xref: /freebsd/sys/dev/ioat/ioat.c (revision bae28eaa8d26ddd67e9af51979f1a54599e9b3a8)
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/sysctl.h>
42 #include <sys/time.h>
43 #include <dev/pci/pcireg.h>
44 #include <dev/pci/pcivar.h>
45 #include <machine/bus.h>
46 #include <machine/resource.h>
47 #include <machine/stdarg.h>
48 
49 #include "ioat.h"
50 #include "ioat_hw.h"
51 #include "ioat_internal.h"
52 
53 #define	IOAT_INTR_TIMO	(hz / 10)
54 #define	IOAT_REFLK	(&ioat->submit_lock)
55 
56 static int ioat_probe(device_t device);
57 static int ioat_attach(device_t device);
58 static int ioat_detach(device_t device);
59 static int ioat_setup_intr(struct ioat_softc *ioat);
60 static int ioat_teardown_intr(struct ioat_softc *ioat);
61 static int ioat3_attach(device_t device);
62 static int ioat_start_channel(struct ioat_softc *ioat);
63 static int ioat_map_pci_bar(struct ioat_softc *ioat);
64 static void ioat_dmamap_cb(void *arg, bus_dma_segment_t *segs, int nseg,
65     int error);
66 static void ioat_interrupt_handler(void *arg);
67 static boolean_t ioat_model_resets_msix(struct ioat_softc *ioat);
68 static void ioat_process_events(struct ioat_softc *ioat);
69 static inline uint32_t ioat_get_active(struct ioat_softc *ioat);
70 static inline uint32_t ioat_get_ring_space(struct ioat_softc *ioat);
71 static void ioat_free_ring(struct ioat_softc *, uint32_t size,
72     struct ioat_descriptor **);
73 static void ioat_free_ring_entry(struct ioat_softc *ioat,
74     struct ioat_descriptor *desc);
75 static struct ioat_descriptor *ioat_alloc_ring_entry(struct ioat_softc *,
76     int mflags);
77 static int ioat_reserve_space(struct ioat_softc *, uint32_t, int mflags);
78 static struct ioat_descriptor *ioat_get_ring_entry(struct ioat_softc *ioat,
79     uint32_t index);
80 static struct ioat_descriptor **ioat_prealloc_ring(struct ioat_softc *,
81     uint32_t size, boolean_t need_dscr, int mflags);
82 static int ring_grow(struct ioat_softc *, uint32_t oldorder,
83     struct ioat_descriptor **);
84 static int ring_shrink(struct ioat_softc *, uint32_t oldorder,
85     struct ioat_descriptor **);
86 static void ioat_timer_callback(void *arg);
87 static void dump_descriptor(void *hw_desc);
88 static void ioat_submit_single(struct ioat_softc *ioat);
89 static void ioat_comp_update_map(void *arg, bus_dma_segment_t *seg, int nseg,
90     int error);
91 static int ioat_reset_hw(struct ioat_softc *ioat);
92 static void ioat_setup_sysctl(device_t device);
93 static int sysctl_handle_reset(SYSCTL_HANDLER_ARGS);
94 static inline struct ioat_softc *ioat_get(struct ioat_softc *,
95     enum ioat_ref_kind);
96 static inline void ioat_put(struct ioat_softc *, enum ioat_ref_kind);
97 static inline void ioat_putn(struct ioat_softc *, uint32_t,
98     enum ioat_ref_kind);
99 static void ioat_drain(struct ioat_softc *);
100 
101 #define	ioat_log_message(v, ...) do {					\
102 	if ((v) <= g_ioat_debug_level) {				\
103 		device_printf(ioat->device, __VA_ARGS__);		\
104 	}								\
105 } while (0)
106 
107 MALLOC_DEFINE(M_IOAT, "ioat", "ioat driver memory allocations");
108 SYSCTL_NODE(_hw, OID_AUTO, ioat, CTLFLAG_RD, 0, "ioat node");
109 
110 static int g_force_legacy_interrupts;
111 SYSCTL_INT(_hw_ioat, OID_AUTO, force_legacy_interrupts, CTLFLAG_RDTUN,
112     &g_force_legacy_interrupts, 0, "Set to non-zero to force MSI-X disabled");
113 
114 int g_ioat_debug_level = 0;
115 SYSCTL_INT(_hw_ioat, OID_AUTO, debug_level, CTLFLAG_RWTUN, &g_ioat_debug_level,
116     0, "Set log level (0-3) for ioat(4). Higher is more verbose.");
117 
118 /*
119  * OS <-> Driver interface structures
120  */
121 static device_method_t ioat_pci_methods[] = {
122 	/* Device interface */
123 	DEVMETHOD(device_probe,     ioat_probe),
124 	DEVMETHOD(device_attach,    ioat_attach),
125 	DEVMETHOD(device_detach,    ioat_detach),
126 	{ 0, 0 }
127 };
128 
129 static driver_t ioat_pci_driver = {
130 	"ioat",
131 	ioat_pci_methods,
132 	sizeof(struct ioat_softc),
133 };
134 
135 static devclass_t ioat_devclass;
136 DRIVER_MODULE(ioat, pci, ioat_pci_driver, ioat_devclass, 0, 0);
137 
138 /*
139  * Private data structures
140  */
141 static struct ioat_softc *ioat_channel[IOAT_MAX_CHANNELS];
142 static int ioat_channel_index = 0;
143 SYSCTL_INT(_hw_ioat, OID_AUTO, channels, CTLFLAG_RD, &ioat_channel_index, 0,
144     "Number of IOAT channels attached");
145 
146 static struct _pcsid
147 {
148 	u_int32_t   type;
149 	const char  *desc;
150 } pci_ids[] = {
151 	{ 0x34308086, "TBG IOAT Ch0" },
152 	{ 0x34318086, "TBG IOAT Ch1" },
153 	{ 0x34328086, "TBG IOAT Ch2" },
154 	{ 0x34338086, "TBG IOAT Ch3" },
155 	{ 0x34298086, "TBG IOAT Ch4" },
156 	{ 0x342a8086, "TBG IOAT Ch5" },
157 	{ 0x342b8086, "TBG IOAT Ch6" },
158 	{ 0x342c8086, "TBG IOAT Ch7" },
159 
160 	{ 0x37108086, "JSF IOAT Ch0" },
161 	{ 0x37118086, "JSF IOAT Ch1" },
