xref: /freebsd/sys/dev/ioat/ioat.c (revision 30c90f019fae2636046f3789dab92794a138d19a)
1 /*-
2  * Copyright (C) 2012 Intel Corporation
3  * All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions
7  * are met:
8  * 1. Redistributions of source code must retain the above copyright
9  *    notice, this list of conditions and the following disclaimer.
10  * 2. Redistributions in binary form must reproduce the above copyright
11  *    notice, this list of conditions and the following disclaimer in the
12  *    documentation and/or other materials provided with the distribution.
13  *
14  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
15  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
16  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
17  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
18  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
19  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
20  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
21  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
22  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
23  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
24  * SUCH DAMAGE.
25  */
26 
27 #include <sys/cdefs.h>
28 __FBSDID("$FreeBSD$");
29 
30 #include "opt_ddb.h"
31 
32 #include <sys/param.h>
33 #include <sys/systm.h>
34 #include <sys/bus.h>
35 #include <sys/conf.h>
36 #include <sys/fail.h>
37 #include <sys/ioccom.h>
38 #include <sys/kernel.h>
39 #include <sys/lock.h>
40 #include <sys/malloc.h>
41 #include <sys/module.h>
42 #include <sys/mutex.h>
43 #include <sys/rman.h>
44 #include <sys/sbuf.h>
45 #include <sys/sysctl.h>
46 #include <sys/taskqueue.h>
47 #include <sys/time.h>
48 #include <dev/pci/pcireg.h>
49 #include <dev/pci/pcivar.h>
50 #include <machine/bus.h>
51 #include <machine/resource.h>
52 #include <machine/stdarg.h>
53 
54 #ifdef DDB
55 #include <ddb/ddb.h>
56 #endif
57 
58 #include "ioat.h"
59 #include "ioat_hw.h"
60 #include "ioat_internal.h"
61 
62 #ifndef	BUS_SPACE_MAXADDR_40BIT
63 #define	BUS_SPACE_MAXADDR_40BIT	0xFFFFFFFFFFULL
64 #endif
65 #define	IOAT_REFLK	(&ioat->submit_lock)
66 
67 static int ioat_probe(device_t device);
68 static int ioat_attach(device_t device);
69 static int ioat_detach(device_t device);
70 static int ioat_setup_intr(struct ioat_softc *ioat);
71 static int ioat_teardown_intr(struct ioat_softc *ioat);
72 static int ioat3_attach(device_t device);
73 static int ioat_start_channel(struct ioat_softc *ioat);
74 static int ioat_map_pci_bar(struct ioat_softc *ioat);
75 static void ioat_dmamap_cb(void *arg, bus_dma_segment_t *segs, int nseg,
76     int error);
77 static void ioat_interrupt_handler(void *arg);
78 static boolean_t ioat_model_resets_msix(struct ioat_softc *ioat);
79 static int chanerr_to_errno(uint32_t);
80 static void ioat_process_events(struct ioat_softc *ioat);
81 static inline uint32_t ioat_get_active(struct ioat_softc *ioat);
82 static inline uint32_t ioat_get_ring_space(struct ioat_softc *ioat);
83 static void ioat_free_ring(struct ioat_softc *, uint32_t size,
84     struct ioat_descriptor *);
85 static int ioat_reserve_space(struct ioat_softc *, uint32_t, int mflags);
86 static union ioat_hw_descriptor *ioat_get_descriptor(struct ioat_softc *,
87     uint32_t index);
88 static struct ioat_descriptor *ioat_get_ring_entry(struct ioat_softc *,
89     uint32_t index);
90 static void ioat_halted_debug(struct ioat_softc *, uint32_t);
91 static void ioat_poll_timer_callback(void *arg);
92 static void dump_descriptor(void *hw_desc);
93 static void ioat_submit_single(struct ioat_softc *ioat);
94 static void ioat_comp_update_map(void *arg, bus_dma_segment_t *seg, int nseg,
95     int error);
96 static int ioat_reset_hw(struct ioat_softc *ioat);
97 static void ioat_reset_hw_task(void *, int);
98 static void ioat_setup_sysctl(device_t device);
99 static int sysctl_handle_reset(SYSCTL_HANDLER_ARGS);
100 static inline struct ioat_softc *ioat_get(struct ioat_softc *,
101     enum ioat_ref_kind);
102 static inline void ioat_put(struct ioat_softc *, enum ioat_ref_kind);
103 static inline void _ioat_putn(struct ioat_softc *, uint32_t,
104     enum ioat_ref_kind, boolean_t);
105 static inline void ioat_putn(struct ioat_softc *, uint32_t,
106     enum ioat_ref_kind);
107 static inline void ioat_putn_locked(struct ioat_softc *, uint32_t,
108     enum ioat_ref_kind);
109 static void ioat_drain_locked(struct ioat_softc *);
110 
111 #define	ioat_log_message(v, ...) do {					\
112 	if ((v) <= g_ioat_debug_level) {				\
113 		device_printf(ioat->device, __VA_ARGS__);		\
114 	}								\
115 } while (0)
116 
117 MALLOC_DEFINE(M_IOAT, "ioat", "ioat driver memory allocations");
118 SYSCTL_NODE(_hw, OID_AUTO, ioat, CTLFLAG_RD, 0, "ioat node");
119 
120 static int g_force_legacy_interrupts;
121 SYSCTL_INT(_hw_ioat, OID_AUTO, force_legacy_interrupts, CTLFLAG_RDTUN,
122     &g_force_legacy_interrupts, 0, "Set to non-zero to force MSI-X disabled");
123 
124 int g_ioat_debug_level = 0;
125 SYSCTL_INT(_hw_ioat, OID_AUTO, debug_level, CTLFLAG_RWTUN, &g_ioat_debug_level,
126     0, "Set log level (0-3) for ioat(4). Higher is more verbose.");
127 
128 unsigned g_ioat_ring_order = 13;
129 SYSCTL_UINT(_hw_ioat, OID_AUTO, ring_order, CTLFLAG_RDTUN, &g_ioat_ring_order,
130     0, "Set IOAT ring order.  (1 << this) == ring size.");
131 
132 /*
133  * OS <-> Driver interface structures
134  */
135 static device_method_t ioat_pci_methods[] = {
136 	/* Device interface */
137 	DEVMETHOD(device_probe,     ioat_probe),
138 	DEVMETHOD(device_attach,    ioat_attach),
139 	DEVMETHOD(device_detach,    ioat_detach),
140 	DEVMETHOD_END
141 };
142 
143 static driver_t ioat_pci_driver = {
144 	"ioat",
145 	ioat_pci_methods,
146 	sizeof(struct ioat_softc),
147 };
148 
149 static devclass_t ioat_devclass;
150 DRIVER_MODULE(ioat, pci, ioat_pci_driver, ioat_devclass, 0, 0);
151 MODULE_VERSION(ioat, 1);
152 
153 /*
154  * Private data structures
155  */
156 static struct ioat_softc *ioat_channel[IOAT_MAX_CHANNELS];
157 static unsigned ioat_channel_index = 0;
158 SYSCTL_UINT(_hw_ioat, OID_AUTO, channels, CTLFLAG_RD, &ioat_channel_index, 0,
159     "Number of IOAT channels attached");
160 
161 static struct _pcsid
162 {
163 	u_int32_t   type;
164 	const char  *desc;
165 } pci_ids[] = {
166 	{ 0x34308086, "TBG IOAT Ch0" },
167 	{ 0x34318086, "TBG IOAT Ch1" },
168 	{ 0x34328086, "TBG IOAT Ch2" },
169 	{ 0x34338086, "TBG IOAT Ch3" },
170 	{ 0x34298086, "TBG IOAT Ch4" },
171 	{ 0x342a8086, "TBG IOAT Ch5" },
172 	{ 0x342b8086, "TBG IOAT Ch6" },
173 	{ 0x342c8086, "TBG IOAT Ch7" },
174 
175 	{ 0x37108086, "JSF IOAT Ch0" },
176 	{ 0x37118086, "JSF IOAT Ch1" },
177 	{ 0x37128086, "JSF IOAT Ch2" },
178 	{ 0x37138086, "JSF IOAT Ch3" },
179 	{ 0x37148086, "JSF IOAT Ch4" },
180 	{ 0x37158086, "JSF IOAT Ch5" },
181 	{ 0x37168086, "JSF IOAT Ch6" },
182 	{ 0x37178086, "JSF IOAT Ch7" },
183 	{ 0x37188086, "JSF IOAT Ch0 (RAID)" },
184 	{ 0x37198086, "JSF IOAT Ch1 (RAID)" },
185 
186 	{ 0x3c208086, "SNB IOAT Ch0" },
187 	{ 0x3c218086, "SNB IOAT Ch1" },
188 	{ 0x3c228086, "SNB IOAT Ch2" },
189 	{ 0x3c238086, "SNB IOAT Ch3" },
190 	{ 0x3c248086, "SNB IOAT Ch4" },
191 	{ 0x3c258086, "SNB IOAT Ch5" },
192 	{ 0x3c268086, "SNB IOAT Ch6" },
193 	{ 0x3c278086, "SNB IOAT Ch7" },
194 	{ 0x3c2e8086, "SNB IOAT Ch0 (RAID)" },
195 	{ 0x3c2f8086, "SNB IOAT Ch1 (RAID)" },
196 
197 	{ 0x0e208086, "IVB IOAT Ch0" },
198 	{ 0x0e218086, "IVB IOAT Ch1" },
199 	{ 0x0e228086, "IVB IOAT Ch2" },
200 	{ 0x0e238086, "IVB IOAT Ch3" },
201 	{ 0x0e248086, "IVB IOAT Ch4" },
202 	{ 0x0e258086, "IVB IOAT Ch5" },
203 	{ 0x0e268086, "IVB IOAT Ch6" },
204 	{ 0x0e278086, "IVB IOAT Ch7" },
205 	{ 0x0e2e8086, "IVB IOAT Ch0 (RAID)" },
206 	{ 0x0e2f8086, "IVB IOAT Ch1 (RAID)" },
207 
208 	{ 0x2f208086, "HSW IOAT Ch0" },
209 	{ 0x2f218086, "HSW IOAT Ch1" },
210 	{ 0x2f228086, "HSW IOAT Ch2" },
