xref: /freebsd/sys/dev/ioat/ioat.c (revision 94bb24b38701354b6a34607982483227823de864)
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 unsigned ioat_channel_index = 0;
156 SYSCTL_UINT(_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(0, "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 
468 	ring = ioat->ring;
469 	for (i = 0; i < num_descriptors; i++) {
470 		ring[i] = ioat_alloc_ring_entry(ioat, M_WAITOK);
471 		if (ring[i] == NULL)
472 			return (ENOMEM);
473 
474 		ring[i]->id = i;
475 	}
476 
477 	for (i = 0; i < num_descriptors - 1; i++) {
478 		next = ring[i + 1];
479 		dma_hw_desc = ring[i]->u.dma;
480 
481 		dma_hw_desc->next = next->hw_desc_bus_addr;
482 	}
483 
484 	ring[i]->u.dma->next = ring[0]->hw_desc_bus_addr;
485 
486 	ioat->head = ioat->hw_head = 0;
487 	ioat->tail = 0;
488 	ioat->last_seen = 0;
489 	return (0);
490 }
491 
492 static int
493 ioat_map_pci_bar(struct ioat_softc *ioat)
494 {
495 
496 	ioat->pci_resource_id = PCIR_BAR(0);
497 	ioat->pci_resource = bus_alloc_resource_any(ioat->device,
498 	    SYS_RES_MEMORY, &ioat->pci_resource_id, RF_ACTIVE);
499 
500 	if (ioat->pci_resource == NULL) {
501 		ioat_log_message(0, "unable to allocate pci resource\n");
502 		return (ENODEV);
503 	}
504 
505 	ioat->pci_bus_tag = rman_get_bustag(ioat->pci_resource);
506 	ioat->pci_bus_handle = rman_get_bushandle(ioat->pci_resource);
507 	return (0);
508 }
509 
510 static void
511 ioat_comp_update_map(void *arg, bus_dma_segment_t *seg, int nseg, int error)
512 {
513 	struct ioat_softc *ioat = arg;
514 
515 	KASSERT(error == 0, ("%s: error:%d", __func__, error));
516 	ioat->comp_update_bus_addr = seg[0].ds_addr;
517 }
518 
519 static void
520 ioat_dmamap_cb(void *arg, bus_dma_segment_t *segs, int nseg, int error)
521 {
522 	bus_addr_t *baddr;
523 
524 	KASSERT(error == 0, ("%s: error:%d", __func__, error));
525 	baddr = arg;
526 	*baddr = segs->ds_addr;
527 }
528 
529 /*
530  * Interrupt setup and handlers
531  */
532 static int
533 ioat_setup_intr(struct ioat_softc *ioat)
534 {
535 	uint32_t num_vectors;
536 	int error;
537 	boolean_t use_msix;
538 	boolean_t force_legacy_interrupts;
539 
540 	use_msix = FALSE;
541 	force_legacy_interrupts = FALSE;
542 
543 	if (!g_force_legacy_interrupts && pci_msix_count(ioat->device) >= 1) {
544 		num_vectors = 1;
545 		pci_alloc_msix(ioat->device, &num_vectors);
546 		if (num_vectors == 1)
547 			use_msix = TRUE;
548 	}
549 
550 	if (use_msix) {
551 		ioat->rid = 1;
552 		ioat->res = bus_alloc_resource_any(ioat->device, SYS_RES_IRQ,
553 		    &ioat->rid, RF_ACTIVE);
554 	} else {
555 		ioat->rid = 0;
556 		ioat->res = bus_alloc_resource_any(ioat->device, SYS_RES_IRQ,
557 		    &ioat->rid, RF_SHAREABLE | RF_ACTIVE);
558 	}
559 	if (ioat->res == NULL) {
560 		ioat_log_message(0, "bus_alloc_resource failed\n");
561 		return (ENOMEM);
562 	}
563 
564 	ioat->tag = NULL;
565 	error = bus_setup_intr(ioat->device, ioat->res, INTR_MPSAFE |
566 	    INTR_TYPE_MISC, NULL, ioat_interrupt_handler, ioat, &ioat->tag);
567 	if (error != 0) {
568 		ioat_log_message(0, "bus_setup_intr failed\n");
569 		return (error);
570 	}
571 
572 	ioat_write_intrctrl(ioat, IOAT_INTRCTRL_MASTER_INT_EN);
573 	return (0);
574 }
575 
576 static boolean_t
577 ioat_model_resets_msix(struct ioat_softc *ioat)
578 {
579 	u_int32_t pciid;
580 
581 	pciid = pci_get_devid(ioat->device);
582 	switch (pciid) {
583 		/* BWD: */
584 	case 0x0c508086:
585 	case 0x0c518086:
586 	case 0x0c528086:
587 	case 0x0c538086:
588 		/* BDXDE: */
589 	case 0x6f508086:
590 	case 0x6f518086:
591 	case 0x6f528086:
592 	case 0x6f538086:
593 		return (TRUE);
594 	}
595 
596 	return (FALSE);
597 }
598 
599 static void
600 ioat_interrupt_handler(void *arg)
601 {
602 	struct ioat_softc *ioat = arg;
603 
604 	ioat->stats.interrupts++;
605 	ioat_process_events(ioat);
606 }
607 
608 static int
609 chanerr_to_errno(uint32_t chanerr)
610 {
611 
612 	if (chanerr == 0)
613 		return (0);
614 	if ((chanerr & (IOAT_CHANERR_XSADDERR | IOAT_CHANERR_XDADDERR)) != 0)
615 		return (EFAULT);
616 	if ((chanerr & (IOAT_CHANERR_RDERR | IOAT_CHANERR_WDERR)) != 0)
617 		return (EIO);
618 	/* This one is probably our fault: */
619 	if ((chanerr & IOAT_CHANERR_NDADDERR) != 0)
620 		return (EIO);
621 	return (EIO);
622 }
623 
624 static void
625 ioat_process_events(struct ioat_softc *ioat)
626 {
627 	struct ioat_descriptor *desc;
628 	struct bus_dmadesc *dmadesc;
629 	uint64_t comp_update, status;
630 	uint32_t completed, chanerr;
631 	int error;
632 
633 	mtx_lock(&ioat->cleanup_lock);
634 
635 	completed = 0;
636 	comp_update = *ioat->comp_update;
637 	status = comp_update & IOAT_CHANSTS_COMPLETED_DESCRIPTOR_MASK;
638 
639 	CTR0(KTR_IOAT, __func__);
640 
641 	if (status == ioat->last_seen) {
642 		/*
643 		 * If we landed in process_events and nothing has been
644 		 * completed, check for a timeout due to channel halt.
645 		 */
646 		comp_update = ioat_get_chansts(ioat);
647 		goto out;
648 	}
649 
650 	while (1) {
651 		desc = ioat_get_ring_entry(ioat, ioat->tail);
652 		dmadesc = &desc->bus_dmadesc;
653 		CTR1(KTR_IOAT, "completing desc %d", ioat->tail);
654 
655 		if (dmadesc->callback_fn != NULL)
656 			dmadesc->callback_fn(dmadesc->callback_arg, 0);
657 
658 		completed++;
659 		ioat->tail++;
660 		if (desc->hw_desc_bus_addr == status)
661 			break;
662 	}
663 
664 	ioat->last_seen = desc->hw_desc_bus_addr;
665 
666 	if (ioat->head == ioat->tail) {
667 		ioat->is_completion_pending = FALSE;
668 		callout_reset(&ioat->timer, IOAT_INTR_TIMO,
669 		    ioat_timer_callback, ioat);
670 	}
671 
672 	ioat->stats.descriptors_processed += completed;
673 
674 out:
675 	ioat_write_chanctrl(ioat, IOAT_CHANCTRL_RUN);
676 	mtx_unlock(&ioat->cleanup_lock);
677 
678 	if (completed != 0) {
679 		ioat_putn(ioat, completed, IOAT_ACTIVE_DESCR_REF);
680 		wakeup(&ioat->tail);
681 	}
682 
683 	if (!is_ioat_halted(comp_update) && !is_ioat_suspended(comp_update))
684 		return;
685 
686 	ioat->stats.channel_halts++;
687 
688 	/*
689 	 * Fatal programming error on this DMA channel.  Flush any outstanding
690 	 * work with error status and restart the engine.
