xref: /freebsd/sys/arm64/arm64/gicv3_its.c (revision 1f474190)
1 /*-
2  * Copyright (c) 2015-2016 The FreeBSD Foundation
3  * All rights reserved.
4  *
5  * This software was developed by Andrew Turner under
6  * the sponsorship of the FreeBSD Foundation.
7  *
8  * This software was developed by Semihalf under
9  * the sponsorship of the FreeBSD Foundation.
10  *
11  * Redistribution and use in source and binary forms, with or without
12  * modification, are permitted provided that the following conditions
13  * are met:
14  * 1. Redistributions of source code must retain the above copyright
15  *    notice, this list of conditions and the following disclaimer.
16  * 2. Redistributions in binary form must reproduce the above copyright
17  *    notice, this list of conditions and the following disclaimer in the
18  *    documentation and/or other materials provided with the distribution.
19  *
20  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
21  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
22  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
23  * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
24  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
25  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
26  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
27  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
28  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
29  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
30  * SUCH DAMAGE.
31  */
32 
33 #include "opt_acpi.h"
34 #include "opt_platform.h"
35 #include "opt_iommu.h"
36 
37 #include <sys/cdefs.h>
38 __FBSDID("$FreeBSD$");
39 
40 #include <sys/param.h>
41 #include <sys/systm.h>
42 #include <sys/bus.h>
43 #include <sys/cpuset.h>
44 #include <sys/endian.h>
45 #include <sys/kernel.h>
46 #include <sys/lock.h>
47 #include <sys/malloc.h>
48 #include <sys/module.h>
49 #include <sys/mutex.h>
50 #include <sys/proc.h>
51 #include <sys/taskqueue.h>
52 #include <sys/tree.h>
53 #include <sys/queue.h>
54 #include <sys/rman.h>
55 #include <sys/sbuf.h>
56 #include <sys/smp.h>
57 #include <sys/sysctl.h>
58 #include <sys/vmem.h>
59 
60 #include <vm/vm.h>
61 #include <vm/pmap.h>
62 #include <vm/vm_page.h>
63 
64 #include <machine/bus.h>
65 #include <machine/intr.h>
66 
67 #include <arm/arm/gic_common.h>
68 #include <arm64/arm64/gic_v3_reg.h>
69 #include <arm64/arm64/gic_v3_var.h>
70 
71 #ifdef FDT
72 #include <dev/ofw/openfirm.h>
73 #include <dev/ofw/ofw_bus.h>
74 #include <dev/ofw/ofw_bus_subr.h>
75 #endif
76 #include <dev/pci/pcireg.h>
77 #include <dev/pci/pcivar.h>
78 
79 #ifdef IOMMU
80 #include <dev/iommu/iommu.h>
81 #include <dev/iommu/iommu_gas.h>
82 #endif
83 
84 #include "pcib_if.h"
85 #include "pic_if.h"
86 #include "msi_if.h"
87 
88 MALLOC_DEFINE(M_GICV3_ITS, "GICv3 ITS",
89     "ARM GICv3 Interrupt Translation Service");
90 
91 #define	LPI_NIRQS		(64 * 1024)
92 
93 /* The size and alignment of the command circular buffer */
94 #define	ITS_CMDQ_SIZE		(64 * 1024)	/* Must be a multiple of 4K */
95 #define	ITS_CMDQ_ALIGN		(64 * 1024)
96 
97 #define	LPI_CONFTAB_SIZE	LPI_NIRQS
98 #define	LPI_CONFTAB_ALIGN	(64 * 1024)
99 #define	LPI_CONFTAB_MAX_ADDR	((1ul << 48) - 1) /* We need a 47 bit PA */
100 
101 /* 1 bit per SPI, PPI, and SGI (8k), and 1 bit per LPI (LPI_CONFTAB_SIZE) */
102 #define	LPI_PENDTAB_SIZE	((LPI_NIRQS + GIC_FIRST_LPI) / 8)
103 #define	LPI_PENDTAB_ALIGN	(64 * 1024)
104 #define	LPI_PENDTAB_MAX_ADDR	((1ul << 48) - 1) /* We need a 47 bit PA */
105 
106 #define	LPI_INT_TRANS_TAB_ALIGN	256
107 #define	LPI_INT_TRANS_TAB_MAX_ADDR ((1ul << 48) - 1)
108 
109 /* ITS commands encoding */
110 #define	ITS_CMD_MOVI		(0x01)
111 #define	ITS_CMD_SYNC		(0x05)
112 #define	ITS_CMD_MAPD		(0x08)
113 #define	ITS_CMD_MAPC		(0x09)
114 #define	ITS_CMD_MAPTI		(0x0a)
115 #define	ITS_CMD_MAPI		(0x0b)
116 #define	ITS_CMD_INV		(0x0c)
117 #define	ITS_CMD_INVALL		(0x0d)
118 /* Command */
119 #define	CMD_COMMAND_MASK	(0xFFUL)
120 /* PCI device ID */
121 #define	CMD_DEVID_SHIFT		(32)
122 #define	CMD_DEVID_MASK		(0xFFFFFFFFUL << CMD_DEVID_SHIFT)
123 /* Size of IRQ ID bitfield */
124 #define	CMD_SIZE_MASK		(0xFFUL)
125 /* Virtual LPI ID */
126 #define	CMD_ID_MASK		(0xFFFFFFFFUL)
127 /* Physical LPI ID */
128 #define	CMD_PID_SHIFT		(32)
129 #define	CMD_PID_MASK		(0xFFFFFFFFUL << CMD_PID_SHIFT)
130 /* Collection */
131 #define	CMD_COL_MASK		(0xFFFFUL)
132 /* Target (CPU or Re-Distributor) */
133 #define	CMD_TARGET_SHIFT	(16)
134 #define	CMD_TARGET_MASK		(0xFFFFFFFFUL << CMD_TARGET_SHIFT)
135 /* Interrupt Translation Table address */
136 #define	CMD_ITT_MASK		(0xFFFFFFFFFF00UL)
137 /* Valid command bit */
138 #define	CMD_VALID_SHIFT		(63)
139 #define	CMD_VALID_MASK		(1UL << CMD_VALID_SHIFT)
140 
141 #define	ITS_TARGET_NONE		0xFBADBEEF
142 
143 /* LPI chunk owned by ITS device */
144 struct lpi_chunk {
145 	u_int	lpi_base;
146 	u_int	lpi_free;	/* First free LPI in set */
147 	u_int	lpi_num;	/* Total number of LPIs in chunk */
148 	u_int	lpi_busy;	/* Number of busy LPIs in chink */
149 };
150 
151 /* ITS device */
152 struct its_dev {
153 	TAILQ_ENTRY(its_dev)	entry;
154 	/* PCI device */
155 	device_t		pci_dev;
156 	/* Device ID (i.e. PCI device ID) */
157 	uint32_t		devid;
158 	/* List of assigned LPIs */
159 	struct lpi_chunk	lpis;
160 	/* Virtual address of ITT */
161 	vm_offset_t		itt;
162 	size_t			itt_size;
163 };
164 
165 /*
166  * ITS command descriptor.
167  * Idea for command description passing taken from Linux.
168  */
169 struct its_cmd_desc {
170 	uint8_t cmd_type;
171 
172 	union {
173 		struct {
174 			struct its_dev *its_dev;
175 			struct its_col *col;
176 			uint32_t id;
177 		} cmd_desc_movi;
178 
179 		struct {
180 			struct its_col *col;
181 		} cmd_desc_sync;
182 
183 		struct {
184 			struct its_col *col;
185 			uint8_t valid;
186 		} cmd_desc_mapc;
187 
188 		struct {
189 			struct its_dev *its_dev;
190 			struct its_col *col;
191 			uint32_t pid;
192 			uint32_t id;
193 		} cmd_desc_mapvi;
194 
195 		struct {
196 			struct its_dev *its_dev;
197 			struct its_col *col;
198 			uint32_t pid;
199 		} cmd_desc_mapi;
200 
201 		struct {
202 			struct its_dev *its_dev;
203 			uint8_t valid;
204 		} cmd_desc_mapd;
205 
206 		struct {
207 			struct its_dev *its_dev;
208 			struct its_col *col;
209 			uint32_t pid;
210 		} cmd_desc_inv;
211 
212 		struct {
213 			struct its_col *col;
214 		} cmd_desc_invall;
215 	};
216 };
217 
218 /* ITS command. Each command is 32 bytes long */
219 struct its_cmd {
220 	uint64_t	cmd_dword[4];	/* ITS command double word */
221 };
222 
223 /* An ITS private table */
224 struct its_ptable {
225 	vm_offset_t	ptab_vaddr;
226 	unsigned long	ptab_size;
227 };
228 
229 /* ITS collection description. */
230 struct its_col {
231 	uint64_t	col_target;	/* Target Re-Distributor */
232 	uint64_t	col_id;		/* Collection ID */
233 };
234 
235 struct gicv3_its_irqsrc {
236 	struct intr_irqsrc	gi_isrc;
237 	u_int			gi_id;
238 	u_int			gi_lpi;
239 	struct its_dev		*gi_its_dev;
240 	TAILQ_ENTRY(gicv3_its_irqsrc) gi_link;
241 };
242 
243 struct gicv3_its_softc {
244 	device_t	dev;
245 	struct intr_pic *sc_pic;
246 	struct resource *sc_its_res;
247 
248 	cpuset_t	sc_cpus;
249 	u_int		gic_irq_cpu;
250 
251 	struct its_ptable sc_its_ptab[GITS_BASER_NUM];
252 	struct its_col *sc_its_cols[MAXCPU];	/* Per-CPU collections */
253 
254 	/*
255 	 * TODO: We should get these from the parent as we only want a
256 	 * single copy of each across the interrupt controller.