211 	{ 0x2f238086, "HSW IOAT Ch3" },
212 	{ 0x2f248086, "HSW IOAT Ch4" },
213 	{ 0x2f258086, "HSW IOAT Ch5" },
214 	{ 0x2f268086, "HSW IOAT Ch6" },
215 	{ 0x2f278086, "HSW IOAT Ch7" },
216 	{ 0x2f2e8086, "HSW IOAT Ch0 (RAID)" },
217 	{ 0x2f2f8086, "HSW IOAT Ch1 (RAID)" },
218 
219 	{ 0x0c508086, "BWD IOAT Ch0" },
220 	{ 0x0c518086, "BWD IOAT Ch1" },
221 	{ 0x0c528086, "BWD IOAT Ch2" },
222 	{ 0x0c538086, "BWD IOAT Ch3" },
223 
224 	{ 0x6f508086, "BDXDE IOAT Ch0" },
225 	{ 0x6f518086, "BDXDE IOAT Ch1" },
226 	{ 0x6f528086, "BDXDE IOAT Ch2" },
227 	{ 0x6f538086, "BDXDE IOAT Ch3" },
228 
229 	{ 0x6f208086, "BDX IOAT Ch0" },
230 	{ 0x6f218086, "BDX IOAT Ch1" },
231 	{ 0x6f228086, "BDX IOAT Ch2" },
232 	{ 0x6f238086, "BDX IOAT Ch3" },
233 	{ 0x6f248086, "BDX IOAT Ch4" },
234 	{ 0x6f258086, "BDX IOAT Ch5" },
235 	{ 0x6f268086, "BDX IOAT Ch6" },
236 	{ 0x6f278086, "BDX IOAT Ch7" },
237 	{ 0x6f2e8086, "BDX IOAT Ch0 (RAID)" },
238 	{ 0x6f2f8086, "BDX IOAT Ch1 (RAID)" },
239 
240 	{ 0x00000000, NULL           }
241 };
242 
243 /*
244  * OS <-> Driver linkage functions
245  */
246 static int
247 ioat_probe(device_t device)
248 {
249 	struct _pcsid *ep;
250 	u_int32_t type;
251 
252 	type = pci_get_devid(device);
253 	for (ep = pci_ids; ep->type; ep++) {
254 		if (ep->type == type) {
255 			device_set_desc(device, ep->desc);
256 			return (0);
257 		}
258 	}
259 	return (ENXIO);
260 }
261 
262 static int
263 ioat_attach(device_t device)
264 {
265 	struct ioat_softc *ioat;
266 	int error;
267 
268 	ioat = DEVICE2SOFTC(device);
269 	ioat->device = device;
270 
271 	error = ioat_map_pci_bar(ioat);
272 	if (error != 0)
273 		goto err;
274 
275 	ioat->version = ioat_read_cbver(ioat);
276 	if (ioat->version < IOAT_VER_3_0) {
277 		error = ENODEV;
278 		goto err;
279 	}
280 
281 	error = ioat3_attach(device);
282 	if (error != 0)
283 		goto err;
284 
285 	error = pci_enable_busmaster(device);
286 	if (error != 0)
287 		goto err;
288 
289 	error = ioat_setup_intr(ioat);
290 	if (error != 0)
291 		goto err;
292 
293 	error = ioat_reset_hw(ioat);
294 	if (error != 0)
295 		goto err;
296 
297 	ioat_process_events(ioat);
298 	ioat_setup_sysctl(device);
299 
300 	ioat->chan_idx = ioat_channel_index;
301 	ioat_channel[ioat_channel_index++] = ioat;
302 	ioat_test_attach();
303 
304 err:
305 	if (error != 0)
306 		ioat_detach(device);
307 	return (error);
308 }
309 
310 static int
311 ioat_detach(device_t device)
312 {
313 	struct ioat_softc *ioat;
314 
315 	ioat = DEVICE2SOFTC(device);
316 
317 	ioat_test_detach();
318 	taskqueue_drain(taskqueue_thread, &ioat->reset_task);
319 
320 	mtx_lock(IOAT_REFLK);
321 	ioat->quiescing = TRUE;
322 	ioat->destroying = TRUE;
323 	wakeup(&ioat->quiescing);
324 	wakeup(&ioat->resetting);
325 
326 	ioat_channel[ioat->chan_idx] = NULL;
327 
328 	ioat_drain_locked(ioat);
329 	mtx_unlock(IOAT_REFLK);
330 
331 	ioat_teardown_intr(ioat);
332 	callout_drain(&ioat->poll_timer);
333 
334 	pci_disable_busmaster(device);
335 
336 	if (ioat->pci_resource != NULL)
337 		bus_release_resource(device, SYS_RES_MEMORY,
338 		    ioat->pci_resource_id, ioat->pci_resource);
339 
340 	if (ioat->ring != NULL)
341 		ioat_free_ring(ioat, 1 << ioat->ring_size_order, ioat->ring);
342 
343 	if (ioat->comp_update != NULL) {
344 		bus_dmamap_unload(ioat->comp_update_tag, ioat->comp_update_map);
345 		bus_dmamem_free(ioat->comp_update_tag, ioat->comp_update,
346 		    ioat->comp_update_map);
347 		bus_dma_tag_destroy(ioat->comp_update_tag);
348 	}
349 
350 	if (ioat->hw_desc_ring != NULL) {
351 		bus_dmamap_unload(ioat->hw_desc_tag, ioat->hw_desc_map);
352 		bus_dmamem_free(ioat->hw_desc_tag, ioat->hw_desc_ring,
353 		    ioat->hw_desc_map);
354 		bus_dma_tag_destroy(ioat->hw_desc_tag);
355 	}
356 
357 	return (0);
358 }
359 
360 static int
361 ioat_teardown_intr(struct ioat_softc *ioat)
362 {
363 
364 	if (ioat->tag != NULL)
365 		bus_teardown_intr(ioat->device, ioat->res, ioat->tag);
366 
367 	if (ioat->res != NULL)
368 		bus_release_resource(ioat->device, SYS_RES_IRQ,
369 		    rman_get_rid(ioat->res), ioat->res);
370 
371 	pci_release_msi(ioat->device);
372 	return (0);
373 }
374 
375 static int
376 ioat_start_channel(struct ioat_softc *ioat)
377 {
378 	struct ioat_dma_hw_descriptor *hw_desc;
379 	struct ioat_descriptor *desc;
380 	struct bus_dmadesc *dmadesc;
381 	uint64_t status;
382 	uint32_t chanerr;
383 	int i;
384 
385 	ioat_acquire(&ioat->dmaengine);
386 
387 	/* Submit 'NULL' operation manually to avoid quiescing flag */
388 	desc = ioat_get_ring_entry(ioat, ioat->head);
389 	hw_desc = &ioat_get_descriptor(ioat, ioat->head)->dma;
390 	dmadesc = &desc->bus_dmadesc;
391 
392 	dmadesc->callback_fn = NULL;
393 	dmadesc->callback_arg = NULL;
394 
395 	hw_desc->u.control_raw = 0;
396 	hw_desc->u.control_generic.op = IOAT_OP_COPY;
397 	hw_desc->u.control_generic.completion_update = 1;
398 	hw_desc->size = 8;
399 	hw_desc->src_addr = 0;
400 	hw_desc->dest_addr = 0;
401 	hw_desc->u.control.null = 1;
402 
403 	ioat_submit_single(ioat);
404 	ioat_release(&ioat->dmaengine);
405 
406 	for (i = 0; i < 100; i++) {
407 		DELAY(1);
408 		status = ioat_get_chansts(ioat);
409 		if (is_ioat_idle(status))
410 			return (0);
411 	}
412 
413 	chanerr = ioat_read_4(ioat, IOAT_CHANERR_OFFSET);
414 	ioat_log_message(0, "could not start channel: "
415 	    "status = %#jx error = %b\n", (uintmax_t)status, (int)chanerr,
416 	    IOAT_CHANERR_STR);
417 	return (ENXIO);
418 }
419 
420 /*
421  * Initialize Hardware
422  */
423 static int
424 ioat3_attach(device_t device)
425 {
426 	struct ioat_softc *ioat;
427 	struct ioat_descriptor *ring;
428 	struct ioat_dma_hw_descriptor *dma_hw_desc;
429 	void *hw_desc;
430 	size_t ringsz;
431 	int i, num_descriptors;
432 	int error;
433 	uint8_t xfercap;
434 
435 	error = 0;
436 	ioat = DEVICE2SOFTC(device);
437 	ioat->capabilities = ioat_read_dmacapability(ioat);
438 
439 	ioat_log_message(0, "Capabilities: %b\n", (int)ioat->capabilities,
440 	    IOAT_DMACAP_STR);
441 
442 	xfercap = ioat_read_xfercap(ioat);
443 	ioat->max_xfer_size = 1 << xfercap;
444 
445 	ioat->intrdelay_supported = (ioat_read_2(ioat, IOAT_INTRDELAY_OFFSET) &
446 	    IOAT_INTRDELAY_SUPPORTED) != 0;
447 	if (ioat->intrdelay_supported)
448 		ioat->intrdelay_max = IOAT_INTRDELAY_US_MASK;
449 
450 	/* TODO: need to check DCA here if we ever do XOR/PQ */
451 
452 	mtx_init(&ioat->submit_lock, "ioat_submit", NULL, MTX_DEF);
453 	mtx_init(&ioat->cleanup_lock, "ioat_cleanup", NULL, MTX_DEF);
454 	callout_init(&ioat->poll_timer, 1);
455 	TASK_INIT(&ioat->reset_task, 0, ioat_reset_hw_task, ioat);
456 
457 	/* Establish lock order for Witness */
458 	mtx_lock(&ioat->submit_lock);
459 	mtx_lock(&ioat->cleanup_lock);
460 	mtx_unlock(&ioat->cleanup_lock);
461 	mtx_unlock(&ioat->submit_lock);
462 
463 	ioat->is_submitter_processing = FALSE;
464 	ioat->is_completion_pending = FALSE;
465 	ioat->is_reset_pending = FALSE;
466 	ioat->is_channel_running = FALSE;
467 
468 	bus_dma_tag_create(bus_get_dma_tag(ioat->device), sizeof(uint64_t), 0x0,
469 	    BUS_SPACE_MAXADDR, BUS_SPACE_MAXADDR, NULL, NULL,
470 	    sizeof(uint64_t), 1, sizeof(uint64_t), 0, NULL, NULL,
471 	    &ioat->comp_update_tag);
472 
473 	error = bus_dmamem_alloc(ioat->comp_update_tag,
474 	    (void **)&ioat->comp_update, BUS_DMA_ZERO, &ioat->comp_update_map);
475 	if (ioat->comp_update == NULL)
476 		return (ENOMEM);
477 
478 	error = bus_dmamap_load(ioat->comp_update_tag, ioat->comp_update_map,
479 	    ioat->comp_update, sizeof(uint64_t), ioat_comp_update_map, ioat,
480 	    0);
481 	if (error != 0)
482 		return (error);
483 
484 	ioat->ring_size_order = g_ioat_ring_order;
485 	num_descriptors = 1 << ioat->ring_size_order;
486 	ringsz = sizeof(struct ioat_dma_hw_descriptor) * num_descriptors;
487 
488 	error = bus_dma_tag_create(bus_get_dma_tag(ioat->device),