691 	 */
692 	ioat_log_message(0, "Channel halted due to fatal programming error\n");
693 	mtx_lock(&ioat->submit_lock);
694 	mtx_lock(&ioat->cleanup_lock);
695 	ioat->quiescing = TRUE;
696 
697 	chanerr = ioat_read_4(ioat, IOAT_CHANERR_OFFSET);
698 	ioat_halted_debug(ioat, chanerr);
699 	ioat->stats.last_halt_chanerr = chanerr;
700 
701 	while (ioat_get_active(ioat) > 0) {
702 		desc = ioat_get_ring_entry(ioat, ioat->tail);
703 		dmadesc = &desc->bus_dmadesc;
704 		CTR1(KTR_IOAT, "completing err desc %d", ioat->tail);
705 
706 		if (dmadesc->callback_fn != NULL)
707 			dmadesc->callback_fn(dmadesc->callback_arg,
708 			    chanerr_to_errno(chanerr));
709 
710 		ioat_putn_locked(ioat, 1, IOAT_ACTIVE_DESCR_REF);
711 		ioat->tail++;
712 		ioat->stats.descriptors_processed++;
713 		ioat->stats.descriptors_error++;
714 	}
715 
716 	/* Clear error status */
717 	ioat_write_4(ioat, IOAT_CHANERR_OFFSET, chanerr);
718 
719 	mtx_unlock(&ioat->cleanup_lock);
720 	mtx_unlock(&ioat->submit_lock);
721 
722 	ioat_log_message(0, "Resetting channel to recover from error\n");
723 	error = taskqueue_enqueue(taskqueue_thread, &ioat->reset_task);
724 	KASSERT(error == 0,
725 	    ("%s: taskqueue_enqueue failed: %d", __func__, error));
726 }
727 
728 static void
729 ioat_reset_hw_task(void *ctx, int pending __unused)
730 {
731 	struct ioat_softc *ioat;
732 	int error;
733 
734 	ioat = ctx;
735 	ioat_log_message(1, "%s: Resetting channel\n", __func__);
736 
737 	error = ioat_reset_hw(ioat);
738 	KASSERT(error == 0, ("%s: reset failed: %d", __func__, error));
739 	(void)error;
740 }
741 
742 /*
743  * User API functions
744  */
745 unsigned
746 ioat_get_nchannels(void)
747 {
748 
749 	return (ioat_channel_index);
750 }
751 
752 bus_dmaengine_t
753 ioat_get_dmaengine(uint32_t index, int flags)
754 {
755 	struct ioat_softc *ioat;
756 
757 	KASSERT((flags & ~(M_NOWAIT | M_WAITOK)) == 0,
758 	    ("invalid flags: 0x%08x", flags));
759 	KASSERT((flags & (M_NOWAIT | M_WAITOK)) != (M_NOWAIT | M_WAITOK),
760 	    ("invalid wait | nowait"));
761 
762 	if (index >= ioat_channel_index)
763 		return (NULL);
764 
765 	ioat = ioat_channel[index];
766 	if (ioat == NULL || ioat->destroying)
767 		return (NULL);
768 
769 	if (ioat->quiescing) {
770 		if ((flags & M_NOWAIT) != 0)
771 			return (NULL);
772 
773 		mtx_lock(IOAT_REFLK);
774 		while (ioat->quiescing && !ioat->destroying)
775 			msleep(&ioat->quiescing, IOAT_REFLK, 0, "getdma", 0);
776 		mtx_unlock(IOAT_REFLK);
777 
778 		if (ioat->destroying)
779 			return (NULL);
780 	}
781 
782 	/*
783 	 * There's a race here between the quiescing check and HW reset or
784 	 * module destroy.
785 	 */
786 	return (&ioat_get(ioat, IOAT_DMAENGINE_REF)->dmaengine);
787 }
788 
789 void
790 ioat_put_dmaengine(bus_dmaengine_t dmaengine)
791 {
792 	struct ioat_softc *ioat;
793 
794 	ioat = to_ioat_softc(dmaengine);
795 	ioat_put(ioat, IOAT_DMAENGINE_REF);
796 }
797 
798 int
799 ioat_get_hwversion(bus_dmaengine_t dmaengine)
800 {
801 	struct ioat_softc *ioat;
802 
803 	ioat = to_ioat_softc(dmaengine);
804 	return (ioat->version);
805 }
806 
807 size_t
808 ioat_get_max_io_size(bus_dmaengine_t dmaengine)
809 {
810 	struct ioat_softc *ioat;
811 
812 	ioat = to_ioat_softc(dmaengine);
813 	return (ioat->max_xfer_size);
814 }
815 
816 int
817 ioat_set_interrupt_coalesce(bus_dmaengine_t dmaengine, uint16_t delay)
818 {
819 	struct ioat_softc *ioat;
820 
821 	ioat = to_ioat_softc(dmaengine);
822 	if (!ioat->intrdelay_supported)
823 		return (ENODEV);
824 	if (delay > ioat->intrdelay_max)
825 		return (ERANGE);
826 
827 	ioat_write_2(ioat, IOAT_INTRDELAY_OFFSET, delay);
828 	ioat->cached_intrdelay =
829 	    ioat_read_2(ioat, IOAT_INTRDELAY_OFFSET) & IOAT_INTRDELAY_US_MASK;
830 	return (0);
831 }
832 
833 uint16_t
834 ioat_get_max_coalesce_period(bus_dmaengine_t dmaengine)
835 {
836 	struct ioat_softc *ioat;
837 
838 	ioat = to_ioat_softc(dmaengine);
839 	return (ioat->intrdelay_max);
840 }
841 
842 void
843 ioat_acquire(bus_dmaengine_t dmaengine)
844 {
845 	struct ioat_softc *ioat;
846 
847 	ioat = to_ioat_softc(dmaengine);
848 	mtx_lock(&ioat->submit_lock);
849 	CTR0(KTR_IOAT, __func__);
850 }
851 
852 int
853 ioat_acquire_reserve(bus_dmaengine_t dmaengine, unsigned n, int mflags)
854 {
855 	struct ioat_softc *ioat;
856 	int error;
857 
858 	ioat = to_ioat_softc(dmaengine);
859 	ioat_acquire(dmaengine);
860 
861 	error = ioat_reserve_space(ioat, n, mflags);
862 	if (error != 0)
863 		ioat_release(dmaengine);
864 	return (error);
865 }
866 
867 void
868 ioat_release(bus_dmaengine_t dmaengine)
869 {
870 	struct ioat_softc *ioat;
871 
872 	ioat = to_ioat_softc(dmaengine);
873 	CTR0(KTR_IOAT, __func__);
874 	ioat_write_2(ioat, IOAT_DMACOUNT_OFFSET, (uint16_t)ioat->hw_head);
875 	mtx_unlock(&ioat->submit_lock);
876 }
877 
878 static struct ioat_descriptor *
879 ioat_op_generic(struct ioat_softc *ioat, uint8_t op,
880     uint32_t size, uint64_t src, uint64_t dst,
881     bus_dmaengine_callback_t callback_fn, void *callback_arg,
882     uint32_t flags)
883 {
884 	struct ioat_generic_hw_descriptor *hw_desc;
885 	struct ioat_descriptor *desc;
886 	int mflags;
887 
888 	mtx_assert(&ioat->submit_lock, MA_OWNED);
889 
890 	KASSERT((flags & ~_DMA_GENERIC_FLAGS) == 0,
891 	    ("Unrecognized flag(s): %#x", flags & ~_DMA_GENERIC_FLAGS));
892 	if ((flags & DMA_NO_WAIT) != 0)
893 		mflags = M_NOWAIT;
894 	else
895 		mflags = M_WAITOK;
896 
897 	if (size > ioat->max_xfer_size) {
898 		ioat_log_message(0, "%s: max_xfer_size = %d, requested = %u\n",
899 		    __func__, ioat->max_xfer_size, (unsigned)size);
900 		return (NULL);
901 	}
902 
903 	if (ioat_reserve_space(ioat, 1, mflags) != 0)
904 		return (NULL);
905 
906 	desc = ioat_get_ring_entry(ioat, ioat->head);
907 	hw_desc = desc->u.generic;
908 
909 	hw_desc->u.control_raw = 0;
910 	hw_desc->u.control_generic.op = op;
