xref: /freebsd/sys/arm64/arm64/gicv3_its.c (revision 4b9d6057)
1 /*-
2  * Copyright (c) 2015-2016 The FreeBSD Foundation
3  * Copyright (c) 2023 Arm Ltd
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/param.h>
38 #include <sys/systm.h>
39 #include <sys/bus.h>
40 #include <sys/cpuset.h>
41 #include <sys/domainset.h>
42 #include <sys/endian.h>
43 #include <sys/kernel.h>
44 #include <sys/lock.h>
45 #include <sys/malloc.h>
46 #include <sys/module.h>
47 #include <sys/mutex.h>
48 #include <sys/proc.h>
49 #include <sys/taskqueue.h>
50 #include <sys/tree.h>
51 #include <sys/queue.h>
52 #include <sys/rman.h>
53 #include <sys/sbuf.h>
54 #include <sys/smp.h>
55 #include <sys/sysctl.h>
56 #include <sys/vmem.h>
57 
58 #include <vm/vm.h>
59 #include <vm/pmap.h>
60 #include <vm/vm_page.h>
61 
62 #include <machine/bus.h>
63 #include <machine/intr.h>
64 
65 #include <arm/arm/gic_common.h>
66 #include <arm64/arm64/gic_v3_reg.h>
67 #include <arm64/arm64/gic_v3_var.h>
68 
69 #ifdef FDT
70 #include <dev/ofw/openfirm.h>
71 #include <dev/ofw/ofw_bus.h>
72 #include <dev/ofw/ofw_bus_subr.h>
73 #endif
74 #include <dev/pci/pcireg.h>
75 #include <dev/pci/pcivar.h>
76 
77 #ifdef IOMMU
78 #include <dev/iommu/iommu.h>
79 #include <dev/iommu/iommu_gas.h>
80 #endif
81 
82 #include "pcib_if.h"
83 #include "pic_if.h"
84 #include "msi_if.h"
85 
86 MALLOC_DEFINE(M_GICV3_ITS, "GICv3 ITS",
87     "ARM GICv3 Interrupt Translation Service");
88 
89 #define	LPI_NIRQS		(64 * 1024)
90 
91 /* The size and alignment of the command circular buffer */
92 #define	ITS_CMDQ_SIZE		(64 * 1024)	/* Must be a multiple of 4K */
93 #define	ITS_CMDQ_ALIGN		(64 * 1024)
94 
95 #define	LPI_CONFTAB_SIZE	LPI_NIRQS
96 #define	LPI_CONFTAB_ALIGN	(64 * 1024)
97 #define	LPI_CONFTAB_MAX_ADDR	((1ul << 48) - 1) /* We need a 47 bit PA */
98 
99 /* 1 bit per SPI, PPI, and SGI (8k), and 1 bit per LPI (LPI_CONFTAB_SIZE) */
100 #define	LPI_PENDTAB_SIZE	((LPI_NIRQS + GIC_FIRST_LPI) / 8)
101 #define	LPI_PENDTAB_ALIGN	(64 * 1024)
102 #define	LPI_PENDTAB_MAX_ADDR	((1ul << 48) - 1) /* We need a 47 bit PA */
103 
104 #define	LPI_INT_TRANS_TAB_ALIGN	256
105 #define	LPI_INT_TRANS_TAB_MAX_ADDR ((1ul << 48) - 1)
106 
107 /* ITS commands encoding */
108 #define	ITS_CMD_MOVI		(0x01)
109 #define	ITS_CMD_SYNC		(0x05)
110 #define	ITS_CMD_MAPD		(0x08)
111 #define	ITS_CMD_MAPC		(0x09)
112 #define	ITS_CMD_MAPTI		(0x0a)
113 #define	ITS_CMD_MAPI		(0x0b)
114 #define	ITS_CMD_INV		(0x0c)
115 #define	ITS_CMD_INVALL		(0x0d)
116 /* Command */
117 #define	CMD_COMMAND_MASK	(0xFFUL)
118 /* PCI device ID */
119 #define	CMD_DEVID_SHIFT		(32)
120 #define	CMD_DEVID_MASK		(0xFFFFFFFFUL << CMD_DEVID_SHIFT)
121 /* Size of IRQ ID bitfield */
122 #define	CMD_SIZE_MASK		(0xFFUL)
123 /* Virtual LPI ID */
124 #define	CMD_ID_MASK		(0xFFFFFFFFUL)
125 /* Physical LPI ID */
126 #define	CMD_PID_SHIFT		(32)
127 #define	CMD_PID_MASK		(0xFFFFFFFFUL << CMD_PID_SHIFT)
128 /* Collection */
129 #define	CMD_COL_MASK		(0xFFFFUL)
130 /* Target (CPU or Re-Distributor) */
131 #define	CMD_TARGET_SHIFT	(16)
132 #define	CMD_TARGET_MASK		(0xFFFFFFFFUL << CMD_TARGET_SHIFT)
133 /* Interrupt Translation Table address */
134 #define	CMD_ITT_MASK		(0xFFFFFFFFFF00UL)
