_bushandle(ioat->pci_resource);
509 	return (0);
510 }
511 
512 static void
513 ioat_comp_update_map(void *arg, bus_dma_segment_t *seg, int nseg, int error)
514 {
515 	struct ioat_softc *ioat = arg;
516 
517 	KASSERT(error == 0, ("%s: error:%d", __func__, error));
518 	ioat->comp_update_bus_addr = seg[0].ds_addr;
519 }
520 
521 static void
522 ioat_dmamap_cb(void *arg, bus_dma_segment_t *segs, int nseg, int error)
523 {
524 	bus_addr_t *baddr;
525 
526 	KASSERT(error == 0, ("%s: error:%d", __func__, error));
527 	baddr = arg;
528 	*baddr = segs->ds_addr;
529 }
530 
531 /*
532  * Interrupt setup and handlers
533  */
534 static int
535 ioat_setup_intr(struct ioat_softc *ioat)
536 {
537 	uint32_t num_vectors;
538 	int error;
539 	boolean_t use_msix;
540 	boolean_t force_legacy_interrupts;
541 
542 	use_msix = FALSE;
543 	force_legacy_interrupts = FALSE;
544 
545 	if (!g_force_legacy_interrupts && pci_msix_count(ioat->device) >= 1) {
546 		num_vectors = 1;
547 		pci_alloc_msix(ioat->device, &num_vectors);
548 		if (num_vectors == 1)
549 			use_msix = TRUE;
550 	}
551 
552 	if (use_msix) {
553 		ioat->rid = 1;
554 		ioat->res = bus_alloc_resource_any(ioat->device, SYS_RES_IRQ,
555 		    &ioat->rid, RF_ACTIVE);
556 	} else {
557 		ioat->rid = 0;
558 		ioat->res = bus_alloc_resource_any(ioat->device, SYS_RES_IRQ,
559 		    &ioat->rid, RF_SHAREABLE | RF_ACTIVE);
560 	}
561 	if (ioat->res == NULL) {
562 		ioat_log_message(0, "bus_alloc_resource failed\n");
563 		return (ENOMEM);
564 	}
565 
566 	ioat->tag = NULL;
567 	error = bus_setup_intr(ioat->device, ioat->res, INTR_MPSAFE |
568 	    INTR_TYPE_MISC, NULL, ioat_interrupt_handler, ioat, &ioat->tag);
569 	if (error != 0) {
570 		ioat_log_message(0, "bus_setup_intr failed\n");
571 		return (error);
572 	}
573 
574 	ioat_write_intrctrl(ioat, IOAT_INTRCTRL_MASTER_INT_EN);
575 	return (0);
576 }
577 
578 static boolean_t
579 ioat_model_resets_msix(struct ioat_softc *ioat)
580 {
581 	u_int32_t pciid;
582 
583 	pciid = pci_get_devid(ioat->device);
584 	switch (pciid) {
585 		/* BWD: */
586 	case 0x0c508086:
587 	case 0x0c518086:
588 	case 0x0c528086:
589 	case 0x0c538086:
590 		/* BDXDE: */
591 	case 0x6f508086:
592 	case 0x6f518086:
593 	case 0x6f528086:
594 	case 0x6f538086:
595 		return (TRUE);
596 	}
597 
598 	return (FALSE);
599 }
600 
601 static void
602 ioat_interrupt_handler(void *arg)
603 {
604 	struct ioat_softc *ioat = arg;
605 
606 	ioat->stats.interrupts++;
607 	ioat_process_events(ioat);
608 }
609 
610 static int
611 chanerr_to_errno(uint32_t chanerr)
612 {
613 
614 	if (chanerr == 0)
615 		return (0);
616 	if ((chanerr & (IOAT_CHANERR_XSADDERR | IOAT_CHANERR_XDADDERR)) != 0)
617 		return (EFAULT);
618 	if ((chanerr & (IOAT_CHANERR_RDERR | IOAT_CHANERR_WDERR)) != 0)
619 		return (EIO);
620 	/* This one is probably our fault: */
621 	if ((chanerr & IOAT_CHANERR_NDADDERR) != 0)
622 		return (EIO);
623 	return (EIO);
624 }
625 
626 static void
627 ioat_process_events(struct ioat_softc *ioat)
628 {
629 	struct ioat_descriptor *desc;
630 	struct bus_dmadesc *dmadesc;
631 	uint64_t comp_update, status;
632 	uint32_t completed, chanerr;
633 	int error;
634 
635 	mtx_lock(&ioat->cleanup_lock);
636 
637 	completed = 0;
638 	comp_update = *ioat->comp_update;
639 	status = comp_update & IOAT_CHANSTS_COMPLETED_DESCRIPTOR_MASK;
640 
641 	CTR0(KTR_IOAT, __func__);
642 
643 	if (status == ioat->last_seen) {
644 		/*
645 		 * If we landed in process_events and nothing has been
646 		 * completed, check for a timeout due to channel halt.
647 		 */
648 		comp_update = ioat_get_chansts(ioat);
649 		goto out;
650 	}
651 
652 	while (1) {
653 		desc = ioat_get_ring_entry(ioat, ioat->tail);
654 		dmadesc = &desc->bus_dmadesc;
655 		CTR1(KTR_IOAT, "completing desc %d", ioat->tail);
656 
657 		if (dmadesc->callback_fn != NULL)
658 			dmadesc->callback_fn(dmadesc->callback_arg, 0);
659 
660 		completed++;
661 		ioat->tail++;
662 		if (desc->hw_desc_bus_addr == status)
663 			break;
664 	}
665 
666 	ioat->last_seen = desc->hw_desc_bus_addr;
667 
668 	if (ioat->head == ioat->tail) {
669 		ioat->is_completion_pending = FALSE;
670 		callout_reset(&ioat->timer, IOAT_INTR_TIMO,
671 		    ioat_timer_callback, ioat);
672 	}
673 
674 	ioat->stats.descriptors_processed += completed;
675 
676 out:
677 	ioat_write_chanctrl(ioat, IOAT_CHANCTRL_RUN);
678 	mtx_unlock(&ioat->cleanup_lock);
679 
680 	if (completed != 0) {
681 		ioat_putn(ioat, completed, IOAT_ACTIVE_DESCR_REF);
682 		wakeup(&ioat->tail);
683 	}
684 
685 	if (!is_ioat_halted(comp_update) && !is_ioat_suspended(comp_update))
686 		return;
687 
688 	ioat->stats.channel_halts++;
689 
690 	/*
691 	 * Fatal programming error on this DMA channel.  Flush any outstanding
692 	 * work with error status and restart the engine.