911 	hw_desc->u.control_generic.completion_update = 1;
912 
913 	if ((flags & DMA_INT_EN) != 0)
914 		hw_desc->u.control_generic.int_enable = 1;
915 	if ((flags & DMA_FENCE) != 0)
916 		hw_desc->u.control_generic.fence = 1;
917 
918 	hw_desc->size = size;
919 	hw_desc->src_addr = src;
920 	hw_desc->dest_addr = dst;
921 
922 	desc->bus_dmadesc.callback_fn = callback_fn;
923 	desc->bus_dmadesc.callback_arg = callback_arg;
924 	return (desc);
925 }
926 
927 struct bus_dmadesc *
928 ioat_null(bus_dmaengine_t dmaengine, bus_dmaengine_callback_t callback_fn,
929     void *callback_arg, uint32_t flags)
930 {
931 	struct ioat_dma_hw_descriptor *hw_desc;
932 	struct ioat_descriptor *desc;
933 	struct ioat_softc *ioat;
934 
935 	CTR0(KTR_IOAT, __func__);
936 	ioat = to_ioat_softc(dmaengine);
937 
938 	desc = ioat_op_generic(ioat, IOAT_OP_COPY, 8, 0, 0, callback_fn,
939 	    callback_arg, flags);
940 	if (desc == NULL)
941 		return (NULL);
942 
943 	hw_desc = desc->u.dma;
944 	hw_desc->u.control.null = 1;
945 	ioat_submit_single(ioat);
946 	return (&desc->bus_dmadesc);
947 }
948 
949 struct bus_dmadesc *
950 ioat_copy(bus_dmaengine_t dmaengine, bus_addr_t dst,
951     bus_addr_t src, bus_size_t len, bus_dmaengine_callback_t callback_fn,
952     void *callback_arg, uint32_t flags)
953 {
954 	struct ioat_dma_hw_descriptor *hw_desc;
955 	struct ioat_descriptor *desc;
956 	struct ioat_softc *ioat;
957 
958 	CTR0(KTR_IOAT, __func__);
959 	ioat = to_ioat_softc(dmaengine);
960 
961 	if (((src | dst) & (0xffffull << 48)) != 0) {
962 		ioat_log_message(0, "%s: High 16 bits of src/dst invalid\n",
963 		    __func__);
964 		return (NULL);
965 	}
966 
967 	desc = ioat_op_generic(ioat, IOAT_OP_COPY, len, src, dst, callback_fn,
968 	    callback_arg, flags);
969 	if (desc == NULL)
970 		return (NULL);
971 
972 	hw_desc = desc->u.dma;
973 	if (g_ioat_debug_level >= 3)
974 		dump_descriptor(hw_desc);
975 
976 	ioat_submit_single(ioat);
977 	return (&desc->bus_dmadesc);
978 }
979 
980 struct bus_dmadesc *
981 ioat_copy_8k_aligned(bus_dmaengine_t dmaengine, bus_addr_t dst1,
982     bus_addr_t dst2, bus_addr_t src1, bus_addr_t src2,
983     bus_dmaengine_callback_t callback_fn, void *callback_arg, uint32_t flags)
984 {
985 	struct ioat_dma_hw_descriptor *hw_desc;
986 	struct ioat_descriptor *desc;
987 	struct ioat_softc *ioat;
988 
989 	CTR0(KTR_IOAT, __func__);
990 	ioat = to_ioat_softc(dmaengine);
991 
992 	if (((src1 | src2 | dst1 | dst2) & (0xffffull << 48)) != 0) {
993 		ioat_log_message(0, "%s: High 16 bits of src/dst invalid\n",
994 		    __func__);
995 		return (NULL);
996 	}
997 	if (((src1 | src2 | dst1 | dst2) & PAGE_MASK) != 0) {
998 		ioat_log_message(0, "%s: Addresses must be page-aligned\n",
999 		    __func__);
1000 		return (NULL);
1001 	}
1002 
1003 	desc = ioat_op_generic(ioat, IOAT_OP_COPY, 2 * PAGE_SIZE, src1, dst1,
1004 	    callback_fn, callback_arg, flags);
1005 	if (desc == NULL)
1006 		return (NULL);
1007 
1008 	hw_desc = desc->u.dma;
1009 	if (src2 != src1 + PAGE_SIZE) {
1010 		hw_desc->u.control.src_page_break = 1;
1011 		hw_desc->next_src_addr = src2;
1012 	}
1013 	if (dst2 != dst1 + PAGE_SIZE) {
1014 		hw_desc->u.control.dest_page_break = 1;
1015 		hw_desc->next_dest_addr = dst2;
1016 	}
1017 
1018 	if (g_ioat_debug_level >= 3)
1019 		dump_descriptor(hw_desc);
1020 
1021 	ioat_submit_single(ioat);
1022 	return (&desc->bus_dmadesc);
1023 }
1024 
1025 struct bus_dmadesc *
1026 ioat_copy_crc(bus_dmaengine_t dmaengine, bus_addr_t dst, bus_addr_t src,
1027     bus_size_t len, uint32_t *initialseed, bus_addr_t crcptr,
1028     bus_dmaengine_callback_t callback_fn, void *callback_arg, uint32_t flags)
1029 {
1030 	struct ioat_crc32_hw_descriptor *hw_desc;
1031 	struct ioat_descriptor *desc;
1032 	struct ioat_softc *ioat;
1033 	uint32_t teststore;
1034 	uint8_t op;
1035 
1036 	CTR0(KTR_IOAT, __func__);
1037 	ioat = to_ioat_softc(dmaengine);
1038 
1039 	if ((ioat->capabilities & IOAT_DMACAP_MOVECRC) == 0) {
1040 		ioat_log_message(0, "%s: Device lacks MOVECRC capability\n",
1041 		    __func__);
1042 		return (NULL);
1043 	}
1044 	if (((src | dst) & (0xffffffull << 40)) != 0) {
1045 		ioat_log_message(0, "%s: High 24 bits of src/dst invalid\n",
1046 		    __func__);
1047 		return (NULL);
1048 	}
1049 	teststore = (flags & _DMA_CRC_TESTSTORE);
1050 	if (teststore == _DMA_CRC_TESTSTORE) {
1051 		ioat_log_message(0, "%s: TEST and STORE invalid\n", __func__);
1052 		return (NULL);
1053 	}
1054 	if (teststore == 0 && (flags & DMA_CRC_INLINE) != 0) {
1055 		ioat_log_message(0, "%s: INLINE invalid without TEST or STORE\n",
1056 		    __func__);
1057 		return (NULL);
1058 	}
1059 
1060 	switch (teststore) {
1061 	case DMA_CRC_STORE:
1062 		op = IOAT_OP_MOVECRC_STORE;
1063 		break;
1064 	case DMA_CRC_TEST:
1065 		op = IOAT_OP_MOVECRC_TEST;
1066 		break;
1067 	default:
1068 		KASSERT(teststore == 0, ("bogus"));
1069 		op = IOAT_OP_MOVECRC;
1070 		break;
1071 	}
1072 
1073 	if ((flags & DMA_CRC_INLINE) == 0 &&
1074 	    (crcptr & (0xffffffull << 40)) != 0) {
1075 		ioat_log_message(0,
1076 		    "%s: High 24 bits of crcptr invalid\n", __func__);
1077 		return (NULL);
1078 	}
1079 
1080 	desc = ioat_op_generic(ioat, op, len, src, dst, callback_fn,
1081 	    callback_arg, flags & ~_DMA_CRC_FLAGS);
1082 	if (desc == NULL)
1083 		return (NULL);
1084 
1085 	hw_desc = desc->u.crc32;
1086 
1087 	if ((flags & DMA_CRC_INLINE) == 0)
1088 		hw_desc->crc_address = crcptr;
1089 	else
1090 		hw_desc->u.control.crc_location = 1;
1091 
1092 	if (initialseed != NULL) {
1093 		hw_desc->u.control.use_seed = 1;
1094 		hw_desc->seed = *initialseed;
1095 	}
1096 
1097 	if (g_ioat_debug_level >= 3)
1098 		dump_descriptor(hw_desc);
1099 
1100 	ioat_submit_single(ioat);
1101 	return (&desc->bus_dmadesc);
1102 }
1103 
1104 struct bus_dmadesc *
1105 ioat_crc(bus_dmaengine_t dmaengine, bus_addr_t src, bus_size_t len,