135 /* Valid command bit */
136 #define	CMD_VALID_SHIFT		(63)
137 #define	CMD_VALID_MASK		(1UL << CMD_VALID_SHIFT)
138 
139 #define	ITS_TARGET_NONE		0xFBADBEEF
140 
141 /* LPI chunk owned by ITS device */
142 struct lpi_chunk {
143 	u_int	lpi_base;
144 	u_int	lpi_free;	/* First free LPI in set */
145 	u_int	lpi_num;	/* Total number of LPIs in chunk */
146 	u_int	lpi_busy;	/* Number of busy LPIs in chink */
147 };
148 
149 /* ITS device */
150 struct its_dev {
151 	TAILQ_ENTRY(its_dev)	entry;
152 	/* PCI device */
153 	device_t		pci_dev;
154 	/* Device ID (i.e. PCI device ID) */
155 	uint32_t		devid;
156 	/* List of assigned LPIs */
157 	struct lpi_chunk	lpis;
158 	/* Virtual address of ITT */
159 	vm_offset_t		itt;
160 	size_t			itt_size;
161 };
162 
163 /*
164  * ITS command descriptor.
165  * Idea for command description passing taken from Linux.
166  */
167 struct its_cmd_desc {
168 	uint8_t cmd_type;
169 
170 	union {
171 		struct {
172 			struct its_dev *its_dev;
173 			struct its_col *col;
174 			uint32_t id;
175 		} cmd_desc_movi;
176 
177 		struct {
178 			struct its_col *col;
179 		} cmd_desc_sync;
180 
181 		struct {
182 			struct its_col *col;
183 			uint8_t valid;
184 		} cmd_desc_mapc;
185 
186 		struct {
187 			struct its_dev *its_dev;
188 			struct its_col *col;
189 			uint32_t pid;
190 			uint32_t id;
191 		} cmd_desc_mapvi;
192 
193 		struct {
194 			struct its_dev *its_dev;
195 			struct its_col *col;
196 			uint32_t pid;
197 		} cmd_desc_mapi;
198 
199 		struct {
200 			struct its_dev *its_dev;
201 			uint8_t valid;
202 		} cmd_desc_mapd;
203 
204 		struct {
205 			struct its_dev *its_dev;
206 			struct its_col *col;
207 			uint32_t pid;
208 		} cmd_desc_inv;
209 
210 		struct {
211 			struct its_col *col;
212 		} cmd_desc_invall;
213 	};
214 };
215 
216 /* ITS command. Each command is 32 bytes long */
217 struct its_cmd {
218 	uint64_t	cmd_dword[4];	/* ITS command double word */
219 };
220 
221 /* An ITS private table */
222 struct its_ptable {
223 	vm_offset_t	ptab_vaddr;
224 	/* Size of the L1 and L2 tables */
225 	size_t		ptab_l1_size;
226 	size_t		ptab_l2_size;
227 	/* Number of L1 and L2 entries */
228 	int		ptab_l1_nidents;
229 	int		ptab_l2_nidents;
230 
231 	int		ptab_page_size;
232 	int		ptab_share;
233 	bool		ptab_indirect;
234 };
235 
236 /* ITS collection description. */
237 struct its_col {
238 	uint64_t	col_target;	/* Target Re-Distributor */
239 	uint64_t	col_id;		/* Collection ID */
240 };
241 
242 struct gicv3_its_irqsrc {
243 	struct intr_irqsrc	gi_isrc;
244 	u_int			gi_id;
245 	u_int			gi_lpi;
246 	struct its_dev		*gi_its_dev;
247 	TAILQ_ENTRY(gicv3_its_irqsrc) gi_link;
248 };
249 
250 struct gicv3_its_softc {
251 	device_t	dev;
252 	struct intr_pic *sc_pic;
253 	struct resource *sc_its_res;
254 
255 	cpuset_t	sc_cpus;
256 	struct domainset *sc_ds;
257 	u_int		gic_irq_cpu;
258 	int		sc_devbits;
259 	int		sc_dev_table_idx;
260 
261 	struct its_ptable sc_its_ptab[GITS_BASER_NUM];
262 	struct its_col *sc_its_cols[MAXCPU];	/* Per-CPU collections */
263 
264 	/*
265 	 * TODO: We should get these from the parent as we only want a
266 	 * single copy of each across the interrupt controller.