162 	{ 0x37128086, "JSF IOAT Ch2" },
163 	{ 0x37138086, "JSF IOAT Ch3" },
164 	{ 0x37148086, "JSF IOAT Ch4" },
165 	{ 0x37158086, "JSF IOAT Ch5" },
166 	{ 0x37168086, "JSF IOAT Ch6" },
167 	{ 0x37178086, "JSF IOAT Ch7" },
168 	{ 0x37188086, "JSF IOAT Ch0 (RAID)" },
169 	{ 0x37198086, "JSF IOAT Ch1 (RAID)" },
170 
171 	{ 0x3c208086, "SNB IOAT Ch0" },
172 	{ 0x3c218086, "SNB IOAT Ch1" },
173 	{ 0x3c228086, "SNB IOAT Ch2" },
174 	{ 0x3c238086, "SNB IOAT Ch3" },
175 	{ 0x3c248086, "SNB IOAT Ch4" },
176 	{ 0x3c258086, "SNB IOAT Ch5" },
177 	{ 0x3c268086, "SNB IOAT Ch6" },
178 	{ 0x3c278086, "SNB IOAT Ch7" },
179 	{ 0x3c2e8086, "SNB IOAT Ch0 (RAID)" },
180 	{ 0x3c2f8086, "SNB IOAT Ch1 (RAID)" },
181 
182 	{ 0x0e208086, "IVB IOAT Ch0" },
183 	{ 0x0e218086, "IVB IOAT Ch1" },
184 	{ 0x0e228086, "IVB IOAT Ch2" },
185 	{ 0x0e238086, "IVB IOAT Ch3" },
186 	{ 0x0e248086, "IVB IOAT Ch4" },
187 	{ 0x0e258086, "IVB IOAT Ch5" },
188 	{ 0x0e268086, "IVB IOAT Ch6" },
189 	{ 0x0e278086, "IVB IOAT Ch7" },
190 	{ 0x0e2e8086, "IVB IOAT Ch0 (RAID)" },
191 	{ 0x0e2f8086, "IVB IOAT Ch1 (RAID)" },
192 
193 	{ 0x2f208086, "HSW IOAT Ch0" },
194 	{ 0x2f218086, "HSW IOAT Ch1" },
195 	{ 0x2f228086, "HSW IOAT Ch2" },
196 	{ 0x2f238086, "HSW IOAT Ch3" },
197 	{ 0x2f248086, "HSW IOAT Ch4" },
198 	{ 0x2f258086, "HSW IOAT Ch5" },
199 	{ 0x2f268086, "HSW IOAT Ch6" },
200 	{ 0x2f278086, "HSW IOAT Ch7" },
201 	{ 0x2f2e8086, "HSW IOAT Ch0 (RAID)" },
202 	{ 0x2f2f8086, "HSW IOAT Ch1 (RAID)" },
203 
204 	{ 0x0c508086, "BWD IOAT Ch0" },
205 	{ 0x0c518086, "BWD IOAT Ch1" },
206 	{ 0x0c528086, "BWD IOAT Ch2" },
207 	{ 0x0c538086, "BWD IOAT Ch3" },
208 
209 	{ 0x6f508086, "BDXDE IOAT Ch0" },
210 	{ 0x6f518086, "BDXDE IOAT Ch1" },
211 	{ 0x6f528086, "BDXDE IOAT Ch2" },
212 	{ 0x6f538086, "BDXDE IOAT Ch3" },
213 
214 	{ 0x00000000, NULL           }
215 };
216 
217 /*
218  * OS <-> Driver linkage functions
219  */
220 static int
221 ioat_probe(device_t device)
222 {
223 	struct _pcsid *ep;
224 	u_int32_t type;
225 
226 	type = pci_get_devid(device);
227 	for (ep = pci_ids; ep->type; ep++) {
228 		if (ep->type == type) {
229 			device_set_desc(device, ep->desc);
230 			return (0);
231 		}
232 	}
233 	return (ENXIO);
234 }
235 
236 static int
237 ioat_attach(device_t device)
238 {
239 	struct ioat_softc *ioat;
240 	int error;
241 
242 	ioat = DEVICE2SOFTC(device);
243 	ioat->device = device;
244 
245 	error = ioat_map_pci_bar(ioat);
246 	if (error != 0)
247 		goto err;
248 
249 	ioat->version = ioat_read_cbver(ioat);
250 	if (ioat->version < IOAT_VER_3_0) {
251 		error = ENODEV;
252 		goto err;
253 	}
254 
255 	error = ioat3_attach(device);
256 	if (error != 0)
257 		goto err;
258 
259 	error = pci_enable_busmaster(device);
260 	if (error != 0)
261 		goto err;
262 
263 	error = ioat_setup_intr(ioat);
264 	if (error != 0)
265 		goto err;
266 
267 	error = ioat_reset_hw(ioat);
268 	if (error != 0)
269 		goto err;
270 
271 	ioat_process_events(ioat);
272 	ioat_setup_sysctl(device);
273 
274 	ioat_channel[ioat_channel_index++] = ioat;
275 	ioat_test_attach();
276 
277 err:
278 	if (error != 0)
279 		ioat_detach(device);
280 	return (error);
281 }
282 
283 static int
284 ioat_detach(device_t device)
285 {
286 	struct ioat_softc *ioat;
287 
288 	ioat = DEVICE2SOFTC(device);
289 
290 	ioat_test_detach();
291 	ioat_drain(ioat);
292 
293 	ioat_teardown_intr(ioat);
294 	callout_drain(&ioat->timer);
295 
296 	pci_disable_busmaster(device);
297 
298 	if (ioat->pci_resource != NULL)
299 		bus_release_resource(device, SYS_RES_MEMORY,
300 		    ioat->pci_resource_id, ioat->pci_resource);
301 
302 	if (ioat->ring != NULL)
303 		ioat_free_ring(ioat, 1 << ioat->ring_size_order, ioat->ring);
304 
305 	if (ioat->comp_update != NULL) {
306 		bus_dmamap_unload(ioat->comp_update_tag, ioat->comp_update_map);
307 		bus_dmamem_free(ioat->comp_update_tag, ioat->comp_update,
308 		    ioat->comp_update_map);
309 		bus_dma_tag_destroy(ioat->comp_update_tag);
310 	}
311 
312 	bus_dma_tag_destroy(ioat->hw_desc_tag);
313 
314 	return (0);
315 }
316 
317 static int
318 ioat_teardown_intr(struct ioat_softc *ioat)
319 {
320 
321 	if (ioat->tag != NULL)
322 		bus_teardown_intr(ioat->device, ioat->res, ioat->tag);
323 
324 	if (ioat->res != NULL)
325 		bus_release_resource(ioat->device, SYS_RES_IRQ,
326 		    rman_get_rid(ioat->res), ioat->res);
327 
328 	pci_release_msi(ioat->device);
329 	return (0);
330 }
331 
332 static int
333 ioat_start_channel(struct ioat_softc *ioat)
334 {
335 	uint64_t status;
336 	uint32_t chanerr;
337 	int i;
338 
339 	ioat_acquire(&ioat->dmaengine);
340 	ioat_null(&ioat->dmaengine, NULL, NULL, 0);
341 	ioat_release(&ioat->dmaengine);
342 
343 	for (i = 0; i < 100; i++) {
344 		DELAY(1);
345 		status = ioat_get_chansts(ioat);
346 		if (is_ioat_idle(status))
347 			return (0);
348 	}
349 