1106     uint32_t *initialseed, bus_addr_t crcptr,
1107     bus_dmaengine_callback_t callback_fn, void *callback_arg, uint32_t flags)
1108 {
1109 	struct ioat_crc32_hw_descriptor *hw_desc;
1110 	struct ioat_descriptor *desc;
1111 	struct ioat_softc *ioat;
1112 	uint32_t teststore;
1113 	uint8_t op;
1114 
1115 	CTR0(KTR_IOAT, __func__);
1116 	ioat = to_ioat_softc(dmaengine);
1117 
1118 	if ((ioat->capabilities & IOAT_DMACAP_CRC) == 0) {
1119 		ioat_log_message(0, "%s: Device lacks CRC capability\n",
1120 		    __func__);
1121 		return (NULL);
1122 	}
1123 	if ((src & (0xffffffull << 40)) != 0) {
1124 		ioat_log_message(0, "%s: High 24 bits of src invalid\n",
1125 		    __func__);
1126 		return (NULL);
1127 	}
1128 	teststore = (flags & _DMA_CRC_TESTSTORE);
1129 	if (teststore == _DMA_CRC_TESTSTORE) {
1130 		ioat_log_message(0, "%s: TEST and STORE invalid\n", __func__);
1131 		return (NULL);
1132 	}
1133 	if (teststore == 0 && (flags & DMA_CRC_INLINE) != 0) {
1134 		ioat_log_message(0, "%s: INLINE invalid without TEST or STORE\n",
1135 		    __func__);
1136 		return (NULL);
1137 	}
1138 
1139 	switch (teststore) {
1140 	case DMA_CRC_STORE:
1141 		op = IOAT_OP_CRC_STORE;
1142 		break;
1143 	case DMA_CRC_TEST:
1144 		op = IOAT_OP_CRC_TEST;
1145 		break;
1146 	default:
1147 		KASSERT(teststore == 0, ("bogus"));
1148 		op = IOAT_OP_CRC;
1149 		break;
1150 	}
1151 
1152 	if ((flags & DMA_CRC_INLINE) == 0 &&
1153 	    (crcptr & (0xffffffull << 40)) != 0) {
1154 		ioat_log_message(0,
1155 		    "%s: High 24 bits of crcptr invalid\n", __func__);
1156 		return (NULL);
1157 	}
1158 
1159 	desc = ioat_op_generic(ioat, op, len, src, 0, callback_fn,
1160 	    callback_arg, flags & ~_DMA_CRC_FLAGS);
1161 	if (desc == NULL)
1162 		return (NULL);
1163 
1164 	hw_desc = desc->u.crc32;
1165 
1166 	if ((flags & DMA_CRC_INLINE) == 0)
1167 		hw_desc->crc_address = crcptr;
1168 	else
1169 		hw_desc->u.control.crc_location = 1;
1170 
1171 	if (initialseed != NULL) {
1172 		hw_desc->u.control.use_seed = 1;
1173 		hw_desc->seed = *initialseed;
1174 	}
1175 
1176 	if (g_ioat_debug_level >= 3)
1177 		dump_descriptor(hw_desc);
1178 
1179 	ioat_submit_single(ioat);
1180 	return (&desc->bus_dmadesc);
1181 }
1182 
1183 struct bus_dmadesc *
1184 ioat_blockfill(bus_dmaengine_t dmaengine, bus_addr_t dst, uint64_t fillpattern,
1185     bus_size_t len, bus_dmaengine_callback_t callback_fn, void *callback_arg,
1186     uint32_t flags)
1187 {
1188 	struct ioat_fill_hw_descriptor *hw_desc;
1189 	struct ioat_descriptor *desc;
1190 	struct ioat_softc *ioat;
1191 
1192 	CTR0(KTR_IOAT, __func__);
1193 	ioat = to_ioat_softc(dmaengine);
1194 
1195 	if ((ioat->capabilities & IOAT_DMACAP_BFILL) == 0) {
1196 		ioat_log_message(0, "%s: Device lacks BFILL capability\n",
1197 		    __func__);
1198 		return (NULL);
1199 	}
1200 
1201 	if ((dst & (0xffffull << 48)) != 0) {
1202 		ioat_log_message(0, "%s: High 16 bits of dst invalid\n",
1203 		    __func__);
1204 		return (NULL);
1205 	}
1206 
1207 	desc = ioat_op_generic(ioat, IOAT_OP_FILL, len, fillpattern, dst,
1208 	    callback_fn, callback_arg, flags);
1209 	if (desc == NULL)
1210 		return (NULL);
1211 
1212 	hw_desc = desc->u.fill;
1213 	if (g_ioat_debug_level >= 3)
1214 		dump_descriptor(hw_desc);
1215 
1216 	ioat_submit_single(ioat);
1217 	return (&desc->bus_dmadesc);
1218 }
1219 
1220 /*
1221  * Ring Management
1222  */
1223 static inline uint32_t
1224 ioat_get_active(struct ioat_softc *ioat)
1225 {
1226 
1227 	return ((ioat->head - ioat->tail) & ((1 << ioat->ring_size_order) - 1));
1228 }
1229 
1230 static inline uint32_t
1231 ioat_get_ring_space(struct ioat_softc *ioat)
1232 {
1233 
1234 	return ((1 << ioat->ring_size_order) - ioat_get_active(ioat) - 1);
1235 }
1236 
1237 static struct ioat_descriptor *
1238 ioat_alloc_ring_entry(struct ioat_softc *ioat, int mflags)
1239 {
1240 	struct ioat_generic_hw_descriptor *hw_desc;
1241 	struct ioat_descriptor *desc;
1242 	int error, busdmaflag;
1243 
1244 	error = ENOMEM;
1245 	hw_desc = NULL;
1246 
1247 	if ((mflags & M_WAITOK) != 0)
1248 		busdmaflag = BUS_DMA_WAITOK;
1249 	else
1250 		busdmaflag = BUS_DMA_NOWAIT;
1251 
1252 	desc = malloc(sizeof(*desc), M_IOAT, mflags);
1253 	if (desc == NULL)
1254 		goto out;
1255 
1256 	bus_dmamem_alloc(ioat->hw_desc_tag, (void **)&hw_desc,
1257 	    BUS_DMA_ZERO | busdmaflag, &ioat->hw_desc_map);
1258 	if (hw_desc == NULL)
1259 		goto out;
1260 
1261 	memset(&desc->bus_dmadesc, 0, sizeof(desc->bus_dmadesc));
1262 	desc->u.generic = hw_desc;
1263 
1264 	error = bus_dmamap_load(ioat->hw_desc_tag, ioat->hw_desc_map, hw_desc,
1265 	    sizeof(*hw_desc), ioat_dmamap_cb, &desc->hw_desc_bus_addr,
1266 	    busdmaflag);
1267 	if (error)
1268 		goto out;
1269 
1270 out:
1271 	if (error) {
1272 		ioat_free_ring_entry(ioat, desc);
1273 		return (NULL);
1274 	}
1275 	return (desc);
1276 }
1277 
1278 static void
1279 ioat_free_ring_entry(struct ioat_softc *ioat, struct ioat_descriptor *desc)
1280 {
1281 
1282 	if (desc == NULL)
1283 		return;
1284 
1285 	if (desc->u.generic)
1286 		bus_dmamem_free(ioat->hw_desc_tag, desc->u.generic,
1287 		    ioat->hw_desc_map);
1288 	free(desc, M_IOAT);
1289 }
1290 
1291 /*
1292  * Reserves space in this IOAT descriptor ring by ensuring enough slots remain
1293  * for 'num_descs'.
1294  *
1295  * If mflags contains M_WAITOK, blocks until enough space is available.
1296  *
1297  * Returns zero on success, or an errno on error.  If num_descs is beyond the
1298  * maximum ring size, returns EINVAl; if allocation would block and mflags
1299  * contains M_NOWAIT, returns EAGAIN.
1300  *
1301  * Must be called with the submit_lock held; returns with the lock held.  The
1302  * lock may be dropped to allocate the ring.
1303  *
1304  * (The submit_lock is needed to add any entries to the ring, so callers are
1305  * assured enough room is available.)