489 	    2 * 1024 * 1024, 0x0, BUS_SPACE_MAXADDR_40BIT, BUS_SPACE_MAXADDR,
490 	    NULL, NULL, ringsz, 1, ringsz, 0, NULL, NULL, &ioat->hw_desc_tag);
491 	if (error != 0)
492 		return (error);
493 
494 	error = bus_dmamem_alloc(ioat->hw_desc_tag, &hw_desc,
495 	    BUS_DMA_ZERO | BUS_DMA_WAITOK, &ioat->hw_desc_map);
496 	if (error != 0)
497 		return (error);
498 
499 	error = bus_dmamap_load(ioat->hw_desc_tag, ioat->hw_desc_map, hw_desc,
500 	    ringsz, ioat_dmamap_cb, &ioat->hw_desc_bus_addr, BUS_DMA_WAITOK);
501 	if (error)
502 		return (error);
503 
504 	ioat->hw_desc_ring = hw_desc;
505 
506 	ioat->ring = malloc(num_descriptors * sizeof(*ring), M_IOAT,
507 	    M_ZERO | M_WAITOK);
508 
509 	ring = ioat->ring;
510 	for (i = 0; i < num_descriptors; i++) {
511 		memset(&ring[i].bus_dmadesc, 0, sizeof(ring[i].bus_dmadesc));
512 		ring[i].id = i;
513 	}
514 
515 	for (i = 0; i < num_descriptors; i++) {
516 		dma_hw_desc = &ioat->hw_desc_ring[i].dma;
517 		dma_hw_desc->next = RING_PHYS_ADDR(ioat, i + 1);
518 	}
519 
520 	ioat->head = ioat->hw_head = 0;
521 	ioat->tail = 0;
522 	ioat->last_seen = 0;
523 	*ioat->comp_update = 0;
524 	return (0);
525 }
526 
527 static int
528 ioat_map_pci_bar(struct ioat_softc *ioat)
529 {
530 
531 	ioat->pci_resource_id = PCIR_BAR(0);
532 	ioat->pci_resource = bus_alloc_resource_any(ioat->device,
533 	    SYS_RES_MEMORY, &ioat->pci_resource_id, RF_ACTIVE);
534 
535 	if (ioat->pci_resource == NULL) {
536 		ioat_log_message(0, "unable to allocate pci resource\n");
537 		return (ENODEV);
538 	}
539 
540 	ioat->pci_bus_tag = rman_get_bustag(ioat->pci_resource);
541 	ioat->pci_bus_handle = rman_get_bushandle(ioat->pci_resource);
542 	return (0);
543 }
544 
545 static void
546 ioat_comp_update_map(void *arg, bus_dma_segment_t *seg, int nseg, int error)
547 {
548 	struct ioat_softc *ioat = arg;
549 
550 	KASSERT(error == 0, ("%s: error:%d", __func__, error));
551 	ioat->comp_update_bus_addr = seg[0].ds_addr;
552 }
553 
554 static void
555 ioat_dmamap_cb(void *arg, bus_dma_segment_t *segs, int nseg, int error)
556 {
557 	bus_addr_t *baddr;
558 
559 	KASSERT(error == 0, ("%s: error:%d", __func__, error));
560 	baddr = arg;
561 	*baddr = segs->ds_addr;
562 }
563 
564 /*
565  * Interrupt setup and handlers
566  */
567 static int
568 ioat_setup_intr(struct ioat_softc *ioat)
569 {
570 	uint32_t num_vectors;
571 	int error;
572 	boolean_t use_msix;
573 	boolean_t force_legacy_interrupts;
574 
575 	use_msix = FALSE;
576 	force_legacy_interrupts = FALSE;
577 
578 	if (!g_force_legacy_interrupts && pci_msix_count(ioat->device) >= 1) {
579 		num_vectors = 1;
580 		pci_alloc_msix(ioat->device, &num_vectors);
581 		if (num_vectors == 1)
582 			use_msix = TRUE;
583 	}
584 
585 	if (use_msix) {
586 		ioat->rid = 1;
587 		ioat->res = bus_alloc_resource_any(ioat->device, SYS_RES_IRQ,
588 		    &ioat->rid, RF_ACTIVE);
589 	} else {
590 		ioat->rid = 0;
591 		ioat->res = bus_alloc_resource_any(ioat->device, SYS_RES_IRQ,
592 		    &ioat->rid, RF_SHAREABLE | RF_ACTIVE);
593 	}
594 	if (ioat->res == NULL) {
595 		ioat_log_message(0, "bus_alloc_resource failed\n");
596 		return (ENOMEM);
597 	}
598 
599 	ioat->tag = NULL;
600 	error = bus_setup_intr(ioat->device, ioat->res, INTR_MPSAFE |
601 	    INTR_TYPE_MISC, NULL, ioat_interrupt_handler, ioat, &ioat->tag);
602 	if (error != 0) {
603 		ioat_log_message(0, "bus_setup_intr failed\n");
604 		return (error);
605 	}
606 
607 	ioat_write_intrctrl(ioat, IOAT_INTRCTRL_MASTER_INT_EN);
608 	return (0);
609 }
610 
611 static boolean_t
612 ioat_model_resets_msix(struct ioat_softc *ioat)
613 {
614 	u_int32_t pciid;
615 
616 	pciid = pci_get_devid(ioat->device);
617 	switch (pciid) {
618 		/* BWD: */
619 	case 0x0c508086:
620 	case 0x0c518086:
621 	case 0x0c528086:
622 	case 0x0c538086:
623 		/* BDXDE: */
624 	case 0x6f508086:
625 	case 0x6f518086:
626 	case 0x6f528086:
627 	case 0x6f538086:
628 		return (TRUE);
629 	}
630 
631 	return (FALSE);
632 }
633 
634 static void
635 ioat_interrupt_handler(void *arg)
636 {
637 	struct ioat_softc *ioat = arg;
638 
639 	ioat->stats.interrupts++;
640 	ioat_process_events(ioat);
641 }
642 
643 static int
644 chanerr_to_errno(uint32_t chanerr)
645 {
646 
647 	if (chanerr == 0)
648 		return (0);
649 	if ((chanerr & (IOAT_CHANERR_XSADDERR | IOAT_CHANERR_XDADDERR)) != 0)
650 		return (EFAULT);
651 	if ((chanerr & (IOAT_CHANERR_RDERR | IOAT_CHANERR_WDERR)) != 0)
652 		return (EIO);
653 	/* This one is probably our fault: */
654 	if ((chanerr & IOAT_CHANERR_NDADDERR) != 0)
655 		return (EIO);
656 	return (EIO);
657 }
658 
659 static void
660 ioat_process_events(struct ioat_softc *ioat)
661 {
662 	struct ioat_descriptor *desc;
663 	struct bus_dmadesc *dmadesc;
664 	uint64_t comp_update, status;
665 	uint32_t completed, chanerr;
666 	boolean_t pending;
667 	int error;
668 
669 	mtx_lock(&ioat->cleanup_lock);
670 
671 	/*
672 	 * Don't run while the hardware is being reset.  Reset is responsible
673 	 * for blocking new work and draining & completing existing work, so
674 	 * there is nothing to do until new work is queued after reset anyway.
675 	 */
676 	if (ioat->resetting_cleanup) {
677 		mtx_unlock(&ioat->cleanup_lock);
678 		return;
679 	}
680 
681 	completed = 0;
682 	comp_update = *ioat->comp_update;
683 	status = comp_update & IOAT_CHANSTS_COMPLETED_DESCRIPTOR_MASK;
684 
685 	if (status < ioat->hw_desc_bus_addr ||
686 	    status >= ioat->hw_desc_bus_addr + (1 << ioat->ring_size_order) *
687 	    sizeof(struct ioat_generic_hw_descriptor))
688 		panic("Bogus completion address %jx (channel %u)",
689 		    (uintmax_t)status, ioat->chan_idx);
690 
691 	if (status == ioat->last_seen) {
692 		/*
693 		 * If we landed in process_events and nothing has been
694 		 * completed, check for a timeout due to channel halt.
695 		 */
696 		goto out;
697 	}
698 	CTR4(KTR_IOAT, "%s channel=%u hw_status=0x%lx last_seen=0x%lx",
699 	    __func__, ioat->chan_idx, comp_update, ioat->last_seen);
700 
701 	while (RING_PHYS_ADDR(ioat, ioat->tail - 1) != status) {
702 		desc = ioat_get_ring_entry(ioat, ioat->tail);
703 		dmadesc = &desc->bus_dmadesc;
704 		CTR5(KTR_IOAT, "channel=%u completing desc idx %u (%p) ok  cb %p(%p)",
705 		    ioat->chan_idx, ioat->tail, dmadesc, dmadesc->callback_fn,
706 		    dmadesc->callback_arg);
707 
708 		if (dmadesc->callback_fn != NULL)
709 			dmadesc->callback_fn(dmadesc->callback_arg, 0);
710 
711 		completed++;
712 		ioat->tail++;
713 	}
714 	CTR5(KTR_IOAT, "%s channel=%u head=%u tail=%u active=%u", __func__,
715 	    ioat->chan_idx, ioat->head, ioat->tail, ioat_get_active(ioat));
716 
717 	if (completed != 0) {
718 		ioat->last_seen = RING_PHYS_ADDR(ioat, ioat->tail - 1);
719 		ioat->stats.descriptors_processed += completed;
720 	}
721 
722 out:
723 	ioat_write_chanctrl(ioat, IOAT_CHANCTRL_RUN);
724 
725 	/* Perform a racy check first; only take the locks if it passes. */
726 	pending = (ioat_get_active(ioat) != 0);
727 	if (!pending && ioat->is_completion_pending) {
728 		mtx_unlock(&ioat->cleanup_lock);
729 		mtx_lock(&ioat->submit_lock);
730 		mtx_lock(&ioat->cleanup_lock);
731 
732 		pending = (ioat_get_active(ioat) != 0);
733 		if (!pending && ioat->is_completion_pending) {
734 			ioat->is_completion_pending = FALSE;
735 			callout_stop(&ioat->poll_timer);
736 		}
737 		mtx_unlock(&ioat->submit_lock);
738 	}
739 	mtx_unlock(&ioat->cleanup_lock);
740 
741 	if (pending)
742 		callout_reset(&ioat->poll_timer, 1, ioat_poll_timer_callback,
743 		    ioat);
744 
745 	if (completed != 0) {
746 		ioat_putn(ioat, completed, IOAT_ACTIVE_DESCR_REF);
747 		wakeup(&ioat->tail);
748 	}
749 
750 	if (!is_ioat_halted(comp_update) && !is_ioat_suspended(comp_update))
751 		return;
752 
753 	ioat->stats.channel_halts++;
754 
755 	/*
756 	 * Fatal programming error on this DMA channel.  Flush any outstanding
757 	 * work with error status and restart the engine.