693 	 */
694 	ioat_log_message(0, "Channel halted due to fatal programming error\n");
695 	mtx_lock(&ioat->submit_lock);
696 	mtx_lock(&ioat->cleanup_lock);
697 	ioat->quiescing = TRUE;
698 
699 	chanerr = ioat_read_4(ioat, IOAT_CHANERR_OFFSET);
700 	ioat_halted_debug(ioat, chanerr);
701 	ioat->stats.last_halt_chanerr = chanerr;
702 
703 	while (ioat_get_active(ioat) > 0) {
704 		desc = ioat_get_ring_entry(ioat, ioat->tail);
705 		dmadesc = &desc->bus_dmadesc;
706 		CTR1(KTR_IOAT, "completing err desc %d", ioat->tail);
707 
708 		if (dmadesc->callback_fn != NULL)
709 			dmadesc->callback_fn(dmadesc->callback_arg,
710 			    chanerr_to_errno(chanerr));
711 
712 		ioat_putn_locked(ioat, 1, IOAT_ACTIVE_DESCR_REF);
713 		ioat->tail++;
714 		ioat->stats.descriptors_processed++;
715 		ioat->stats.descriptors_error++;
716 	}
717 
718 	/* Clear error status */
719 	ioat_write_4(ioat, IOAT_CHANERR_OFFSET, chanerr);
720 
721 	mtx_unlock(&ioat->cleanup_lock);
722 	mtx_unlock(&ioat->submit_lock);
723 
724 	ioat_log_message(0, "Resetting channel to recover from error\n");
725 	error = taskqueue_enqueue(taskqueue_thread, &ioat->reset_task);
726 	KASSERT(error == 0,
727 	    ("%s: taskqueue_enqueue failed: %d", __func__, error));
728 }
729 
730 static void
731 ioat_reset_hw_task(void *ctx, int pending __unused)
732 {
733 	struct ioat_softc *ioat;
734 	int error;
735 
736 	ioat = ctx;
737 	ioat_log_message(1, "%s: Resetting channel\n", __func__);
738 
739 	error = ioat_reset_hw(ioat);
740 	KASSERT(error == 0, ("%s: reset failed: %d", __func__, error));
741 	(void)error;
742 }
743 
744 /*
745  * User API functions
746  */
747 bus_dmaengine_t
748 ioat_get_dmaengine(uint32_t index, int flags)
749 {
750 	struct ioat_softc *ioat;
751 
752 	KASSERT((flags & ~(M_NOWAIT | M_WAITOK)) == 0,
753 	    ("invalid flags: 0x%08x", flags));
754 	KASSERT((flags & (M_NOWAIT | M_WAITOK)) != (M_NOWAIT | M_WAITOK),
755 	    ("invalid wait | nowait"));
756 
757 	if (index >= ioat_channel_index)
758 		return (NULL);
759 
760 	ioat = ioat_channel[index];
761 	if (ioat == NULL || ioat->destroying)
762 		return (NULL);
763 
764 	if (ioat->quiescing) {
765 		if ((flags & M_NOWAIT) != 0)
766 			return (NULL);
767 
768 		mtx_lock(IOAT_REFLK);
769 		while (ioat->quiescing && !ioat->destroying)
770 			msleep(&ioat->quiescing, IOAT_REFLK, 0, "getdma", 0);
771 		mtx_unlock(IOAT_REFLK);
772 
773 		if (ioat->destroying)
774 			return (NULL);
775 	}
776 
777 	/*
778 	 * There's a race here between the quiescing check and HW reset or
779 	 * module destroy.
780 	 */
781 	return (&ioat_get(ioat, IOAT_DMAENGINE_REF)->dmaengine);
782 }
783 
784 void
785 ioat_put_dmaengine(bus_dmaengine_t dmaengine)
786 {
787 	struct ioat_softc *ioat;
788 
789 	ioat = to_ioat_softc(dmaengine);
790 	ioat_put(ioat, IOAT_DMAENGINE_REF);
791 }
792 
793 int
794 ioat_get_hwversion(bus_dmaengine_t dmaengine)
795 {
796 	struct ioat_softc *ioat;
797 
798 	ioat = to_ioat_softc(dmaengine);
799 	return (ioat->version);
800 }
801 
802 size_t
803 ioat_get_max_io_size(bus_dmaengine_t dmaengine)
804 {
805 	struct ioat_softc *ioat;
806 
807 	ioat = to_ioat_softc(dmaengine);
808 	return (ioat->max_xfer_size);
809 }
810 
811 int
812 ioat_set_interrupt_coalesce(bus_dmaengine_t dmaengine, uint16_t delay)
813 {
814 	struct ioat_softc *ioat;
815 
816 	ioat = to_ioat_softc(dmaengine);
817 	if (!ioat->intrdelay_supported)
818 		return (ENODEV);
819 	if (delay > ioat->intrdelay_max)
820 		return (ERANGE);
821 
822 	ioat_write_2(ioat, IOAT_INTRDELAY_OFFSET, delay);
823 	ioat->cached_intrdelay =
824 	    ioat_read_2(ioat, IOAT_INTRDELAY_OFFSET) & IOAT_INTRDELAY_US_MASK;
825 	return (0);
826 }
827 
828 uint16_t
829 ioat_get_max_coalesce_period(bus_dmaengine_t dmaengine)
830 {
831 	struct ioat_softc *ioat;
832 
833 	ioat = to_ioat_softc(dmaengine);
834 	return (ioat->intrdelay_max);
835 }
836 
837 void
838 ioat_acquire(bus_dmaengine_t dmaengine)
839 {
840 	struct ioat_softc *ioat;
841 
842 	ioat = to_ioat_softc(dmaengine);
843 	mtx_lock(&ioat->submit_lock);
844 	CTR0(KTR_IOAT, __func__);
845 }
846 
847 int
848 ioat_acquire_reserve(bus_dmaengine_t dmaengine, unsigned n, int mflags)
849 {
850 	struct ioat_softc *ioat;
851 	int error;
852 
853 	ioat = to_ioat_softc(dmaengine);
854 	ioat_acquire(dmaengine);
855 
856 	error = ioat_reserve_space(ioat, n, mflags);
857 	if (error != 0)
858 		ioat_release(dmaengine);
859 	return (error);
860 }
861 
862 void
863 ioat_release(bus_dmaengine_t dmaengine)
864 {
865 	struct ioat_softc *ioat;
866 
867 	ioat = to_ioat_softc(dmaengine);
868 	CTR0(KTR_IOAT, __func__);
869 	ioat_write_2(ioat, IOAT_DMACOUNT_OFFSET, (uint16_t)ioat->hw_head);
870 	mtx_unlock(&ioat->submit_lock);
871 }
872 
873 static struct ioat_descriptor *
874 ioat_op_generic(struct ioat_softc *ioat, uint8_t op,
875     uint32_t size, uint64_t src, uint64_t dst,
876     bus_dmaengine_callback_t callback_fn, void *callback_arg,
877     uint32_t flags)
878 {
879 	struct ioat_generic_hw_descriptor *hw_desc;
880 	struct ioat_descriptor *desc;
881 	int mflags;
882 
883 	mtx_assert(&ioat->submit_lock, MA_OWNED);
884 
885 	KASSERT((flags & ~_DMA_GENERIC_FLAGS) == 0,
886 	    ("Unrecognized flag(s): %#x", flags & ~_DMA_GENERIC_FLAGS));
887 	if ((flags & DMA_NO_WAIT) != 0)
888 		mflags = M_NOWAIT;
889 	else
890 		mflags = M_WAITOK;
891 
892 	if (size > ioat->max_xfer_size) {
893 		ioat_log_message(0, "%s: max_xfer_size = %d, requested = %u\n",
894 		    __func__, ioat->max_xfer_size, (unsigned)size);
895 		return (NULL);
896 	}
897 
898 	if (ioat_reserve_space(ioat, 1, mflags) != 0)
899 		return (NULL);
900 
901 	desc = ioat_get_ring_entry(ioat, ioat->head);
902 	hw_desc = desc->u.generic;
903 
904 	hw_desc->u.control_raw = 0;
905 	hw_desc->u.control_generic.op = op;
906 	hw_desc->u.control_generic.completion_update = 1;
907 