1306  */
1307 static int
1308 ioat_reserve_space(struct ioat_softc *ioat, uint32_t num_descs, int mflags)
1309 {
1310 	struct ioat_descriptor **new_ring;
1311 	uint32_t order;
1312 	int error;
1313 
1314 	mtx_assert(&ioat->submit_lock, MA_OWNED);
1315 	error = 0;
1316 
1317 	if (num_descs < 1 || num_descs > (1 << IOAT_MAX_ORDER)) {
1318 		error = EINVAL;
1319 		goto out;
1320 	}
1321 	if (ioat->quiescing) {
1322 		error = ENXIO;
1323 		goto out;
1324 	}
1325 
1326 	for (;;) {
1327 		if (ioat_get_ring_space(ioat) >= num_descs)
1328 			goto out;
1329 
1330 		order = ioat->ring_size_order;
1331 		if (ioat->is_resize_pending || order == IOAT_MAX_ORDER) {
1332 			if ((mflags & M_WAITOK) != 0) {
1333 				msleep(&ioat->tail, &ioat->submit_lock, 0,
1334 				    "ioat_rsz", 0);
1335 				continue;
1336 			}
1337 
1338 			error = EAGAIN;
1339 			break;
1340 		}
1341 
1342 		ioat->is_resize_pending = TRUE;
1343 		for (;;) {
1344 			mtx_unlock(&ioat->submit_lock);
1345 
1346 			new_ring = ioat_prealloc_ring(ioat, 1 << (order + 1),
1347 			    TRUE, mflags);
1348 
1349 			mtx_lock(&ioat->submit_lock);
1350 			KASSERT(ioat->ring_size_order == order,
1351 			    ("is_resize_pending should protect order"));
1352 
1353 			if (new_ring == NULL) {
1354 				KASSERT((mflags & M_WAITOK) == 0,
1355 				    ("allocation failed"));
1356 				error = EAGAIN;
1357 				break;
1358 			}
1359 
1360 			error = ring_grow(ioat, order, new_ring);
1361 			if (error == 0)
1362 				break;
1363 		}
1364 		ioat->is_resize_pending = FALSE;
1365 		wakeup(&ioat->tail);
1366 		if (error)
1367 			break;
1368 	}
1369 
1370 out:
1371 	mtx_assert(&ioat->submit_lock, MA_OWNED);
1372 	return (error);
1373 }
1374 
1375 static struct ioat_descriptor **
1376 ioat_prealloc_ring(struct ioat_softc *ioat, uint32_t size, boolean_t need_dscr,
1377     int mflags)
1378 {
1379 	struct ioat_descriptor **ring;
1380 	uint32_t i;
1381 	int error;
1382 
1383 	KASSERT(size > 0 && powerof2(size), ("bogus size"));
1384 
1385 	ring = malloc(size * sizeof(*ring), M_IOAT, M_ZERO | mflags);
1386 	if (ring == NULL)
1387 		return (NULL);
1388 
1389 	if (need_dscr) {
1390 		error = ENOMEM;
1391 		for (i = size / 2; i < size; i++) {
1392 			ring[i] = ioat_alloc_ring_entry(ioat, mflags);
1393 			if (ring[i] == NULL)
1394 				goto out;
1395 			ring[i]->id = i;
1396 		}
1397 	}
1398 	error = 0;
1399 
1400 out:
1401 	if (error != 0 && ring != NULL) {
1402 		ioat_free_ring(ioat, size, ring);
1403 		ring = NULL;
1404 	}
1405 	return (ring);
1406 }
1407 
1408 static void
1409 ioat_free_ring(struct ioat_softc *ioat, uint32_t size,
1410     struct ioat_descriptor **ring)
1411 {
1412 	uint32_t i;
1413 
1414 	for (i = 0; i < size; i++) {
1415 		if (ring[i] != NULL)
1416 			ioat_free_ring_entry(ioat, ring[i]);
1417 	}
1418 	free(ring, M_IOAT);
1419 }
1420 
1421 static struct ioat_descriptor *
1422 ioat_get_ring_entry(struct ioat_softc *ioat, uint32_t index)
1423 {
1424 
1425 	return (ioat->ring[index % (1 << ioat->ring_size_order)]);
1426 }
1427 
1428 static int
1429 ring_grow(struct ioat_softc *ioat, uint32_t oldorder,
1430     struct ioat_descriptor **newring)
1431 {
1432 	struct ioat_descriptor *tmp, *next;
1433 	struct ioat_dma_hw_descriptor *hw;
1434 	uint32_t oldsize, newsize, head, tail, i, end;
1435 	int error;
1436 
1437 	CTR0(KTR_IOAT, __func__);
1438 
1439 	mtx_assert(&ioat->submit_lock, MA_OWNED);
1440 
1441 	if (oldorder != ioat->ring_size_order || oldorder >= IOAT_MAX_ORDER) {
1442 		error = EINVAL;
1443 		goto out;
1444 	}
1445 
1446 	oldsize = (1 << oldorder);
1447 	newsize = (1 << (oldorder + 1));
1448 
1449 	mtx_lock(&ioat->cleanup_lock);
1450 
1451 	head = ioat->head & (oldsize - 1);
1452 	tail = ioat->tail & (oldsize - 1);
1453 
1454 	/* Copy old descriptors to new ring */
1455 	for (i = 0; i < oldsize; i++)
1456 		newring[i] = ioat->ring[i];
1457 
1458 	/*
1459 	 * If head has wrapped but tail hasn't, we must swap some descriptors
1460 	 * around so that tail can increment directly to head.
1461 	 */
1462 	if (head < tail) {
1463 		for (i = 0; i <= head; i++) {
1464 			tmp = newring[oldsize + i];
1465 
1466 			newring[oldsize + i] = newring[i];
1467 			newring[oldsize + i]->id = oldsize + i;
1468 
1469 			newring[i] = tmp;
1470 			newring[i]->id = i;
1471 		}
1472 		head += oldsize;
1473 	}
1474 
1475 	KASSERT(head >= tail, ("invariants"));
1476 
1477 	/* Head didn't wrap; we only need to link in oldsize..newsize */
1478 	if (head < oldsize) {
1479 		i = oldsize - 1;
1480 		end = newsize;
1481 	} else {
1482 		/* Head did wrap; link newhead..newsize and 0..oldhead */
1483 		i = head;
1484 		end = newsize + (head - oldsize) + 1;
1485 	}
1486 
1487 	/*
1488 	 * Fix up hardware ring, being careful not to trample the active
1489 	 * section (tail -> head).
1490 	 */
1491 	for (; i < end; i++) {
1492 		KASSERT((i & (newsize - 1)) < tail ||
1493 		    (i & (newsize - 1)) >= head, ("trampling snake"));
1494 
1495 		next = newring[(i + 1) & (newsize - 1)];
1496 		hw = newring[i & (newsize - 1)]->u.dma;
1497 		hw->next = next->hw_desc_bus_addr;
1498 	}
1499 
1500 	free(ioat->ring, M_IOAT);
1501 	ioat->ring = newring;
1502 	ioat->ring_size_order = oldorder + 1;
1503 	ioat->tail = tail;
1504 	ioat->head = head;
1505 	error = 0;
1506 
1507 	mtx_unlock(&ioat->cleanup_lock);
1508 out:
1509 	if (error)
1510 		ioat_free_ring(ioat, (1 << (oldorder + 1)), newring);
1511 	return (error);
1512 }
1513 
1514 static int
1515 ring_shrink(struct ioat_softc *ioat, uint32_t oldorder,
1516     struct ioat_descriptor **newring)
1517 {
1518 	struct ioat_dma_hw_descriptor *hw;
1519 	struct ioat_descriptor *ent, *next;
1520 	uint32_t oldsize, newsize, current_idx, new_idx, i;
1521 	int error;
1522 
1523 	CTR0(KTR_IOAT, __func__);
1524 
1525 	mtx_assert(&ioat->submit_lock, MA_OWNED);
1526 
1527 	if (oldorder != ioat->ring_size_order || oldorder <= IOAT_MIN_ORDER) {
1528 		error = EINVAL;
1529 		goto out_unlocked;
1530 	}
1531 
1532 	oldsize = (1 << oldorder);
1533 	newsize = (1 << (oldorder - 1));
1534 
1535 	mtx_lock(&ioat->cleanup_lock);
1536 
1537 	/* Can't shrink below current active set! */
1538 	if (ioat_get_active(ioat) >= newsize) {
1539 		error = ENOMEM;
1540 		goto out;
1541 	}
1542 
1543 	/*
1544 	 * Copy current descriptors to the new ring, dropping the removed
1545 	 * descriptors.