267 	 */
268 	uint8_t		*sc_conf_base;
269 	vm_offset_t sc_pend_base[MAXCPU];
270 
271 	/* Command handling */
272 	struct mtx sc_its_cmd_lock;
273 	struct its_cmd *sc_its_cmd_base; /* Command circular buffer address */
274 	size_t sc_its_cmd_next_idx;
275 
276 	vmem_t *sc_irq_alloc;
277 	struct gicv3_its_irqsrc	**sc_irqs;
278 	u_int	sc_irq_base;
279 	u_int	sc_irq_length;
280 	u_int	sc_irq_count;
281 
282 	struct mtx sc_its_dev_lock;
283 	TAILQ_HEAD(its_dev_list, its_dev) sc_its_dev_list;
284 	TAILQ_HEAD(free_irqs, gicv3_its_irqsrc) sc_free_irqs;
285 
286 #define	ITS_FLAGS_CMDQ_FLUSH		0x00000001
287 #define	ITS_FLAGS_LPI_CONF_FLUSH	0x00000002
288 #define	ITS_FLAGS_ERRATA_CAVIUM_22375	0x00000004
289 	u_int sc_its_flags;
290 	bool	trace_enable;
291 	vm_page_t ma; /* fake msi page */
292 };
293 
294 static void *conf_base;
295 
296 typedef void (its_quirk_func_t)(device_t);
297 static its_quirk_func_t its_quirk_cavium_22375;
298 
299 static const struct {
300 	const char *desc;
301 	uint32_t iidr;
302 	uint32_t iidr_mask;
303 	its_quirk_func_t *func;
304 } its_quirks[] = {
305 	{
306 		/* Cavium ThunderX Pass 1.x */
307 		.desc = "Cavium ThunderX errata: 22375, 24313",
308 		.iidr = GITS_IIDR_RAW(GITS_IIDR_IMPL_CAVIUM,
309 		    GITS_IIDR_PROD_THUNDER, GITS_IIDR_VAR_THUNDER_1, 0),
310 		.iidr_mask = ~GITS_IIDR_REVISION_MASK,
311 		.func = its_quirk_cavium_22375,
312 	},
313 };
314 
315 #define	gic_its_read_4(sc, reg)			\
316     bus_read_4((sc)->sc_its_res, (reg))
317 #define	gic_its_read_8(sc, reg)			\
318     bus_read_8((sc)->sc_its_res, (reg))
319 
320 #define	gic_its_write_4(sc, reg, val)		\
321     bus_write_4((sc)->sc_its_res, (reg), (val))
322 #define	gic_its_write_8(sc, reg, val)		\
323     bus_write_8((sc)->sc_its_res, (reg), (val))
324 
325 static device_attach_t gicv3_its_attach;
326 static device_detach_t gicv3_its_detach;
327 
328 static pic_disable_intr_t gicv3_its_disable_intr;
329 static pic_enable_intr_t gicv3_its_enable_intr;
330 static pic_map_intr_t gicv3_its_map_intr;
331 static pic_setup_intr_t gicv3_its_setup_intr;
332 static pic_post_filter_t gicv3_its_post_filter;
333 static pic_post_ithread_t gicv3_its_post_ithread;
334 static pic_pre_ithread_t gicv3_its_pre_ithread;
335 static pic_bind_intr_t gicv3_its_bind_intr;
336 #ifdef SMP
337 static pic_init_secondary_t gicv3_its_init_secondary;
338 #endif
339 static msi_alloc_msi_t gicv3_its_alloc_msi;
340 static msi_release_msi_t gicv3_its_release_msi;
341 static msi_alloc_msix_t gicv3_its_alloc_msix;
342 static msi_release_msix_t gicv3_its_release_msix;
343 static msi_map_msi_t gicv3_its_map_msi;
344 #ifdef IOMMU