908 	if ((flags & DMA_INT_EN) != 0)
909 		hw_desc->u.control_generic.int_enable = 1;
910 	if ((flags & DMA_FENCE) != 0)
911 		hw_desc->u.control_generic.fence = 1;
912 
913 	hw_desc->size = size;
914 	hw_desc->src_addr = src;
915 	hw_desc->dest_addr = dst;
916 
917 	desc->bus_dmadesc.callback_fn = callback_fn;
918 	desc->bus_dmadesc.callback_arg = callback_arg;
919 	return (desc);
920 }
921 
922 struct bus_dmadesc *
923 ioat_null(bus_dmaengine_t dmaengine, bus_dmaengine_callback_t callback_fn,
924     void *callback_arg, uint32_t flags)
925 {
926 	struct ioat_dma_hw_descriptor *hw_desc;
927 	struct ioat_descriptor *desc;
928 	struct ioat_softc *ioat;
929 
930 	CTR0(KTR_IOAT, __func__);
931 	ioat = to_ioat_softc(dmaengine);
932 
933 	desc = ioat_op_generic(ioat, IOAT_OP_COPY, 8, 0, 0, callback_fn,
934 	    callback_arg, flags);
935 	if (desc == NULL)
936 		return (NULL);
937 
938 	hw_desc = desc->u.dma;
939 	hw_desc->u.control.null = 1;
940 	ioat_submit_single(ioat);
941 	return (&desc->bus_dmadesc);
942 }
943 
944 struct bus_dmadesc *
945 ioat_copy(bus_dmaengine_t dmaengine, bus_addr_t dst,
946     bus_addr_t src, bus_size_t len, bus_dmaengine_callback_t callback_fn,
947     void *callback_arg, uint32_t flags)
948 {
949 	struct ioat_dma_hw_descriptor *hw_desc;
950 	struct ioat_descriptor *desc;
951 	struct ioat_softc *ioat;
952 
953 	CTR0(KTR_IOAT, __func__);
954 	ioat = to_ioat_softc(dmaengine);
955 
956 	if (((src | dst) & (0xffffull << 48)) != 0) {
957 		ioat_log_message(0, "%s: High 16 bits of src/dst invalid\n",
958 		    __func__);
959 		return (NULL);
960 	}
961 
962 	desc = ioat_op_generic(ioat, IOAT_OP_COPY, len, src, dst, callback_fn,
963 	    callback_arg, flags);
964 	if (desc == NULL)
965 		return (NULL);
966 
967 	hw_desc = desc->u.dma;
968 	if (g_ioat_debug_level >= 3)
969 		dump_descriptor(hw_desc);
970 
971 	ioat_submit_single(ioat);
972 	return (&desc->bus_dmadesc);
973 }
974 
975 struct bus_dmadesc *
976 ioat_copy_8k_aligned(bus_dmaengine_t dmaengine, bus_addr_t dst1,
977     bus_addr_t dst2, bus_addr_t src1, bus_addr_t src2,
978     bus_dmaengine_callback_t callback_fn, void *callback_arg, uint32_t flags)
979 {
980 	struct ioat_dma_hw_descriptor *hw_desc;
981 	struct ioat_descriptor *desc;
982 	struct ioat_softc *ioat;
983 
984 	CTR0(KTR_IOAT, __func__);
985 	ioat = to_ioat_softc(dmaengine);
986 
987 	if (((src1 | src2 | dst1 | dst2) & (0xffffull << 48)) != 0) {
988 		ioat_log_message(0, "%s: High 16 bits of src/dst invalid\n",
989 		    __func__);
990 		return (NULL);
991 	}
992 	if (((src1 | src2 | dst1 | dst2) & PAGE_MASK) != 0) {
993 		ioat_log_message(0, "%s: Addresses must be page-aligned\n",
994 		    __func__);
995 		return (NULL);
996 	}
997 
998 	desc = ioat_op_generic(ioat, IOAT_OP_COPY, 2 * PAGE_SIZE, src1, dst1,
999 	    callback_fn, callback_arg, flags);
1000 	if (desc == NULL)
1001 		return (NULL);
1002 
1003 	hw_desc = desc->u.dma;
1004 	if (src2 != src1 + PAGE_SIZE) {
1005 		hw_desc->u.control.src_page_break = 1;
1006 		hw_desc->next_src_addr = src2;
1007 	}
1008 	if (dst2 != dst1 + PAGE_SIZE) {
1009 		hw_desc->u.control.dest_page_break = 1;
1010 		hw_desc->next_dest_addr = dst2;
1011 	}
1012 
1013 	if (g_ioat_debug_level >= 3)
1014 		dump_descriptor(hw_desc);
1015 
1016 	ioat_submit_single(ioat);
1017 	return (&desc->bus_dmadesc);
1018 }
1019 
1020 struct bus_dmadesc *
1021 ioat_copy_crc(bus_dmaengine_t dmaengine, bus_addr_t dst, bus_addr_t src,
1022     bus_size_t len, uint32_t *initialseed, bus_addr_t crcptr,
1023     bus_dmaengine_callback_t callback_fn, void *callback_arg, uint32_t flags)
1024 {
1025 	struct ioat_crc32_hw_descriptor *hw_desc;
1026 	struct ioat_descriptor *desc;
1027 	struct ioat_softc *ioat;
1028 	uint32_t teststore;
1029 	uint8_t op;
1030 
1031 	CTR0(KTR_IOAT, __func__);
1032 	ioat = to_ioat_softc(dmaengine);
1033 
1034 	if ((ioat->capabilities & IOAT_DMACAP_MOVECRC) == 0) {
1035 		ioat_log_message(0, "%s: Device lacks MOVECRC capability\n",
1036 		    __func__);
1037 		return (NULL);
1038 	}
1039 	if (((src | dst) & (0xffffffull << 40)) != 0) {
1040 		ioat_log_message(0, "%s: High 24 bits of src/dst invalid\n",
1041 		    __func__);
1042 		return (NULL);
1043 	}
1044 	teststore = (flags & _DMA_CRC_TESTSTORE);
1045 	if (teststore == _DMA_CRC_TESTSTORE) {
1046 		ioat_log_message(0, "%s: TEST and STORE invalid\n", __func__);
1047 		return (NULL);
1048 	}
1049 	if (teststore == 0 && (flags & DMA_CRC_INLINE) != 0) {
1050 		ioat_log_message(0, "%s: INLINE invalid without TEST or STORE\n",
1051 		    __func__);
1052 		return (NULL);
1053 	}
1054 
1055 	switch (teststore) {
1056 	case DMA_CRC_STORE:
1057 		op = IOAT_OP_MOVECRC_STORE;
1058 		break;
1059 	case DMA_CRC_TEST:
1060 		op = IOAT_OP_MOVECRC_TEST;
1061 		break;
1062 	default:
1063 		KASSERT(teststore == 0, ("bogus"));
1064 		op = IOAT_OP_MOVECRC;
1065 		break;
1066 	}
1067 
1068 	if ((flags & DMA_CRC_INLINE) == 0 &&
1069 	    (crcptr & (0xffffffull << 40)) != 0) {
1070 		ioat_log_message(0,
1071 		    "%s: High 24 bits of crcptr invalid\n", __func__);
1072 		return (NULL);
1073 	}
1074 
1075 	desc = ioat_op_generic(ioat, op, len, src, dst, callback_fn,
1076 	    callback_arg, flags & ~_DMA_CRC_FLAGS);
1077 	if (desc == NULL)
1078 		return (NULL);
1079 
1080 	hw_desc = desc->u.crc32;
1081 
1082 	if ((flags & DMA_CRC_INLINE) == 0)
1083 		hw_desc->crc_address = crcptr;
1084 	else
1085 		hw_desc->u.control.crc_location = 1;
1086 
1087 	if (initialseed != NULL) {
1088 		hw_desc->u.control.use_seed = 1;
1089 		hw_desc->seed = *initialseed;
1090 	}
1091 
1092 	if (g_ioat_debug_level >= 3)
1093 		dump_descriptor(hw_desc);
1094 
1095 	ioat_submit_single(ioat);
1096 	return (&desc->bus_dmadesc);
1097 }
1098 
1099 struct bus_dmadesc *
1100 ioat_crc(bus_dmaengine_t dmaengine, bus_addr_t src, bus_size_t len,
1101     uint32_t *initialseed, bus_addr_t crcptr,