257 	 */
258 	uint8_t		*sc_conf_base;
259 	vm_offset_t sc_pend_base[MAXCPU];
260 
261 	/* Command handling */
262 	struct mtx sc_its_cmd_lock;
263 	struct its_cmd *sc_its_cmd_base; /* Command circular buffer address */
264 	size_t sc_its_cmd_next_idx;
265 
266 	vmem_t *sc_irq_alloc;
267 	struct gicv3_its_irqsrc	**sc_irqs;
268 	u_int	sc_irq_base;
269 	u_int	sc_irq_length;
270 	u_int	sc_irq_count;
271 
272 	struct mtx sc_its_dev_lock;
273 	TAILQ_HEAD(its_dev_list, its_dev) sc_its_dev_list;
274 	TAILQ_HEAD(free_irqs, gicv3_its_irqsrc) sc_free_irqs;
275 
276 #define	ITS_FLAGS_CMDQ_FLUSH		0x00000001
277 #define	ITS_FLAGS_LPI_CONF_FLUSH	0x00000002
278 #define	ITS_FLAGS_ERRATA_CAVIUM_22375	0x00000004
279 	u_int sc_its_flags;
280 	bool	trace_enable;
281 	vm_page_t ma; /* fake msi page */
282 };
283 
284 static void *conf_base;
285 
286 typedef void (its_quirk_func_t)(device_t);
287 static its_quirk_func_t its_quirk_cavium_22375;
288 
289 static const struct {
290 	const char *desc;
291 	uint32_t iidr;
292 	uint32_t iidr_mask;
293 	its_quirk_func_t *func;
294 } its_quirks[] = {
295 	{
296 		/* Cavium ThunderX Pass 1.x */
297 		.desc = "Cavium ThunderX errata: 22375, 24313",
298 		.iidr = GITS_IIDR_RAW(GITS_IIDR_IMPL_CAVIUM,
299 		    GITS_IIDR_PROD_THUNDER, GITS_IIDR_VAR_THUNDER_1, 0),
300 		.iidr_mask = ~GITS_IIDR_REVISION_MASK,
301 		.func = its_quirk_cavium_22375,
302 	},
303 };
304 
305 #define	gic_its_read_4(sc, reg)			\
306     bus_read_4((sc)->sc_its_res, (reg))
307 #define	gic_its_read_8(sc, reg)			\
308     bus_read_8((sc)->sc_its_res, (reg))
309 
310 #define	gic_its_write_4(sc, reg, val)		\
311     bus_write_4((sc)->sc_its_res, (reg), (val))
312 #define	gic_its_write_8(sc, reg, val)		\
313     bus_write_8((sc)->sc_its_res, (reg), (val))
314 
315 static device_attach_t gicv3_its_attach;
316 static device_detach_t gicv3_its_detach;
317 
318 static pic_disable_intr_t gicv3_its_disable_intr;
319 static pic_enable_intr_t gicv3_its_enable_intr;
320 static pic_map_intr_t gicv3_its_map_intr;
321 static pic_setup_intr_t gicv3_its_setup_intr;
322 static pic_post_filter_t gicv3_its_post_filter;
323 static pic_post_ithread_t gicv3_its_post_ithread;
324 static pic_pre_ithread_t gicv3_its_pre_ithread;
325 static pic_bind_intr_t gicv3_its_bind_intr;
326 #ifdef SMP
327 static pic_init_secondary_t gicv3_its_init_secondary;
328 #endif
329 static msi_alloc_msi_t gicv3_its_alloc_msi;
330 static msi_release_msi_t gicv3_its_release_msi;
331 static msi_alloc_msix_t gicv3_its_alloc_msix;
332 static msi_release_msix_t gicv3_its_release_msix;
333 static msi_map_msi_t gicv3_its_map_msi;
334 #ifdef IOMMU
335 static msi_iommu_init_t gicv3_iommu_init;
336 static msi_iommu_deinit_t gicv3_iommu_deinit;
337 #endif
338 
339 static void its_cmd_movi(device_t, struct gicv3_its_irqsrc *);
340 static void its_cmd_mapc(device_t, struct its_col *, uint8_t);
341 static void its_cmd_mapti(device_t, struct gicv3_its_irqsrc *);
342 static void its_cmd_mapd(device_t, struct its_dev *, uint8_t);
343 static void its_cmd_inv(device_t, struct its_dev *, struct gicv3_its_irqsrc *);
344 static void its_cmd_invall(device_t, struct its_col *);
345 
346 static device_method_t gicv3_its_methods[] = {
347 	/* Device interface */
348 	DEVMETHOD(device_detach,	gicv3_its_detach),
349 
350 	/* Interrupt controller interface */
351 	DEVMETHOD(pic_disable_intr,	gicv3_its_disable_intr),
352 	DEVMETHOD(pic_enable_intr,	gicv3_its_enable_intr),
353 	DEVMETHOD(pic_map_intr,		gicv3_its_map_intr),
354 	DEVMETHOD(pic_setup_intr,	gicv3_its_setup_intr),
355 	DEVMETHOD(pic_post_filter,	gicv3_its_post_filter),
356 	DEVMETHOD(pic_post_ithread,	gicv3_its_post_ithread),
357 	DEVMETHOD(pic_pre_ithread,	gicv3_its_pre_ithread),
358 #ifdef SMP
359 	DEVMETHOD(pic_bind_intr,	gicv3_its_bind_intr),
360 	DEVMETHOD(pic_init_secondary,	gicv3_its_init_secondary),
361 #endif
362 
363 	/* MSI/MSI-X */
364 	DEVMETHOD(msi_alloc_msi,	gicv3_its_alloc_msi),
365 	DEVMETHOD(msi_release_msi,	gicv3_its_release_msi),
366 	DEVMETHOD(msi_alloc_msix,	gicv3_its_alloc_msix),
367 	DEVMETHOD(msi_release_msix,	gicv3_its_release_msix),
368 	DEVMETHOD(msi_map_msi,		gicv3_its_map_msi),
369 #ifdef IOMMU
370 	DEVMETHOD(msi_iommu_init,	gicv3_iommu_init),
371 	DEVMETHOD(msi_iommu_deinit,	gicv3_iommu_deinit),
372 #endif
373 
374 	/* End */
375 	DEVMETHOD_END
376 };
377 
378 static DEFINE_CLASS_0(gic, gicv3_its_driver, gicv3_its_methods,
379     sizeof(struct gicv3_its_softc));
380 
381 static void
382 gicv3_its_cmdq_init(struct gicv3_its_softc *sc)
383 {
384 	vm_paddr_t cmd_paddr;
385 	uint64_t reg, tmp;
386 
387 	/* Set up the command circular buffer */
388 	sc->sc_its_cmd_base = contigmalloc(ITS_CMDQ_SIZE, M_GICV3_ITS,
389 	    M_WAITOK | M_ZERO, 0, (1ul << 48) - 1, ITS_CMDQ_ALIGN, 0);
390 	sc->sc_its_cmd_next_idx = 0;
391 
392 	cmd_paddr = vtophys(sc->sc_its_cmd_base);
393 
394 	/* Set the base of the command buffer */
395 	reg = GITS_CBASER_VALID |
396 	    (GITS_CBASER_CACHE_NIWAWB << GITS_CBASER_CACHE_SHIFT) |
397 	    cmd_paddr | (GITS_CBASER_SHARE_IS << GITS_CBASER_SHARE_SHIFT) |
398 	    (ITS_CMDQ_SIZE / 4096 - 1);
399 	gic_its_write_8(sc, GITS_CBASER, reg);
400 
401 	/* Read back to check for fixed value fields */
402 	tmp = gic_its_read_8(sc, GITS_CBASER);
403 
404 	if ((tmp & GITS_CBASER_SHARE_MASK) !=
405 	    (GITS_CBASER_SHARE_IS << GITS_CBASER_SHARE_SHIFT)) {
406 		/* Check if the hardware reported non-shareable */
407 		if ((tmp & GITS_CBASER_SHARE_MASK) ==
408 		    (GITS_CBASER_SHARE_NS << GITS_CBASER_SHARE_SHIFT)) {
409 			/* If so remove the cache attribute */
410 			reg &= ~GITS_CBASER_CACHE_MASK;
411 			reg &= ~GITS_CBASER_SHARE_MASK;
412 			/* Set to Non-cacheable, Non-shareable */
413 			reg |= GITS_CBASER_CACHE_NIN << GITS_CBASER_CACHE_SHIFT;
414 			reg |= GITS_CBASER_SHARE_NS << GITS_CBASER_SHARE_SHIFT;
415 
416 			gic_its_write_8(sc, GITS_CBASER, reg);
417 		}
418 
419 		/* The command queue has to be flushed after each command */
420 		sc->sc_its_flags |= ITS_FLAGS_CMDQ_FLUSH;
421 	}
422 
423 	/* Get the next command from the start of the buffer */
424 	gic_its_write_8(sc, GITS_CWRITER, 0x0);
425 }
426 
427 static int
428 gicv3_its_table_init(device_t dev, struct gicv3_its_softc *sc)
429 {
430 	vm_offset_t table;
431 	vm_paddr_t paddr;
432 	uint64_t cache, reg, share, tmp, type;
433 	size_t esize, its_tbl_size, nidents, nitspages, npages;
434 	int i, page_size;
435 	int devbits;
436 
437 	if ((sc->sc_its_flags & ITS_FLAGS_ERRATA_CAVIUM_22375) != 0) {
438 		/*
439 		 * GITS_TYPER[17:13] of ThunderX reports that device IDs
440 		 * are to be 21 bits in length. The entry size of the ITS
441 		 * table can be read from GITS_BASERn[52:48] and on ThunderX
442 		 * is supposed to be 8 bytes in length (for device table).
443 		 * Finally the page size that is to be used by ITS to access
444 		 * this table will be set to 64KB.
445 		 *
446 		 * This gives 0x200000 entries of size 0x8 bytes covered by
447 		 * 256 pages each of which 64KB in size. The number of pages
448 		 * (minus 1) should then be written to GITS_BASERn[7:0]. In
449 		 * that case this value would be 0xFF but on ThunderX the
450 		 * maximum value that HW accepts is 0xFD.
451 		 *
452 		 * Set an arbitrary number of device ID bits to 20 in order
453 		 * to limit the number of entries in ITS device table to
454 		 * 0x100000 and the table size to 8MB.