345 static msi_iommu_init_t gicv3_iommu_init;
346 static msi_iommu_deinit_t gicv3_iommu_deinit;
347 #endif
348 
349 static void its_cmd_movi(device_t, struct gicv3_its_irqsrc *);
350 static void its_cmd_mapc(device_t, struct its_col *, uint8_t);
351 static void its_cmd_mapti(device_t, struct gicv3_its_irqsrc *);
352 static void its_cmd_mapd(device_t, struct its_dev *, uint8_t);
353 static void its_cmd_inv(device_t, struct its_dev *, struct gicv3_its_irqsrc *);
354 static void its_cmd_invall(device_t, struct its_col *);
355 
356 static device_method_t gicv3_its_methods[] = {
357 	/* Device interface */
358 	DEVMETHOD(device_detach,	gicv3_its_detach),
359 
360 	/* Interrupt controller interface */
361 	DEVMETHOD(pic_disable_intr,	gicv3_its_disable_intr),
362 	DEVMETHOD(pic_enable_intr,	gicv3_its_enable_intr),
363 	DEVMETHOD(pic_map_intr,		gicv3_its_map_intr),
364 	DEVMETHOD(pic_setup_intr,	gicv3_its_setup_intr),
365 	DEVMETHOD(pic_post_filter,	gicv3_its_post_filter),
366 	DEVMETHOD(pic_post_ithread,	gicv3_its_post_ithread),
367 	DEVMETHOD(pic_pre_ithread,	gicv3_its_pre_ithread),
368 #ifdef SMP
369 	DEVMETHOD(pic_bind_intr,	gicv3_its_bind_intr),
370 	DEVMETHOD(pic_init_secondary,	gicv3_its_init_secondary),
371 #endif
372 
373 	/* MSI/MSI-X */
374 	DEVMETHOD(msi_alloc_msi,	gicv3_its_alloc_msi),
375 	DEVMETHOD(msi_release_msi,	gicv3_its_release_msi),
376 	DEVMETHOD(msi_alloc_msix,	gicv3_its_alloc_msix),
377 	DEVMETHOD(msi_release_msix,	gicv3_its_release_msix),
378 	DEVMETHOD(msi_map_msi,		gicv3_its_map_msi),
379 #ifdef IOMMU
380 	DEVMETHOD(msi_iommu_init,	gicv3_iommu_init),
381 	DEVMETHOD(msi_iommu_deinit,	gicv3_iommu_deinit),
382 #endif
383 
384 	/* End */
385 	DEVMETHOD_END
386 };
387 
388 static DEFINE_CLASS_0(gic, gicv3_its_driver, gicv3_its_methods,
389     sizeof(struct gicv3_its_softc));
390 
391 static void
392 gicv3_its_cmdq_init(struct gicv3_its_softc *sc)
393 {
394 	vm_paddr_t cmd_paddr;
395 	uint64_t reg, tmp;
396 
397 	/* Set up the command circular buffer */
398 	sc->sc_its_cmd_base = contigmalloc_domainset(ITS_CMDQ_SIZE, M_GICV3_ITS,
399 	    sc->sc_ds, M_WAITOK | M_ZERO, 0, (1ul << 48) - 1, ITS_CMDQ_ALIGN,
400 	    0);
401 	sc->sc_its_cmd_next_idx = 0;
402 
403 	cmd_paddr = vtophys(sc->sc_its_cmd_base);
404 
405 	/* Set the base of the command buffer */
406 	reg = GITS_CBASER_VALID |
407 	    (GITS_CBASER_CACHE_NIWAWB << GITS_CBASER_CACHE_SHIFT) |
408 	    cmd_paddr | (GITS_CBASER_SHARE_IS << GITS_CBASER_SHARE_SHIFT) |
409 	    (ITS_CMDQ_SIZE / 4096 - 1);
410 	gic_its_write_8(sc, GITS_CBASER, reg);
411 
412 	/* Read back to check for fixed value fields */
413 	tmp = gic_its_read_8(sc, GITS_CBASER);
414 