1102     bus_dmaengine_callback_t callback_fn, void *callback_arg, uint32_t flags)
1103 {
1104 	struct ioat_crc32_hw_descriptor *hw_desc;
1105 	struct ioat_descriptor *desc;
1106 	struct ioat_softc *ioat;
1107 	uint32_t teststore;
1108 	uint8_t op;
1109 
1110 	CTR0(KTR_IOAT, __func__);
1111 	ioat = to_ioat_softc(dmaengine);
1112 
1113 	if ((ioat->capabilities & IOAT_DMACAP_CRC) == 0) {
1114 		ioat_log_message(0, "%s: Device lacks CRC capability\n",
1115 		    __func__);
1116 		return (NULL);
1117 	}
1118 	if ((src & (0xffffffull << 40)) != 0) {
1119 		ioat_log_message(0, "%s: High 24 bits of src invalid\n",
1120 		    __func__);
1121 		return (NULL);
1122 	}
1123 	teststore = (flags & _DMA_CRC_TESTSTORE);
1124 	if (teststore == _DMA_CRC_TESTSTORE) {
1125 		ioat_log_message(0, "%s: TEST and STORE invalid\n", __func__);
1126 		return (NULL);
1127 	}
1128 	if (teststore == 0 && (flags & DMA_CRC_INLINE) != 0) {
1129 		ioat_log_message(0, "%s: INLINE invalid without TEST or STORE\n",
1130 		    __func__);
1131 		return (NULL);
1132 	}
1133 
1134 	switch (teststore) {
1135 	case DMA_CRC_STORE:
1136 		op = IOAT_OP_CRC_STORE;
1137 		break;
1138 	case DMA_CRC_TEST:
1139 		op = IOAT_OP_CRC_TEST;
1140 		break;
1141 	default:
1142 		KASSERT(teststore == 0, ("bogus"));
1143 		op = IOAT_OP_CRC;
1144 		break;
1145 	}
1146 
1147 	if ((flags & DMA_CRC_INLINE) == 0 &&
1148 	    (crcptr & (0xffffffull << 40)) != 0) {
1149 		ioat_log_message(0,
1150 		    "%s: High 24 bits of crcptr invalid\n", __func__);
1151 		return (NULL);
1152 	}
1153 
1154 	desc = ioat_op_generic(ioat, op, len, src, 0, callback_fn,
1155 	    callback_arg, flags & ~_DMA_CRC_FLAGS);
1156 	if (desc == NULL)
1157 		return (NULL);
1158 
1159 	hw_desc = desc->u.crc32;
1160 
1161 	if ((flags & DMA_CRC_INLINE) == 0)
1162 		hw_desc->crc_address = crcptr;
1163 	else
1164 		hw_desc->u.control.crc_location = 1;
1165 
1166 	if (initialseed != NULL) {
1167 		hw_desc->u.control.use_seed = 1;
1168 		hw_desc->seed = *initialseed;
1169 	}
1170 
1171 	if (g_ioat_debug_level >= 3)
1172 		dump_descriptor(hw_desc);
1173 
1174 	ioat_submit_single(ioat);
1175 	return (&desc->bus_dmadesc);
1176 }
1177 
1178 struct bus_dmadesc *
1179 ioat_blockfill(bus_dmaengine_t dmaengine, bus_addr_t dst, uint64_t fillpattern,
1180     bus_size_t len, bus_dmaengine_callback_t callback_fn, void *callback_arg,
1181     uint32_t flags)
1182 {
1183 	struct ioat_fill_hw_descriptor *hw_desc;
1184 	struct ioat_descriptor *desc;
1185 	struct ioat_softc *ioat;
1186 
1187 	CTR0(KTR_IOAT, __func__);
1188 	ioat = to_ioat_softc(dmaengine);
1189 
1190 	if ((ioat->capabilities & IOAT_DMACAP_BFILL) == 0) {
1191 		ioat_log_message(0, "%s: Device lacks BFILL capability\n",
1192 		    __func__);
1193 		return (NULL);
1194 	}
1195 
1196 	if ((dst & (0xffffull << 48)) != 0) {
1197 		ioat_log_message(0, "%s: High 16 bits of dst invalid\n",
1198 		    __func__);
1199 		return (NULL);
1200 	}
1201 
1202 	desc = ioat_op_generic(ioat, IOAT_OP_FILL, len, fillpattern, dst,
1203 	    callback_fn, callback_arg, flags);
1204 	if (desc == NULL)
1205 		return (NULL);
1206 
1207 	hw_desc = desc->u.fill;
1208 	if (g_ioat_debug_level >= 3)
1209 		dump_descriptor(hw_desc);
1210 
1211 	ioat_submit_single(ioat);
1212 	return (&desc->bus_dmadesc);
1213 }
1214 
1215 /*
1216  * Ring Management
1217  */
1218 static inline uint32_t
1219 ioat_get_active(struct ioat_softc *ioat)
1220 {
1221 
1222 	return ((ioat->head - ioat->tail) & ((1 << ioat->ring_size_order) - 1));
1223 }
1224 
1225 static inline uint32_t
1226 ioat_get_ring_space(struct ioat_softc *ioat)
1227 {
1228 
1229 	return ((1 << ioat->ring_size_order) - ioat_get_active(ioat) - 1);
1230 }
1231 
1232 static struct ioat_descriptor *
1233 ioat_alloc_ring_entry(struct ioat_softc *ioat, int mflags)
1234 {
1235 	struct ioat_generic_hw_descriptor *hw_desc;
1236 	struct ioat_descriptor *desc;
1237 	int error, busdmaflag;
1238 
1239 	error = ENOMEM;
1240 	hw_desc = NULL;
1241 
1242 	if ((mflags & M_WAITOK) != 0)
1243 		busdmaflag = BUS_DMA_WAITOK;
1244 	else
1245 		busdmaflag = BUS_DMA_NOWAIT;
1246 
1247 	desc = malloc(sizeof(*desc), M_IOAT, mflags);
1248 	if (desc == NULL)
1249 		goto out;
1250 
1251 	bus_dmamem_alloc(ioat->hw_desc_tag, (void **)&hw_desc,
1252 	    BUS_DMA_ZERO | busdmaflag, &ioat->hw_desc_map);
1253 	if (hw_desc == NULL)
1254 		goto out;
1255 
1256 	memset(&desc->bus_dmadesc, 0, sizeof(desc->bus_dmadesc));
1257 	desc->u.generic = hw_desc;
1258 
1259 	error = bus_dmamap_load(ioat->hw_desc_tag, ioat->hw_desc_map, hw_desc,
1260 	    sizeof(*hw_desc), ioat_dmamap_cb, &desc->hw_desc_bus_addr,
1261 	    busdmaflag);
1262 	if (error)
1263 		goto out;
1264 
1265 out:
1266 	if (error) {
1267 		ioat_free_ring_entry(ioat, desc);
1268 		return (NULL);
1269 	}
1270 	return (desc);
1271 }
1272 
1273 static void
1274 ioat_free_ring_entry(struct ioat_softc *ioat, struct ioat_descriptor *desc)
1275 {
1276 
1277 	if (desc == NULL)
1278 		return;
1279 
1280 	if (desc->u.generic)
1281 		bus_dmamem_free(ioat->hw_desc_tag, desc->u.generic,
1282 		    ioat->hw_desc_map);
1283 	free(desc, M_IOAT);
1284 }
1285 
1286 /*
1287  * Reserves space in this IOAT descriptor ring by ensuring enough slots remain
1288  * for 'num_descs'.
1289  *
1290  * If mflags contains M_WAITOK, blocks until enough space is available.
1291  *
1292  * Returns zero on success, or an errno on error.  If num_descs is beyond the
1293  * maximum ring size, returns EINVAl; if allocation would block and mflags
1294  * contains M_NOWAIT, returns EAGAIN.
1295  *
1296  * Must be called with the submit_lock held; returns with the lock held.  The
1297  * lock may be dropped to allocate the ring.
1298  *
1299  * (The submit_lock is needed to add any entries to the ring, so callers are
1300  * assured enough room is available.)