1546 	 */
1547 	for (i = 0; i < newsize; i++) {
1548 		current_idx = (ioat->tail + i) & (oldsize - 1);
1549 		new_idx = (ioat->tail + i) & (newsize - 1);
1550 
1551 		newring[new_idx] = ioat->ring[current_idx];
1552 		newring[new_idx]->id = new_idx;
1553 	}
1554 
1555 	/* Free deleted descriptors */
1556 	for (i = newsize; i < oldsize; i++) {
1557 		ent = ioat_get_ring_entry(ioat, ioat->tail + i);
1558 		ioat_free_ring_entry(ioat, ent);
1559 	}
1560 
1561 	/* Fix up hardware ring. */
1562 	hw = newring[(ioat->tail + newsize - 1) & (newsize - 1)]->u.dma;
1563 	next = newring[(ioat->tail + newsize) & (newsize - 1)];
1564 	hw->next = next->hw_desc_bus_addr;
1565 
1566 	free(ioat->ring, M_IOAT);
1567 	ioat->ring = newring;
1568 	ioat->ring_size_order = oldorder - 1;
1569 	error = 0;
1570 
1571 out:
1572 	mtx_unlock(&ioat->cleanup_lock);
1573 out_unlocked:
1574 	if (error)
1575 		ioat_free_ring(ioat, (1 << (oldorder - 1)), newring);
1576 	return (error);
1577 }
1578 
1579 static void
1580 ioat_halted_debug(struct ioat_softc *ioat, uint32_t chanerr)
1581 {
1582 	struct ioat_descriptor *desc;
1583 
1584 	ioat_log_message(0, "Channel halted (%b)\n", (int)chanerr,
1585 	    IOAT_CHANERR_STR);
1586 	if (chanerr == 0)
1587 		return;
1588 
1589 	mtx_assert(&ioat->cleanup_lock, MA_OWNED);
1590 
1591 	desc = ioat_get_ring_entry(ioat, ioat->tail + 0);
1592 	dump_descriptor(desc->u.raw);
1593 
1594 	desc = ioat_get_ring_entry(ioat, ioat->tail + 1);
1595 	dump_descriptor(desc->u.raw);
1596 }
1597 
1598 static void
1599 ioat_timer_callback(void *arg)
1600 {
1601 	struct ioat_descriptor **newring;
1602 	struct ioat_softc *ioat;
1603 	uint32_t order;
1604 
1605 	ioat = arg;
1606 	ioat_log_message(1, "%s\n", __func__);
1607 
1608 	if (ioat->is_completion_pending) {
1609 		ioat_process_events(ioat);
1610 		return;
1611 	}
1612 
1613 	/* Slowly scale the ring down if idle. */
1614 	mtx_lock(&ioat->submit_lock);
1615 	order = ioat->ring_size_order;
1616 	if (ioat->is_resize_pending || order == IOAT_MIN_ORDER) {
1617 		mtx_unlock(&ioat->submit_lock);
1618 		goto out;
1619 	}
1620 	ioat->is_resize_pending = TRUE;
1621 	mtx_unlock(&ioat->submit_lock);
1622 
1623 	newring = ioat_prealloc_ring(ioat, 1 << (order - 1), FALSE,
1624 	    M_NOWAIT);
1625 
1626 	mtx_lock(&ioat->submit_lock);
1627 	KASSERT(ioat->ring_size_order == order,
1628 	    ("resize_pending protects order"));
1629 
1630 	if (newring != NULL)
1631 		ring_shrink(ioat, order, newring);
1632 
1633 	ioat->is_resize_pending = FALSE;
1634 	mtx_unlock(&ioat->submit_lock);
1635 
1636 out:
1637 	if (ioat->ring_size_order > IOAT_MIN_ORDER)
1638 		callout_reset(&ioat->timer, 10 * hz,
1639 		    ioat_timer_callback, ioat);
1640 }
1641 
1642 /*
1643  * Support Functions
1644  */
1645 static void
1646 ioat_submit_single(struct ioat_softc *ioat)
1647 {
1648 
1649 	ioat_get(ioat, IOAT_ACTIVE_DESCR_REF);
1650 	atomic_add_rel_int(&ioat->head, 1);
1651 	atomic_add_rel_int(&ioat->hw_head, 1);
1652 
1653 	if (!ioat->is_completion_pending) {
1654 		ioat->is_completion_pending = TRUE;
1655 		callout_reset(&ioat->timer, IOAT_INTR_TIMO,
1656 		    ioat_timer_callback, ioat);
1657 	}
1658 
1659 	ioat->stats.descriptors_submitted++;
1660 }
1661 
1662 static int
1663 ioat_reset_hw(struct ioat_softc *ioat)
1664 {
1665 	uint64_t status;
1666 	uint32_t chanerr;
1667 	unsigned timeout;
1668 	int error;
1669 
1670 	mtx_lock(IOAT_REFLK);
1671 	ioat->quiescing = TRUE;
1672 	ioat_drain_locked(ioat);
1673 	mtx_unlock(IOAT_REFLK);
1674 
1675 	status = ioat_get_chansts(ioat);
1676 	if (is_ioat_active(status) || is_ioat_idle(status))
1677 		ioat_suspend(ioat);
1678 
1679 	/* Wait at most 20 ms */
1680 	for (timeout = 0; (is_ioat_active(status) || is_ioat_idle(status)) &&
1681 	    timeout < 20; timeout++) {
1682 		DELAY(1000);
1683 		status = ioat_get_chansts(ioat);
1684 	}
1685 	if (timeout == 20) {
1686 		error = ETIMEDOUT;
1687 		goto out;
1688 	}
1689 
1690 	KASSERT(ioat_get_active(ioat) == 0, ("active after quiesce"));
1691 
1692 	chanerr = ioat_read_4(ioat, IOAT_CHANERR_OFFSET);
1693 	ioat_write_4(ioat, IOAT_CHANERR_OFFSET, chanerr);
1694 
1695 	/*
1696 	 * IOAT v3 workaround - CHANERRMSK_INT with 3E07h to masks out errors
1697 	 *  that can cause stability issues for IOAT v3.
1698 	 */
1699 	pci_write_config(ioat->device, IOAT_CFG_CHANERRMASK_INT_OFFSET, 0x3e07,
1700 	    4);
1701 	chanerr = pci_read_config(ioat->device, IOAT_CFG_CHANERR_INT_OFFSET, 4);
1702 	pci_write_config(ioat->device, IOAT_CFG_CHANERR_INT_OFFSET, chanerr, 4);
1703 
1704 	/*
1705 	 * BDXDE and BWD models reset MSI-X registers on device reset.
1706 	 * Save/restore their contents manually.
1707 	 */
1708 	if (ioat_model_resets_msix(ioat)) {
1709 		ioat_log_message(1, "device resets MSI-X registers; saving\n");
1710 		pci_save_state(ioat->device);
1711 	}
1712 
1713 	ioat_reset(ioat);
1714 
1715 	/* Wait at most 20 ms */
1716 	for (timeout = 0; ioat_reset_pending(ioat) && timeout < 20; timeout++)
1717 		DELAY(1000);
1718 	if (timeout == 20) {
1719 		error = ETIMEDOUT;
1720 		goto out;
1721 	}
1722 
1723 	if (ioat_model_resets_msix(ioat)) {
1724 		ioat_log_message(1, "device resets registers; restored\n");
1725 		pci_restore_state(ioat->device);
1726 	}
1727 
1728 	/* Reset attempts to return the hardware to "halted." */
1729 	status = ioat_get_chansts(ioat);
1730 	if (is_ioat_active(status) || is_ioat_idle(status)) {
1731 		/* So this really shouldn't happen... */
1732 		ioat_log_message(0, "Device is active after a reset?\n");
1733 		ioat_write_chanctrl(ioat, IOAT_CHANCTRL_RUN);
1734 		error = 0;
1735 		goto out;
1736 	}
1737 
1738 	chanerr = ioat_read_4(ioat, IOAT_CHANERR_OFFSET);
1739 	if (chanerr != 0) {
1740 		mtx_lock(&ioat->cleanup_lock);
1741 		ioat_halted_debug(ioat, chanerr);
1742 		mtx_unlock(&ioat->cleanup_lock);
1743 		error = EIO;
1744 		goto out;
1745 	}
1746 
1747 	/*
1748 	 * Bring device back online after reset.  Writing CHAINADDR brings the
1749 	 * device back to active.