415 	if ((tmp & GITS_CBASER_SHARE_MASK) !=
416 	    (GITS_CBASER_SHARE_IS << GITS_CBASER_SHARE_SHIFT)) {
417 		/* Check if the hardware reported non-shareable */
418 		if ((tmp & GITS_CBASER_SHARE_MASK) ==
419 		    (GITS_CBASER_SHARE_NS << GITS_CBASER_SHARE_SHIFT)) {
420 			/* If so remove the cache attribute */
421 			reg &= ~GITS_CBASER_CACHE_MASK;
422 			reg &= ~GITS_CBASER_SHARE_MASK;
423 			/* Set to Non-cacheable, Non-shareable */
424 			reg |= GITS_CBASER_CACHE_NIN << GITS_CBASER_CACHE_SHIFT;
425 			reg |= GITS_CBASER_SHARE_NS << GITS_CBASER_SHARE_SHIFT;
426 
427 			gic_its_write_8(sc, GITS_CBASER, reg);
428 		}
429 
430 		/* The command queue has to be flushed after each command */
431 		sc->sc_its_flags |= ITS_FLAGS_CMDQ_FLUSH;
432 	}
433 
434 	/* Get the next command from the start of the buffer */
435 	gic_its_write_8(sc, GITS_CWRITER, 0x0);
436 }
437 
438 static int
439 gicv3_its_table_page_size(struct gicv3_its_softc *sc, int table)
440 {
441 	uint64_t reg, tmp;
442 	int page_size;
443 
444 	page_size = PAGE_SIZE_64K;
445 	reg = gic_its_read_8(sc, GITS_BASER(table));
446 
447 	while (1) {
448 		reg &= GITS_BASER_PSZ_MASK;
449 		switch (page_size) {
450 		case PAGE_SIZE_4K:	/* 4KB */
451 			reg |= GITS_BASER_PSZ_4K << GITS_BASER_PSZ_SHIFT;
452 			break;
453 		case PAGE_SIZE_16K:	/* 16KB */
454 			reg |= GITS_BASER_PSZ_16K << GITS_BASER_PSZ_SHIFT;
455 			break;
456 		case PAGE_SIZE_64K:	/* 64KB */
457 			reg |= GITS_BASER_PSZ_64K << GITS_BASER_PSZ_SHIFT;
458 			break;
459 		}
460 
461 		/* Write the new page size */
462 		gic_its_write_8(sc, GITS_BASER(table), reg);
463 
464 		/* Read back to check */
465 		tmp = gic_its_read_8(sc, GITS_BASER(table));
466 
467 		/* The page size is correct */
468 		if ((tmp & GITS_BASER_PSZ_MASK) == (reg & GITS_BASER_PSZ_MASK))
469 			return (page_size);
470 
471 		switch (page_size) {
472 		default:
473 			return (-1);
474 		case PAGE_SIZE_16K:
475 			page_size = PAGE_SIZE_4K;
476 			break;
477 		case PAGE_SIZE_64K:
478 			page_size = PAGE_SIZE_16K;
479 			break;
480 		}
481 	}
482 }
483 
484 static bool
485 gicv3_its_table_supports_indirect(struct gicv3_its_softc *sc, int table)
486 {
487 	uint64_t reg;
488 
489 	reg = gic_its_read_8(sc, GITS_BASER(table));
490 
491 	/* Try setting the indirect flag */
492 	reg |= GITS_BASER_INDIRECT;
493 	gic_its_write_8(sc, GITS_BASER(table), reg);
494 
495 	/* Read back to check */
496 	reg = gic_its_read_8(sc, GITS_BASER(table));
497 	return ((reg & GITS_BASER_INDIRECT) != 0);
498 }
499 
500 
501 static int
502 gicv3_its_table_init(device_t dev, struct gicv3_its_softc *sc)
503 {
504 	vm_offset_t table;