758 	 */
759 	mtx_lock(&ioat->submit_lock);
760 	mtx_lock(&ioat->cleanup_lock);
761 	ioat->quiescing = TRUE;
762 
763 	chanerr = ioat_read_4(ioat, IOAT_CHANERR_OFFSET);
764 	if (1 <= g_ioat_debug_level)
765 		ioat_halted_debug(ioat, chanerr);
766 	ioat->stats.last_halt_chanerr = chanerr;
767 
768 	while (ioat_get_active(ioat) > 0) {
769 		desc = ioat_get_ring_entry(ioat, ioat->tail);
770 		dmadesc = &desc->bus_dmadesc;
771 		CTR5(KTR_IOAT, "channel=%u completing desc idx %u (%p) err cb %p(%p)",
772 		    ioat->chan_idx, ioat->tail, dmadesc, dmadesc->callback_fn,
773 		    dmadesc->callback_arg);
774 
775 		if (dmadesc->callback_fn != NULL)
776 			dmadesc->callback_fn(dmadesc->callback_arg,
777 			    chanerr_to_errno(chanerr));
778 
779 		ioat_putn_locked(ioat, 1, IOAT_ACTIVE_DESCR_REF);
780 		ioat->tail++;
781 		ioat->stats.descriptors_processed++;
782 		ioat->stats.descriptors_error++;
783 	}
784 	CTR5(KTR_IOAT, "%s channel=%u head=%u tail=%u active=%u", __func__,
785 	    ioat->chan_idx, ioat->head, ioat->tail, ioat_get_active(ioat));
786 
787 	if (ioat->is_completion_pending) {
788 		ioat->is_completion_pending = FALSE;
789 		callout_stop(&ioat->poll_timer);
790 	}
791 
792 	/* Clear error status */
793 	ioat_write_4(ioat, IOAT_CHANERR_OFFSET, chanerr);
794 
795 	mtx_unlock(&ioat->cleanup_lock);
796 	mtx_unlock(&ioat->submit_lock);
797 
798 	ioat_log_message(0, "Resetting channel to recover from error\n");
799 	error = taskqueue_enqueue(taskqueue_thread, &ioat->reset_task);
800 	KASSERT(error == 0,
801 	    ("%s: taskqueue_enqueue failed: %d", __func__, error));
802 }
803 
804 static void
805 ioat_reset_hw_task(void *ctx, int pending __unused)
806 {
807 	struct ioat_softc *ioat;
808 	int error;
809 
810 	ioat = ctx;
811 	ioat_log_message(1, "%s: Resetting channel\n", __func__);
812 
813 	error = ioat_reset_hw(ioat);
814 	KASSERT(error == 0, ("%s: reset failed: %d", __func__, error));
815 	(void)error;
816 }
817 
818 /*
819  * User API functions
820  */
821 unsigned
822 ioat_get_nchannels(void)
823 {
824 
825 	return (ioat_channel_index);
826 }
827 
828 bus_dmaengine_t
829 ioat_get_dmaengine(uint32_t index, int flags)
830 {
831 	struct ioat_softc *ioat;
832 
833 	KASSERT((flags & ~(M_NOWAIT | M_WAITOK)) == 0,
834 	    ("invalid flags: 0x%08x", flags));
835 	KASSERT((flags & (M_NOWAIT | M_WAITOK)) != (M_NOWAIT | M_WAITOK),
836 	    ("invalid wait | nowait"));
837 
838 	if (index >= ioat_channel_index)
839 		return (NULL);
840 
841 	ioat = ioat_channel[index];
842 	if (ioat == NULL || ioat->destroying)
843 		return (NULL);
844 
845 	if (ioat->quiescing) {
846 		if ((flags & M_NOWAIT) != 0)
847 			return (NULL);
848 
849 		mtx_lock(IOAT_REFLK);
850 		while (ioat->quiescing && !ioat->destroying)
851 			msleep(&ioat->quiescing, IOAT_REFLK, 0, "getdma", 0);
852 		mtx_unlock(IOAT_REFLK);
853 
854 		if (ioat->destroying)
855 			return (NULL);
856 	}
857 
858 	/*
859 	 * There's a race here between the quiescing check and HW reset or
860 	 * module destroy.
861 	 */
862 	return (&ioat_get(ioat, IOAT_DMAENGINE_REF)->dmaengine);
863 }
864 
865 void
866 ioat_put_dmaengine(bus_dmaengine_t dmaengine)
867 {
868 	struct ioat_softc *ioat;
869 
870 	ioat = to_ioat_softc(dmaengine);
871 	ioat_put(ioat, IOAT_DMAENGINE_REF);
872 }
873 
874 int
875 ioat_get_hwversion(bus_dmaengine_t dmaengine)
876 {
877 	struct ioat_softc *ioat;
878 
879 	ioat = to_ioat_softc(dmaengine);
880 	return (ioat->version);
881 }
882 
883 size_t
884 ioat_get_max_io_size(bus_dmaengine_t dmaengine)
885 {
886 	struct ioat_softc *ioat;
887 
888 	ioat = to_ioat_softc(dmaengine);
889 	return (ioat->max_xfer_size);
890 }
891 
892 uint32_t
893 ioat_get_capabilities(bus_dmaengine_t dmaengine)
894 {
895 	struct ioat_softc *ioat;
896 
897 	ioat = to_ioat_softc(dmaengine);
898 	return (ioat->capabilities);
899 }
900 
901 int
902 ioat_set_interrupt_coalesce(bus_dmaengine_t dmaengine, uint16_t delay)
903 {
904 	struct ioat_softc *ioat;
905 
906 	ioat = to_ioat_softc(dmaengine);
907 	if (!ioat->intrdelay_supported)
908 		return (ENODEV);
909 	if (delay > ioat->intrdelay_max)
910 		return (ERANGE);
911 
912 	ioat_write_2(ioat, IOAT_INTRDELAY_OFFSET, delay);
913 	ioat->cached_intrdelay =
914 	    ioat_read_2(ioat, IOAT_INTRDELAY_OFFSET) & IOAT_INTRDELAY_US_MASK;
915 	return (0);
916 }
917 
918 uint16_t
919 ioat_get_max_coalesce_period(bus_dmaengine_t dmaengine)
920 {
921 	struct ioat_softc *ioat;
922 
923 	ioat = to_ioat_softc(dmaengine);
924 	return (ioat->intrdelay_max);
925 }
926 
927 void
928 ioat_acquire(bus_dmaengine_t dmaengine)
929 {
930 	struct ioat_softc *ioat;
931 
932 	ioat = to_ioat_softc(dmaengine);
933 	mtx_lock(&ioat->submit_lock);
934 	CTR2(KTR_IOAT, "%s channel=%u", __func__, ioat->chan_idx);
935 }
936 
937 int
938 ioat_acquire_reserve(bus_dmaengine_t dmaengine, unsigned n, int mflags)
939 {
940 	struct ioat_softc *ioat;
941 	int error;
942 
943 	ioat = to_ioat_softc(dmaengine);
944 	ioat_acquire(dmaengine);
945 
946 	error = ioat_reserve_space(ioat, n, mflags);
947 	if (error != 0)
948 		ioat_release(dmaengine);
949 	return (error);
950 }
951 
952 void
953 ioat_release(bus_dmaengine_t dmaengine)
954 {
955 	struct ioat_softc *ioat;
956 
957 	ioat = to_ioat_softc(dmaengine);
958 	CTR4(KTR_IOAT, "%s channel=%u dispatch1 hw_head=%u head=%u", __func__,
959 	    ioat->chan_idx, ioat->hw_head & UINT16_MAX, ioat->head);
960 	KFAIL_POINT_CODE(DEBUG_FP, ioat_release, /* do nothing */);
961 	CTR4(KTR_IOAT, "%s channel=%u dispatch2 hw_head=%u head=%u", __func__,
962 	    ioat->chan_idx, ioat->hw_head & UINT16_MAX, ioat->head);
963 
964 	ioat_write_2(ioat, IOAT_DMACOUNT_OFFSET, (uint16_t)ioat->hw_head);
965 
966 	if (!ioat->is_completion_pending) {
967 		ioat->is_completion_pending = TRUE;
968 		callout_reset(&ioat->poll_timer, 1, ioat_poll_timer_callback,
969 		    ioat);
970 	}
971 	mtx_unlock(&ioat->submit_lock);
972 }
973 
974 static struct ioat_descriptor *
975 ioat_op_generic(struct ioat_softc *ioat, uint8_t op,
976     uint32_t size, uint64_t src, uint64_t dst,
977     bus_dmaengine_callback_t callback_fn, void *callback_arg,
978     uint32_t flags)
979 {
980 	struct ioat_generic_hw_descriptor *hw_desc;
981 	struct ioat_descriptor *desc;
982 	int mflags;
983 
984 	mtx_assert(&ioat->submit_lock, MA_OWNED);
985 
986 	KASSERT((flags & ~_DMA_GENERIC_FLAGS) == 0,
987 	    ("Unrecognized flag(s): %#x", flags & ~_DMA_GENERIC_FLAGS));
988 	if ((flags & DMA_NO_WAIT) != 0)
989 		mflags = M_NOWAIT;
990 	else
991 		mflags = M_WAITOK;
992 
993 	if (size > ioat->max_xfer_size) {
994 		ioat_log_message(0, "%s: max_xfer_size = %d, requested = %u\n",
995 		    __func__, ioat->max_xfer_size, (unsigned)size);
996 		return (NULL);
997 	}
998 
999 	if (ioat_reserve_space(ioat, 1, mflags) != 0)
1000 		return (NULL);
1001 
1002 	desc = ioat_get_ring_entry(ioat, ioat->head);
1003 	hw_desc = &ioat_get_descriptor(ioat, ioat->head)->generic;
1004 
1005 	hw_desc->u.control_raw = 0;
1006 	hw_desc->u.control_generic.op = op;
1007 	hw_desc->u.control_generic.completion_update = 1;
1008 
1009 	if ((flags & DMA_INT_EN) != 0)
1010 		hw_desc->u.control_generic.int_enable = 1;
1011 	if ((flags & DMA_FENCE) != 0)
1012 		hw_desc->u.control_generic.fence = 1;
1013 
1014 	hw_desc->size = size;
1015 	hw_desc->src_addr = src;
1016 	hw_desc->dest_addr = dst;
1017 
1018 	desc->bus_dmadesc.callback_fn = callback_fn;
1019 	desc->bus_dmadesc.callback_arg = callback_arg;
1020 	return (desc);
1021 }
1022 
1023 struct bus_dmadesc *
1024 ioat_null(bus_dmaengine_t dmaengine, bus_dmaengine_callback_t callback_fn,
1025     void *callback_arg, uint32_t flags)
1026 {
1027 	struct ioat_dma_hw_descriptor *hw_desc;
1028 	struct ioat_descriptor *desc;
1029 	struct ioat_softc *ioat;
1030 
1031 	ioat = to_ioat_softc(dmaengine);
1032 	CTR2(KTR_IOAT, "%s channel=%u", __func__, ioat->chan_idx);
1033 
1034 	desc = ioat_op_generic(ioat, IOAT_OP_COPY, 8, 0, 0, callback_fn,
1035 	    callback_arg, flags);
1036 	if (desc == NULL)
1037 		return (NULL);
1038 
1039 	hw_desc = &ioat_get_descriptor(ioat, desc->id)->dma;
1040 	hw_desc->u.control.null = 1;
1041 	ioat_submit_single(ioat);
1042 	return (&desc->bus_dmadesc);
1043 }
1044 
1045 struct bus_dmadesc *
1046 ioat_copy(bus_dmaengine_t dmaengine, bus_addr_t dst,
1047     bus_addr_t src, bus_size_t len, bus_dmaengine_callback_t callback_fn,
1048     void *callback_arg, uint32_t flags)
1049 {
1050 	struct ioat_dma_hw_descriptor *hw_desc;
1051 	struct ioat_descriptor *desc;
1052 	struct ioat_softc *ioat;
1053 
1054 	ioat = to_ioat_softc(dmaengine);
1055 
1056 	if (((src | dst) & (0xffffull << 48)) != 0) {