1301  */
1302 static int
1303 ioat_reserve_space(struct ioat_softc *ioat, uint32_t num_descs, int mflags)
1304 {
1305 	struct ioat_descriptor **new_ring;
1306 	uint32_t order;
1307 	int error;
1308 
1309 	mtx_assert(&ioat->submit_lock, MA_OWNED);
1310 	error = 0;
1311 
1312 	if (num_descs < 1 || num_descs > (1 << IOAT_MAX_ORDER)) {
1313 		error = EINVAL;
1314 		goto out;
1315 	}
1316 	if (ioat->quiescing) {
1317 		error = ENXIO;
1318 		goto out;
1319 	}
1320 
1321 	for (;;) {
1322 		if (ioat_get_ring_space(ioat) >= num_descs)
1323 			goto out;
1324 
1325 		order = ioat->ring_size_order;
1326 		if (ioat->is_resize_pending || order == IOAT_MAX_ORDER) {
1327 			if ((mflags & M_WAITOK) != 0) {
1328 				msleep(&ioat->tail, &ioat->submit_lock, 0,
1329 				    "ioat_rsz", 0);
1330 				continue;
1331 			}
1332 
1333 			error = EAGAIN;
1334 			break;
1335 		}
1336 
1337 		ioat->is_resize_pending = TRUE;
1338 		for (;;) {
1339 			mtx_unlock(&ioat->submit_lock);
1340 
1341 			new_ring = ioat_prealloc_ring(ioat, 1 << (order + 1),
1342 			    TRUE, mflags);
1343 
1344 			mtx_lock(&ioat->submit_lock);
1345 			KASSERT(ioat->ring_size_order == order,
1346 			    ("is_resize_pending should protect order"));
1347 
1348 			if (new_ring == NULL) {
1349 				KASSERT((mflags & M_WAITOK) == 0,
1350 				    ("allocation failed"));
1351 				error = EAGAIN;
1352 				break;
1353 			}
1354 
1355 			error = ring_grow(ioat, order, new_ring);
1356 			if (error == 0)
1357 				break;
1358 		}
1359 		ioat->is_resize_pending = FALSE;
1360 		wakeup(&ioat->tail);
1361 		if (error)
1362 			break;
1363 	}
1364 
1365 out:
1366 	mtx_assert(&ioat->submit_lock, MA_OWNED);
1367 	return (error);
1368 }
1369 
1370 static struct ioat_descriptor **
1371 ioat_prealloc_ring(struct ioat_softc *ioat, uint32_t size, boolean_t need_dscr,
1372     int mflags)
1373 {
1374 	struct ioat_descriptor **ring;
1375 	uint32_t i;
1376 	int error;
1377 
1378 	KASSERT(size > 0 && powerof2(size), ("bogus size"));
1379 
1380 	ring = malloc(size * sizeof(*ring), M_IOAT, M_ZERO | mflags);
1381 	if (ring == NULL)
1382 		return (NULL);
1383 
1384 	if (need_dscr) {
1385 		error = ENOMEM;
1386 		for (i = size / 2; i < size; i++) {
1387 			ring[i] = ioat_alloc_ring_entry(ioat, mflags);
1388 			if (ring[i] == NULL)
1389 				goto out;
1390 			ring[i]->id = i;
1391 		}
1392 	}
1393 	error = 0;
1394 
1395 out:
1396 	if (error != 0 && ring != NULL) {
1397 		ioat_free_ring(ioat, size, ring);
1398 		ring = NULL;
1399 	}
1400 	return (ring);
1401 }
1402 
1403 static void
1404 ioat_free_ring(struct ioat_softc *ioat, uint32_t size,
1405     struct ioat_descriptor **ring)
1406 {
1407 	uint32_t i;
1408 
1409 	for (i = 0; i < size; i++) {
1410 		if (ring[i] != NULL)
1411 			ioat_free_ring_entry(ioat, ring[i]);
1412 	}
1413 	free(ring, M_IOAT);
1414 }
1415 
1416 static struct ioat_descriptor *
1417 ioat_get_ring_entry(struct ioat_softc *ioat, uint32_t index)
1418 {
1419 
1420 	return (ioat->ring[index % (1 << ioat->ring_size_order)]);
1421 }
1422 
1423 static int
1424 ring_grow(struct ioat_softc *ioat, uint32_t oldorder,
1425     struct ioat_descriptor **newring)
1426 {
1427 	struct ioat_descriptor *tmp, *next;
1428 	struct ioat_dma_hw_descriptor *hw;
1429 	uint32_t oldsize, newsize, head, tail, i, end;
1430 	int error;
1431 
1432 	CTR0(KTR_IOAT, __func__);
1433 
1434 	mtx_assert(&ioat->submit_lock, MA_OWNED);
1435 
1436 	if (oldorder != ioat->ring_size_order || oldorder >= IOAT_MAX_ORDER) {
1437 		error = EINVAL;
1438 		goto out;
1439 	}
1440 
1441 	oldsize = (1 << oldorder);
1442 	newsize = (1 << (oldorder + 1));
1443 
1444 	mtx_lock(&ioat->cleanup_lock);
1445 
1446 	head = ioat->head & (oldsize - 1);
1447 	tail = ioat->tail & (oldsize - 1);
1448 
1449 	/* Copy old descriptors to new ring */
1450 	for (i = 0; i < oldsize; i++)
1451 		newring[i] = ioat->ring[i];
1452 
1453 	/*
1454 	 * If head has wrapped but tail hasn't, we must swap some descriptors
1455 	 * around so that tail can increment directly to head.
1456 	 */
1457 	if (head < tail) {
1458 		for (i = 0; i <= head; i++) {
1459 			tmp = newring[oldsize + i];
1460 
1461 			newring[oldsize + i] = newring[i];
1462 			newring[oldsize + i]->id = oldsize + i;
1463 
1464 			newring[i] = tmp;
1465 			newring[i]->id = i;
1466 		}
1467 		head += oldsize;
1468 	}
1469 
1470 	KASSERT(head >= tail, ("invariants"));
1471 
1472 	/* Head didn't wrap; we only need to link in oldsize..newsize */
1473 	if (head < oldsize) {
1474 		i = oldsize - 1;
1475 		end = newsize;
1476 	} else {
1477 		/* Head did wrap; link newhead..newsize and 0..oldhead */
1478 		i = head;
1479 		end = newsize + (head - oldsize) + 1;
1480 	}
1481 
1482 	/*
1483 	 * Fix up hardware ring, being careful not to trample the active
1484 	 * section (tail -> head).
1485 	 */
1486 	for (; i < end; i++) {
1487 		KASSERT((i & (newsize - 1)) < tail ||
1488 		    (i & (newsize - 1)) >= head, ("trampling snake"));
1489 
1490 		next = newring[(i + 1) & (newsize - 1)];
1491 		hw = newring[i & (newsize - 1)]->u.dma;
1492 		hw->next = next->hw_desc_bus_addr;
1493 	}
1494 
1495 	free(ioat->ring, M_IOAT);
1496 	ioat->ring = newring;
1497 	ioat->ring_size_order = oldorder + 1;
1498 	ioat->tail = tail;
1499 	ioat->head = head;
1500 	error = 0;
1501 
1502 	mtx_unlock(&ioat->cleanup_lock);
1503 out:
1504 	if (error)
1505 		ioat_free_ring(ioat, (1 << (oldorder + 1)), newring);
1506 	return (error);
1507 }
1508 
1509 static int
1510 ring_shrink(struct ioat_softc *ioat, uint32_t oldorder,
1511     struct ioat_descriptor **newring)
1512 {
1513 	struct ioat_dma_hw_descriptor *hw;
1514 	struct ioat_descriptor *ent, *next;
1515 	uint32_t oldsize, newsize, current_idx, new_idx, i;
1516 	int error;
1517 
1518 	CTR0(KTR_IOAT, __func__);
1519 
1520 	mtx_assert(&ioat->submit_lock, MA_OWNED);
1521 
1522 	if (oldorder != ioat->ring_size_order || oldorder <= IOAT_MIN_ORDER) {
1523 		error = EINVAL;
1524 		goto out_unlocked;
1525 	}
1526 
1527 	oldsize = (1 << oldorder);
1528 	newsize = (1 << (oldorder - 1));
1529 
1530 	mtx_lock(&ioat->cleanup_lock);
1531 
1532 	/* Can't shrink below current active set! */
1533 	if (ioat_get_active(ioat) >= newsize) {
1534 		error = ENOMEM;
1535 		goto out;
1536 	}
1537 
1538 	/*
1539 	 * Copy current descriptors to the new ring, dropping the removed
1540 	 * descriptors.