455 		 */
456 		devbits = 20;
457 		cache = 0;
458 	} else {
459 		devbits = GITS_TYPER_DEVB(gic_its_read_8(sc, GITS_TYPER));
460 		cache = GITS_BASER_CACHE_WAWB;
461 	}
462 	share = GITS_BASER_SHARE_IS;
463 	page_size = PAGE_SIZE_64K;
464 
465 	for (i = 0; i < GITS_BASER_NUM; i++) {
466 		reg = gic_its_read_8(sc, GITS_BASER(i));
467 		/* The type of table */
468 		type = GITS_BASER_TYPE(reg);
469 		/* The table entry size */
470 		esize = GITS_BASER_ESIZE(reg);
471 
472 		switch(type) {
473 		case GITS_BASER_TYPE_DEV:
474 			nidents = (1 << devbits);
475 			its_tbl_size = esize * nidents;
476 			its_tbl_size = roundup2(its_tbl_size, PAGE_SIZE_64K);
477 			break;
478 		case GITS_BASER_TYPE_VP:
479 		case GITS_BASER_TYPE_PP: /* Undocumented? */
480 		case GITS_BASER_TYPE_IC:
481 			its_tbl_size = page_size;
482 			break;
483 		default:
484 			continue;
485 		}
486 		npages = howmany(its_tbl_size, PAGE_SIZE);
487 
488 		/* Allocate the table */
489 		table = (vm_offset_t)contigmalloc(npages * PAGE_SIZE,
490 		    M_GICV3_ITS, M_WAITOK | M_ZERO, 0, (1ul << 48) - 1,
491 		    PAGE_SIZE_64K, 0);
492 
493 		sc->sc_its_ptab[i].ptab_vaddr = table;
494 		sc->sc_its_ptab[i].ptab_size = npages * PAGE_SIZE;
495 
496 		paddr = vtophys(table);
497 
498 		while (1) {
499 			nitspages = howmany(its_tbl_size, page_size);
500 
501 			/* Clear the fields we will be setting */
502 			reg &= ~(GITS_BASER_VALID |
503 			    GITS_BASER_CACHE_MASK | GITS_BASER_TYPE_MASK |
504 			    GITS_BASER_ESIZE_MASK | GITS_BASER_PA_MASK |
505 			    GITS_BASER_SHARE_MASK | GITS_BASER_PSZ_MASK |
506 			    GITS_BASER_SIZE_MASK);
507 			/* Set the new values */
508 			reg |= GITS_BASER_VALID |
509 			    (cache << GITS_BASER_CACHE_SHIFT) |
510 			    (type << GITS_BASER_TYPE_SHIFT) |
511 			    ((esize - 1) << GITS_BASER_ESIZE_SHIFT) |
512 			    paddr | (share << GITS_BASER_SHARE_SHIFT) |
513 			    (nitspages - 1);
514 
515 			switch (page_size) {
516 			case PAGE_SIZE:		/* 4KB */
517 				reg |=
518 				    GITS_BASER_PSZ_4K << GITS_BASER_PSZ_SHIFT;
519 				break;
520 			case PAGE_SIZE_16K:	/* 16KB */
521 				reg |=
522 				    GITS_BASER_PSZ_16K << GITS_BASER_PSZ_SHIFT;
523 				break;
524 			case PAGE_SIZE_64K:	/* 64KB */
525 				reg |=
526 				    GITS_BASER_PSZ_64K << GITS_BASER_PSZ_SHIFT;
527 				break;
528 			}
529 
530 			gic_its_write_8(sc, GITS_BASER(i), reg);
531 
532 			/* Read back to check */
533 			tmp = gic_its_read_8(sc, GITS_BASER(i));
534 
535 			/* Do the shareability masks line up? */
536 			if ((tmp & GITS_BASER_SHARE_MASK) !=
537 			    (reg & GITS_BASER_SHARE_MASK)) {
538 				share = (tmp & GITS_BASER_SHARE_MASK) >>
539 				    GITS_BASER_SHARE_SHIFT;
540 				continue;
541 			}
542 
543 			if ((tmp & GITS_BASER_PSZ_MASK) !=
544 			    (reg & GITS_BASER_PSZ_MASK)) {
545 				switch (page_size) {
546 				case PAGE_SIZE_16K:
547 					page_size = PAGE_SIZE;
548 					continue;
549 				case PAGE_SIZE_64K:
550 					page_size = PAGE_SIZE_16K;
551 					continue;
552 				}
553 			}
554 
555 			if (tmp != reg) {
556 				device_printf(dev, "GITS_BASER%d: "
557 				    "unable to be updated: %lx != %lx\n",
558 				    i, reg, tmp);
559 				return (ENXIO);
560 			}
561 
562 			/* We should have made all needed changes */
563 			break;
564 		}
565 	}
566 
567 	return (0);
568 }
569 
570 static void
571 gicv3_its_conftable_init(struct gicv3_its_softc *sc)
572 {
573 	void *conf_table;
574 
575 	conf_table = atomic_load_ptr(&conf_base);
576 	if (conf_table == NULL) {
577 		conf_table = contigmalloc(LPI_CONFTAB_SIZE,
578 		    M_GICV3_ITS, M_WAITOK, 0, LPI_CONFTAB_MAX_ADDR,
579 		    LPI_CONFTAB_ALIGN, 0);
580 
581 		if (atomic_cmpset_ptr((uintptr_t *)&conf_base,
582 		    (uintptr_t)NULL, (uintptr_t)conf_table) == 0) {
583 			contigfree(conf_table, LPI_CONFTAB_SIZE, M_GICV3_ITS);
584 			conf_table = atomic_load_ptr(&conf_base);
585 		}
586 	}
587 	sc->sc_conf_base = conf_table;
588 
589 	/* Set the default configuration */
590 	memset(sc->sc_conf_base, GIC_PRIORITY_MAX | LPI_CONF_GROUP1,
591 	    LPI_CONFTAB_SIZE);
592 
593 	/* Flush the table to memory */
594 	cpu_dcache_wb_range((vm_offset_t)sc->sc_conf_base, LPI_CONFTAB_SIZE);
595 }
596 
597 static void
598 gicv3_its_pendtables_init(struct gicv3_its_softc *sc)
599 {
600 	int i;
601 
602 	for (i = 0; i <= mp_maxid; i++) {
603 		if (CPU_ISSET(i, &sc->sc_cpus) == 0)
604 			continue;
605 
606 		sc->sc_pend_base[i] = (vm_offset_t)contigmalloc(
607 		    LPI_PENDTAB_SIZE, M_GICV3_ITS, M_WAITOK | M_ZERO,
608 		    0, LPI_PENDTAB_MAX_ADDR, LPI_PENDTAB_ALIGN, 0);
609 
610 		/* Flush so the ITS can see the memory */
611 		cpu_dcache_wb_range((vm_offset_t)sc->sc_pend_base[i],
612 		    LPI_PENDTAB_SIZE);
613 	}
614 }
615 
616 static void
617 its_init_cpu_lpi(device_t dev, struct gicv3_its_softc *sc)
618 {
619 	device_t gicv3;
620 	uint64_t xbaser, tmp;
621 	uint32_t ctlr;
622 	u_int cpuid;
623 
624 	gicv3 = device_get_parent(dev);
625 	cpuid = PCPU_GET(cpuid);
626 
627 	/* Disable LPIs */
628 	ctlr = gic_r_read_4(gicv3, GICR_CTLR);
629 	ctlr &= ~GICR_CTLR_LPI_ENABLE;
630 	gic_r_write_4(gicv3, GICR_CTLR, ctlr);
631 
632 	/* Make sure changes are observable my the GIC */
633 	dsb(sy);
634 
635 	/*
636 	 * Set the redistributor base
637 	 */
638 	xbaser = vtophys(sc->sc_conf_base) |
639 	    (GICR_PROPBASER_SHARE_IS << GICR_PROPBASER_SHARE_SHIFT) |
640 	    (GICR_PROPBASER_CACHE_NIWAWB << GICR_PROPBASER_CACHE_SHIFT) |
641 	    (flsl(LPI_CONFTAB_SIZE | GIC_FIRST_LPI) - 1);
642 	gic_r_write_8(gicv3, GICR_PROPBASER, xbaser);
643 
644 	/* Check the cache attributes we set */
645 	tmp = gic_r_read_8(gicv3, GICR_PROPBASER);
646 
647 	if ((tmp & GICR_PROPBASER_SHARE_MASK) !=
648 	    (xbaser & GICR_PROPBASER_SHARE_MASK)) {
649 		if ((tmp & GICR_PROPBASER_SHARE_MASK) ==
650 		    (GICR_PROPBASER_SHARE_NS << GICR_PROPBASER_SHARE_SHIFT)) {
651 			/* We need to mark as non-cacheable */
652 			xbaser &= ~(GICR_PROPBASER_SHARE_MASK |
653 			    GICR_PROPBASER_CACHE_MASK);
654 			/* Non-cacheable */
655 			xbaser |= GICR_PROPBASER_CACHE_NIN <<
656 			    GICR_PROPBASER_CACHE_SHIFT;
657 			/* Non-sareable */
658 			xbaser |= GICR_PROPBASER_SHARE_NS <<
659 			    GICR_PROPBASER_SHARE_SHIFT;
660 			gic_r_write_8(gicv3, GICR_PROPBASER, xbaser);
661 		}
662 		sc->sc_its_flags |= ITS_FLAGS_LPI_CONF_FLUSH;
663 	}
664 
665 	/*
666 	 * Set the LPI pending table base
667 	 */
668 	xbaser = vtophys(sc->sc_pend_base[cpuid]) |
669 	    (GICR_PENDBASER_CACHE_NIWAWB << GICR_PENDBASER_CACHE_SHIFT) |
670 	    (GICR_PENDBASER_SHARE_IS << GICR_PENDBASER_SHARE_SHIFT);
671 
672 	gic_r_write_8(gicv3, GICR_PENDBASER, xbaser);
673 
674 	tmp = gic_r_read_8(gicv3, GICR_PENDBASER);
675 
676 	if ((tmp & GICR_PENDBASER_SHARE_MASK) ==
677 	    (GICR_PENDBASER_SHARE_NS << GICR_PENDBASER_SHARE_SHIFT)) {
678 		/* Clear the cahce and shareability bits */
679 		xbaser &= ~(GICR_PENDBASER_CACHE_MASK |
680 		    GICR_PENDBASER_SHARE_MASK);