1750 	 *
1751 	 * The internal ring counter resets to zero, so we have to start over
1752 	 * at zero as well.
1753 	 */
1754 	ioat->tail = ioat->head = ioat->hw_head = 0;
1755 	ioat->last_seen = 0;
1756 
1757 	ioat_write_chanctrl(ioat, IOAT_CHANCTRL_RUN);
1758 	ioat_write_chancmp(ioat, ioat->comp_update_bus_addr);
1759 	ioat_write_chainaddr(ioat, ioat->ring[0]->hw_desc_bus_addr);
1760 	error = 0;
1761 
1762 out:
1763 	mtx_lock(IOAT_REFLK);
1764 	ioat->quiescing = FALSE;
1765 	wakeup(&ioat->quiescing);
1766 	mtx_unlock(IOAT_REFLK);
1767 
1768 	if (error == 0)
1769 		error = ioat_start_channel(ioat);
1770 
1771 	return (error);
1772 }
1773 
1774 static int
1775 sysctl_handle_chansts(SYSCTL_HANDLER_ARGS)
1776 {
1777 	struct ioat_softc *ioat;
1778 	struct sbuf sb;
1779 	uint64_t status;
1780 	int error;
1781 
1782 	ioat = arg1;
1783 
1784 	status = ioat_get_chansts(ioat) & IOAT_CHANSTS_STATUS;
1785 
1786 	sbuf_new_for_sysctl(&sb, NULL, 256, req);
1787 	switch (status) {
1788 	case IOAT_CHANSTS_ACTIVE:
1789 		sbuf_printf(&sb, "ACTIVE");
1790 		break;
1791 	case IOAT_CHANSTS_IDLE:
1792 		sbuf_printf(&sb, "IDLE");
1793 		break;
1794 	case IOAT_CHANSTS_SUSPENDED:
1795 		sbuf_printf(&sb, "SUSPENDED");
1796 		break;
1797 	case IOAT_CHANSTS_HALTED:
1798 		sbuf_printf(&sb, "HALTED");
1799 		break;
1800 	case IOAT_CHANSTS_ARMED:
1801 		sbuf_printf(&sb, "ARMED");
1802 		break;
1803 	default:
1804 		sbuf_printf(&sb, "UNKNOWN");
1805 		break;
1806 	}
1807 	error = sbuf_finish(&sb);
1808 	sbuf_delete(&sb);
1809 
1810 	if (error != 0 || req->newptr == NULL)
1811 		return (error);
1812 	return (EINVAL);
1813 }
1814 
1815 static int
1816 sysctl_handle_dpi(SYSCTL_HANDLER_ARGS)
1817 {
1818 	struct ioat_softc *ioat;
1819 	struct sbuf sb;
1820 #define	PRECISION	"1"
1821 	const uintmax_t factor = 10;
1822 	uintmax_t rate;
1823 	int error;
1824 
1825 	ioat = arg1;
1826 	sbuf_new_for_sysctl(&sb, NULL, 16, req);
1827 
1828 	if (ioat->stats.interrupts == 0) {
1829 		sbuf_printf(&sb, "NaN");
1830 		goto out;
1831 	}
1832 	rate = ioat->stats.descriptors_processed * factor /
1833 	    ioat->stats.interrupts;
1834 	sbuf_printf(&sb, "%ju.%." PRECISION "ju", rate / factor,
1835 	    rate % factor);
1836 #undef	PRECISION
1837 out:
1838 	error = sbuf_finish(&sb);
1839 	sbuf_delete(&sb);
1840 	if (error != 0 || req->newptr == NULL)
1841 		return (error);
1842 	return (EINVAL);
1843 }
1844 
1845 static int
1846 sysctl_handle_error(SYSCTL_HANDLER_ARGS)
1847 {
1848 	struct ioat_descriptor *desc;
1849 	struct ioat_softc *ioat;
1850 	int error, arg;
1851 
1852 	ioat = arg1;
1853 
1854 	arg = 0;
1855 	error = SYSCTL_OUT(req, &arg, sizeof(arg));
1856 	if (error != 0 || req->newptr == NULL)
1857 		return (error);
1858 
1859 	error = SYSCTL_IN(req, &arg, sizeof(arg));
1860 	if (error != 0)
1861 		return (error);
1862 
1863 	if (arg != 0) {
1864 		ioat_acquire(&ioat->dmaengine);
1865 		desc = ioat_op_generic(ioat, IOAT_OP_COPY, 1,
1866 		    0xffff000000000000ull, 0xffff000000000000ull, NULL, NULL,
1867 		    0);
1868 		if (desc == NULL)
1869 			error = ENOMEM;
1870 		else
1871 			ioat_submit_single(ioat);
1872 		ioat_release(&ioat->dmaengine);
1873 	}
1874 	return (error);
1875 }
1876 
1877 static int
1878 sysctl_handle_reset(SYSCTL_HANDLER_ARGS)
1879 {
1880 	struct ioat_softc *ioat;
1881 	int error, arg;
1882 
1883 	ioat = arg1;
1884 
1885 	arg = 0;
1886 	error = SYSCTL_OUT(req, &arg, sizeof(arg));
1887 	if (error != 0 || req->newptr == NULL)
1888 		return (error);
1889 
1890 	error = SYSCTL_IN(req, &arg, sizeof(arg));
1891 	if (error != 0)
1892 		return (error);
1893 
1894 	if (arg != 0)
1895 		error = ioat_reset_hw(ioat);
1896 
1897 	return (error);
1898 }
1899 
1900 static void
1901 dump_descriptor(void *hw_desc)
1902 {
1903 	int i, j;
1904 
1905 	for (i = 0; i < 2; i++) {
1906 		for (j = 0; j < 8; j++)
1907 			printf("%08x ", ((uint32_t *)hw_desc)[i * 8 + j]);
1908 		printf("\n");
1909 	}
1910 }
1911 
1912 static void
1913 ioat_setup_sysctl(device_t device)
1914 {
1915 	struct sysctl_oid_list *par, *statpar, *state, *hammer;
1916 	struct sysctl_ctx_list *ctx;
1917 	struct sysctl_oid *tree, *tmp;
1918 	struct ioat_softc *ioat;
1919 
1920 	ioat = DEVICE2SOFTC(device);
1921 	ctx = device_get_sysctl_ctx(device);
1922 	tree = device_get_sysctl_tree(device);
1923 	par = SYSCTL_CHILDREN(tree);
1924 
1925 	SYSCTL_ADD_INT(ctx, par, OID_AUTO, "version", CTLFLAG_RD,
1926 	    &ioat->version, 0, "HW version (0xMM form)");
1927 	SYSCTL_ADD_UINT(ctx, par, OID_AUTO, "max_xfer_size", CTLFLAG_RD,
1928 	    &ioat->max_xfer_size, 0, "HW maximum transfer size");
1929 	SYSCTL_ADD_INT(ctx, par, OID_AUTO, "intrdelay_supported", CTLFLAG_RD,
1930 	    &ioat->intrdelay_supported, 0, "Is INTRDELAY supported");
1931 	SYSCTL_ADD_U16(ctx, par, OID_AUTO, "intrdelay_max", CTLFLAG_RD,
1932 	    &ioat->intrdelay_max, 0,
1933 	    "Maximum configurable INTRDELAY on this channel (microseconds)");
1934 
1935 	tmp = SYSCTL_ADD_NODE(ctx, par, OID_AUTO, "state", CTLFLAG_RD, NULL,
1936 	    "IOAT channel internal state");
1937 	state = SYSCTL_CHILDREN(tmp);
1938 
1939 	SYSCTL_ADD_UINT(ctx, state, OID_AUTO, "ring_size_order", CTLFLAG_RD,
1940 	    &ioat->ring_size_order, 0, "SW descriptor ring size order");
1941 	SYSCTL_ADD_UINT(ctx, state, OID_AUTO, "head", CTLFLAG_RD, &ioat->head,
1942 	    0, "SW descriptor head pointer index");
1943 	SYSCTL_ADD_UINT(ctx, state, OID_AUTO, "tail", CTLFLAG_RD, &ioat->tail,