505 	vm_paddr_t paddr;
506 	uint64_t cache, reg, share, tmp, type;
507 	size_t its_tbl_size, nitspages, npages;
508 	size_t l1_esize, l2_esize, l1_nidents, l2_nidents;
509 	int i, page_size;
510 	int devbits;
511 	bool indirect;
512 
513 	if ((sc->sc_its_flags & ITS_FLAGS_ERRATA_CAVIUM_22375) != 0) {
514 		/*
515 		 * GITS_TYPER[17:13] of ThunderX reports that device IDs
516 		 * are to be 21 bits in length. The entry size of the ITS
517 		 * table can be read from GITS_BASERn[52:48] and on ThunderX
518 		 * is supposed to be 8 bytes in length (for device table).
519 		 * Finally the page size that is to be used by ITS to access
520 		 * this table will be set to 64KB.
521 		 *
522 		 * This gives 0x200000 entries of size 0x8 bytes covered by
523 		 * 256 pages each of which 64KB in size. The number of pages
524 		 * (minus 1) should then be written to GITS_BASERn[7:0]. In
525 		 * that case this value would be 0xFF but on ThunderX the
526 		 * maximum value that HW accepts is 0xFD.
527 		 *
528 		 * Set an arbitrary number of device ID bits to 20 in order
529 		 * to limit the number of entries in ITS device table to
530 		 * 0x100000 and the table size to 8MB.
531 		 */
532 		devbits = 20;
533 		cache = 0;
534 	} else {
535 		devbits = GITS_TYPER_DEVB(gic_its_read_8(sc, GITS_TYPER));
536 		cache = GITS_BASER_CACHE_WAWB;
537 	}
538 	sc->sc_devbits = devbits;
539 	share = GITS_BASER_SHARE_IS;
540 
541 	for (i = 0; i < GITS_BASER_NUM; i++) {
542 		reg = gic_its_read_8(sc, GITS_BASER(i));
543 		/* The type of table */
544 		type = GITS_BASER_TYPE(reg);
545 		if (type == GITS_BASER_TYPE_UNIMPL)
546 			continue;
547 
548 		/* The table entry size */
549 		l1_esize = GITS_BASER_ESIZE(reg);
550 
551 		/* Find the tables page size */
552 		page_size = gicv3_its_table_page_size(sc, i);
553 		if (page_size == -1) {
554 			device_printf(dev, "No valid page size for table %d\n",
555 			    i);
556 			return (EINVAL);
557 		}
558 
559 		indirect = false;
560 		l2_nidents = 0;
561 		l2_esize = 0;
562 		switch(type) {
563 		case GITS_BASER_TYPE_DEV:
564 			if (sc->sc_dev_table_idx != -1)
565 				device_printf(dev,
566 				    "Warning: Multiple device tables found\n");
567 
568 			sc->sc_dev_table_idx = i;
569 			l1_nidents = (1 << devbits);
570 			if ((l1_esize * l1_nidents) > (page_size * 2)) {
571 				indirect =
572 				    gicv3_its_table_supports_indirect(sc, i);
573 				if (indirect) {
574 					/*
575 					 * Each l1 entry is 8 bytes and points
576 					 * to an l2 table of size page_size.
577 					 * Calculate how many entries this is
578 					 * and use this to find how many
579 					 * 8 byte l1 idents we need.