1057 		ioat_log_message(0, "%s: High 16 bits of src/dst invalid\n",
1058 		    __func__);
1059 		return (NULL);
1060 	}
1061 
1062 	desc = ioat_op_generic(ioat, IOAT_OP_COPY, len, src, dst, callback_fn,
1063 	    callback_arg, flags);
1064 	if (desc == NULL)
1065 		return (NULL);
1066 
1067 	hw_desc = &ioat_get_descriptor(ioat, desc->id)->dma;
1068 	if (g_ioat_debug_level >= 3)
1069 		dump_descriptor(hw_desc);
1070 
1071 	ioat_submit_single(ioat);
1072 	CTR6(KTR_IOAT, "%s channel=%u desc=%p dest=%lx src=%lx len=%lx",
1073 	    __func__, ioat->chan_idx, &desc->bus_dmadesc, dst, src, len);
1074 	return (&desc->bus_dmadesc);
1075 }
1076 
1077 struct bus_dmadesc *
1078 ioat_copy_8k_aligned(bus_dmaengine_t dmaengine, bus_addr_t dst1,
1079     bus_addr_t dst2, bus_addr_t src1, bus_addr_t src2,
1080     bus_dmaengine_callback_t callback_fn, void *callback_arg, uint32_t flags)
1081 {
1082 	struct ioat_dma_hw_descriptor *hw_desc;
1083 	struct ioat_descriptor *desc;
1084 	struct ioat_softc *ioat;
1085 
1086 	ioat = to_ioat_softc(dmaengine);
1087 	CTR2(KTR_IOAT, "%s channel=%u", __func__, ioat->chan_idx);
1088 
1089 	if (((src1 | src2 | dst1 | dst2) & (0xffffull << 48)) != 0) {
1090 		ioat_log_message(0, "%s: High 16 bits of src/dst invalid\n",
1091 		    __func__);
1092 		return (NULL);
1093 	}
1094 	if (((src1 | src2 | dst1 | dst2) & PAGE_MASK) != 0) {
1095 		ioat_log_message(0, "%s: Addresses must be page-aligned\n",
1096 		    __func__);
1097 		return (NULL);
1098 	}
1099 
1100 	desc = ioat_op_generic(ioat, IOAT_OP_COPY, 2 * PAGE_SIZE, src1, dst1,
1101 	    callback_fn, callback_arg, flags);
1102 	if (desc == NULL)
1103 		return (NULL);
1104 
1105 	hw_desc = &ioat_get_descriptor(ioat, desc->id)->dma;
1106 	if (src2 != src1 + PAGE_SIZE) {
1107 		hw_desc->u.control.src_page_break = 1;
1108 		hw_desc->next_src_addr = src2;
1109 	}
1110 	if (dst2 != dst1 + PAGE_SIZE) {
1111 		hw_desc->u.control.dest_page_break = 1;
1112 		hw_desc->next_dest_addr = dst2;
1113 	}
1114 
1115 	if (g_ioat_debug_level >= 3)
1116 		dump_descriptor(hw_desc);
1117 
1118 	ioat_submit_single(ioat);
1119 	return (&desc->bus_dmadesc);
1120 }
1121 
1122 struct bus_dmadesc *
1123 ioat_copy_crc(bus_dmaengine_t dmaengine, bus_addr_t dst, bus_addr_t src,
1124     bus_size_t len, uint32_t *initialseed, bus_addr_t crcptr,
1125     bus_dmaengine_callback_t callback_fn, void *callback_arg, uint32_t flags)
1126 {
1127 	struct ioat_crc32_hw_descriptor *hw_desc;
1128 	struct ioat_descriptor *desc;
1129 	struct ioat_softc *ioat;
1130 	uint32_t teststore;
1131 	uint8_t op;
1132 
1133 	ioat = to_ioat_softc(dmaengine);
1134 	CTR2(KTR_IOAT, "%s channel=%u", __func__, ioat->chan_idx);
1135 
1136 	if ((ioat->capabilities & IOAT_DMACAP_MOVECRC) == 0) {
1137 		ioat_log_message(0, "%s: Device lacks MOVECRC capability\n",
1138 		    __func__);
1139 		return (NULL);
1140 	}
1141 	if (((src | dst) & (0xffffffull << 40)) != 0) {
1142 		ioat_log_message(0, "%s: High 24 bits of src/dst invalid\n",
1143 		    __func__);
1144 		return (NULL);
1145 	}
1146 	teststore = (flags & _DMA_CRC_TESTSTORE);
1147 	if (teststore == _DMA_CRC_TESTSTORE) {
1148 		ioat_log_message(0, "%s: TEST and STORE invalid\n", __func__);
1149 		return (NULL);
1150 	}
1151 	if (teststore == 0 && (flags & DMA_CRC_INLINE) != 0) {
1152 		ioat_log_message(0, "%s: INLINE invalid without TEST or STORE\n",
1153 		    __func__);
1154 		return (NULL);
1155 	}
1156 
1157 	switch (teststore) {
1158 	case DMA_CRC_STORE:
1159 		op = IOAT_OP_MOVECRC_STORE;
1160 		break;
1161 	case DMA_CRC_TEST:
1162 		op = IOAT_OP_MOVECRC_TEST;
1163 		break;
1164 	default:
1165 		KASSERT(teststore == 0, ("bogus"));
1166 		op = IOAT_OP_MOVECRC;
1167 		break;
1168 	}
1169 
1170 	if ((flags & DMA_CRC_INLINE) == 0 &&
1171 	    (crcptr & (0xffffffull << 40)) != 0) {
1172 		ioat_log_message(0,
1173 		    "%s: High 24 bits of crcptr invalid\n", __func__);
1174 		return (NULL);
1175 	}
1176 
1177 	desc = ioat_op_generic(ioat, op, len, src, dst, callback_fn,
1178 	    callback_arg, flags & ~_DMA_CRC_FLAGS);
1179 	if (desc == NULL)
1180 		return (NULL);
1181 
1182 	hw_desc = &ioat_get_descriptor(ioat, desc->id)->crc32;
1183 
1184 	if ((flags & DMA_CRC_INLINE) == 0)
1185 		hw_desc->crc_address = crcptr;
1186 	else
1187 		hw_desc->u.control.crc_location = 1;
1188 
1189 	if (initialseed != NULL) {
1190 		hw_desc->u.control.use_seed = 1;
1191 		hw_desc->seed = *initialseed;
1192 	}
1193 
1194 	if (g_ioat_debug_level >= 3)
1195 		dump_descriptor(hw_desc);
1196 
1197 	ioat_submit_single(ioat);
1198 	return (&desc->bus_dmadesc);
1199 }
1200 
1201 struct bus_dmadesc *
1202 ioat_crc(bus_dmaengine_t dmaengine, bus_addr_t src, bus_size_t len,
1203     uint32_t *initialseed, bus_addr_t crcptr,
1204     bus_dmaengine_callback_t callback_fn, void *callback_arg, uint32_t flags)
1205 {
1206 	struct ioat_crc32_hw_descriptor *hw_desc;
1207 	struct ioat_descriptor *desc;
1208 	struct ioat_softc *ioat;
1209 	uint32_t teststore;
1210 	uint8_t op;
1211 
1212 	ioat = to_ioat_softc(dmaengine);
1213 	CTR2(KTR_IOAT, "%s channel=%u", __func__, ioat->chan_idx);
1214 
1215 	if ((ioat->capabilities & IOAT_DMACAP_CRC) == 0) {
1216 		ioat_log_message(0, "%s: Device lacks CRC capability\n",
1217 		    __func__);
1218 		return (NULL);
1219 	}
1220 	if ((src & (0xffffffull << 40)) != 0) {
1221 		ioat_log_message(0, "%s: High 24 bits of src invalid\n",
1222 		    __func__);
1223 		return (NULL);
1224 	}
1225 	teststore = (flags & _DMA_CRC_TESTSTORE);
1226 	if (teststore == _DMA_CRC_TESTSTORE) {
1227 		ioat_log_message(0, "%s: TEST and STORE invalid\n", __func__);
1228 		return (NULL);
1229 	}
1230 	if (teststore == 0 && (flags & DMA_CRC_INLINE) != 0) {
1231 		ioat_log_message(0, "%s: INLINE invalid without TEST or STORE\n",
1232 		    __func__);
1233 		return (NULL);
1234 	}
1235 
1236 	switch (teststore) {
1237 	case DMA_CRC_STORE:
1238 		op = IOAT_OP_CRC_STORE;
1239 		break;
1240 	case DMA_CRC_TEST:
1241 		op = IOAT_OP_CRC_TEST;
1242 		break;
1243 	default:
1244 		KASSERT(teststore == 0, ("bogus"));
1245 		op = IOAT_OP_CRC;
1246 		break;
1247 	}
1248 
1249 	if ((flags & DMA_CRC_INLINE) == 0 &&
1250 	    (crcptr & (0xffffffull << 40)) != 0) {
1251 		ioat_log_message(0,
1252 		    "%s: High 24 bits of crcptr invalid\n", __func__);
1253 		return (NULL);
1254 	}
1255 
1256 	desc = ioat_op_generic(ioat, op, len, src, 0, callback_fn,
1257 	    callback_arg, flags & ~_DMA_CRC_FLAGS);
1258 	if (desc == NULL)
1259 		return (NULL);
1260 
1261 	hw_desc = &ioat_get_descriptor(ioat, desc->id)->crc32;
1262 
1263 	if ((flags & DMA_CRC_INLINE) == 0)
1264 		hw_desc->crc_address = crcptr;
1265 	else
1266 		hw_desc->u.control.crc_location = 1;
1267 
1268 	if (initialseed != NULL) {
1269 		hw_desc->u.control.use_seed = 1;
1270 		hw_desc->seed = *initialseed;
1271 	}
1272 
1273 	if (g_ioat_debug_level >= 3)
1274 		dump_descriptor(hw_desc);
1275 
1276 	ioat_submit_single(ioat);
1277 	return (&desc->bus_dmadesc);
1278 }
1279 
1280 struct bus_dmadesc *
1281 ioat_blockfill(bus_dmaengine_t dmaengine, bus_addr_t dst, uint64_t fillpattern,
1282     bus_size_t len, bus_dmaengine_callback_t callback_fn, void *callback_arg,
1283     uint32_t flags)
1284 {
1285 	struct ioat_fill_hw_descriptor *hw_desc;
1286 	struct ioat_descriptor *desc;
1287 	struct ioat_softc *ioat;
1288 
1289 	ioat = to_ioat_softc(dmaengine);
1290 	CTR2(KTR_IOAT, "%s channel=%u", __func__, ioat->chan_idx);
1291 
1292 	if ((ioat->capabilities & IOAT_DMACAP_BFILL) == 0) {
1293 		ioat_log_message(0, "%s: Device lacks BFILL capability\n",
1294 		    __func__);
1295 		return (NULL);
1296 	}
1297 
1298 	if ((dst & (0xffffull << 48)) != 0) {
1299 		ioat_log_message(0, "%s: High 16 bits of dst invalid\n",
1300 		    __func__);
1301 		return (NULL);
1302 	}
1303 
1304 	desc = ioat_op_generic(ioat, IOAT_OP_FILL, len, fillpattern, dst,
1305 	    callback_fn, callback_arg, flags);
1306 	if (desc == NULL)
1307 		return (NULL);
1308 
1309 	hw_desc = &ioat_get_descriptor(ioat, desc->id)->fill;
1310 	if (g_ioat_debug_level >= 3)
1311 		dump_descriptor(hw_desc);
1312 
1313 	ioat_submit_single(ioat);
1314 	return (&desc->bus_dmadesc);
1315 }
1316 
1317 /*
1318  * Ring Management
1319  */
1320 static inline uint32_t
1321 ioat_get_active(struct ioat_softc *ioat)
1322 {
1323 
1324 	return ((ioat->head - ioat->tail) & ((1 << ioat->ring_size_order) - 1));
1325 }
1326 
1327 static inline uint32_t
1328 ioat_get_ring_space(struct ioat_softc *ioat)
1329 {
1330 
1331 	return ((1 << ioat->ring_size_order) - ioat_get_active(ioat) - 1);
1332 }
1333 
1334 /*
1335  * Reserves space in this IOAT descriptor ring by ensuring enough slots remain
1336  * for 'num_descs'.