1541 	 */
1542 	for (i = 0; i < newsize; i++) {
1543 		current_idx = (ioat->tail + i) & (oldsize - 1);
1544 		new_idx = (ioat->tail + i) & (newsize - 1);
1545 
1546 		newring[new_idx] = ioat->ring[current_idx];
1547 		newring[new_idx]->id = new_idx;
1548 	}
1549 
1550 	/* Free deleted descriptors */
1551 	for (i = newsize; i < oldsize; i++) {
1552 		ent = ioat_get_ring_entry(ioat, ioat->tail + i);
1553 		ioat_free_ring_entry(ioat, ent);
1554 	}
1555 
1556 	/* Fix up hardware ring. */
1557 	hw = newring[(ioat->tail + newsize - 1) & (newsize - 1)]->u.dma;
1558 	next = newring[(ioat->tail + newsize) & (newsize - 1)];
1559 	hw->next = next->hw_desc_bus_addr;
1560 
1561 	free(ioat->ring, M_IOAT);
1562 	ioat->ring = newring;
1563 	ioat->ring_size_order = oldorder - 1;
1564 	error = 0;
1565 
1566 out:
1567 	mtx_unlock(&ioat->cleanup_lock);
1568 out_unlocked:
1569 	if (error)
1570 		ioat_free_ring(ioat, (1 << (oldorder - 1)), newring);
1571 	return (error);
1572 }
1573 
1574 static void
1575 ioat_halted_debug(struct ioat_softc *ioat, uint32_t chanerr)
1576 {
1577 	struct ioat_descriptor *desc;
1578 
1579 	ioat_log_message(0, "Channel halted (%b)\n", (int)chanerr,
1580 	    IOAT_CHANERR_STR);
1581 	if (chanerr == 0)
1582 		return;
1583 
1584 	mtx_assert(&ioat->cleanup_lock, MA_OWNED);
1585 
1586 	desc = ioat_get_ring_entry(ioat, ioat->tail + 0);
1587 	dump_descriptor(desc->u.raw);
1588 
1589 	desc = ioat_get_ring_entry(ioat, ioat->tail + 1);
1590 	dump_descriptor(desc->u.raw);
1591 }
1592 
1593 static void
1594 ioat_timer_callback(void *arg)
1595 {
1596 	struct ioat_descriptor **newring;
1597 	struct ioat_softc *ioat;
1598 	uint32_t order;
1599 
1600 	ioat = arg;
1601 	ioat_log_message(1, "%s\n", __func__);
1602 
1603 	if (ioat->is_completion_pending) {
1604 		ioat_process_events(ioat);
1605 		return;
1606 	}
1607 
1608 	/* Slowly scale the ring down if idle. */
1609 	mtx_lock(&ioat->submit_lock);
1610 	order = ioat->ring_size_order;
1611 	if (ioat->is_resize_pending || order == IOAT_MIN_ORDER) {
1612 		mtx_unlock(&ioat->submit_lock);
1613 		goto out;
1614 	}
1615 	ioat->is_resize_pending = TRUE;
1616 	mtx_unlock(&ioat->submit_lock);
1617 
1618 	newring = ioat_prealloc_ring(ioat, 1 << (order - 1), FALSE,
1619 	    M_NOWAIT);
1620 
1621 	mtx_lock(&ioat->submit_lock);
1622 	KASSERT(ioat->ring_size_order == order,
1623 	    ("resize_pending protects order"));
1624 
1625 	if (newring != NULL)
1626 		ring_shrink(ioat, order, newring);
1627 
1628 	ioat->is_resize_pending = FALSE;
1629 	mtx_unlock(&ioat->submit_lock);
1630 
1631 out:
1632 	if (ioat->ring_size_order > IOAT_MIN_ORDER)
1633 		callout_reset(&ioat->timer, 10 * hz,
1634 		    ioat_timer_callback, ioat);
1635 }
1636 
1637 /*
1638  * Support Functions
1639  */
1640 static void
1641 ioat_submit_single(struct ioat_softc *ioat)
1642 {
1643 
1644 	ioat_get(ioat, IOAT_ACTIVE_DESCR_REF);
1645 	atomic_add_rel_int(&ioat->head, 1);
1646 	atomic_add_rel_int(&ioat->hw_head, 1);
1647 
1648 	if (!ioat->is_completion_pending) {
1649 		ioat->is_completion_pending = TRUE;
1650 		callout_reset(&ioat->timer, IOAT_INTR_TIMO,
1651 		    ioat_timer_callback, ioat);
1652 	}
1653 
1654 	ioat->stats.descriptors_submitted++;
1655 }
1656 
1657 static int
1658 ioat_reset_hw(struct ioat_softc *ioat)
1659 {
1660 	uint64_t status;
1661 	uint32_t chanerr;
1662 	unsigned timeout;
1663 	int error;
1664 
1665 	mtx_lock(IOAT_REFLK);
1666 	ioat->quiescing = TRUE;
1667 	ioat_drain_locked(ioat);
1668 	mtx_unlock(IOAT_REFLK);
1669 
1670 	status = ioat_get_chansts(ioat);
1671 	if (is_ioat_active(status) || is_ioat_idle(status))
1672 		ioat_suspend(ioat);
1673 
1674 	/* Wait at most 20 ms */
1675 	for (timeout = 0; (is_ioat_active(status) || is_ioat_idle(status)) &&
1676 	    timeout < 20; timeout++) {
1677 		DELAY(1000);
1678 		status = ioat_get_chansts(ioat);
1679 	}
1680 	if (timeout == 20) {
1681 		error = ETIMEDOUT;
1682 		goto out;
1683 	}
1684 
1685 	KASSERT(ioat_get_active(ioat) == 0, ("active after quiesce"));
1686 
1687 	chanerr = ioat_read_4(ioat, IOAT_CHANERR_OFFSET);
1688 	ioat_write_4(ioat, IOAT_CHANERR_OFFSET, chanerr);
1689 
1690 	/*
1691 	 * IOAT v3 workaround - CHANERRMSK_INT with 3E07h to masks out errors
1692 	 *  that can cause stability issues for IOAT v3.
1693 	 */
1694 	pci_write_config(ioat->device, IOAT_CFG_CHANERRMASK_INT_OFFSET, 0x3e07,
1695 	    4);
1696 	chanerr = pci_read_config(ioat->device, IOAT_CFG_CHANERR_INT_OFFSET, 4);
1697 	pci_write_config(ioat->device, IOAT_CFG_CHANERR_INT_OFFSET, chanerr, 4);
1698 
1699 	/*
1700 	 * BDXDE and BWD models reset MSI-X registers on device reset.
1701 	 * Save/restore their contents manually.
1702 	 */
1703 	if (ioat_model_resets_msix(ioat)) {
1704 		ioat_log_message(1, "device resets MSI-X registers; saving\n");
1705 		pci_save_state(ioat->device);
1706 	}
1707 
1708 	ioat_reset(ioat);
1709 
1710 	/* Wait at most 20 ms */
1711 	for (timeout = 0; ioat_reset_pending(ioat) && timeout < 20; timeout++)
1712 		DELAY(1000);
1713 	if (timeout == 20) {
1714 		error = ETIMEDOUT;
1715 		goto out;
1716 	}
1717 
1718 	if (ioat_model_resets_msix(ioat)) {
1719 		ioat_log_message(1, "device resets registers; restored\n");
1720 		pci_restore_state(ioat->device);
1721 	}
1722 
1723 	/* Reset attempts to return the hardware to "halted." */
1724 	status = ioat_get_chansts(ioat);
1725 	if (is_ioat_active(status) || is_ioat_idle(status)) {
1726 		/* So this really shouldn't happen... */
1727 		ioat_log_message(0, "Device is active after a reset?\n");
1728 		ioat_write_chanctrl(ioat, IOAT_CHANCTRL_RUN);
1729 		error = 0;
1730 		goto out;
1731 	}
1732 
1733 	chanerr = ioat_read_4(ioat, IOAT_CHANERR_OFFSET);
1734 	if (chanerr != 0) {
1735 		mtx_lock(&ioat->cleanup_lock);
1736 		ioat_halted_debug(ioat, chanerr);
1737 		mtx_unlock(&ioat->cleanup_lock);
1738 		error = EIO;
1739 		goto out;
1740 	}
1741 
1742 	/*
1743 	 * Bring device back online after reset.  Writing CHAINADDR brings the
1744 	 * device back to active.