681 		/* Mark as non-shareable */
682 		xbaser |= GICR_PENDBASER_SHARE_NS << GICR_PENDBASER_SHARE_SHIFT;
683 		/* And non-cacheable */
684 		xbaser |= GICR_PENDBASER_CACHE_NIN <<
685 		    GICR_PENDBASER_CACHE_SHIFT;
686 	}
687 
688 	/* Enable LPIs */
689 	ctlr = gic_r_read_4(gicv3, GICR_CTLR);
690 	ctlr |= GICR_CTLR_LPI_ENABLE;
691 	gic_r_write_4(gicv3, GICR_CTLR, ctlr);
692 
693 	/* Make sure the GIC has seen everything */
694 	dsb(sy);
695 }
696 
697 static int
698 its_init_cpu(device_t dev, struct gicv3_its_softc *sc)
699 {
700 	device_t gicv3;
701 	vm_paddr_t target;
702 	u_int cpuid;
703 	struct redist_pcpu *rpcpu;
704 
705 	gicv3 = device_get_parent(dev);
706 	cpuid = PCPU_GET(cpuid);
707 	if (!CPU_ISSET(cpuid, &sc->sc_cpus))
708 		return (0);
709 
710 	/* Check if the ITS is enabled on this CPU */
711 	if ((gic_r_read_4(gicv3, GICR_TYPER) & GICR_TYPER_PLPIS) == 0)
712 		return (ENXIO);
713 
714 	rpcpu = gicv3_get_redist(dev);
715 
716 	/* Do per-cpu LPI init once */
717 	if (!rpcpu->lpi_enabled) {
718 		its_init_cpu_lpi(dev, sc);
719 		rpcpu->lpi_enabled = true;
720 	}
721 
722 	if ((gic_its_read_8(sc, GITS_TYPER) & GITS_TYPER_PTA) != 0) {
723 		/* This ITS wants the redistributor physical address */
724 		target = vtophys(rman_get_virtual(&rpcpu->res));
725 	} else {
726 		/* This ITS wants the unique processor number */
727 		target = GICR_TYPER_CPUNUM(gic_r_read_8(gicv3, GICR_TYPER)) <<
728 		    CMD_TARGET_SHIFT;
729 	}
730 
731 	sc->sc_its_cols[cpuid]->col_target = target;
732 	sc->sc_its_cols[cpuid]->col_id = cpuid;
733 
734 	its_cmd_mapc(dev, sc->sc_its_cols[cpuid], 1);
735 	its_cmd_invall(dev, sc->sc_its_cols[cpuid]);
736 
737 	return (0);
738 }
739 
740 static int
741 gicv3_its_sysctl_trace_enable(SYSCTL_HANDLER_ARGS)
742 {
743 	struct gicv3_its_softc *sc;
744 	int rv;
745 
746 	sc = arg1;
747 
748 	rv = sysctl_handle_bool(oidp, &sc->trace_enable, 0, req);
749 	if (rv != 0 || req->newptr == NULL)
750 		return (rv);
751 	if (sc->trace_enable)
752 		gic_its_write_8(sc, GITS_TRKCTLR, 3);
753 	else
754 		gic_its_write_8(sc, GITS_TRKCTLR, 0);
755 
756 	return (0);
757 }
758 
759 static int
760 gicv3_its_sysctl_trace_regs(SYSCTL_HANDLER_ARGS)
761 {
762 	struct gicv3_its_softc *sc;
763 	struct sbuf *sb;
764 	int err;
765 
766 	sc = arg1;
767 	sb = sbuf_new_for_sysctl(NULL, NULL, 128, req);
768 	if (sb == NULL) {
769 		device_printf(sc->dev, "Could not allocate sbuf for output.\n");
770 		return (ENOMEM);
771 	}
772 	sbuf_cat(sb, "\n");
773 	sbuf_printf(sb, "GITS_TRKCTLR: 0x%08X\n",
774 	    gic_its_read_4(sc, GITS_TRKCTLR));
775 	sbuf_printf(sb, "GITS_TRKR:    0x%08X\n",
776 	    gic_its_read_4(sc, GITS_TRKR));
777 	sbuf_printf(sb, "GITS_TRKDIDR: 0x%08X\n",
778 	    gic_its_read_4(sc, GITS_TRKDIDR));
779 	sbuf_printf(sb, "GITS_TRKPIDR: 0x%08X\n",
780 	    gic_its_read_4(sc, GITS_TRKPIDR));
781 	sbuf_printf(sb, "GITS_TRKVIDR: 0x%08X\n",
782 	    gic_its_read_4(sc, GITS_TRKVIDR));
783 	sbuf_printf(sb, "GITS_TRKTGTR: 0x%08X\n",
784 	   gic_its_read_4(sc, GITS_TRKTGTR));
785 
786 	err = sbuf_finish(sb);
787 	if (err)
788 		device_printf(sc->dev, "Error finishing sbuf: %d\n", err);
789 	sbuf_delete(sb);
790 	return(err);
791 }
792 
793 static int
794 gicv3_its_init_sysctl(struct gicv3_its_softc *sc)
795 {
796 	struct sysctl_oid *oid, *child;
797 	struct sysctl_ctx_list *ctx_list;
798 
799 	ctx_list = device_get_sysctl_ctx(sc->dev);
800 	child = device_get_sysctl_tree(sc->dev);
801 	oid = SYSCTL_ADD_NODE(ctx_list,
802 	    SYSCTL_CHILDREN(child), OID_AUTO, "tracing",
803 	    CTLFLAG_RD| CTLFLAG_MPSAFE, NULL, "Messages tracing");
804 	if (oid == NULL)
805 		return (ENXIO);
806 
807 	/* Add registers */
808 	SYSCTL_ADD_PROC(ctx_list,
809 	    SYSCTL_CHILDREN(oid), OID_AUTO, "enable",
810 	    CTLTYPE_U8 | CTLFLAG_RW | CTLFLAG_MPSAFE, sc, 0,
811 	    gicv3_its_sysctl_trace_enable, "CU", "Enable tracing");
812 	SYSCTL_ADD_PROC(ctx_list,
813 	    SYSCTL_CHILDREN(oid), OID_AUTO, "capture",
814 	    CTLTYPE_STRING | CTLFLAG_RW | CTLFLAG_MPSAFE, sc, 0,
815 	    gicv3_its_sysctl_trace_regs, "", "Captured tracing registers.");
816 
817 	return (0);
818 }
819 
820 static int
821 gicv3_its_attach(device_t dev)
822 {
823 	struct gicv3_its_softc *sc;
824 	int domain, err, i, rid;
825 	uint64_t phys;
826 	uint32_t iidr;
827 
828 	sc = device_get_softc(dev);
829 
830 	sc->sc_irq_length = gicv3_get_nirqs(dev);
831 	sc->sc_irq_base = GIC_FIRST_LPI;
832 	sc->sc_irq_base += device_get_unit(dev) * sc->sc_irq_length;
833 
834 	rid = 0;
835 	sc->sc_its_res = bus_alloc_resource_any(dev, SYS_RES_MEMORY, &rid,
836 	    RF_ACTIVE);
837 	if (sc->sc_its_res == NULL) {
838 		device_printf(dev, "Could not allocate memory\n");
839 		return (ENXIO);
840 	}
841 
842 	phys = rounddown2(vtophys(rman_get_virtual(sc->sc_its_res)) +
843 	    GITS_TRANSLATER, PAGE_SIZE);
844 	sc->ma = malloc(sizeof(struct vm_page), M_DEVBUF, M_WAITOK | M_ZERO);
845 	vm_page_initfake(sc->ma, phys, VM_MEMATTR_DEFAULT);
846 
847 	iidr = gic_its_read_4(sc, GITS_IIDR);
848 	for (i = 0; i < nitems(its_quirks); i++) {
849 		if ((iidr & its_quirks[i].iidr_mask) == its_quirks[i].iidr) {
850 			if (bootverbose) {
851 				device_printf(dev, "Applying %s\n",
852 				    its_quirks[i].desc);
853 			}
854 			its_quirks[i].func(dev);
855 			break;
856 		}
857 	}
858 
859 	/* Allocate the private tables */
860 	err = gicv3_its_table_init(dev, sc);
861 	if (err != 0)
862 		return (err);
863 
864 	/* Protects access to the device list */
865 	mtx_init(&sc->sc_its_dev_lock, "ITS device lock", NULL, MTX_SPIN);
866 
867 	/* Protects access to the ITS command circular buffer. */
868 	mtx_init(&sc->sc_its_cmd_lock, "ITS cmd lock", NULL, MTX_SPIN);
869 
870 	CPU_ZERO(&sc->sc_cpus);
871 	if (bus_get_domain(dev, &domain) == 0) {
872 		if (domain < MAXMEMDOM)
873 			CPU_COPY(&cpuset_domain[domain], &sc->sc_cpus);
874 	} else {
875 		CPU_COPY(&all_cpus, &sc->sc_cpus);
876 	}
877 
878 	/* Allocate the command circular buffer */
879 	gicv3_its_cmdq_init(sc);
880 
881 	/* Allocate the per-CPU collections */
882 	for (int cpu = 0; cpu <= mp_maxid; cpu++)
883 		if (CPU_ISSET(cpu, &sc->sc_cpus) != 0)
884 			sc->sc_its_cols[cpu] = malloc(
885 			    sizeof(*sc->sc_its_cols[0]), M_GICV3_ITS,
886 			    M_WAITOK | M_ZERO);
887 		else
888 			sc->sc_its_cols[cpu] = NULL;
889 
890 	/* Enable the ITS */
891 	gic_its_write_4(sc, GITS_CTLR,
892 	    gic_its_read_4(sc, GITS_CTLR) | GITS_CTLR_EN);
893 
894 	/* Create the LPI configuration table */
895 	gicv3_its_conftable_init(sc);
896 
897 	/* And the pending tebles */
898 	gicv3_its_pendtables_init(sc);
899 
900 	/* Enable LPIs on this CPU */
901 	its_init_cpu(dev, sc);
902 
903 	TAILQ_INIT(&sc->sc_its_dev_list);
904 	TAILQ_INIT(&sc->sc_free_irqs);
905 
906 	/*
907 	 * Create the vmem object to allocate INTRNG IRQs from. We try to
908 	 * use all IRQs not already used by the GICv3.
909 	 * XXX: This assumes there are no other interrupt controllers in the
910 	 * system.