350 	chanerr = ioat_read_4(ioat, IOAT_CHANERR_OFFSET);
351 	ioat_log_message(0, "could not start channel: "
352 	    "status = %#jx error = %b\n", (uintmax_t)status, (int)chanerr,
353 	    IOAT_CHANERR_STR);
354 	return (ENXIO);
355 }
356 
357 /*
358  * Initialize Hardware
359  */
360 static int
361 ioat3_attach(device_t device)
362 {
363 	struct ioat_softc *ioat;
364 	struct ioat_descriptor **ring;
365 	struct ioat_descriptor *next;
366 	struct ioat_dma_hw_descriptor *dma_hw_desc;
367 	int i, num_descriptors;
368 	int error;
369 	uint8_t xfercap;
370 
371 	error = 0;
372 	ioat = DEVICE2SOFTC(device);
373 	ioat->capabilities = ioat_read_dmacapability(ioat);
374 
375 	ioat_log_message(1, "Capabilities: %b\n", (int)ioat->capabilities,
376 	    IOAT_DMACAP_STR);
377 
378 	xfercap = ioat_read_xfercap(ioat);
379 	ioat->max_xfer_size = 1 << xfercap;
380 
381 	/* TODO: need to check DCA here if we ever do XOR/PQ */
382 
383 	mtx_init(&ioat->submit_lock, "ioat_submit", NULL, MTX_DEF);
384 	mtx_init(&ioat->cleanup_lock, "ioat_process_events", NULL, MTX_DEF);
385 	callout_init(&ioat->timer, 1);
386 
387 	ioat->is_resize_pending = FALSE;
388 	ioat->is_completion_pending = FALSE;
389 	ioat->is_reset_pending = FALSE;
390 	ioat->is_channel_running = FALSE;
391 
392 	bus_dma_tag_create(bus_get_dma_tag(ioat->device), sizeof(uint64_t), 0x0,
393 	    BUS_SPACE_MAXADDR, BUS_SPACE_MAXADDR, NULL, NULL,
394 	    sizeof(uint64_t), 1, sizeof(uint64_t), 0, NULL, NULL,
395 	    &ioat->comp_update_tag);
396 
397 	error = bus_dmamem_alloc(ioat->comp_update_tag,
398 	    (void **)&ioat->comp_update, BUS_DMA_ZERO, &ioat->comp_update_map);
399 	if (ioat->comp_update == NULL)
400 		return (ENOMEM);
401 
402 	error = bus_dmamap_load(ioat->comp_update_tag, ioat->comp_update_map,
403 	    ioat->comp_update, sizeof(uint64_t), ioat_comp_update_map, ioat,
404 	    0);
405 	if (error != 0)
406 		return (error);
407 
408 	ioat->ring_size_order = IOAT_MIN_ORDER;
409 
410 	num_descriptors = 1 << ioat->ring_size_order;
411 
412 	bus_dma_tag_create(bus_get_dma_tag(ioat->device), 0x40, 0x0,
413 	    BUS_SPACE_MAXADDR, BUS_SPACE_MAXADDR, NULL, NULL,
414 	    sizeof(struct ioat_dma_hw_descriptor), 1,
415 	    sizeof(struct ioat_dma_hw_descriptor), 0, NULL, NULL,
416 	    &ioat->hw_desc_tag);
417 
418 	ioat->ring = malloc(num_descriptors * sizeof(*ring), M_IOAT,
419 	    M_ZERO | M_WAITOK);
420 	if (ioat->ring == NULL)
421 		return (ENOMEM);
422 
423 	ring = ioat->ring;
424 	for (i = 0; i < num_descriptors; i++) {
425 		ring[i] = ioat_alloc_ring_entry(ioat, M_WAITOK);
426 		if (ring[i] == NULL)
427 			return (ENOMEM);
428 
429 		ring[i]->id = i;
430 	}
431 
432 	for (i = 0; i < num_descriptors - 1; i++) {
433 		next = ring[i + 1];
434 		dma_hw_desc = ring[i]->u.dma;
435 
436 		dma_hw_desc->next = next->hw_desc_bus_addr;
437 	}
438 
439 	ring[i]->u.dma->next = ring[0]->hw_desc_bus_addr;
440 
441 	ioat->head = ioat->hw_head = 0;
442 	ioat->tail = 0;
443 	ioat->last_seen = 0;
444 	return (0);
445 }
446 
447 static int
448 ioat_map_pci_bar(struct ioat_softc *ioat)
449 {
450 
451 	ioat->pci_resource_id = PCIR_BAR(0);
452 	ioat->pci_resource = bus_alloc_resource_any(ioat->device,
453 	    SYS_RES_MEMORY, &ioat->pci_resource_id, RF_ACTIVE);
454 
455 	if (ioat->pci_resource == NULL) {
456 		ioat_log_message(0, "unable to allocate pci resource\n");
457 		return (ENODEV);
458 	}
459 
460 	ioat->pci_bus_tag = rman_get_bustag(ioat->pci_resource);
461 	ioat->pci_bus_handle = rman_get_bushandle(ioat->pci_resource);
462 	return (0);
463 }
464 
465 static void
466 ioat_comp_update_map(void *arg, bus_dma_segment_t *seg, int nseg, int error)
467 {
468 	struct ioat_softc *ioat = arg;
469 
470 	KASSERT(error == 0, ("%s: error:%d", __func__, error));
471 	ioat->comp_update_bus_addr = seg[0].ds_addr;
472 }
473 
474 static void
475 ioat_dmamap_cb(void *arg, bus_dma_segment_t *segs, int nseg, int error)
476 {
477 	bus_addr_t *baddr;
478 
479 	KASSERT(error == 0, ("%s: error:%d", __func__, error));
480 	baddr = arg;
481 	*baddr = segs->ds_addr;
482 }
483 
484 /*
485  * Interrupt setup and handlers
486  */
487 static int
488 ioat_setup_intr(struct ioat_softc *ioat)
489 {
490 	uint32_t num_vectors;
491 	int error;
492 	boolean_t use_msix;
493 	boolean_t force_legacy_interrupts;
494 
495 	use_msix = FALSE;
496 	force_legacy_interrupts = FALSE;
497 
498 	if (!g_force_legacy_interrupts && pci_msix_count(ioat->device) >= 1) {
499 		num_vectors = 1;
500 		pci_alloc_msix(ioat->device, &num_vectors);
501 		if (num_vectors == 1)
502 			use_msix = TRUE;
503 	}
504 
505 	if (use_msix) {
506 		ioat->rid = 1;
507 		ioat->res = bus_alloc_resource_any(ioat->device, SYS_RES_IRQ,
508 		    &ioat->rid, RF_ACTIVE);
509 	} else {
510 		ioat->rid = 0;
511 		ioat->res = bus_alloc_resource_any(ioat->device, SYS_RES_IRQ,
512 		    &ioat->rid, RF_SHAREABLE | RF_ACTIVE);
513 	}