1745 	 *
1746 	 * The internal ring counter resets to zero, so we have to start over
1747 	 * at zero as well.
1748 	 */
1749 	ioat->tail = ioat->head = ioat->hw_head = 0;
1750 	ioat->last_seen = 0;
1751 
1752 	ioat_write_chanctrl(ioat, IOAT_CHANCTRL_RUN);
1753 	ioat_write_chancmp(ioat, ioat->comp_update_bus_addr);
1754 	ioat_write_chainaddr(ioat, ioat->ring[0]->hw_desc_bus_addr);
1755 	error = 0;
1756 
1757 out:
1758 	mtx_lock(IOAT_REFLK);
1759 	ioat->quiescing = FALSE;
1760 	wakeup(&ioat->quiescing);
1761 	mtx_unlock(IOAT_REFLK);
1762 
1763 	if (error == 0)
1764 		error = ioat_start_channel(ioat);
1765 
1766 	return (error);
1767 }
1768 
1769 static int
1770 sysctl_handle_chansts(SYSCTL_HANDLER_ARGS)
1771 {
1772 	struct ioat_softc *ioat;
1773 	struct sbuf sb;
1774 	uint64_t status;
1775 	int error;
1776 
1777 	ioat = arg1;
1778 
1779 	status = ioat_get_chansts(ioat) & IOAT_CHANSTS_STATUS;
1780 
1781 	sbuf_new_for_sysctl(&sb, NULL, 256, req);
1782 	switch (status) {
1783 	case IOAT_CHANSTS_ACTIVE:
1784 		sbuf_printf(&sb, "ACTIVE");
1785 		break;
1786 	case IOAT_CHANSTS_IDLE:
1787 		sbuf_printf(&sb, "IDLE");
1788 		break;
1789 	case IOAT_CHANSTS_SUSPENDED:
1790 		sbuf_printf(&sb, "SUSPENDED");
1791 		break;
1792 	case IOAT_CHANSTS_HALTED:
1793 		sbuf_printf(&sb, "HALTED");
1794 		break;
1795 	case IOAT_CHANSTS_ARMED:
1796 		sbuf_printf(&sb, "ARMED");
1797 		break;
1798 	default:
1799 		sbuf_printf(&sb, "UNKNOWN");
1800 		break;
1801 	}
1802 	error = sbuf_finish(&sb);
1803 	sbuf_delete(&sb);
1804 
1805 	if (error != 0 || req->newptr == NULL)
1806 		return (error);
1807 	return (EINVAL);
1808 }
1809 
1810 static int
1811 sysctl_handle_dpi(SYSCTL_HANDLER_ARGS)
1812 {
1813 	struct ioat_softc *ioat;
1814 	struct sbuf sb;
1815 #define	PRECISION	"1"
1816 	const uintmax_t factor = 10;
1817 	uintmax_t rate;
1818 	int error;
1819 
1820 	ioat = arg1;
1821 	sbuf_new_for_sysctl(&sb, NULL, 16, req);
1822 
1823 	if (ioat->stats.interrupts == 0) {
1824 		sbuf_printf(&sb, "NaN");
1825 		goto out;
1826 	}
1827 	rate = ioat->stats.descriptors_processed * factor /
1828 	    ioat->stats.interrupts;
1829 	sbuf_printf(&sb, "%ju.%." PRECISION "ju", rate / factor,
1830 	    rate % factor);
1831 #undef	PRECISION
1832 out:
1833 	error = sbuf_finish(&sb);
1834 	sbuf_delete(&sb);
1835 	if (error != 0 || req->newptr == NULL)
1836 		return (error);
1837 	return (EINVAL);
1838 }
1839 
1840 static int
1841 sysctl_handle_error(SYSCTL_HANDLER_ARGS)
1842 {
1843 	struct ioat_descriptor *desc;
1844 	struct ioat_softc *ioat;
1845 	int error, arg;
1846 
1847 	ioat = arg1;
1848 
1849 	arg = 0;
1850 	error = SYSCTL_OUT(req, &arg, sizeof(arg));
1851 	if (error != 0 || req->newptr == NULL)
1852 		return (error);
1853 
1854 	error = SYSCTL_IN(req, &arg, sizeof(arg));
1855 	if (error != 0)
1856 		return (error);
1857 
1858 	if (arg != 0) {
1859 		ioat_acquire(&ioat->dmaengine);
1860 		desc = ioat_op_generic(ioat, IOAT_OP_COPY, 1,
1861 		    0xffff000000000000ull, 0xffff000000000000ull, NULL, NULL,
1862 		    0);
1863 		if (desc == NULL)
1864 			error = ENOMEM;
1865 		else
1866 			ioat_submit_single(ioat);
1867 		ioat_release(&ioat->dmaengine);
1868 	}
1869 	return (error);
1870 }
1871 
1872 static int
1873 sysctl_handle_reset(SYSCTL_HANDLER_ARGS)
1874 {
1875 	struct ioat_softc *ioat;
1876 	int error, arg;
1877 
1878 	ioat = arg1;
1879 
1880 	arg = 0;
1881 	error = SYSCTL_OUT(req, &arg, sizeof(arg));
1882 	if (error != 0 || req->newptr == NULL)
1883 		return (error);
1884 
1885 	error = SYSCTL_IN(req, &arg, sizeof(arg));
1886 	if (error != 0)
1887 		return (error);
1888 
1889 	if (arg != 0)
1890 		error = ioat_reset_hw(ioat);
1891 
1892 	return (error);
1893 }
1894 
1895 static void
1896 dump_descriptor(void *hw_desc)
1897 {
1898 	int i, j;
1899 
1900 	for (i = 0; i < 2; i++) {
1901 		for (j = 0; j < 8; j++)
1902 			printf("%08x ", ((uint32_t *)hw_desc)[i * 8 + j]);
1903 		printf("\n");
1904 	}
1905 }
1906 
1907 static void
1908 ioat_setup_sysctl(device_t device)
1909 {
1910 	struct sysctl_oid_list *par, *statpar, *state, *hammer;
1911 	struct sysctl_ctx_list *ctx;
1912 	struct sysctl_oid *tree, *tmp;
1913 	struct ioat_softc *ioat;
1914 
1915 	ioat = DEVICE2SOFTC(device);
1916 	ctx = device_get_sysctl_ctx(device);
1917 	tree = device_get_sysctl_tree(device);
1918 	par = SYSCTL_CHILDREN(tree);
1919 
1920 	SYSCTL_ADD_INT(ctx, par, OID_AUTO, "version", CTLFLAG_RD,
1921 	    &ioat->version, 0, "HW version (0xMM form)");
1922 	SYSCTL_ADD_UINT(ctx, par, OID_AUTO, "max_xfer_size", CTLFLAG_RD,
1923 	    &ioat->max_xfer_size, 0, "HW maximum transfer size");
1924 	SYSCTL_ADD_INT(ctx, par, OID_AUTO, "intrdelay_supported", CTLFLAG_RD,
1925 	    &ioat->intrdelay_supported, 0, "Is INTRDELAY supported");
1926 	SYSCTL_ADD_U16(ctx, par, OID_AUTO, "intrdelay_max", CTLFLAG_RD,
1927 	    &ioat->intrdelay_max, 0,
1928 	    "Maximum configurable INTRDELAY on this channel (microseconds)");
1929 
1930 	tmp = SYSCTL_ADD_NODE(ctx, par, OID_AUTO, "state", CTLFLAG_RD, NULL,
1931 	    "IOAT channel internal state");
1932 	state = SYSCTL_CHILDREN(tmp);
1933 
1934 	SYSCTL_ADD_UINT(ctx, state, OID_AUTO, "ring_size_order", CTLFLAG_RD,
1935 	    &ioat->ring_size_order, 0, "SW descriptor ring size order");
1936 	SYSCTL_ADD_UINT(ctx, state, OID_AUTO, "head", CTLFLAG_RD, &ioat->head,