911 	 */
912 	sc->sc_irq_alloc = vmem_create(device_get_nameunit(dev), 0,
913 	    gicv3_get_nirqs(dev), 1, 0, M_FIRSTFIT | M_WAITOK);
914 
915 	sc->sc_irqs = malloc(sizeof(*sc->sc_irqs) * sc->sc_irq_length,
916 	    M_GICV3_ITS, M_WAITOK | M_ZERO);
917 
918 	/* For GIC-500 install tracking sysctls. */
919 	if ((iidr & (GITS_IIDR_PRODUCT_MASK | GITS_IIDR_IMPLEMENTOR_MASK)) ==
920 	    GITS_IIDR_RAW(GITS_IIDR_IMPL_ARM, GITS_IIDR_PROD_GIC500, 0, 0))
921 		gicv3_its_init_sysctl(sc);
922 
923 	return (0);
924 }
925 
926 static int
927 gicv3_its_detach(device_t dev)
928 {
929 
930 	return (ENXIO);
931 }
932 
933 static void
934 its_quirk_cavium_22375(device_t dev)
935 {
936 	struct gicv3_its_softc *sc;
937 
938 	sc = device_get_softc(dev);
939 	sc->sc_its_flags |= ITS_FLAGS_ERRATA_CAVIUM_22375;
940 }
941 
942 static void
943 gicv3_its_disable_intr(device_t dev, struct intr_irqsrc *isrc)
944 {
945 	struct gicv3_its_softc *sc;
946 	struct gicv3_its_irqsrc *girq;
947 	uint8_t *conf;
948 
949 	sc = device_get_softc(dev);
950 	girq = (struct gicv3_its_irqsrc *)isrc;
951 	conf = sc->sc_conf_base;
952 
953 	conf[girq->gi_lpi] &= ~LPI_CONF_ENABLE;
954 
955 	if ((sc->sc_its_flags & ITS_FLAGS_LPI_CONF_FLUSH) != 0) {
956 		/* Clean D-cache under command. */
957 		cpu_dcache_wb_range((vm_offset_t)&conf[girq->gi_lpi], 1);
958 	} else {
959 		/* DSB inner shareable, store */
960 		dsb(ishst);
961 	}
962 
963 	its_cmd_inv(dev, girq->gi_its_dev, girq);
964 }
965 
966 static void
967 gicv3_its_enable_intr(device_t dev, struct intr_irqsrc *isrc)
968 {
969 	struct gicv3_its_softc *sc;
970 	struct gicv3_its_irqsrc *girq;
971 	uint8_t *conf;
972 
973 	sc = device_get_softc(dev);
974 	girq = (struct gicv3_its_irqsrc *)isrc;
975 	conf = sc->sc_conf_base;
976 
977 	conf[girq->gi_lpi] |= LPI_CONF_ENABLE;
978 
979 	if ((sc->sc_its_flags & ITS_FLAGS_LPI_CONF_FLUSH) != 0) {
980 		/* Clean D-cache under command. */
981 		cpu_dcache_wb_range((vm_offset_t)&conf[girq->gi_lpi], 1);
982 	} else {
983 		/* DSB inner shareable, store */
984 		dsb(ishst);
985 	}
986 
987 	its_cmd_inv(dev, girq->gi_its_dev, girq);
988 }
989 
990 static int
991 gicv3_its_intr(void *arg, uintptr_t irq)
992 {
993 	struct gicv3_its_softc *sc = arg;
994 	struct gicv3_its_irqsrc *girq;
995 	struct trapframe *tf;
996 
997 	irq -= sc->sc_irq_base;
998 	girq = sc->sc_irqs[irq];
999 	if (girq == NULL)
1000 		panic("gicv3_its_intr: Invalid interrupt %ld",
1001 		    irq + sc->sc_irq_base);
1002 
1003 	tf = curthread->td_intr_frame;
1004 	intr_isrc_dispatch(&girq->gi_isrc, tf);
1005 	return (FILTER_HANDLED);
1006 }
1007 
1008 static void
1009 gicv3_its_pre_ithread(device_t dev, struct intr_irqsrc *isrc)
1010 {
1011 	struct gicv3_its_irqsrc *girq;
1012 	struct gicv3_its_softc *sc;
1013 
1014 	sc = device_get_softc(dev);
1015 	girq = (struct gicv3_its_irqsrc *)isrc;
1016 	gicv3_its_disable_intr(dev, isrc);
1017 	gic_icc_write(EOIR1, girq->gi_lpi + GIC_FIRST_LPI);
1018 }
1019 
1020 static void
1021 gicv3_its_post_ithread(device_t dev, struct intr_irqsrc *isrc)
1022 {
1023 
1024 	gicv3_its_enable_intr(dev, isrc);
1025 }
1026 
1027 static void
1028 gicv3_its_post_filter(device_t dev, struct intr_irqsrc *isrc)
1029 {
1030 	struct gicv3_its_irqsrc *girq;
1031 	struct gicv3_its_softc *sc;
1032 
1033 	sc = device_get_softc(dev);
1034 	girq = (struct gicv3_its_irqsrc *)isrc;
1035 	gic_icc_write(EOIR1, girq->gi_lpi + GIC_FIRST_LPI);
1036 }
1037 
1038 static int
1039 gicv3_its_select_cpu(device_t dev, struct intr_irqsrc *isrc)
1040 {
1041 	struct gicv3_its_softc *sc;
1042 
1043 	sc = device_get_softc(dev);
1044 	if (CPU_EMPTY(&isrc->isrc_cpu)) {
1045 		sc->gic_irq_cpu = intr_irq_next_cpu(sc->gic_irq_cpu,
1046 		    &sc->sc_cpus);
1047 		CPU_SETOF(sc->gic_irq_cpu, &isrc->isrc_cpu);
1048 	}
1049 
1050 	return (0);
1051 }
1052 
1053 static int
1054 gicv3_its_bind_intr(device_t dev, struct intr_irqsrc *isrc)
1055 {
1056 	struct gicv3_its_irqsrc *girq;
1057 
1058 	gicv3_its_select_cpu(dev, isrc);
1059 
1060 	girq = (struct gicv3_its_irqsrc *)isrc;
1061 	its_cmd_movi(dev, girq);
1062 	return (0);
1063 }
1064 
1065 static int
1066 gicv3_its_map_intr(device_t dev, struct intr_map_data *data,
1067     struct intr_irqsrc **isrcp)
1068 {
1069 
1070 	/*
1071 	 * This should never happen, we only call this function to map
1072 	 * interrupts found before the controller driver is ready.
1073 	 */
1074 	panic("gicv3_its_map_intr: Unable to map a MSI interrupt");
1075 }
1076 
1077 static int
1078 gicv3_its_setup_intr(device_t dev, struct intr_irqsrc *isrc,
1079     struct resource *res, struct intr_map_data *data)
1080 {
1081 
1082 	/* Bind the interrupt to a CPU */
1083 	gicv3_its_bind_intr(dev, isrc);
1084 
1085 	return (0);
1086 }
1087 
1088 #ifdef SMP
1089 static void
1090 gicv3_its_init_secondary(device_t dev)
1091 {
1092 	struct gicv3_its_softc *sc;
1093 
1094 	sc = device_get_softc(dev);
1095 
1096 	/*
1097 	 * This is fatal as otherwise we may bind interrupts to this CPU.
1098 	 * We need a way to tell the interrupt framework to only bind to a
1099 	 * subset of given CPUs when it performs the shuffle.
1100 	 */
1101 	if (its_init_cpu(dev, sc) != 0)
1102 		panic("gicv3_its_init_secondary: No usable ITS on CPU%d",
1103 		    PCPU_GET(cpuid));
1104 }
1105 #endif
1106 
1107 static uint32_t
1108 its_get_devid(device_t pci_dev)
1109 {
1110 	uintptr_t id;
1111 
1112 	if (pci_get_id(pci_dev, PCI_ID_MSI, &id) != 0)
1113 		panic("its_get_devid: Unable to get the MSI DeviceID");
1114 
1115 	return (id);
1116 }
1117 
1118 static struct its_dev *
1119 its_device_find(device_t dev, device_t child)
1120 {
1121 	struct gicv3_its_softc *sc;
1122 	struct its_dev *its_dev = NULL;
1123 
1124 	sc = device_get_softc(dev);
1125 
1126 	mtx_lock_spin(&sc->sc_its_dev_lock);
1127 	TAILQ_FOREACH(its_dev, &sc->sc_its_dev_list, entry) {
1128 		if (its_dev->pci_dev == child)
1129 			break;
1130 	}
1131 	mtx_unlock_spin(&sc->sc_its_dev_lock);
1132 
1133 	return (its_dev);
1134 }
1135 
1136 static struct its_dev *
1137 its_device_get(device_t dev, device_t child, u_int nvecs)
1138 {
1139 	struct gicv3_its_softc *sc;
1140 	struct its_dev *its_dev;
1141 	vmem_addr_t irq_base;
1142 	size_t esize;
1143 
1144 	sc = device_get_softc(dev);
1145 
1146 	its_dev = its_device_find(dev, child);
1147 	if (its_dev != NULL)
1148 		return (its_dev);
1149 
1150 	its_dev = malloc(sizeof(*its_dev), M_GICV3_ITS, M_NOWAIT | M_ZERO);
1151 	if (its_dev == NULL)
1152 		return (NULL);
1153 
1154 	its_dev->pci_dev = child;
1155 	its_dev->devid = its_get_devid(child);
1156 
1157 	its_dev->lpis.lpi_busy = 0;
1158 	its_dev->lpis.lpi_num = nvecs;
1159 	its_dev->lpis.lpi_free = nvecs;
1160 
1161 	if (vmem_alloc(sc->sc_irq_alloc, nvecs, M_FIRSTFIT | M_NOWAIT,
1162 	    &irq_base) != 0) {
1163 		free(its_dev, M_GICV3_ITS);
1164 		return (NULL);
1165 	}
1166 	its_dev->lpis.lpi_base = irq_base;
1167 
1168 	/* Get ITT entry size */
1169 	esize = GITS_TYPER_ITTES(gic_its_read_8(sc, GITS_TYPER));
1170 
1171 	/*
1172 	 * Allocate ITT for this device.
1173 	 * PA has to be 256 B aligned. At least two entries for device.
1174 	 */
1175 	its_dev->itt_size = roundup2(MAX(nvecs, 2) * esize, 256);
1176 	its_dev->itt = (vm_offset_t)contigmalloc(its_dev->itt_size,
1177 	    M_GICV3_ITS, M_NOWAIT | M_ZERO, 0, LPI_INT_TRANS_TAB_MAX_ADDR,
1178 	    LPI_INT_TRANS_TAB_ALIGN, 0);
1179 	if (its_dev->itt == 0) {
1180 		vmem_free(sc->sc_irq_alloc, its_dev->lpis.lpi_base, nvecs);
1181 		free(its_dev, M_GICV3_ITS);
1182 		return (NULL);
1183 	}
1184 
1185 	mtx_lock_spin(&sc->sc_its_dev_lock);
1186 	TAILQ_INSERT_TAIL(&sc->sc_its_dev_list, its_dev, entry);
1187 	mtx_unlock_spin(&sc->sc_its_dev_lock);
1188 
1189 	/* Map device to its ITT */
1190 	its_cmd_mapd(dev, its_dev, 1);
1191 
1192 	return (its_dev);
1193 }
1194 
1195 static void
1196 its_device_release(device_t dev, struct its_dev *its_dev)
1197 {
1198 	struct gicv3_its_softc *sc;
1199 
1200 	KASSERT(its_dev->lpis.lpi_busy == 0,
1201 	    ("its_device_release: Trying to release an inuse ITS device"));
1202 
1203 	/* Unmap device in ITS */