514 	if (ioat->res == NULL) {
515 		ioat_log_message(0, "bus_alloc_resource failed\n");
516 		return (ENOMEM);
517 	}
518 
519 	ioat->tag = NULL;
520 	error = bus_setup_intr(ioat->device, ioat->res, INTR_MPSAFE |
521 	    INTR_TYPE_MISC, NULL, ioat_interrupt_handler, ioat, &ioat->tag);
522 	if (error != 0) {
523 		ioat_log_message(0, "bus_setup_intr failed\n");
524 		return (error);
525 	}
526 
527 	ioat_write_intrctrl(ioat, IOAT_INTRCTRL_MASTER_INT_EN);
528 	return (0);
529 }
530 
531 static boolean_t
532 ioat_model_resets_msix(struct ioat_softc *ioat)
533 {
534 	u_int32_t pciid;
535 
536 	pciid = pci_get_devid(ioat->device);
537 	switch (pciid) {
538 		/* BWD: */
539 	case 0x0c508086:
540 	case 0x0c518086:
541 	case 0x0c528086:
542 	case 0x0c538086:
543 		/* BDXDE: */
544 	case 0x6f508086:
545 	case 0x6f518086:
546 	case 0x6f528086:
547 	case 0x6f538086:
548 		return (TRUE);
549 	}
550 
551 	return (FALSE);
552 }
553 
554 static void
555 ioat_interrupt_handler(void *arg)
556 {
557 	struct ioat_softc *ioat = arg;
558 
559 	ioat_process_events(ioat);
560 }
561 
562 static void
563 ioat_process_events(struct ioat_softc *ioat)
564 {
565 	struct ioat_descriptor *desc;
566 	struct bus_dmadesc *dmadesc;
567 	uint64_t comp_update, status;
568 	uint32_t completed;
569 
570 	mtx_lock(&ioat->cleanup_lock);
571 
572 	completed = 0;
573 	comp_update = *ioat->comp_update;
574 	status = comp_update & IOAT_CHANSTS_COMPLETED_DESCRIPTOR_MASK;
575 
576 	CTR0(KTR_IOAT, __func__);
577 
578 	if (status == ioat->last_seen)
579 		goto out;
580 
581 	while (1) {
582 		desc = ioat_get_ring_entry(ioat, ioat->tail);
583 		dmadesc = &desc->bus_dmadesc;
584 		CTR1(KTR_IOAT, "completing desc %d", ioat->tail);
585 
586 		if (dmadesc->callback_fn)
587 			(*dmadesc->callback_fn)(dmadesc->callback_arg);
588 
589 		completed++;
590 		ioat->tail++;
591 		if (desc->hw_desc_bus_addr == status)
592 			break;
593 	}
594 
595 	ioat->last_seen = desc->hw_desc_bus_addr;
596 
597 	if (ioat->head == ioat->tail) {
598 		ioat->is_completion_pending = FALSE;
599 		callout_reset(&ioat->timer, IOAT_INTR_TIMO,
600 		    ioat_timer_callback, ioat);
601 	}
602 
603 out:
604 	ioat_write_chanctrl(ioat, IOAT_CHANCTRL_RUN);
605 	mtx_unlock(&ioat->cleanup_lock);
606 
607 	ioat_putn(ioat, completed, IOAT_ACTIVE_DESCR_REF);
608 	wakeup(&ioat->tail);
609 }
610 
611 /*
612  * User API functions
613  */
614 bus_dmaengine_t
615 ioat_get_dmaengine(uint32_t index)
616 {
617 
618 	if (index >= ioat_channel_index)
619 		return (NULL);
620 	return (&ioat_get(ioat_channel[index], IOAT_DMAENGINE_REF)->dmaengine);
621 }
622 
623 void
624 ioat_put_dmaengine(bus_dmaengine_t dmaengine)
625 {
626 	struct ioat_softc *ioat;
627 
628 	ioat = to_ioat_softc(dmaengine);
629 	ioat_put(ioat, IOAT_DMAENGINE_REF);
630 }
631 
632 void
633 ioat_acquire(bus_dmaengine_t dmaengine)
634 {
635 	struct ioat_softc *ioat;
636 
637 	ioat = to_ioat_softc(dmaengine);
638 	mtx_lock(&ioat->submit_lock);
639 	CTR0(KTR_IOAT, __func__);
640 }
641 
642 void
643 ioat_release(bus_dmaengine_t dmaengine)
644 {
645 	struct ioat_softc *ioat;
646 
647 	ioat = to_ioat_softc(dmaengine);
648 	CTR0(KTR_IOAT, __func__);
649 	ioat_write_2(ioat, IOAT_DMACOUNT_OFFSET, (uint16_t)ioat->hw_head);
650 	mtx_unlock(&ioat->submit_lock);
651 }
652 
653 static struct ioat_descriptor *
654 ioat_op_generic(struct ioat_softc *ioat, uint8_t op,
655     uint32_t size, uint64_t src, uint64_t dst,
656     bus_dmaengine_callback_t callback_fn, void *callback_arg,
657     uint32_t flags)
658 {
659 	struct ioat_generic_hw_descriptor *hw_desc;
660 	struct ioat_descriptor *desc;
661 	int mflags;
662 
663 	mtx_assert(&ioat->submit_lock, MA_OWNED);
664 
665 	KASSERT((flags & ~DMA_ALL_FLAGS) == 0, ("Unrecognized flag(s): %#x",
666 		flags & ~DMA_ALL_FLAGS));
667 	if ((flags & DMA_NO_WAIT) != 0)
668 		mflags = M_NOWAIT;
669 	else
670 		mflags = M_WAITOK;
671 
672 	if (size > ioat->max_xfer_size) {
673 		ioat_log_message(0, "%s: max_xfer_size = %d, requested = %u\n",
674 		    __func__, ioat->max_xfer_size, (unsigned)size);
675 		return (NULL);
676 	}
677 
678 	if (ioat_reserve_space(ioat, 1, mflags) != 0)
679 		return (NULL);
680 
681 	desc = ioat_get_ring_entry(ioat, ioat->head);
682 	hw_desc = desc->u.generic;
683 
684 	hw_desc->u.control_raw = 0;
685 	hw_desc->u.control_generic.op = op;
686 	hw_desc->u.control_generic.completion_update = 1;
687 
688 	if ((flags & DMA_INT_EN) != 0)
689 		hw_desc->u.control_generic.int_enable = 1;
690 
691 	hw_desc->size = size;
692 	hw_desc->src_addr = src;
693 	hw_desc->dest_addr = dst;
694 
695 	desc->bus_dmadesc.callback_fn = callback_fn;
696 	desc->bus_dmadesc.callback_arg = callback_arg;
697 	return (desc);
698 }
699 
700 struct bus_dmadesc *
701 ioat_null(bus_dmaengine_t dmaengine, bus_dmaengine_callback_t callback_fn,
702     void *callback_arg, uint32_t flags)