1337  *
1338  * If mflags contains M_WAITOK, blocks until enough space is available.
1339  *
1340  * Returns zero on success, or an errno on error.  If num_descs is beyond the
1341  * maximum ring size, returns EINVAl; if allocation would block and mflags
1342  * contains M_NOWAIT, returns EAGAIN.
1343  *
1344  * Must be called with the submit_lock held; returns with the lock held.  The
1345  * lock may be dropped to allocate the ring.
1346  *
1347  * (The submit_lock is needed to add any entries to the ring, so callers are
1348  * assured enough room is available.)
1349  */
1350 static int
1351 ioat_reserve_space(struct ioat_softc *ioat, uint32_t num_descs, int mflags)
1352 {
1353 	boolean_t dug;
1354 	int error;
1355 
1356 	mtx_assert(&ioat->submit_lock, MA_OWNED);
1357 	error = 0;
1358 	dug = FALSE;
1359 
1360 	if (num_descs < 1 || num_descs >= (1 << ioat->ring_size_order)) {
1361 		error = EINVAL;
1362 		goto out;
1363 	}
1364 
1365 	for (;;) {
1366 		if (ioat->quiescing) {
1367 			error = ENXIO;
1368 			goto out;
1369 		}
1370 
1371 		if (ioat_get_ring_space(ioat) >= num_descs)
1372 			goto out;
1373 
1374 		CTR3(KTR_IOAT, "%s channel=%u starved (%u)", __func__,
1375 		    ioat->chan_idx, num_descs);
1376 
1377 		if (!dug && !ioat->is_submitter_processing) {
1378 			ioat->is_submitter_processing = TRUE;
1379 			mtx_unlock(&ioat->submit_lock);
1380 
1381 			CTR2(KTR_IOAT, "%s channel=%u attempting to process events",
1382 			    __func__, ioat->chan_idx);
1383 			ioat_process_events(ioat);
1384 
1385 			mtx_lock(&ioat->submit_lock);
1386 			dug = TRUE;
1387 			KASSERT(ioat->is_submitter_processing == TRUE,
1388 			    ("is_submitter_processing"));
1389 			ioat->is_submitter_processing = FALSE;
1390 			wakeup(&ioat->tail);
1391 			continue;
1392 		}
1393 
1394 		if ((mflags & M_WAITOK) == 0) {
1395 			error = EAGAIN;
1396 			break;
1397 		}
1398 		CTR2(KTR_IOAT, "%s channel=%u blocking on completions",
1399 		    __func__, ioat->chan_idx);
1400 		msleep(&ioat->tail, &ioat->submit_lock, 0,
1401 		    "ioat_full", 0);
1402 		continue;
1403 	}
1404 
1405 out:
1406 	mtx_assert(&ioat->submit_lock, MA_OWNED);
1407 	KASSERT(!ioat->quiescing || error == ENXIO,
1408 	    ("reserved during quiesce"));
1409 	return (error);
1410 }
1411 
1412 static void
1413 ioat_free_ring(struct ioat_softc *ioat, uint32_t size,
1414     struct ioat_descriptor *ring)
1415 {
1416 
1417 	free(ring, M_IOAT);
1418 }
1419 
1420 static struct ioat_descriptor *
1421 ioat_get_ring_entry(struct ioat_softc *ioat, uint32_t index)
1422 {
1423 
1424 	return (&ioat->ring[index % (1 << ioat->ring_size_order)]);
1425 }
1426 
1427 static union ioat_hw_descriptor *
1428 ioat_get_descriptor(struct ioat_softc *ioat, uint32_t index)
1429 {
1430 
1431 	return (&ioat->hw_desc_ring[index % (1 << ioat->ring_size_order)]);
1432 }
1433 
1434 static void
1435 ioat_halted_debug(struct ioat_softc *ioat, uint32_t chanerr)
1436 {
1437 	union ioat_hw_descriptor *desc;
1438 
1439 	ioat_log_message(0, "Channel halted (%b)\n", (int)chanerr,
1440 	    IOAT_CHANERR_STR);
1441 	if (chanerr == 0)
1442 		return;
1443 
1444 	mtx_assert(&ioat->cleanup_lock, MA_OWNED);
1445 
1446 	desc = ioat_get_descriptor(ioat, ioat->tail + 0);
1447 	dump_descriptor(desc);
1448 
1449 	desc = ioat_get_descriptor(ioat, ioat->tail + 1);
1450 	dump_descriptor(desc);
1451 }
1452 
1453 static void
1454 ioat_poll_timer_callback(void *arg)
1455 {
1456 	struct ioat_softc *ioat;
1457 
1458 	ioat = arg;
1459 	ioat_log_message(3, "%s\n", __func__);
1460 
1461 	ioat_process_events(ioat);
1462 }
1463 
1464 /*
1465  * Support Functions
1466  */
1467 static void
1468 ioat_submit_single(struct ioat_softc *ioat)
1469 {
1470 
1471 	mtx_assert(&ioat->submit_lock, MA_OWNED);
1472 
1473 	ioat_get(ioat, IOAT_ACTIVE_DESCR_REF);
1474 	atomic_add_rel_int(&ioat->head, 1);
1475 	atomic_add_rel_int(&ioat->hw_head, 1);
1476 	CTR5(KTR_IOAT, "%s channel=%u head=%u hw_head=%u tail=%u", __func__,
1477 	    ioat->chan_idx, ioat->head, ioat->hw_head & UINT16_MAX,
1478 	    ioat->tail);
1479 
1480 	ioat->stats.descriptors_submitted++;
1481 }
1482 
1483 static int
1484 ioat_reset_hw(struct ioat_softc *ioat)
1485 {
1486 	uint64_t status;
1487 	uint32_t chanerr;
1488 	unsigned timeout;
1489 	int error;
1490 
1491 	CTR2(KTR_IOAT, "%s channel=%u", __func__, ioat->chan_idx);
1492 
1493 	mtx_lock(IOAT_REFLK);
1494 	while (ioat->resetting && !ioat->destroying)
1495 		msleep(&ioat->resetting, IOAT_REFLK, 0, "IRH_drain", 0);
1496 	if (ioat->destroying) {
1497 		mtx_unlock(IOAT_REFLK);
1498 		return (ENXIO);
1499 	}
1500 	ioat->resetting = TRUE;
1501 
1502 	ioat->quiescing = TRUE;
1503 	ioat_drain_locked(ioat);
1504 	mtx_unlock(IOAT_REFLK);
1505 
1506 	/*
1507 	 * Suspend ioat_process_events while the hardware and softc are in an
1508 	 * indeterminate state.
1509 	 */
1510 	mtx_lock(&ioat->cleanup_lock);
1511 	ioat->resetting_cleanup = TRUE;
1512 	mtx_unlock(&ioat->cleanup_lock);
1513 
1514 	CTR2(KTR_IOAT, "%s channel=%u quiesced and drained", __func__,
1515 	    ioat->chan_idx);
1516 
1517 	status = ioat_get_chansts(ioat);
1518 	if (is_ioat_active(status) || is_ioat_idle(status))
1519 		ioat_suspend(ioat);
1520 
1521 	/* Wait at most 20 ms */
1522 	for (timeout = 0; (is_ioat_active(status) || is_ioat_idle(status)) &&
1523 	    timeout < 20; timeout++) {
1524 		DELAY(1000);
1525 		status = ioat_get_chansts(ioat);
1526 	}
1527 	if (timeout == 20) {
1528 		error = ETIMEDOUT;
1529 		goto out;
1530 	}
1531 
1532 	KASSERT(ioat_get_active(ioat) == 0, ("active after quiesce"));
1533 
1534 	chanerr = ioat_read_4(ioat, IOAT_CHANERR_OFFSET);
1535 	ioat_write_4(ioat, IOAT_CHANERR_OFFSET, chanerr);
1536 
1537 	CTR2(KTR_IOAT, "%s channel=%u hardware suspended", __func__,
1538 	    ioat->chan_idx);
1539 
1540 	/*
1541 	 * IOAT v3 workaround - CHANERRMSK_INT with 3E07h to masks out errors
1542 	 *  that can cause stability issues for IOAT v3.
1543 	 */
1544 	pci_write_config(ioat->device, IOAT_CFG_CHANERRMASK_INT_OFFSET, 0x3e07,
1545 	    4);
1546 	chanerr = pci_read_config(ioat->device, IOAT_CFG_CHANERR_INT_OFFSET, 4);
1547 	pci_write_config(ioat->device, IOAT_CFG_CHANERR_INT_OFFSET, chanerr, 4);
1548 
1549 	/*
1550 	 * BDXDE and BWD models reset MSI-X registers on device reset.
1551 	 * Save/restore their contents manually.
1552 	 */
1553 	if (ioat_model_resets_msix(ioat)) {
1554 		ioat_log_message(1, "device resets MSI-X registers; saving\n");
1555 		pci_save_state(ioat->device);
1556 	}
1557 
1558 	ioat_reset(ioat);
1559 	CTR2(KTR_IOAT, "%s channel=%u hardware reset", __func__,
1560 	    ioat->chan_idx);
1561 
1562 	/* Wait at most 20 ms */
1563 	for (timeout = 0; ioat_reset_pending(ioat) && timeout < 20; timeout++)
1564 		DELAY(1000);
1565 	if (timeout == 20) {
1566 		error = ETIMEDOUT;
1567 		goto out;
1568 	}
1569 
1570 	if (ioat_model_resets_msix(ioat)) {
1571 		ioat_log_message(1, "device resets registers; restored\n");
1572 		pci_restore_state(ioat->device);
1573 	}
1574 
1575 	/* Reset attempts to return the hardware to "halted." */
1576 	status = ioat_get_chansts(ioat);
1577 	if (is_ioat_active(status) || is_ioat_idle(status)) {
1578 		/* So this really shouldn't happen... */
1579 		ioat_log_message(0, "Device is active after a reset?\n");
1580 		ioat_write_chanctrl(ioat, IOAT_CHANCTRL_RUN);
1581 		error = 0;
1582 		goto out;
1583 	}
1584 
1585 	chanerr = ioat_read_4(ioat, IOAT_CHANERR_OFFSET);
1586 	if (chanerr != 0) {
1587 		mtx_lock(&ioat->cleanup_lock);
1588 		ioat_halted_debug(ioat, chanerr);
1589 		mtx_unlock(&ioat->cleanup_lock);
1590 		error = EIO;
1591 		goto out;
1592 	}
1593 
1594 	/*
1595 	 * Bring device back online after reset.  Writing CHAINADDR brings the
1596 	 * device back to active.
1597 	 *
1598 	 * The internal ring counter resets to zero, so we have to start over
1599 	 * at zero as well.