1204 	its_cmd_mapd(dev, its_dev, 0);
1205 
1206 	sc = device_get_softc(dev);
1207 
1208 	/* Remove the device from the list of devices */
1209 	mtx_lock_spin(&sc->sc_its_dev_lock);
1210 	TAILQ_REMOVE(&sc->sc_its_dev_list, its_dev, entry);
1211 	mtx_unlock_spin(&sc->sc_its_dev_lock);
1212 
1213 	/* Free ITT */
1214 	KASSERT(its_dev->itt != 0, ("Invalid ITT in valid ITS device"));
1215 	contigfree((void *)its_dev->itt, its_dev->itt_size, M_GICV3_ITS);
1216 
1217 	/* Free the IRQ allocation */
1218 	vmem_free(sc->sc_irq_alloc, its_dev->lpis.lpi_base,
1219 	    its_dev->lpis.lpi_num);
1220 
1221 	free(its_dev, M_GICV3_ITS);
1222 }
1223 
1224 static struct gicv3_its_irqsrc *
1225 gicv3_its_alloc_irqsrc(device_t dev, struct gicv3_its_softc *sc, u_int irq)
1226 {
1227 	struct gicv3_its_irqsrc *girq = NULL;
1228 
1229 	KASSERT(sc->sc_irqs[irq] == NULL,
1230 	    ("%s: Interrupt %u already allocated", __func__, irq));
1231 	mtx_lock_spin(&sc->sc_its_dev_lock);
1232 	if (!TAILQ_EMPTY(&sc->sc_free_irqs)) {
1233 		girq = TAILQ_FIRST(&sc->sc_free_irqs);
1234 		TAILQ_REMOVE(&sc->sc_free_irqs, girq, gi_link);
1235 	}
1236 	mtx_unlock_spin(&sc->sc_its_dev_lock);
1237 	if (girq == NULL) {
1238 		girq = malloc(sizeof(*girq), M_GICV3_ITS,
1239 		    M_NOWAIT | M_ZERO);
1240 		if (girq == NULL)
1241 			return (NULL);
1242 		girq->gi_id = -1;
1243 		if (intr_isrc_register(&girq->gi_isrc, dev, 0,
1244 		    "%s,%u", device_get_nameunit(dev), irq) != 0) {
1245 			free(girq, M_GICV3_ITS);
1246 			return (NULL);
1247 		}
1248 	}
1249 	girq->gi_lpi = irq + sc->sc_irq_base - GIC_FIRST_LPI;
1250 	sc->sc_irqs[irq] = girq;
1251 
1252 	return (girq);
1253 }
1254 
1255 static void
1256 gicv3_its_release_irqsrc(struct gicv3_its_softc *sc,
1257     struct gicv3_its_irqsrc *girq)
1258 {
1259 	u_int irq;
1260 
1261 	mtx_assert(&sc->sc_its_dev_lock, MA_OWNED);
1262 
1263 	irq = girq->gi_lpi + GIC_FIRST_LPI - sc->sc_irq_base;
1264 	sc->sc_irqs[irq] = NULL;
1265 
1266 	girq->gi_id = -1;
1267 	girq->gi_its_dev = NULL;
1268 	TAILQ_INSERT_TAIL(&sc->sc_free_irqs, girq, gi_link);
1269 }
1270 
1271 static int
1272 gicv3_its_alloc_msi(device_t dev, device_t child, int count, int maxcount,
1273     device_t *pic, struct intr_irqsrc **srcs)
1274 {
1275 	struct gicv3_its_softc *sc;
1276 	struct gicv3_its_irqsrc *girq;
1277 	struct its_dev *its_dev;
1278 	u_int irq;
1279 	int i;
1280 
1281 	its_dev = its_device_get(dev, child, count);
1282 	if (its_dev == NULL)
1283 		return (ENXIO);
1284 
1285 	KASSERT(its_dev->lpis.lpi_free >= count,
1286 	    ("gicv3_its_alloc_msi: No free LPIs"));
1287 	sc = device_get_softc(dev);
1288 	irq = its_dev->lpis.lpi_base + its_dev->lpis.lpi_num -
1289 	    its_dev->lpis.lpi_free;
1290 
1291 	/* Allocate the irqsrc for each MSI */
1292 	for (i = 0; i < count; i++, irq++) {
1293 		its_dev->lpis.lpi_free--;
1294 		srcs[i] = (struct intr_irqsrc *)gicv3_its_alloc_irqsrc(dev,
1295 		    sc, irq);
1296 		if (srcs[i] == NULL)
1297 			break;
1298 	}
1299 
1300 	/* The allocation failed, release them */
1301 	if (i != count) {
1302 		mtx_lock_spin(&sc->sc_its_dev_lock);
1303 		for (i = 0; i < count; i++) {
1304 			girq = (struct gicv3_its_irqsrc *)srcs[i];
1305 			if (girq == NULL)
1306 				break;
1307 			gicv3_its_release_irqsrc(sc, girq);
1308 			srcs[i] = NULL;
1309 		}
1310 		mtx_unlock_spin(&sc->sc_its_dev_lock);
1311 		return (ENXIO);
1312 	}
1313 
1314 	/* Finish the allocation now we have all MSI irqsrcs */
1315 	for (i = 0; i < count; i++) {
1316 		girq = (struct gicv3_its_irqsrc *)srcs[i];
1317 		girq->gi_id = i;
1318 		girq->gi_its_dev = its_dev;
1319 
1320 		/* Map the message to the given IRQ */
1321 		gicv3_its_select_cpu(dev, (struct intr_irqsrc *)girq);
1322 		its_cmd_mapti(dev, girq);
1323 	}
1324 	its_dev->lpis.lpi_busy += count;
1325 	*pic = dev;
1326 
1327 	return (0);
1328 }
1329 
1330 static int
1331 gicv3_its_release_msi(device_t dev, device_t child, int count,
1332     struct intr_irqsrc **isrc)
1333 {
1334 	struct gicv3_its_softc *sc;
1335 	struct gicv3_its_irqsrc *girq;
1336 	struct its_dev *its_dev;
1337 	int i;
1338 
1339 	its_dev = its_device_find(dev, child);
1340 
1341 	KASSERT(its_dev != NULL,
1342 	    ("gicv3_its_release_msi: Releasing a MSI interrupt with "
1343 	     "no ITS device"));
1344 	KASSERT(its_dev->lpis.lpi_busy >= count,
1345 	    ("gicv3_its_release_msi: Releasing more interrupts than "
1346 	     "were allocated: releasing %d, allocated %d", count,
1347 	     its_dev->lpis.lpi_busy));
1348 
1349 	sc = device_get_softc(dev);
1350 	mtx_lock_spin(&sc->sc_its_dev_lock);
1351 	for (i = 0; i < count; i++) {
1352 		girq = (struct gicv3_its_irqsrc *)isrc[i];
1353 		gicv3_its_release_irqsrc(sc, girq);
1354 	}
1355 	mtx_unlock_spin(&sc->sc_its_dev_lock);
1356 	its_dev->lpis.lpi_busy -= count;
1357 
1358 	if (its_dev->lpis.lpi_busy == 0)
1359 		its_device_release(dev, its_dev);
1360 
1361 	return (0);
1362 }
1363 
1364 static int
1365 gicv3_its_alloc_msix(device_t dev, device_t child, device_t *pic,
1366     struct intr_irqsrc **isrcp)
1367 {
1368 	struct gicv3_its_softc *sc;
1369 	struct gicv3_its_irqsrc *girq;
1370 	struct its_dev *its_dev;
1371 	u_int nvecs, irq;
1372 
1373 	nvecs = pci_msix_count(child);
1374 	its_dev = its_device_get(dev, child, nvecs);
1375 	if (its_dev == NULL)
1376 		return (ENXIO);
1377 
1378 	KASSERT(its_dev->lpis.lpi_free > 0,
1379 	    ("gicv3_its_alloc_msix: No free LPIs"));
1380 	sc = device_get_softc(dev);
1381 	irq = its_dev->lpis.lpi_base + its_dev->lpis.lpi_num -
1382 	    its_dev->lpis.lpi_free;
1383 
1384 	girq = gicv3_its_alloc_irqsrc(dev, sc, irq);
1385 	if (girq == NULL)
1386 		return (ENXIO);
1387 	girq->gi_id = its_dev->lpis.lpi_busy;
1388 	girq->gi_its_dev = its_dev;
1389 
1390 	its_dev->lpis.lpi_free--;
1391 	its_dev->lpis.lpi_busy++;
1392 
1393 	/* Map the message to the given IRQ */
1394 	gicv3_its_select_cpu(dev, (struct intr_irqsrc *)girq);
1395 	its_cmd_mapti(dev, girq);
1396 
1397 	*pic = dev;
1398 	*isrcp = (struct intr_irqsrc *)girq;
1399 
1400 	return (0);
1401 }
1402 
1403 static int
1404 gicv3_its_release_msix(device_t dev, device_t child, struct intr_irqsrc *isrc)
1405 {
1406 	struct gicv3_its_softc *sc;
1407 	struct gicv3_its_irqsrc *girq;
1408 	struct its_dev *its_dev;
1409 
1410 	its_dev = its_device_find(dev, child);
1411 
1412 	KASSERT(its_dev != NULL,
1413 	    ("gicv3_its_release_msix: Releasing a MSI-X interrupt with "
1414 	     "no ITS device"));
1415 	KASSERT(its_dev->lpis.lpi_busy > 0,
1416 	    ("gicv3_its_release_msix: Releasing more interrupts than "
1417 	     "were allocated: allocated %d", its_dev->lpis.lpi_busy));
1418 
1419 	sc = device_get_softc(dev);
1420 	girq = (struct gicv3_its_irqsrc *)isrc;
1421 	gicv3_its_release_irqsrc(sc, girq);
1422 	its_dev->lpis.lpi_busy--;
1423 
1424 	if (its_dev->lpis.lpi_busy == 0)
1425 		its_device_release(dev, its_dev);
1426 
1427 	return (0);
1428 }
1429 
1430 static int
1431 gicv3_its_map_msi(device_t dev, device_t child, struct intr_irqsrc *isrc,
1432     uint64_t *addr, uint32_t *data)
1433 {
1434 	struct gicv3_its_softc *sc;
1435 	struct gicv3_its_irqsrc *girq;
1436 
1437 	sc = device_get_softc(dev);
1438 	girq = (struct gicv3_its_irqsrc *)isrc;
1439 
1440 	*addr = vtophys(rman_get_virtual(sc->sc_its_res)) + GITS_TRANSLATER;
1441 	*data = girq->gi_id;
1442 
1443 	return (0);
1444 }
1445 
1446 #ifdef IOMMU
1447 static int
1448 gicv3_iommu_init(device_t dev, device_t child, struct iommu_domain **domain)
1449 {
1450 	struct gicv3_its_softc *sc;
1451 	struct iommu_ctx *ctx;
1452 	int error;
1453 
1454 	sc = device_get_softc(dev);
1455 	ctx = iommu_get_dev_ctx(child);
1456 	error = iommu_map_msi(ctx, PAGE_SIZE, GITS_TRANSLATER,
1457 	    IOMMU_MAP_ENTRY_WRITE, IOMMU_MF_CANWAIT, &sc->ma);
1458 	*domain = iommu_get_ctx_domain(ctx);
1459 
1460 	return (error);
1461 }
1462 
1463 static void
1464 gicv3_iommu_deinit(device_t dev, device_t child)
1465 {
1466 	struct iommu_ctx *ctx;
1467 
1468 	ctx = iommu_get_dev_ctx(child);
1469 	iommu_unmap_msi(ctx);
1470 }
1471 #endif
1472 
1473 /*
1474  * Commands handling.