703 {
704 	struct ioat_dma_hw_descriptor *hw_desc;
705 	struct ioat_descriptor *desc;
706 	struct ioat_softc *ioat;
707 
708 	CTR0(KTR_IOAT, __func__);
709 	ioat = to_ioat_softc(dmaengine);
710 
711 	desc = ioat_op_generic(ioat, IOAT_OP_COPY, 8, 0, 0, callback_fn,
712 	    callback_arg, flags);
713 	if (desc == NULL)
714 		return (NULL);
715 
716 	hw_desc = desc->u.dma;
717 	hw_desc->u.control.null = 1;
718 	ioat_submit_single(ioat);
719 	return (&desc->bus_dmadesc);
720 }
721 
722 struct bus_dmadesc *
723 ioat_copy(bus_dmaengine_t dmaengine, bus_addr_t dst,
724     bus_addr_t src, bus_size_t len, bus_dmaengine_callback_t callback_fn,
725     void *callback_arg, uint32_t flags)
726 {
727 	struct ioat_dma_hw_descriptor *hw_desc;
728 	struct ioat_descriptor *desc;
729 	struct ioat_softc *ioat;
730 
731 	CTR0(KTR_IOAT, __func__);
732 	ioat = to_ioat_softc(dmaengine);
733 
734 	if (((src | dst) & (0xffffull << 48)) != 0) {
735 		ioat_log_message(0, "%s: High 16 bits of src/dst invalid\n",
736 		    __func__);
737 		return (NULL);
738 	}
739 
740 	desc = ioat_op_generic(ioat, IOAT_OP_COPY, len, src, dst, callback_fn,
741 	    callback_arg, flags);
742 	if (desc == NULL)
743 		return (NULL);
744 
745 	hw_desc = desc->u.dma;
746 	if (g_ioat_debug_level >= 3)
747 		dump_descriptor(hw_desc);
748 
749 	ioat_submit_single(ioat);
750 	return (&desc->bus_dmadesc);
751 }
752 
753 struct bus_dmadesc *
754 ioat_blockfill(bus_dmaengine_t dmaengine, bus_addr_t dst, uint64_t fillpattern,
755     bus_size_t len, bus_dmaengine_callback_t callback_fn, void *callback_arg,
756     uint32_t flags)
757 {
758 	struct ioat_fill_hw_descriptor *hw_desc;
759 	struct ioat_descriptor *desc;
760 	struct ioat_softc *ioat;
761 
762 	CTR0(KTR_IOAT, __func__);
763 	ioat = to_ioat_softc(dmaengine);
764 
765 	if ((ioat->capabilities & IOAT_DMACAP_BFILL) == 0) {
766 		ioat_log_message(0, "%s: Device lacks BFILL capability\n",
767 		    __func__);
768 		return (NULL);
769 	}
770 
771 	if ((dst & (0xffffull << 48)) != 0) {
772 		ioat_log_message(0, "%s: High 16 bits of dst invalid\n",
773 		    __func__);
774 		return (NULL);
775 	}
776 
777 	desc = ioat_op_generic(ioat, IOAT_OP_FILL, len, fillpattern, dst,
778 	    callback_fn, callback_arg, flags);
779 	if (desc == NULL)
780 		return (NULL);
781 
782 	hw_desc = desc->u.fill;
783 	if (g_ioat_debug_level >= 3)
784 		dump_descriptor(hw_desc);
785 
786 	ioat_submit_single(ioat);
787 	return (&desc->bus_dmadesc);
788 }
789 
790 /*
791  * Ring Management
792  */
793 static inline uint32_t
794 ioat_get_active(struct ioat_softc *ioat)
795 {
796 
797 	return ((ioat->head - ioat->tail) & ((1 << ioat->ring_size_order) - 1));
798 }
799 
800 static inline uint32_t
801 ioat_get_ring_space(struct ioat_softc *ioat)
802 {
803 
804 	return ((1 << ioat->ring_size_order) - ioat_get_active(ioat) - 1);
805 }
806 
807 static struct ioat_descriptor *
808 ioat_alloc_ring_entry(struct ioat_softc *ioat, int mflags)
809 {
810 	struct ioat_generic_hw_descriptor *hw_desc;
811 	struct ioat_descriptor *desc;
812 	int error, busdmaflag;
813 
814 	error = ENOMEM;
815 	hw_desc = NULL;
816 
817 	if ((mflags & M_WAITOK) != 0)
818 		busdmaflag = BUS_DMA_WAITOK;
819 	else
820 		busdmaflag = BUS_DMA_NOWAIT;
821 
822 	desc = malloc(sizeof(*desc), M_IOAT, mflags);
823 	if (desc == NULL)
824 		goto out;
825 
826 	bus_dmamem_alloc(ioat->hw_desc_tag, (void **)&hw_desc,
827 	    BUS_DMA_ZERO | busdmaflag, &ioat->hw_desc_map);
828 	if (hw_desc == NULL)
829 		goto out;
830 
831 	desc->u.generic = hw_desc;
832 
833 	error = bus_dmamap_load(ioat->hw_desc_tag, ioat->hw_desc_map, hw_desc,
834 	    sizeof(*hw_desc), ioat_dmamap_cb, &desc->hw_desc_bus_addr,
835 	    busdmaflag);
836 	if (error)
837 		goto out;
838 
839 out:
840 	if (error) {
841 		ioat_free_ring_entry(ioat, desc);
842 		return (NULL);
843 	}
844 	return (desc);
845 }
846 
847 static void
848 ioat_free_ring_entry(struct ioat_softc *ioat, struct ioat_descriptor *desc)
849 {
850 
851 	if (desc == NULL)
852 		return;
853 
854 	if (desc->u.generic)
855 		bus_dmamem_free(ioat->hw_desc_tag, desc->u.generic,
856 		    ioat->hw_desc_map);
857 	free(desc, M_IOAT);
858 }
859 
860 /*
861  * Reserves space in this IOAT descriptor ring by ensuring enough slots remain
862  * for 'num_descs'.
863  *
864  * If mflags contains M_WAITOK, blocks until enough space is available.
865  *
866  * Returns zero on success, or an errno on error.  If num_descs is beyond the
867  * maximum ring size, returns EINVAl; if allocation would block and mflags
868  * contains M_NOWAIT, returns EAGAIN.
869  *
870  * Must be called with the submit_lock held; returns with the lock held.  The
871  * lock may be dropped to allocate the ring.
872  *
873  * (The submit_lock is needed to add any entries to the ring, so callers are
874  * assured enough room is available.)