1600 	 */
1601 	ioat->tail = ioat->head = ioat->hw_head = 0;
1602 	ioat->last_seen = 0;
1603 	*ioat->comp_update = 0;
1604 	KASSERT(!ioat->is_completion_pending, ("bogus completion_pending"));
1605 
1606 	ioat_write_chanctrl(ioat, IOAT_CHANCTRL_RUN);
1607 	ioat_write_chancmp(ioat, ioat->comp_update_bus_addr);
1608 	ioat_write_chainaddr(ioat, RING_PHYS_ADDR(ioat, 0));
1609 	error = 0;
1610 	CTR2(KTR_IOAT, "%s channel=%u configured channel", __func__,
1611 	    ioat->chan_idx);
1612 
1613 out:
1614 	/* Enqueues a null operation and ensures it completes. */
1615 	if (error == 0) {
1616 		error = ioat_start_channel(ioat);
1617 		CTR2(KTR_IOAT, "%s channel=%u started channel", __func__,
1618 		    ioat->chan_idx);
1619 	}
1620 
1621 	/*
1622 	 * Resume completions now that ring state is consistent.
1623 	 */
1624 	mtx_lock(&ioat->cleanup_lock);
1625 	ioat->resetting_cleanup = FALSE;
1626 	mtx_unlock(&ioat->cleanup_lock);
1627 
1628 	/* Unblock submission of new work */
1629 	mtx_lock(IOAT_REFLK);
1630 	ioat->quiescing = FALSE;
1631 	wakeup(&ioat->quiescing);
1632 
1633 	ioat->resetting = FALSE;
1634 	wakeup(&ioat->resetting);
1635 
1636 	if (ioat->is_completion_pending)
1637 		callout_reset(&ioat->poll_timer, 1, ioat_poll_timer_callback,
1638 		    ioat);
1639 	CTR2(KTR_IOAT, "%s channel=%u reset done", __func__, ioat->chan_idx);
1640 	mtx_unlock(IOAT_REFLK);
1641 
1642 	return (error);
1643 }
1644 
1645 static int
1646 sysctl_handle_chansts(SYSCTL_HANDLER_ARGS)
1647 {
1648 	struct ioat_softc *ioat;
1649 	struct sbuf sb;
1650 	uint64_t status;
1651 	int error;
1652 
1653 	ioat = arg1;
1654 
1655 	status = ioat_get_chansts(ioat) & IOAT_CHANSTS_STATUS;
1656 
1657 	sbuf_new_for_sysctl(&sb, NULL, 256, req);
1658 	switch (status) {
1659 	case IOAT_CHANSTS_ACTIVE:
1660 		sbuf_printf(&sb, "ACTIVE");
1661 		break;
1662 	case IOAT_CHANSTS_IDLE:
1663 		sbuf_printf(&sb, "IDLE");
1664 		break;
1665 	case IOAT_CHANSTS_SUSPENDED:
1666 		sbuf_printf(&sb, "SUSPENDED");
1667 		break;
1668 	case IOAT_CHANSTS_HALTED:
1669 		sbuf_printf(&sb, "HALTED");
1670 		break;
1671 	case IOAT_CHANSTS_ARMED:
1672 		sbuf_printf(&sb, "ARMED");
1673 		break;
1674 	default:
1675 		sbuf_printf(&sb, "UNKNOWN");
1676 		break;
1677 	}
1678 	error = sbuf_finish(&sb);
1679 	sbuf_delete(&sb);
1680 
1681 	if (error != 0 || req->newptr == NULL)
1682 		return (error);
1683 	return (EINVAL);
1684 }
1685 
1686 static int
1687 sysctl_handle_dpi(SYSCTL_HANDLER_ARGS)
1688 {
1689 	struct ioat_softc *ioat;
1690 	struct sbuf sb;
1691 #define	PRECISION	"1"
1692 	const uintmax_t factor = 10;
1693 	uintmax_t rate;
1694 	int error;
1695 
1696 	ioat = arg1;
1697 	sbuf_new_for_sysctl(&sb, NULL, 16, req);
1698 
1699 	if (ioat->stats.interrupts == 0) {
1700 		sbuf_printf(&sb, "NaN");
1701 		goto out;
1702 	}
1703 	rate = ioat->stats.descriptors_processed * factor /
1704 	    ioat->stats.interrupts;
1705 	sbuf_printf(&sb, "%ju.%." PRECISION "ju", rate / factor,
1706 	    rate % factor);
1707 #undef	PRECISION
1708 out:
1709 	error = sbuf_finish(&sb);
1710 	sbuf_delete(&sb);
1711 	if (error != 0 || req->newptr == NULL)
1712 		return (error);
1713 	return (EINVAL);
1714 }
1715 
1716 static int
1717 sysctl_handle_reset(SYSCTL_HANDLER_ARGS)
1718 {
1719 	struct ioat_softc *ioat;
1720 	int error, arg;
1721 
1722 	ioat = arg1;
1723 
1724 	arg = 0;
1725 	error = SYSCTL_OUT(req, &arg, sizeof(arg));
1726 	if (error != 0 || req->newptr == NULL)
1727 		return (error);
1728 
1729 	error = SYSCTL_IN(req, &arg, sizeof(arg));
1730 	if (error != 0)
1731 		return (error);
1732 
1733 	if (arg != 0)
1734 		error = ioat_reset_hw(ioat);
1735 
1736 	return (error);
1737 }
1738 
1739 static void
1740 dump_descriptor(void *hw_desc)
1741 {
1742 	int i, j;
1743 
1744 	for (i = 0; i < 2; i++) {
1745 		for (j = 0; j < 8; j++)
1746 			printf("%08x ", ((uint32_t *)hw_desc)[i * 8 + j]);
1747 		printf("\n");
1748 	}
1749 }
1750 
1751 static void
1752 ioat_setup_sysctl(device_t device)
1753 {
1754 	struct sysctl_oid_list *par, *statpar, *state, *hammer;
1755 	struct sysctl_ctx_list *ctx;
1756 	struct sysctl_oid *tree, *tmp;
1757 	struct ioat_softc *ioat;
1758 
1759 	ioat = DEVICE2SOFTC(device);
1760 	ctx = device_get_sysctl_ctx(device);
1761 	tree = device_get_sysctl_tree(device);
1762 	par = SYSCTL_CHILDREN(tree);
1763 
1764 	SYSCTL_ADD_INT(ctx, par, OID_AUTO, "version", CTLFLAG_RD,
1765 	    &ioat->version, 0, "HW version (0xMM form)");
1766 	SYSCTL_ADD_UINT(ctx, par, OID_AUTO, "max_xfer_size", CTLFLAG_RD,
1767 	    &ioat->max_xfer_size, 0, "HW maximum transfer size");
1768 	SYSCTL_ADD_INT(ctx, par, OID_AUTO, "intrdelay_supported", CTLFLAG_RD,
1769 	    &ioat->intrdelay_supported, 0, "Is INTRDELAY supported");
1770 	SYSCTL_ADD_U16(ctx, par, OID_AUTO, "intrdelay_max", CTLFLAG_RD,
1771 	    &ioat->intrdelay_max, 0,
1772 	    "Maximum configurable INTRDELAY on this channel (microseconds)");
1773 
1774 	tmp = SYSCTL_ADD_NODE(ctx, par, OID_AUTO, "state", CTLFLAG_RD, NULL,
1775 	    "IOAT channel internal state");
1776 	state = SYSCTL_CHILDREN(tmp);
1777 
1778 	SYSCTL_ADD_UINT(ctx, state, OID_AUTO, "ring_size_order", CTLFLAG_RD,
1779 	    &ioat->ring_size_order, 0, "SW descriptor ring size order");
1780 	SYSCTL_ADD_UINT(ctx, state, OID_AUTO, "head", CTLFLAG_RD, &ioat->head,
1781 	    0, "SW descriptor head pointer index");
1782 	SYSCTL_ADD_UINT(ctx, state, OID_AUTO, "tail", CTLFLAG_RD, &ioat->tail,
1783 	    0, "SW descriptor tail pointer index");
1784 	SYSCTL_ADD_UINT(ctx, state, OID_AUTO, "hw_head", CTLFLAG_RD,
1785 	    &ioat->hw_head, 0, "HW DMACOUNT");
1786 
1787 	SYSCTL_ADD_UQUAD(ctx, state, OID_AUTO, "last_completion", CTLFLAG_RD,
1788 	    ioat->comp_update, "HW addr of last completion");
1789 
1790 	SYSCTL_ADD_INT(ctx, state, OID_AUTO, "is_submitter_processing",
1791 	    CTLFLAG_RD, &ioat->is_submitter_processing, 0,
1792 	    "submitter processing");
1793 	SYSCTL_ADD_INT(ctx, state, OID_AUTO, "is_completion_pending",
1794 	    CTLFLAG_RD, &ioat->is_completion_pending, 0, "completion pending");
1795 	SYSCTL_ADD_INT(ctx, state, OID_AUTO, "is_reset_pending", CTLFLAG_RD,
1796 	    &ioat->is_reset_pending, 0, "reset pending");
1797 	SYSCTL_ADD_INT(ctx, state, OID_AUTO, "is_channel_running", CTLFLAG_RD,
1798 	    &ioat->is_channel_running, 0, "channel running");
1799 
1800 	SYSCTL_ADD_PROC(ctx, state, OID_AUTO, "chansts",
1801 	    CTLTYPE_STRING | CTLFLAG_RD, ioat, 0, sysctl_handle_chansts, "A",
1802 	    "String of the channel status");
1803 
1804 	SYSCTL_ADD_U16(ctx, state, OID_AUTO, "intrdelay", CTLFLAG_RD,
1805 	    &ioat->cached_intrdelay, 0,
1806 	    "Current INTRDELAY on this channel (cached, microseconds)");
1807 
1808 	tmp = SYSCTL_ADD_NODE(ctx, par, OID_AUTO, "hammer", CTLFLAG_RD, NULL,
1809 	    "Big hammers (mostly for testing)");
1810 	hammer = SYSCTL_CHILDREN(tmp);
1811 
1812 	SYSCTL_ADD_PROC(ctx, hammer, OID_AUTO, "force_hw_reset",
1813 	    CTLTYPE_INT | CTLFLAG_RW, ioat, 0, sysctl_handle_reset, "I",
1814 	    "Set to non-zero to reset the hardware");
1815 
1816 	tmp = SYSCTL_ADD_NODE(ctx, par, OID_AUTO, "stats", CTLFLAG_RD, NULL,
1817 	    "IOAT channel statistics");
1818 	statpar = SYSCTL_CHILDREN(tmp);
1819 
1820 	SYSCTL_ADD_UQUAD(ctx, statpar, OID_AUTO, "interrupts", CTLFLAG_RW,
1821 	    &ioat->stats.interrupts,
1822 	    "Number of interrupts processed on this channel");
1823 	SYSCTL_ADD_UQUAD(ctx, statpar, OID_AUTO, "descriptors", CTLFLAG_RW,
1824 	    &ioat->stats.descriptors_processed,
1825 	    "Number of descriptors processed on this channel");
1826 	SYSCTL_ADD_UQUAD(ctx, statpar, OID_AUTO, "submitted", CTLFLAG_RW,
1827 	    &ioat->stats.descriptors_submitted,
1828 	    "Number of descriptors submitted to this channel");
1829 	SYSCTL_ADD_UQUAD(ctx, statpar, OID_AUTO, "errored", CTLFLAG_RW,
1830 	    &ioat->stats.descriptors_error,
1831 	    "Number of descriptors failed by channel errors");
1832 	SYSCTL_ADD_U32(ctx, statpar, OID_AUTO, "halts", CTLFLAG_RW,