1475  */
1476 
1477 static __inline void
1478 cmd_format_command(struct its_cmd *cmd, uint8_t cmd_type)
1479 {
1480 	/* Command field: DW0 [7:0] */
1481 	cmd->cmd_dword[0] &= htole64(~CMD_COMMAND_MASK);
1482 	cmd->cmd_dword[0] |= htole64(cmd_type);
1483 }
1484 
1485 static __inline void
1486 cmd_format_devid(struct its_cmd *cmd, uint32_t devid)
1487 {
1488 	/* Device ID field: DW0 [63:32] */
1489 	cmd->cmd_dword[0] &= htole64(~CMD_DEVID_MASK);
1490 	cmd->cmd_dword[0] |= htole64((uint64_t)devid << CMD_DEVID_SHIFT);
1491 }
1492 
1493 static __inline void
1494 cmd_format_size(struct its_cmd *cmd, uint16_t size)
1495 {
1496 	/* Size field: DW1 [4:0] */
1497 	cmd->cmd_dword[1] &= htole64(~CMD_SIZE_MASK);
1498 	cmd->cmd_dword[1] |= htole64((size & CMD_SIZE_MASK));
1499 }
1500 
1501 static __inline void
1502 cmd_format_id(struct its_cmd *cmd, uint32_t id)
1503 {
1504 	/* ID field: DW1 [31:0] */
1505 	cmd->cmd_dword[1] &= htole64(~CMD_ID_MASK);
1506 	cmd->cmd_dword[1] |= htole64(id);
1507 }
1508 
1509 static __inline void
1510 cmd_format_pid(struct its_cmd *cmd, uint32_t pid)
1511 {
1512 	/* Physical ID field: DW1 [63:32] */
1513 	cmd->cmd_dword[1] &= htole64(~CMD_PID_MASK);
1514 	cmd->cmd_dword[1] |= htole64((uint64_t)pid << CMD_PID_SHIFT);
1515 }
1516 
1517 static __inline void
1518 cmd_format_col(struct its_cmd *cmd, uint16_t col_id)
1519 {
1520 	/* Collection field: DW2 [16:0] */
1521 	cmd->cmd_dword[2] &= htole64(~CMD_COL_MASK);
1522 	cmd->cmd_dword[2] |= htole64(col_id);
1523 }
1524 
1525 static __inline void
1526 cmd_format_target(struct its_cmd *cmd, uint64_t target)
1527 {
1528 	/* Target Address field: DW2 [47:16] */
1529 	cmd->cmd_dword[2] &= htole64(~CMD_TARGET_MASK);
1530 	cmd->cmd_dword[2] |= htole64(target & CMD_TARGET_MASK);
1531 }
1532 
1533 static __inline void
1534 cmd_format_itt(struct its_cmd *cmd, uint64_t itt)
1535 {
1536 	/* ITT Address field: DW2 [47:8] */
1537 	cmd->cmd_dword[2] &= htole64(~CMD_ITT_MASK);
1538 	cmd->cmd_dword[2] |= htole64(itt & CMD_ITT_MASK);
1539 }
1540 
1541 static __inline void
1542 cmd_format_valid(struct its_cmd *cmd, uint8_t valid)
1543 {
1544 	/* Valid field: DW2 [63] */
1545 	cmd->cmd_dword[2] &= htole64(~CMD_VALID_MASK);
1546 	cmd->cmd_dword[2] |= htole64((uint64_t)valid << CMD_VALID_SHIFT);
1547 }
1548 
1549 static inline bool
1550 its_cmd_queue_full(struct gicv3_its_softc *sc)
1551 {
1552 	size_t read_idx, next_write_idx;
1553 
1554 	/* Get the index of the next command */
1555 	next_write_idx = (sc->sc_its_cmd_next_idx + 1) %
1556 	    (ITS_CMDQ_SIZE / sizeof(struct its_cmd));
1557 	/* And the index of the current command being read */
1558 	read_idx = gic_its_read_4(sc, GITS_CREADR) / sizeof(struct its_cmd);
1559 
1560 	/*
1561 	 * The queue is full when the write offset points
1562 	 * at the command before the current read offset.
1563 	 */
1564 	return (next_write_idx == read_idx);
1565 }
1566 
1567 static inline void
1568 its_cmd_sync(struct gicv3_its_softc *sc, struct its_cmd *cmd)
1569 {
1570 
1571 	if ((sc->sc_its_flags & ITS_FLAGS_CMDQ_FLUSH) != 0) {
1572 		/* Clean D-cache under command. */
1573 		cpu_dcache_wb_range((vm_offset_t)cmd, sizeof(*cmd));
1574 	} else {
1575 		/* DSB inner shareable, store */
1576 		dsb(ishst);
1577 	}
1578 
1579 }
1580 
1581 static inline uint64_t
1582 its_cmd_cwriter_offset(struct gicv3_its_softc *sc, struct its_cmd *cmd)
1583 {
1584 	uint64_t off;
1585 
1586 	off = (cmd - sc->sc_its_cmd_base) * sizeof(*cmd);
1587 
1588 	return (off);
1589 }
1590 
1591 static void
1592 its_cmd_wait_completion(device_t dev, struct its_cmd *cmd_first,
1593     struct its_cmd *cmd_last)
1594 {
1595 	struct gicv3_its_softc *sc;
1596 	uint64_t first, last, read;
1597 	size_t us_left;
1598 
1599 	sc = device_get_softc(dev);
1600 
1601 	/*
1602 	 * XXX ARM64TODO: This is obviously a significant delay.
1603 	 * The reason for that is that currently the time frames for
1604 	 * the command to complete are not known.
1605 	 */
1606 	us_left = 1000000;
1607 
1608 	first = its_cmd_cwriter_offset(sc, cmd_first);
1609 	last = its_cmd_cwriter_offset(sc, cmd_last);
1610 
1611 	for (;;) {
1612 		read = gic_its_read_8(sc, GITS_CREADR);
1613 		if (first < last) {
1614 			if (read < first || read >= last)
1615 				break;
1616 		} else if (read < first && read >= last)
1617 			break;
1618 
1619 		if (us_left-- == 0) {
1620 			/* This means timeout */
1621 			device_printf(dev,
1622 			    "Timeout while waiting for CMD completion.\n");
1623 			return;
1624 		}
1625 		DELAY(1);
1626 	}
1627 }
1628 
1629 static struct its_cmd *
1630 its_cmd_alloc_locked(device_t dev)
1631 {
1632 	struct gicv3_its_softc *sc;
1633 	struct its_cmd *cmd;
1634 	size_t us_left;
1635 
1636 	sc = device_get_softc(dev);
1637 
1638 	/*
1639 	 * XXX ARM64TODO: This is obviously a significant delay.
1640 	 * The reason for that is that currently the time frames for
1641 	 * the command to complete (and therefore free the descriptor)
1642 	 * are not known.
1643 	 */
1644 	us_left = 1000000;
1645 
1646 	mtx_assert(&sc->sc_its_cmd_lock, MA_OWNED);
1647 	while (its_cmd_queue_full(sc)) {
1648 		if (us_left-- == 0) {
1649 			/* Timeout while waiting for free command */
1650 			device_printf(dev,
1651 			    "Timeout while waiting for free command\n");
1652 			return (NULL);
1653 		}
1654 		DELAY(1);
1655 	}
1656 
1657 	cmd = &sc->sc_its_cmd_base[sc->sc_its_cmd_next_idx];
1658 	sc->sc_its_cmd_next_idx++;
1659 	sc->sc_its_cmd_next_idx %= ITS_CMDQ_SIZE / sizeof(struct its_cmd);
1660 
1661 	return (cmd);
1662 }
1663 
1664 static uint64_t
1665 its_cmd_prepare(struct its_cmd *cmd, struct its_cmd_desc *desc)
1666 {
1667 	uint64_t target;
1668 	uint8_t cmd_type;
1669 	u_int size;
1670 
1671 	cmd_type = desc->cmd_type;
1672 	target = ITS_TARGET_NONE;
1673 
1674 	switch (cmd_type) {
1675 	case ITS_CMD_MOVI:	/* Move interrupt ID to another collection */
1676 		target = desc->cmd_desc_movi.col->col_target;
1677 		cmd_format_command(cmd, ITS_CMD_MOVI);
1678 		cmd_format_id(cmd, desc->cmd_desc_movi.id);
1679 		cmd_format_col(cmd, desc->cmd_desc_movi.col->col_id);
1680 		cmd_format_devid(cmd, desc->cmd_desc_movi.its_dev->devid);
1681 		break;
1682 	case ITS_CMD_SYNC:	/* Wait for previous commands completion */
1683 		target = desc->cmd_desc_sync.col->col_target;
1684 		cmd_format_command(cmd, ITS_CMD_SYNC);
1685 		cmd_format_target(cmd, target);
1686 		break;
1687 	case ITS_CMD_MAPD:	/* Assign ITT to device */
1688 		cmd_format_command(cmd, ITS_CMD_MAPD);
1689 		cmd_format_itt(cmd, vtophys(desc->cmd_desc_mapd.its_dev->itt));
1690 		/*
1691 		 * Size describes number of bits to encode interrupt IDs
1692 		 * supported by the device minus one.
1693 		 * When V (valid) bit is zero, this field should be written
1694 		 * as zero.
1695 		 */
1696 		if (desc->cmd_desc_mapd.valid != 0) {
1697 			size = fls(desc->cmd_desc_mapd.its_dev->lpis.lpi_num);
1698 			size = MAX(1, size) - 1;
1699 		} else
1700 			size = 0;
1701 
1702 		cmd_format_size(cmd, size);
1703 		cmd_format_devid(cmd, desc->cmd_desc_mapd.its_dev->devid);
1704 		cmd_format_valid(cmd, desc->cmd_desc_mapd.valid);
1705 		break;
1706 	case ITS_CMD_MAPC:	/* Map collection to Re-Distributor */
1707 		target = desc->cmd_desc_mapc.col->col_target;
1708 		cmd_format_command(cmd, ITS_CMD_MAPC);
1709 		cmd_format_col(cmd, desc->cmd_desc_mapc.col->col_id);
1710 		cmd_format_valid(cmd, desc->cmd_desc_mapc.valid);
1711 		cmd_format_target(cmd, target);
1712 		break;
1713 	case ITS_CMD_MAPTI:
1714 		target = desc->cmd_desc_mapvi.col->col_target;
1715 		cmd_format_command(cmd, ITS_CMD_MAPTI);
1716 		cmd_format_devid(cmd, desc->cmd_desc_mapvi.its_dev->devid);
1717 		cmd_format_id(cmd, desc->cmd_desc_mapvi.id);
1718 		cmd_format_pid(cmd, desc->cmd_desc_mapvi.pid);
1719 		cmd_format_col(cmd, desc->cmd_desc_mapvi.col->col_id);
1720 		break;
1721 	case ITS_CMD_MAPI:
1722 		target = desc->cmd_desc_mapi.col->col_target;
1723 		cmd_format_command(cmd, ITS_CMD_MAPI);
1724 		cmd_format_devid(cmd, desc->cmd_desc_mapi.its_dev->devid);
1725 		cmd_format_id(cmd, desc->cmd_desc_mapi.pid);
1726 		cmd_format_col(cmd, desc->cmd_desc_mapi.col->col_id);
1727 		break;
1728 	case ITS_CMD_INV:
1729 		target = desc->cmd_desc_inv.col->col_target;
1730 		cmd_format_command(cmd, ITS_CMD_INV);
1731 		cmd_format_devid(cmd, desc->cmd_desc_inv.its_dev->devid);
1732 		cmd_format_id(cmd, desc->cmd_desc_inv.pid);
1733 		break;
1734 	case ITS_CMD_INVALL:
1735 		cmd_format_command(cmd, ITS_CMD_INVALL);
1736 		cmd_format_col(cmd, desc->cmd_desc_invall.col->col_id);
1737 		break;
1738 	default:
1739 		panic("its_cmd_prepare: Invalid command: %x", cmd_type);
1740 	}
1741 
1742 	return (target);
1743 }
1744 
1745 static int
1746 its_cmd_send(device_t dev, struct its_cmd_desc *desc)
1747 {
1748 	struct gicv3_its_softc *sc;
1749 	struct its_cmd *cmd, *cmd_sync, *cmd_write;
1750 	struct its_col col_sync;
1751 	struct its_cmd_desc desc_sync;
1752 	uint64_t target, cwriter;
1753 
1754 	sc = device_get_softc(dev);
1755 	mtx_lock_spin(&sc->sc_its_cmd_lock);
1756 	cmd = its_cmd_alloc_locked(dev);
1757 	if (cmd == NULL) {
1758 		device_printf(dev, "could not allocate ITS command\n");
1759 		mtx_unlock_spin(&sc->sc_its_cmd_lock);
1760 		return (EBUSY);
1761 	}
1762 
1763 	target = its_cmd_prepare(cmd, desc);
1764 	its_cmd_sync(sc, cmd);
1765 
1766 	if (target != ITS_TARGET_NONE) {
1767 		cmd_sync = its_cmd_alloc_locked(dev);
1768 		if (cmd_sync != NULL) {
1769 			desc_sync.cmd_type = ITS_CMD_SYNC;
1770 			col_sync.col_target = target;
1771 			desc_sync.cmd_desc_sync.col = &col_sync;
1772 			its_cmd_prepare(cmd_sync, &desc_sync);
1773 			its_cmd_sync(sc, cmd_sync);
1774 		}
1775 	}
1776 
1777 	/* Update GITS_CWRITER */
1778 	cwriter = sc->sc_its_cmd_next_idx * sizeof(struct its_cmd);
1779 	gic_its_write_8(sc, GITS_CWRITER, cwriter);
1780 	cmd_write = &sc->sc_its_cmd_base[sc->sc_its_cmd_next_idx];
1781 	mtx_unlock_spin(&sc->sc_its_cmd_lock);
1782 
1783 	its_cmd_wait_completion(dev, cmd, cmd_write);
1784 
1785 	return (0);
1786 }
1787 
1788 /* Handlers to send commands */
1789 static void
1790 its_cmd_movi(device_t dev, struct gicv3_its_irqsrc *girq)
1791 {
1792 	struct gicv3_its_softc *sc;
1793 	struct its_cmd_desc desc;
1794 	struct its_col *col;
1795 
1796 	sc = device_get_softc(dev);
1797 	col = sc->sc_its_cols[CPU_FFS(&girq->gi_isrc.isrc_cpu) - 1];
1798 
1799 	desc.cmd_type = ITS_CMD_MOVI;
1800 	desc.cmd_desc_movi.its_dev = girq->gi_its_dev;
1801 	desc.cmd_desc_movi.col = col;
1802 	desc.cmd_desc_movi.id = girq->gi_id;
1803 
1804 	its_cmd_send(dev, &desc);
1805 }
1806 
1807 static void
1808 its_cmd_mapc(device_t dev, struct its_col *col, uint8_t valid)
1809 {
1810 	struct its_cmd_desc desc;
1811 
1812 	desc.cmd_type = ITS_CMD_MAPC;
1813 	desc.cmd_desc_mapc.col = col;
1814 	/*
1815 	 * Valid bit set - map the collection.