875  */
876 static int
877 ioat_reserve_space(struct ioat_softc *ioat, uint32_t num_descs, int mflags)
878 {
879 	struct ioat_descriptor **new_ring;
880 	uint32_t order;
881 	int error;
882 
883 	mtx_assert(&ioat->submit_lock, MA_OWNED);
884 	error = 0;
885 
886 	if (num_descs < 1 || num_descs > (1 << IOAT_MAX_ORDER)) {
887 		error = EINVAL;
888 		goto out;
889 	}
890 
891 	for (;;) {
892 		if (ioat_get_ring_space(ioat) >= num_descs)
893 			goto out;
894 
895 		order = ioat->ring_size_order;
896 		if (ioat->is_resize_pending || order == IOAT_MAX_ORDER) {
897 			if ((mflags & M_WAITOK) != 0) {
898 				msleep(&ioat->tail, &ioat->submit_lock, 0,
899 				    "ioat_rsz", 0);
900 				continue;
901 			}
902 
903 			error = EAGAIN;
904 			break;
905 		}
906 
907 		ioat->is_resize_pending = TRUE;
908 		for (;;) {
909 			mtx_unlock(&ioat->submit_lock);
910 
911 			new_ring = ioat_prealloc_ring(ioat, 1 << (order + 1),
912 			    TRUE, mflags);
913 
914 			mtx_lock(&ioat->submit_lock);
915 			KASSERT(ioat->ring_size_order == order,
916 			    ("is_resize_pending should protect order"));
917 
918 			if (new_ring == NULL) {
919 				KASSERT((mflags & M_WAITOK) == 0,
920 				    ("allocation failed"));
921 				error = EAGAIN;
922 				break;
923 			}
924 
925 			error = ring_grow(ioat, order, new_ring);
926 			if (error == 0)
927 				break;
928 		}
929 		ioat->is_resize_pending = FALSE;
930 		wakeup(&ioat->tail);
931 		if (error)
932 			break;
933 	}
934 
935 out:
936 	mtx_assert(&ioat->submit_lock, MA_OWNED);
937 	return (error);
938 }
939 
940 static struct ioat_descriptor **
941 ioat_prealloc_ring(struct ioat_softc *ioat, uint32_t size, boolean_t need_dscr,
942     int mflags)
943 {
944 	struct ioat_descriptor **ring;
945 	uint32_t i;
946 	int error;
947 
948 	KASSERT(size > 0 && powerof2(size), ("bogus size"));
949 
950 	ring = malloc(size * sizeof(*ring), M_IOAT, M_ZERO | mflags);
951 	if (ring == NULL)
952 		return (NULL);
953 
954 	if (need_dscr) {
955 		error = ENOMEM;
956 		for (i = size / 2; i < size; i++) {
957 			ring[i] = ioat_alloc_ring_entry(ioat, mflags);
958 			if (ring[i] == NULL)
959 				goto out;
960 			ring[i]->id = i;
961 		}
962 	}
963 	error = 0;
964 
965 out:
966 	if (error != 0 && ring != NULL) {
967 		ioat_free_ring(ioat, size, ring);
968 		ring = NULL;
969 	}
970 	return (ring);
971 }
972 
973 static void
974 ioat_free_ring(struct ioat_softc *ioat, uint32_t size,
975     struct ioat_descriptor **ring)
976 {
977 	uint32_t i;
978 
979 	for (i = 0; i < size; i++) {
980 		if (ring[i] != NULL)
981 			ioat_free_ring_entry(ioat, ring[i]);
982 	}
983 	free(ring, M_IOAT);
984 }
985 
986 static struct ioat_descriptor *
987 ioat_get_ring_entry(struct ioat_softc *ioat, uint32_t index)
988 {
989 
990 	return (ioat->ring[index % (1 << ioat->ring_size_order)]);
991 }
992 
993 static int
994 ring_grow(struct ioat_softc *ioat, uint32_t oldorder,
995     struct ioat_descriptor **newring)
996 {
997 	struct ioat_descriptor *tmp, *next;
998 	struct ioat_dma_hw_descriptor *hw;
999 	uint32_t oldsize, newsize, head, tail, i, end;
1000 	int error;
1001 
1002 	CTR0(KTR_IOAT, __func__);
1003 
1004 	mtx_assert(&ioat->submit_lock, MA_OWNED);
1005 
1006 	if (oldorder != ioat->ring_size_order || oldorder >= IOAT_MAX_ORDER) {
1007 		error = EINVAL;
1008 		goto out;
1009 	}
1010 
1011 	oldsize = (1 << oldorder);
1012 	newsize = (1 << (oldorder + 1));
1013 
1014 	mtx_lock(&ioat->cleanup_lock);
1015 
1016 	head = ioat->head & (oldsize - 1);
1017 	tail = ioat->tail & (oldsize - 1);
1018 
1019 	/* Copy old descriptors to new ring */
1020 	for (i = 0; i < oldsize; i++)
1021 		newring[i] = ioat->ring[i];
1022 
1023 	/*
1024 	 * If head has wrapped but tail hasn't, we must swap some descriptors
1025 	 * around so that tail can increment directly to head.
1026 	 */
1027 	if (head < tail) {
1028 		for (i = 0; i <= head; i++) {
1029 			tmp = newring[oldsize + i];
1030 
1031 			newring[oldsize + i] = newring[i];
1032 			newring[oldsize + i]->id = oldsize + i;
1033 
1034 			newring[i] = tmp;
1035 			newring[i]->id = i;
1036 		}
1037 		head += oldsize;
1038 	}
1039 
1040 	KASSERT(head >= tail, ("invariants"));
1041 
1042 	/* Head didn't wrap; we only need to link in oldsize..newsize */
1043 	if (head < oldsize) {
1044 		i = oldsize - 1;
1045 		end = newsize;
1046 	} else {
1047 		/* Head did wrap; link newhead..newsize and 0..oldhead */
1048 		i = head;
1049 		end = newsize + (head - oldsize) + 1;
1050 	}
1051 
1052 	/*
1053 	 * Fix up hardware ring, being careful not to trample the active
1054 	 * section (tail -> head).
1055 	 */
1056 	for (; i < end; i++) {
1057 		KASSERT((i & (newsize - 1)) < tail ||
1058 		    (i & (newsize - 1)) >= head, ("trampling snake"));
1059 
1060 		next = newring[(i + 1) & (newsize - 1)];
1061 		hw = newring[i & (newsize - 1)]->u.dma;
1062 		hw->next = next->hw_desc_bus_addr;
1063 	}
1064 
1065 	free(ioat->ring, M_IOAT);
1066 	ioat->ring = newring;
1067 	ioat->ring_size_order = oldorder + 1;
1068 	ioat->tail = tail;
1069 	ioat->head = head;
1070 	error = 0;
1071 
1072 	mtx_unlock(&ioat->cleanup_lock);
1073 out:
1074 	if (error)
1075 		ioat_free_ring(ioat, (1 << (oldorder + 1)), newring);
1076 	return (error);
1077 }
1078 
1079 static int
1080 ring_shrink(struct ioat_softc *ioat, uint32_t oldorder,
1081     struct ioat_descriptor **newring)
1082 {
1083 	struct ioat_dma_hw_descriptor *hw;
1084 	struct ioat_descriptor *ent, *next;
1085 	uint32_t oldsize, newsize, current_idx, new_idx, i;
1086 	int error;
1087 
1088 	CTR0(KTR_IOAT, __func__);
1089 
1090 	mtx_assert(&ioat->submit_lock, MA_OWNED);
1091 
1092 	if (oldorder != ioat->ring_size_order || oldorder <= IOAT_MIN_ORDER) {
1093 		error = EINVAL;
1094 		goto out_unlocked;
1095 	}
1096 
1097 	oldsize = (1 << oldorder);
1098 	newsize = (1 << (oldorder - 1));
1099 
1100 	mtx_lock(&ioat->cleanup_lock);
1101 
1102 	/* Can't shrink below current active set! */
1103 	if (ioat_get_active(ioat) >= newsize) {
1104 		error = ENOMEM;
1105 		goto out;
1106 	}
1107 
1108 	/*
1109 	 * Copy current descriptors to the new ring, dropping the removed
1110 	 * descriptors.