1944 	    0, "SW descriptor tail pointer index");
1945 	SYSCTL_ADD_UINT(ctx, state, OID_AUTO, "hw_head", CTLFLAG_RD,
1946 	    &ioat->hw_head, 0, "HW DMACOUNT");
1947 
1948 	SYSCTL_ADD_UQUAD(ctx, state, OID_AUTO, "last_completion", CTLFLAG_RD,
1949 	    ioat->comp_update, "HW addr of last completion");
1950 
1951 	SYSCTL_ADD_INT(ctx, state, OID_AUTO, "is_resize_pending", CTLFLAG_RD,
1952 	    &ioat->is_resize_pending, 0, "resize pending");
1953 	SYSCTL_ADD_INT(ctx, state, OID_AUTO, "is_completion_pending",
1954 	    CTLFLAG_RD, &ioat->is_completion_pending, 0, "completion pending");
1955 	SYSCTL_ADD_INT(ctx, state, OID_AUTO, "is_reset_pending", CTLFLAG_RD,
1956 	    &ioat->is_reset_pending, 0, "reset pending");
1957 	SYSCTL_ADD_INT(ctx, state, OID_AUTO, "is_channel_running", CTLFLAG_RD,
1958 	    &ioat->is_channel_running, 0, "channel running");
1959 
1960 	SYSCTL_ADD_PROC(ctx, state, OID_AUTO, "chansts",
1961 	    CTLTYPE_STRING | CTLFLAG_RD, ioat, 0, sysctl_handle_chansts, "A",
1962 	    "String of the channel status");
1963 
1964 	SYSCTL_ADD_U16(ctx, state, OID_AUTO, "intrdelay", CTLFLAG_RD,
1965 	    &ioat->cached_intrdelay, 0,
1966 	    "Current INTRDELAY on this channel (cached, microseconds)");
1967 
1968 	tmp = SYSCTL_ADD_NODE(ctx, par, OID_AUTO, "hammer", CTLFLAG_RD, NULL,
1969 	    "Big hammers (mostly for testing)");
1970 	hammer = SYSCTL_CHILDREN(tmp);
1971 
1972 	SYSCTL_ADD_PROC(ctx, hammer, OID_AUTO, "force_hw_reset",
1973 	    CTLTYPE_INT | CTLFLAG_RW, ioat, 0, sysctl_handle_reset, "I",
1974 	    "Set to non-zero to reset the hardware");
1975 	SYSCTL_ADD_PROC(ctx, hammer, OID_AUTO, "force_hw_error",
1976 	    CTLTYPE_INT | CTLFLAG_RW, ioat, 0, sysctl_handle_error, "I",
1977 	    "Set to non-zero to inject a recoverable hardware error");
1978 
1979 	tmp = SYSCTL_ADD_NODE(ctx, par, OID_AUTO, "stats", CTLFLAG_RD, NULL,
1980 	    "IOAT channel statistics");
1981 	statpar = SYSCTL_CHILDREN(tmp);
1982 
1983 	SYSCTL_ADD_UQUAD(ctx, statpar, OID_AUTO, "interrupts", CTLFLAG_RW,
1984 	    &ioat->stats.interrupts,
1985 	    "Number of interrupts processed on this channel");
1986 	SYSCTL_ADD_UQUAD(ctx, statpar, OID_AUTO, "descriptors", CTLFLAG_RW,
1987 	    &ioat->stats.descriptors_processed,
1988 	    "Number of descriptors processed on this channel");
1989 	SYSCTL_ADD_UQUAD(ctx, statpar, OID_AUTO, "submitted", CTLFLAG_RW,
1990 	    &ioat->stats.descriptors_submitted,
1991 	    "Number of descriptors submitted to this channel");
1992 	SYSCTL_ADD_UQUAD(ctx, statpar, OID_AUTO, "errored", CTLFLAG_RW,
1993 	    &ioat->stats.descriptors_error,
1994 	    "Number of descriptors failed by channel errors");
1995 	SYSCTL_ADD_U32(ctx, statpar, OID_AUTO, "halts", CTLFLAG_RW,
1996 	    &ioat->stats.channel_halts, 0,
1997 	    "Number of times the channel has halted");
1998 	SYSCTL_ADD_U32(ctx, statpar, OID_AUTO, "last_halt_chanerr", CTLFLAG_RW,
1999 	    &ioat->stats.last_halt_chanerr, 0,
2000 	    "The raw CHANERR when the channel was last halted");
2001 
2002 	SYSCTL_ADD_PROC(ctx, statpar, OID_AUTO, "desc_per_interrupt",
2003 	    CTLTYPE_STRING | CTLFLAG_RD, ioat, 0, sysctl_handle_dpi, "A",
2004 	    "Descriptors per interrupt");
2005 }
2006 
2007 static inline struct ioat_softc *
2008 ioat_get(struct ioat_softc *ioat, enum ioat_ref_kind kind)
2009 {
2010 	uint32_t old;
2011 
2012 	KASSERT(kind < IOAT_NUM_REF_KINDS, ("bogus"));
2013 
2014 	old = atomic_fetchadd_32(&ioat->refcnt, 1);
2015 	KASSERT(old < UINT32_MAX, ("refcnt overflow"));
2016 
2017 #ifdef INVARIANTS
2018 	old = atomic_fetchadd_32(&ioat->refkinds[kind], 1);
2019 	KASSERT(old < UINT32_MAX, ("refcnt kind overflow"));
2020 #endif
2021 
2022 	return (ioat);
2023 }
2024 
2025 static inline void
2026 ioat_putn(struct ioat_softc *ioat, uint32_t n, enum ioat_ref_kind kind)
2027 {
2028 
2029 	_ioat_putn(ioat, n, kind, FALSE);
2030 }
2031 
2032 static inline void
2033 ioat_putn_locked(struct ioat_softc *ioat, uint32_t n, enum ioat_ref_kind kind)
2034 {
2035 
2036 	_ioat_putn(ioat, n, kind, TRUE);
2037 }
2038 
2039 static inline void
2040 _ioat_putn(struct ioat_softc *ioat, uint32_t n, enum ioat_ref_kind kind,
2041     boolean_t locked)
2042 {
2043 	uint32_t old;
2044 
2045 	KASSERT(kind < IOAT_NUM_REF_KINDS, ("bogus"));
2046 
2047 	if (n == 0)
2048 		return;
2049 
2050 #ifdef INVARIANTS
2051 	old = atomic_fetchadd_32(&ioat->refkinds[kind], -n);
2052 	KASSERT(old >= n, ("refcnt kind underflow"));
2053 #endif
2054 
2055 	/* Skip acquiring the lock if resulting refcnt > 0. */
2056 	for (;;) {
2057 		old = ioat->refcnt;
2058 		if (old <= n)
2059 			break;
2060 		if (atomic_cmpset_32(&ioat->refcnt, old, old - n))
2061 			return;
2062 	}
2063 
2064 	if (locked)
2065 		mtx_assert(IOAT_REFLK, MA_OWNED);
2066 	else
2067 		mtx_lock(IOAT_REFLK);
2068 
2069 	old = atomic_fetchadd_32(&ioat->refcnt, -n);
2070 	KASSERT(old >= n, ("refcnt error"));
2071 
2072 	if (old == n)
2073 		wakeup(IOAT_REFLK);
2074 	if (!locked)
2075 		mtx_unlock(IOAT_REFLK);
2076 }
2077 
2078 static inline void
2079 ioat_put(struct ioat_softc *ioat, enum ioat_ref_kind kind)
2080 {
2081 
2082 	ioat_putn(ioat, 1, kind);
2083 }
2084 
2085 static void
2086 ioat_drain_locked(struct ioat_softc *ioat)
2087 {
2088 
2089 	mtx_assert(IOAT_REFLK, MA_OWNED);
2090 	while (ioat->refcnt > 0)
2091 		msleep(IOAT_REFLK, IOAT_REFLK, 0, "ioat_drain", 0);
2092 }
2093