1816 	 * Valid bit cleared - unmap the collection.
1817 	 */
1818 	desc.cmd_desc_mapc.valid = valid;
1819 
1820 	its_cmd_send(dev, &desc);
1821 }
1822 
1823 static void
1824 its_cmd_mapti(device_t dev, struct gicv3_its_irqsrc *girq)
1825 {
1826 	struct gicv3_its_softc *sc;
1827 	struct its_cmd_desc desc;
1828 	struct its_col *col;
1829 	u_int col_id;
1830 
1831 	sc = device_get_softc(dev);
1832 
1833 	col_id = CPU_FFS(&girq->gi_isrc.isrc_cpu) - 1;
1834 	col = sc->sc_its_cols[col_id];
1835 
1836 	desc.cmd_type = ITS_CMD_MAPTI;
1837 	desc.cmd_desc_mapvi.its_dev = girq->gi_its_dev;
1838 	desc.cmd_desc_mapvi.col = col;
1839 	/* The EventID sent to the device */
1840 	desc.cmd_desc_mapvi.id = girq->gi_id;
1841 	/* The physical interrupt presented to softeware */
1842 	desc.cmd_desc_mapvi.pid = girq->gi_lpi + GIC_FIRST_LPI;
1843 
1844 	its_cmd_send(dev, &desc);
1845 }
1846 
1847 static void
1848 its_cmd_mapd(device_t dev, struct its_dev *its_dev, uint8_t valid)
1849 {
1850 	struct its_cmd_desc desc;
1851 
1852 	desc.cmd_type = ITS_CMD_MAPD;
1853 	desc.cmd_desc_mapd.its_dev = its_dev;
1854 	desc.cmd_desc_mapd.valid = valid;
1855 
1856 	its_cmd_send(dev, &desc);
1857 }
1858 
1859 static void
1860 its_cmd_inv(device_t dev, struct its_dev *its_dev,
1861     struct gicv3_its_irqsrc *girq)
1862 {
1863 	struct gicv3_its_softc *sc;
1864 	struct its_cmd_desc desc;
1865 	struct its_col *col;
1866 
1867 	sc = device_get_softc(dev);
1868 	col = sc->sc_its_cols[CPU_FFS(&girq->gi_isrc.isrc_cpu) - 1];
1869 
1870 	desc.cmd_type = ITS_CMD_INV;
1871 	/* The EventID sent to the device */
1872 	desc.cmd_desc_inv.pid = girq->gi_id;
1873 	desc.cmd_desc_inv.its_dev = its_dev;
1874 	desc.cmd_desc_inv.col = col;
1875 
1876 	its_cmd_send(dev, &desc);
1877 }
1878 
1879 static void
1880 its_cmd_invall(device_t dev, struct its_col *col)
1881 {
1882 	struct its_cmd_desc desc;
1883 
1884 	desc.cmd_type = ITS_CMD_INVALL;
1885 	desc.cmd_desc_invall.col = col;
1886 
1887 	its_cmd_send(dev, &desc);
1888 }
1889 
1890 #ifdef FDT
1891 static device_probe_t gicv3_its_fdt_probe;
1892 static device_attach_t gicv3_its_fdt_attach;
1893 
1894 static device_method_t gicv3_its_fdt_methods[] = {
1895 	/* Device interface */
1896 	DEVMETHOD(device_probe,		gicv3_its_fdt_probe),
1897 	DEVMETHOD(device_attach,	gicv3_its_fdt_attach),
1898 
1899 	/* End */
1900 	DEVMETHOD_END
1901 };
1902 
1903 #define its_baseclasses its_fdt_baseclasses
1904 DEFINE_CLASS_1(its, gicv3_its_fdt_driver, gicv3_its_fdt_methods,
1905     sizeof(struct gicv3_its_softc), gicv3_its_driver);
1906 #undef its_baseclasses
1907 static devclass_t gicv3_its_fdt_devclass;
1908 
1909 EARLY_DRIVER_MODULE(its_fdt, gic, gicv3_its_fdt_driver,
1910     gicv3_its_fdt_devclass, 0, 0, BUS_PASS_INTERRUPT + BUS_PASS_ORDER_MIDDLE);
1911 
1912 static int
1913 gicv3_its_fdt_probe(device_t dev)
1914 {
1915 
1916 	if (!ofw_bus_status_okay(dev))
1917 		return (ENXIO);
1918 
1919 	if (!ofw_bus_is_compatible(dev, "arm,gic-v3-its"))
1920 		return (ENXIO);
1921 
1922 	device_set_desc(dev, "ARM GIC Interrupt Translation Service");
1923 	return (BUS_PROBE_DEFAULT);
1924 }
1925 
1926 static int
1927 gicv3_its_fdt_attach(device_t dev)
1928 {
1929 	struct gicv3_its_softc *sc;
1930 	phandle_t xref;
1931 	int err;
1932 
1933 	sc = device_get_softc(dev);
1934 	sc->dev = dev;
1935 	err = gicv3_its_attach(dev);
1936 	if (err != 0)
1937 		return (err);
1938 
1939 	/* Register this device as a interrupt controller */
1940 	xref = OF_xref_from_node(ofw_bus_get_node(dev));
1941 	sc->sc_pic = intr_pic_register(dev, xref);
1942 	intr_pic_add_handler(device_get_parent(dev), sc->sc_pic,
1943 	    gicv3_its_intr, sc, sc->sc_irq_base, sc->sc_irq_length);
1944 
1945 	/* Register this device to handle MSI interrupts */
1946 	intr_msi_register(dev, xref);
1947 
1948 	return (0);
1949 }
1950 #endif
1951 
1952 #ifdef DEV_ACPI
1953 static device_probe_t gicv3_its_acpi_probe;
1954 static device_attach_t gicv3_its_acpi_attach;
1955 
1956 static device_method_t gicv3_its_acpi_methods[] = {
1957 	/* Device interface */
1958 	DEVMETHOD(device_probe,		gicv3_its_acpi_probe),
1959 	DEVMETHOD(device_attach,	gicv3_its_acpi_attach),
1960 
1961 	/* End */
1962 	DEVMETHOD_END
1963 };
1964 
1965 #define its_baseclasses its_acpi_baseclasses
1966 DEFINE_CLASS_1(its, gicv3_its_acpi_driver, gicv3_its_acpi_methods,
1967     sizeof(struct gicv3_its_softc), gicv3_its_driver);
1968 #undef its_baseclasses
1969 static devclass_t gicv3_its_acpi_devclass;
1970 
1971 EARLY_DRIVER_MODULE(its_acpi, gic, gicv3_its_acpi_driver,
1972     gicv3_its_acpi_devclass, 0, 0, BUS_PASS_INTERRUPT + BUS_PASS_ORDER_MIDDLE);
1973 
1974 static int
1975 gicv3_its_acpi_probe(device_t dev)
1976 {
1977 
1978 	if (gic_get_bus(dev) != GIC_BUS_ACPI)
1979 		return (EINVAL);
1980 
1981 	if (gic_get_hw_rev(dev) < 3)
1982 		return (EINVAL);
1983 
1984 	device_set_desc(dev, "ARM GIC Interrupt Translation Service");
1985 	return (BUS_PROBE_DEFAULT);
1986 }
1987 
1988 static int
1989 gicv3_its_acpi_attach(device_t dev)
1990 {
1991 	struct gicv3_its_softc *sc;
1992 	struct gic_v3_devinfo *di;
1993 	int err;
1994 
1995 	sc = device_get_softc(dev);
1996 	sc->dev = dev;
1997 	err = gicv3_its_attach(dev);
1998 	if (err != 0)
1999 		return (err);
2000 
2001 	di = device_get_ivars(dev);
2002 	sc->sc_pic = intr_pic_register(dev, di->msi_xref);
2003 	intr_pic_add_handler(device_get_parent(dev), sc->sc_pic,
2004 	    gicv3_its_intr, sc, sc->sc_irq_base, sc->sc_irq_length);
2005 
2006 	/* Register this device to handle MSI interrupts */
2007 	intr_msi_register(dev, di->msi_xref);
2008 
2009 	return (0);
2010 }
2011 #endif
2012