1111 	 */
1112 	for (i = 0; i < newsize; i++) {
1113 		current_idx = (ioat->tail + i) & (oldsize - 1);
1114 		new_idx = (ioat->tail + i) & (newsize - 1);
1115 
1116 		newring[new_idx] = ioat->ring[current_idx];
1117 		newring[new_idx]->id = new_idx;
1118 	}
1119 
1120 	/* Free deleted descriptors */
1121 	for (i = newsize; i < oldsize; i++) {
1122 		ent = ioat_get_ring_entry(ioat, ioat->tail + i);
1123 		ioat_free_ring_entry(ioat, ent);
1124 	}
1125 
1126 	/* Fix up hardware ring. */
1127 	hw = newring[(ioat->tail + newsize - 1) & (newsize - 1)]->u.dma;
1128 	next = newring[(ioat->tail + newsize) & (newsize - 1)];
1129 	hw->next = next->hw_desc_bus_addr;
1130 
1131 	free(ioat->ring, M_IOAT);
1132 	ioat->ring = newring;
1133 	ioat->ring_size_order = oldorder - 1;
1134 	error = 0;
1135 
1136 out:
1137 	mtx_unlock(&ioat->cleanup_lock);
1138 out_unlocked:
1139 	if (error)
1140 		ioat_free_ring(ioat, (1 << (oldorder - 1)), newring);
1141 	return (error);
1142 }
1143 
1144 static void
1145 ioat_halted_debug(struct ioat_softc *ioat, uint32_t chanerr)
1146 {
1147 	struct ioat_descriptor *desc;
1148 
1149 	ioat_log_message(0, "Channel halted (%b)\n", (int)chanerr,
1150 	    IOAT_CHANERR_STR);
1151 	if (chanerr == 0)
1152 		return;
1153 
1154 	mtx_lock(&ioat->submit_lock);
1155 	desc = ioat_get_ring_entry(ioat, ioat->tail + 0);
1156 	dump_descriptor(desc->u.raw);
1157 
1158 	desc = ioat_get_ring_entry(ioat, ioat->tail + 1);
1159 	dump_descriptor(desc->u.raw);
1160 	mtx_unlock(&ioat->submit_lock);
1161 }
1162 
1163 static void
1164 ioat_timer_callback(void *arg)
1165 {
1166 	struct ioat_descriptor **newring;
1167 	struct ioat_softc *ioat;
1168 	uint64_t status;
1169 	uint32_t chanerr, order;
1170 
1171 	ioat = arg;
1172 	ioat_log_message(1, "%s\n", __func__);
1173 
1174 	if (ioat->is_completion_pending) {
1175 		status = ioat_get_chansts(ioat);
1176 
1177 		/*
1178 		 * When halted due to errors, check for channel programming
1179 		 * errors before advancing the completion state.
1180 		 */
1181 		if (is_ioat_halted(status)) {
1182 			chanerr = ioat_read_4(ioat, IOAT_CHANERR_OFFSET);
1183 			ioat_halted_debug(ioat, chanerr);
1184 		}
1185 		ioat_process_events(ioat);
1186 	} else {
1187 		mtx_lock(&ioat->submit_lock);
1188 		order = ioat->ring_size_order;
1189 		if (ioat->is_resize_pending || order == IOAT_MIN_ORDER) {
1190 			mtx_unlock(&ioat->submit_lock);
1191 			goto out;
1192 		}
1193 		ioat->is_resize_pending = TRUE;
1194 		mtx_unlock(&ioat->submit_lock);
1195 
1196 		newring = ioat_prealloc_ring(ioat, 1 << (order - 1), FALSE,
1197 		    M_NOWAIT);
1198 
1199 		mtx_lock(&ioat->submit_lock);
1200 		KASSERT(ioat->ring_size_order == order,
1201 		    ("resize_pending protects order"));
1202 
1203 		if (newring != NULL)
1204 			ring_shrink(ioat, order, newring);
1205 
1206 		ioat->is_resize_pending = FALSE;
1207 		mtx_unlock(&ioat->submit_lock);
1208 
1209 out:
1210 		/* Slowly scale the ring down if idle. */
1211 		if (ioat->ring_size_order > IOAT_MIN_ORDER)
1212 			callout_reset(&ioat->timer, 10 * hz,
1213 			    ioat_timer_callback, ioat);
1214 	}
1215 }
1216 
1217 /*
1218  * Support Functions
1219  */
1220 static void
1221 ioat_submit_single(struct ioat_softc *ioat)
1222 {
1223 
1224 	ioat_get(ioat, IOAT_ACTIVE_DESCR_REF);
1225 	atomic_add_rel_int(&ioat->head, 1);
1226 	atomic_add_rel_int(&ioat->hw_head, 1);
1227 
1228 	if (!ioat->is_completion_pending) {
1229 		ioat->is_completion_pending = TRUE;
1230 		callout_reset(&ioat->timer, IOAT_INTR_TIMO,
1231 		    ioat_timer_callback, ioat);
1232 	}
1233 }
1234 
1235 static int
1236 ioat_reset_hw(struct ioat_softc *ioat)
1237 {
1238 	uint64_t status;
1239 	uint32_t chanerr;
1240 	unsigned timeout;
1241 
1242 	status = ioat_get_chansts(ioat);
1243 	if (is_ioat_active(status) || is_ioat_idle(status))
1244 		ioat_suspend(ioat);
1245 
1246 	/* Wait at most 20 ms */
1247 	for (timeout = 0; (is_ioat_active(status) || is_ioat_idle(status)) &&
1248 	    timeout < 20; timeout++) {
1249 		DELAY(1000);
1250 		status = ioat_get_chansts(ioat);
1251 	}
1252 	if (timeout == 20)
1253 		return (ETIMEDOUT);
1254 
1255 	KASSERT(ioat_get_active(ioat) == 0, ("active after quiesce"));
1256 
1257 	chanerr = ioat_read_4(ioat, IOAT_CHANERR_OFFSET);
1258 	ioat_write_4(ioat, IOAT_CHANERR_OFFSET, chanerr);
1259 
1260 	/*
1261 	 * IOAT v3 workaround - CHANERRMSK_INT with 3E07h to masks out errors
1262 	 *  that can cause stability issues for IOAT v3.
1263 	 */
1264 	pci_write_config(ioat->device, IOAT_CFG_CHANERRMASK_INT_OFFSET, 0x3e07,
1265 	    4);
1266 	chanerr = pci_read_config(ioat->device, IOAT_CFG_CHANERR_INT_OFFSET, 4);
1267 	pci_write_config(ioat->device, IOAT_CFG_CHANERR_INT_OFFSET, chanerr, 4);
1268 
1269 	/*
1270 	 * BDXDE and BWD models reset MSI-X registers on device reset.
1271 	 * Save/restore their contents manually.
1272 	 */
1273 	if (ioat_model_resets_msix(ioat)) {
1274 		ioat_log_message(1, "device resets MSI-X registers; saving\n");
1275 		pci_save_state(ioat->device);
1276 	}
1277 
1278 	ioat_reset(ioat);
1279 
1280 	/* Wait at most 20 ms */
1281 	for (timeout = 0; ioat_reset_pending(ioat) && timeout < 20; timeout++)
1282 		DELAY(1000);
1283 	if (timeout == 20)
1284 		return (ETIMEDOUT);
1285 
1286 	if (ioat_model_resets_msix(ioat)) {
1287 		ioat_log_message(1, "device resets registers; restored\n");
1288 		pci_restore_state(ioat->device);
1289 	}
1290 
1291 	/* Reset attempts to return the hardware to "halted." */
1292 	status = ioat_get_chansts(ioat);
1293 	if (is_ioat_active(status) || is_ioat_idle(status)) {
1294 		/* So this really shouldn't happen... */
1295 		ioat_log_message(0, "Device is active after a reset?\n");
1296 		ioat_write_chanctrl(ioat, IOAT_CHANCTRL_RUN);
1297 		return (0);
1298 	}
1299 
1300 	chanerr = ioat_read_4(ioat, IOAT_CHANERR_OFFSET);
1301 	ioat_halted_debug(ioat, chanerr);
1302 	if (chanerr != 0)
1303 		return (EIO);
1304 
1305 	/*
1306 	 * Bring device back online after reset.  Writing CHAINADDR brings the
1307 	 * device back to active.
1308 	 *
1309 	 * The internal ring counter resets to zero, so we have to start over
1310 	 * at zero as well.