1833 	    &ioat->stats.channel_halts, 0,
1834 	    "Number of times the channel has halted");
1835 	SYSCTL_ADD_U32(ctx, statpar, OID_AUTO, "last_halt_chanerr", CTLFLAG_RW,
1836 	    &ioat->stats.last_halt_chanerr, 0,
1837 	    "The raw CHANERR when the channel was last halted");
1838 
1839 	SYSCTL_ADD_PROC(ctx, statpar, OID_AUTO, "desc_per_interrupt",
1840 	    CTLTYPE_STRING | CTLFLAG_RD, ioat, 0, sysctl_handle_dpi, "A",
1841 	    "Descriptors per interrupt");
1842 }
1843 
1844 static inline struct ioat_softc *
1845 ioat_get(struct ioat_softc *ioat, enum ioat_ref_kind kind)
1846 {
1847 	uint32_t old;
1848 
1849 	KASSERT(kind < IOAT_NUM_REF_KINDS, ("bogus"));
1850 
1851 	old = atomic_fetchadd_32(&ioat->refcnt, 1);
1852 	KASSERT(old < UINT32_MAX, ("refcnt overflow"));
1853 
1854 #ifdef INVARIANTS
1855 	old = atomic_fetchadd_32(&ioat->refkinds[kind], 1);
1856 	KASSERT(old < UINT32_MAX, ("refcnt kind overflow"));
1857 #endif
1858 
1859 	return (ioat);
1860 }
1861 
1862 static inline void
1863 ioat_putn(struct ioat_softc *ioat, uint32_t n, enum ioat_ref_kind kind)
1864 {
1865 
1866 	_ioat_putn(ioat, n, kind, FALSE);
1867 }
1868 
1869 static inline void
1870 ioat_putn_locked(struct ioat_softc *ioat, uint32_t n, enum ioat_ref_kind kind)
1871 {
1872 
1873 	_ioat_putn(ioat, n, kind, TRUE);
1874 }
1875 
1876 static inline void
1877 _ioat_putn(struct ioat_softc *ioat, uint32_t n, enum ioat_ref_kind kind,
1878     boolean_t locked)
1879 {
1880 	uint32_t old;
1881 
1882 	KASSERT(kind < IOAT_NUM_REF_KINDS, ("bogus"));
1883 
1884 	if (n == 0)
1885 		return;
1886 
1887 #ifdef INVARIANTS
1888 	old = atomic_fetchadd_32(&ioat->refkinds[kind], -n);
1889 	KASSERT(old >= n, ("refcnt kind underflow"));
1890 #endif
1891 
1892 	/* Skip acquiring the lock if resulting refcnt > 0. */
1893 	for (;;) {
1894 		old = ioat->refcnt;
1895 		if (old <= n)
1896 			break;
1897 		if (atomic_cmpset_32(&ioat->refcnt, old, old - n))
1898 			return;
1899 	}
1900 
1901 	if (locked)
1902 		mtx_assert(IOAT_REFLK, MA_OWNED);
1903 	else
1904 		mtx_lock(IOAT_REFLK);
1905 
1906 	old = atomic_fetchadd_32(&ioat->refcnt, -n);
1907 	KASSERT(old >= n, ("refcnt error"));
1908 
1909 	if (old == n)
1910 		wakeup(IOAT_REFLK);
1911 	if (!locked)
1912 		mtx_unlock(IOAT_REFLK);
1913 }
1914 
1915 static inline void
1916 ioat_put(struct ioat_softc *ioat, enum ioat_ref_kind kind)
1917 {
1918 
1919 	ioat_putn(ioat, 1, kind);
1920 }
1921 
1922 static void
1923 ioat_drain_locked(struct ioat_softc *ioat)
1924 {
1925 
1926 	mtx_assert(IOAT_REFLK, MA_OWNED);
1927 	while (ioat->refcnt > 0)
1928 		msleep(IOAT_REFLK, IOAT_REFLK, 0, "ioat_drain", 0);
1929 }
1930 
1931 #ifdef DDB
1932 #define	_db_show_lock(lo)	LOCK_CLASS(lo)->lc_ddb_show(lo)
1933 #define	db_show_lock(lk)	_db_show_lock(&(lk)->lock_object)
1934 DB_SHOW_COMMAND(ioat, db_show_ioat)
1935 {
1936 	struct ioat_softc *sc;
1937 	unsigned idx;
1938 
1939 	if (!have_addr)
1940 		goto usage;
1941 	idx = (unsigned)addr;
1942 	if (idx >= ioat_channel_index)
1943 		goto usage;
1944 
1945 	sc = ioat_channel[idx];
1946 	db_printf("ioat softc at %p\n", sc);
1947 	if (sc == NULL)
1948 		return;
1949 
1950 	db_printf(" version: %d\n", sc->version);
1951 	db_printf(" chan_idx: %u\n", sc->chan_idx);
1952 	db_printf(" submit_lock: ");
1953 	db_show_lock(&sc->submit_lock);
1954 
1955 	db_printf(" capabilities: %b\n", (int)sc->capabilities,
1956 	    IOAT_DMACAP_STR);
1957 	db_printf(" cached_intrdelay: %u\n", sc->cached_intrdelay);
1958 	db_printf(" *comp_update: 0x%jx\n", (uintmax_t)*sc->comp_update);
1959 
1960 	db_printf(" poll_timer:\n");
1961 	db_printf("  c_time: %ju\n", (uintmax_t)sc->poll_timer.c_time);
1962 	db_printf("  c_arg: %p\n", sc->poll_timer.c_arg);
1963 	db_printf("  c_func: %p\n", sc->poll_timer.c_func);
1964 	db_printf("  c_lock: %p\n", sc->poll_timer.c_lock);
1965 	db_printf("  c_flags: 0x%x\n", (unsigned)sc->poll_timer.c_flags);
1966 
1967 	db_printf(" quiescing: %d\n", (int)sc->quiescing);
1968 	db_printf(" destroying: %d\n", (int)sc->destroying);
1969 	db_printf(" is_submitter_processing: %d\n",
1970 	    (int)sc->is_submitter_processing);
1971 	db_printf(" is_completion_pending: %d\n", (int)sc->is_completion_pending);
1972 	db_printf(" is_reset_pending: %d\n", (int)sc->is_reset_pending);
1973 	db_printf(" is_channel_running: %d\n", (int)sc->is_channel_running);
1974 	db_printf(" intrdelay_supported: %d\n", (int)sc->intrdelay_supported);
1975 	db_printf(" resetting: %d\n", (int)sc->resetting);
1976 
1977 	db_printf(" head: %u\n", sc->head);
1978 	db_printf(" tail: %u\n", sc->tail);
1979 	db_printf(" hw_head: %u\n", sc->hw_head);
1980 	db_printf(" ring_size_order: %u\n", sc->ring_size_order);
1981 	db_printf(" last_seen: 0x%lx\n", sc->last_seen);
1982 	db_printf(" ring: %p\n", sc->ring);
1983 	db_printf(" descriptors: %p\n", sc->hw_desc_ring);
1984 	db_printf(" descriptors (phys): 0x%jx\n",
1985 	    (uintmax_t)sc->hw_desc_bus_addr);
1986 
1987 	db_printf("  ring[%u] (tail):\n", sc->tail %
1988 	    (1 << sc->ring_size_order));
1989 	db_printf("   id: %u\n", ioat_get_ring_entry(sc, sc->tail)->id);
1990 	db_printf("   addr: 0x%lx\n",
1991 	    RING_PHYS_ADDR(sc, sc->tail));
1992 	db_printf("   next: 0x%lx\n",
1993 	     ioat_get_descriptor(sc, sc->tail)->generic.next);
1994 
1995 	db_printf("  ring[%u] (head - 1):\n", (sc->head - 1) %
1996 	    (1 << sc->ring_size_order));
1997 	db_printf("   id: %u\n", ioat_get_ring_entry(sc, sc->head - 1)->id);
1998 	db_printf("   addr: 0x%lx\n",
1999 	    RING_PHYS_ADDR(sc, sc->head - 1));
2000 	db_printf("   next: 0x%lx\n",
2001 	     ioat_get_descriptor(sc, sc->head - 1)->generic.next);
2002 
2003 	db_printf("  ring[%u] (head):\n", (sc->head) %
2004 	    (1 << sc->ring_size_order));
2005 	db_printf("   id: %u\n", ioat_get_ring_entry(sc, sc->head)->id);
2006 	db_printf("   addr: 0x%lx\n",
2007 	    RING_PHYS_ADDR(sc, sc->head));
2008 	db_printf("   next: 0x%lx\n",
2009 	     ioat_get_descriptor(sc, sc->head)->generic.next);
2010 
2011 	for (idx = 0; idx < (1 << sc->ring_size_order); idx++)
2012 		if ((*sc->comp_update & IOAT_CHANSTS_COMPLETED_DESCRIPTOR_MASK)
2013 		    == RING_PHYS_ADDR(sc, idx))
2014 			db_printf("  ring[%u] == hardware tail\n", idx);
2015 
2016 	db_printf(" cleanup_lock: ");
2017 	db_show_lock(&sc->cleanup_lock);
2018 
2019 	db_printf(" refcnt: %u\n", sc->refcnt);
2020 #ifdef INVARIANTS
2021 	CTASSERT(IOAT_NUM_REF_KINDS == 2);
2022 	db_printf(" refkinds: [ENG=%u, DESCR=%u]\n", sc->refkinds[0],
2023 	    sc->refkinds[1]);
2024 #endif
2025 	db_printf(" stats:\n");
2026 	db_printf("  interrupts: %lu\n", sc->stats.interrupts);
2027 	db_printf("  descriptors_processed: %lu\n", sc->stats.descriptors_processed);
2028 	db_printf("  descriptors_error: %lu\n", sc->stats.descriptors_error);
2029 	db_printf("  descriptors_submitted: %lu\n", sc->stats.descriptors_submitted);
2030 
2031 	db_printf("  channel_halts: %u\n", sc->stats.channel_halts);
2032 	db_printf("  last_halt_chanerr: %u\n", sc->stats.last_halt_chanerr);
2033 
2034 	if (db_pager_quit)
2035 		return;
2036 
2037 	db_printf(" hw status:\n");
2038 	db_printf("  status: 0x%lx\n", ioat_get_chansts(sc));
2039 	db_printf("  chanctrl: 0x%x\n",
2040 	    (unsigned)ioat_read_2(sc, IOAT_CHANCTRL_OFFSET));
2041 	db_printf("  chancmd: 0x%x\n",
2042 	    (unsigned)ioat_read_1(sc, IOAT_CHANCMD_OFFSET));
2043 	db_printf("  dmacount: 0x%x\n",
2044 	    (unsigned)ioat_read_2(sc, IOAT_DMACOUNT_OFFSET));
2045 	db_printf("  chainaddr: 0x%lx\n",
2046 	    ioat_read_double_4(sc, IOAT_CHAINADDR_OFFSET_LOW));
2047 	db_printf("  chancmp: 0x%lx\n",
2048 	    ioat_read_double_4(sc, IOAT_CHANCMP_OFFSET_LOW));
2049 	db_printf("  chanerr: %b\n",
2050 	    (int)ioat_read_4(sc, IOAT_CHANERR_OFFSET), IOAT_CHANERR_STR);
2051 	return;
2052 usage:
2053 	db_printf("usage: show ioat <0-%u>\n", ioat_channel_index);
2054 	return;
2055 }
2056 #endif /* DDB */
2057