1937 	    0, "SW descriptor head pointer index");
1938 	SYSCTL_ADD_UINT(ctx, state, OID_AUTO, "tail", CTLFLAG_RD, &ioat->tail,
1939 	    0, "SW descriptor tail pointer index");
1940 	SYSCTL_ADD_UINT(ctx, state, OID_AUTO, "hw_head", CTLFLAG_RD,
1941 	    &ioat->hw_head, 0, "HW DMACOUNT");
1942 
1943 	SYSCTL_ADD_UQUAD(ctx, state, OID_AUTO, "last_completion", CTLFLAG_RD,
1944 	    ioat->comp_update, "HW addr of last completion");
1945 
1946 	SYSCTL_ADD_INT(ctx, state, OID_AUTO, "is_resize_pending", CTLFLAG_RD,
1947 	    &ioat->is_resize_pending, 0, "resize pending");
1948 	SYSCTL_ADD_INT(ctx, state, OID_AUTO, "is_completion_pending",
1949 	    CTLFLAG_RD, &ioat->is_completion_pending, 0, "completion pending");
1950 	SYSCTL_ADD_INT(ctx, state, OID_AUTO, "is_reset_pending", CTLFLAG_RD,
1951 	    &ioat->is_reset_pending, 0, "reset pending");
1952 	SYSCTL_ADD_INT(ctx, state, OID_AUTO, "is_channel_running", CTLFLAG_RD,
1953 	    &ioat->is_channel_running, 0, "channel running");
1954 
1955 	SYSCTL_ADD_PROC(ctx, state, OID_AUTO, "chansts",
1956 	    CTLTYPE_STRING | CTLFLAG_RD, ioat, 0, sysctl_handle_chansts, "A",
1957 	    "String of the channel status");
1958 
1959 	SYSCTL_ADD_U16(ctx, state, OID_AUTO, "intrdelay", CTLFLAG_RD,
1960 	    &ioat->cached_intrdelay, 0,
1961 	    "Current INTRDELAY on this channel (cached, microseconds)");
1962 
1963 	tmp = SYSCTL_ADD_NODE(ctx, par, OID_AUTO, "hammer", CTLFLAG_RD, 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, ioat, 0, sysctl_handle_reset, "I",
1969 	    "Set to non-zero to reset the hardware");
1970 	SYSCTL_ADD_PROC(ctx, hammer, OID_AUTO, "force_hw_error",
1971 	    CTLTYPE_INT | CTLFLAG_RW, ioat, 0, sysctl_handle_error, "I",
1972 	    "Set to non-zero to inject a recoverable hardware error");
1973 
1974 	tmp = SYSCTL_ADD_NODE(ctx, par, OID_AUTO, "stats", CTLFLAG_RD, NULL,
1975 	    "IOAT channel statistics");
1976 	statpar = SYSCTL_CHILDREN(tmp);
1977 
1978 	SYSCTL_ADD_UQUAD(ctx, statpar, OID_AUTO, "interrupts", CTLFLAG_RW,
1979 	    &ioat->stats.interrupts,
1980 	    "Number of interrupts processed on this channel");
1981 	SYSCTL_ADD_UQUAD(ctx, statpar, OID_AUTO, "descriptors", CTLFLAG_RW,
1982 	    &ioat->stats.descriptors_processed,
1983 	    "Number of descriptors processed on this channel");
1984 	SYSCTL_ADD_UQUAD(ctx, statpar, OID_AUTO, "submitted", CTLFLAG_RW,
1985 	    &ioat->stats.descriptors_submitted,
1986 	    "Number of descriptors submitted to this channel");
1987 	SYSCTL_ADD_UQUAD(ctx, statpar, OID_AUTO, "errored", CTLFLAG_RW,
1988 	    &ioat->stats.descriptors_error,
1989 	    "Number of descriptors failed by channel errors");
1990 	SYSCTL_ADD_U32(ctx, statpar, OID_AUTO, "halts", CTLFLAG_RW,
1991 	    &ioat->stats.channel_halts, 0,
1992 	    "Number of times the channel has halted");
1993 	SYSCTL_ADD_U32(ctx, statpar, OID_AUTO, "last_halt_chanerr", CTLFLAG_RW,
1994 	    &ioat->stats.last_halt_chanerr, 0,
1995 	    "The raw CHANERR when the channel was last halted");
1996 
1997 	SYSCTL_ADD_PROC(ctx, statpar, OID_AUTO, "desc_per_interrupt",
1998 	    CTLTYPE_STRING | CTLFLAG_RD, ioat, 0, sysctl_handle_dpi, "A",
1999 	    "Descriptors per interrupt");
2000 }
2001 
2002 static inline struct ioat_softc *
2003 ioat_get(struct ioat_softc *ioat, enum ioat_ref_kind kind)
2004 {
2005 	uint32_t old;
2006 
2007 	KASSERT(kind < IOAT_NUM_REF_KINDS, ("bogus"));
2008 
2009 	old = atomic_fetchadd_32(&ioat->refcnt, 1);
2010 	KASSERT(old < UINT32_MAX, ("refcnt overflow"));
2011 
2012 #ifdef INVARIANTS
2013 	old = atomic_fetchadd_32(&ioat->refkinds[kind], 1);
2014 	KASSERT(old < UINT32_MAX, ("refcnt kind overflow"));
2015 #endif
2016 
2017 	return (ioat);
2018 }
2019 
2020 static inline void
2021 ioat_putn(struct ioat_softc *ioat, uint32_t n, enum ioat_ref_kind kind)
2022 {
2023 
2024 	_ioat_putn(ioat, n, kind, FALSE);
2025 }
2026 
2027 static inline void
2028 ioat_putn_locked(struct ioat_softc *ioat, uint32_t n, enum ioat_ref_kind kind)
2029 {
2030 
2031 	_ioat_putn(ioat, n, kind, TRUE);
2032 }
2033 
2034 static inline void
2035 _ioat_putn(struct ioat_softc *ioat, uint32_t n, enum ioat_ref_kind kind,
2036     boolean_t locked)
2037 {
2038 	uint32_t old;
2039 
2040 	KASSERT(kind < IOAT_NUM_REF_KINDS, ("bogus"));
2041 
2042 	if (n == 0)
2043 		return;
2044 
2045 #ifdef INVARIANTS
2046 	old = atomic_fetchadd_32(&ioat->refkinds[kind], -n);
2047 	KASSERT(old >= n, ("refcnt kind underflow"));
2048 #endif
2049 
2050 	/* Skip acquiring the lock if resulting refcnt > 0. */
2051 	for (;;) {
2052 		old = ioat->refcnt;
2053 		if (old <= n)
2054 			break;
2055 		if (atomic_cmpset_32(&ioat->refcnt, old, old - n))
2056 			return;
2057 	}
2058 
2059 	if (locked)
2060 		mtx_assert(IOAT_REFLK, MA_OWNED);
2061 	else
2062 		mtx_lock(IOAT_REFLK);
2063 
2064 	old = atomic_fetchadd_32(&ioat->refcnt, -n);
2065 	KASSERT(old >= n, ("refcnt error"));
2066 
2067 	if (old == n)
2068 		wakeup(IOAT_REFLK);
2069 	if (!locked)
2070 		mtx_unlock(IOAT_REFLK);
2071 }
2072 
2073 static inline void
2074 ioat_put(struct ioat_softc *ioat, enum ioat_ref_kind kind)
2075 {
2076 
2077 	ioat_putn(ioat, 1, kind);
2078 }
2079 
2080 static void
2081 ioat_drain_locked(struct ioat_softc *ioat)
2082 {
2083 
2084 	mtx_assert(IOAT_REFLK, MA_OWNED);
2085 	while (ioat->refcnt > 0)
2086 		msleep(IOAT_REFLK, IOAT_REFLK, 0, "ioat_drain", 0);
2087 }
2088