1311 	 */
1312 	ioat->tail = ioat->head = ioat->hw_head = 0;
1313 	ioat->last_seen = 0;
1314 
1315 	ioat_write_chanctrl(ioat, IOAT_CHANCTRL_RUN);
1316 	ioat_write_chancmp(ioat, ioat->comp_update_bus_addr);
1317 	ioat_write_chainaddr(ioat, ioat->ring[0]->hw_desc_bus_addr);
1318 	return (ioat_start_channel(ioat));
1319 }
1320 
1321 static int
1322 sysctl_handle_reset(SYSCTL_HANDLER_ARGS)
1323 {
1324 	struct ioat_softc *ioat;
1325 	int error, arg;
1326 
1327 	ioat = arg1;
1328 
1329 	arg = 0;
1330 	error = SYSCTL_OUT(req, &arg, sizeof(arg));
1331 	if (error != 0 || req->newptr == NULL)
1332 		return (error);
1333 
1334 	error = SYSCTL_IN(req, &arg, sizeof(arg));
1335 	if (error != 0)
1336 		return (error);
1337 
1338 	if (arg != 0)
1339 		error = ioat_reset_hw(ioat);
1340 
1341 	return (error);
1342 }
1343 
1344 static void
1345 dump_descriptor(void *hw_desc)
1346 {
1347 	int i, j;
1348 
1349 	for (i = 0; i < 2; i++) {
1350 		for (j = 0; j < 8; j++)
1351 			printf("%08x ", ((uint32_t *)hw_desc)[i * 8 + j]);
1352 		printf("\n");
1353 	}
1354 }
1355 
1356 static void
1357 ioat_setup_sysctl(device_t device)
1358 {
1359 	struct sysctl_oid_list *par;
1360 	struct sysctl_ctx_list *ctx;
1361 	struct sysctl_oid *tree;
1362 	struct ioat_softc *ioat;
1363 
1364 	ioat = DEVICE2SOFTC(device);
1365 	ctx = device_get_sysctl_ctx(device);
1366 	tree = device_get_sysctl_tree(device);
1367 	par = SYSCTL_CHILDREN(tree);
1368 
1369 	SYSCTL_ADD_INT(ctx, par, OID_AUTO, "version", CTLFLAG_RD,
1370 	    &ioat->version, 0, "HW version (0xMM form)");
1371 	SYSCTL_ADD_UINT(ctx, par, OID_AUTO, "max_xfer_size", CTLFLAG_RD,
1372 	    &ioat->max_xfer_size, 0, "HW maximum transfer size");
1373 
1374 	SYSCTL_ADD_UINT(ctx, par, OID_AUTO, "ring_size_order", CTLFLAG_RD,
1375 	    &ioat->ring_size_order, 0, "SW descriptor ring size order");
1376 	SYSCTL_ADD_UINT(ctx, par, OID_AUTO, "head", CTLFLAG_RD, &ioat->head, 0,
1377 	    "SW descriptor head pointer index");
1378 	SYSCTL_ADD_UINT(ctx, par, OID_AUTO, "tail", CTLFLAG_RD, &ioat->tail, 0,
1379 	    "SW descriptor tail pointer index");
1380 	SYSCTL_ADD_UINT(ctx, par, OID_AUTO, "hw_head", CTLFLAG_RD,
1381 	    &ioat->hw_head, 0, "HW DMACOUNT");
1382 
1383 	SYSCTL_ADD_UQUAD(ctx, par, OID_AUTO, "last_completion", CTLFLAG_RD,
1384 	    ioat->comp_update, "HW addr of last completion");
1385 
1386 	SYSCTL_ADD_INT(ctx, par, OID_AUTO, "is_resize_pending", CTLFLAG_RD,
1387 	    &ioat->is_resize_pending, 0, "resize pending");
1388 	SYSCTL_ADD_INT(ctx, par, OID_AUTO, "is_completion_pending", CTLFLAG_RD,
1389 	    &ioat->is_completion_pending, 0, "completion pending");
1390 	SYSCTL_ADD_INT(ctx, par, OID_AUTO, "is_reset_pending", CTLFLAG_RD,
1391 	    &ioat->is_reset_pending, 0, "reset pending");
1392 	SYSCTL_ADD_INT(ctx, par, OID_AUTO, "is_channel_running", CTLFLAG_RD,
1393 	    &ioat->is_channel_running, 0, "channel running");
1394 
1395 	SYSCTL_ADD_PROC(ctx, par, OID_AUTO, "force_hw_reset",
1396 	    CTLTYPE_INT | CTLFLAG_RW, ioat, 0, sysctl_handle_reset, "I",
1397 	    "Set to non-zero to reset the hardware");
1398 }
1399 
1400 static inline struct ioat_softc *
1401 ioat_get(struct ioat_softc *ioat, enum ioat_ref_kind kind)
1402 {
1403 	uint32_t old;
1404 
1405 	KASSERT(kind < IOAT_NUM_REF_KINDS, ("bogus"));
1406 
1407 	old = atomic_fetchadd_32(&ioat->refcnt, 1);
1408 	KASSERT(old < UINT32_MAX, ("refcnt overflow"));
1409 
1410 #ifdef INVARIANTS
1411 	old = atomic_fetchadd_32(&ioat->refkinds[kind], 1);
1412 	KASSERT(old < UINT32_MAX, ("refcnt kind overflow"));
1413 #endif
1414 
1415 	return (ioat);
1416 }
1417 
1418 static inline void
1419 ioat_putn(struct ioat_softc *ioat, uint32_t n, enum ioat_ref_kind kind)
1420 {
1421 	uint32_t old;
1422 
1423 	KASSERT(kind < IOAT_NUM_REF_KINDS, ("bogus"));
1424 
1425 	if (n == 0)
1426 		return;
1427 
1428 #ifdef INVARIANTS
1429 	old = atomic_fetchadd_32(&ioat->refkinds[kind], -n);
1430 	KASSERT(old >= n, ("refcnt kind underflow"));
1431 #endif
1432 
1433 	/* Skip acquiring the lock if resulting refcnt > 0. */
1434 	for (;;) {
1435 		old = ioat->refcnt;
1436 		if (old <= n)
1437 			break;
1438 		if (atomic_cmpset_32(&ioat->refcnt, old, old - n))
1439 			return;
1440 	}
1441 
1442 	mtx_lock(IOAT_REFLK);
1443 	old = atomic_fetchadd_32(&ioat->refcnt, -n);
1444 	KASSERT(old >= n, ("refcnt error"));
1445 
1446 	if (old == n)
1447 		wakeup(IOAT_REFLK);
1448 	mtx_unlock(IOAT_REFLK);
1449 }
1450 
1451 static inline void
1452 ioat_put(struct ioat_softc *ioat, enum ioat_ref_kind kind)
1453 {
1454 
1455 	ioat_putn(ioat, 1, kind);
1456 }
1457 
1458 static void
1459 ioat_drain(struct ioat_softc *ioat)
1460 {
1461 
1462 	mtx_lock(IOAT_REFLK);
1463 	while (ioat->refcnt > 0)
1464 		msleep(IOAT_REFLK, IOAT_REFLK, 0, "ioat_drain", 0);
1465 	mtx_unlock(IOAT_REFLK);
1466 }
1467