580 					 */
581 					l2_esize = l1_esize;
582 					l2_nidents = page_size / l2_esize;
583 					l1_nidents = l1_nidents / l2_nidents;
584 					l1_esize = GITS_INDIRECT_L1_ESIZE;
585 				}
586 			}
587 			its_tbl_size = l1_esize * l1_nidents;
588 			its_tbl_size = roundup2(its_tbl_size, page_size);
589 			break;
590 		case GITS_BASER_TYPE_VP:
591 		case GITS_BASER_TYPE_PP: /* Undocumented? */
592 		case GITS_BASER_TYPE_IC:
593 			its_tbl_size = page_size;
594 			break;
595 		default:
596 			if (bootverbose)
597 				device_printf(dev, "Unhandled table type %lx\n",
598 				    type);
599 			continue;
600 		}
601 		npages = howmany(its_tbl_size, PAGE_SIZE);
602 
603 		/* Allocate the table */
604 		table = (vm_offset_t)contigmalloc_domainset(npages * PAGE_SIZE,
605 		    M_GICV3_ITS, sc->sc_ds, M_WAITOK | M_ZERO, 0,
606 		    (1ul << 48) - 1, PAGE_SIZE_64K, 0);
607 
608 		sc->sc_its_ptab[i].ptab_vaddr = table;
609 		sc->sc_its_ptab[i].ptab_l1_size = its_tbl_size;
610 		sc->sc_its_ptab[i].ptab_l1_nidents = l1_nidents;
611 		sc->sc_its_ptab[i].ptab_l2_size = page_size;
612 		sc->sc_its_ptab[i].ptab_l2_nidents = l2_nidents;
613 
614 		sc->sc_its_ptab[i].ptab_indirect = indirect;
615 		sc->sc_its_ptab[i].ptab_page_size = page_size;
616 
617 		paddr = vtophys(table);
618 
619 		while (1) {
620 			nitspages = howmany(its_tbl_size, page_size);
621 
622 			/* Clear the fields we will be setting */
623 			reg &= ~(GITS_BASER_VALID | GITS_BASER_INDIRECT |
624 			    GITS_BASER_CACHE_MASK | GITS_BASER_TYPE_MASK |
625 			    GITS_BASER_PA_MASK |
626 			    GITS_BASER_SHARE_MASK | GITS_BASER_PSZ_MASK |
627 			    GITS_BASER_SIZE_MASK);
628 			/* Set the new values */
629 			reg |= GITS_BASER_VALID |
630 			    (indirect ? GITS_BASER_INDIRECT : 0) |
631 			    (cache << GITS_BASER_CACHE_SHIFT) |
632 			    (type << GITS_BASER_TYPE_SHIFT) |
633 			    paddr | (share << GITS_BASER_SHARE_SHIFT) |
634 			    (nitspages - 1);
635 
636 			switch (page_size) {
637 			case PAGE_SIZE_4K:	/* 4KB */
638 				reg |=
639 				    GITS_BASER_PSZ_4K << GITS_BASER_PSZ_SHIFT;
640 				break;
641 			case PAGE_SIZE_16K:	/* 16KB */
642 				reg |=
643 				    GITS_BASER_PSZ_16K << GITS_BASER_PSZ_SHIFT;
644 				break;
645 			case PAGE_SIZE_64K:	/* 64KB */
646 				reg |=
647 				    GITS_BASER_PSZ_64K << GITS_BASER_PSZ_SHIFT;
648 				break;
649 			}
650 
651 			gic_its_write_8(sc, GITS_BASER(i), reg);
652 
653 			/* Read back to check */
654 			tmp = gic_its_read_8(sc, GITS_BASER(i));
655 
656 			/* Do the shareability masks line up? */
657 			if ((tmp & GITS_BASER_SHARE_MASK) !=
658 			    (reg & GITS_BASER_SHARE_MASK)) {
659 				share = (tmp & GITS_BASER_SHARE_MASK) >>
660 				    GITS_BASER_SHARE_SHIFT;
661 				continue;
662 			}
663 
664 			if (tmp != reg) {
665 				device_printf(dev, "GITS_BASER%d: "
666 				    "unable to be updated: %lx != %lx\n",
667 				    i, reg, tmp);
668 				return (ENXIO);
669 			}
670 
671 			sc->sc_its_ptab[i].ptab_share = share;
672 			/* We should have made all needed changes */
673 			break;
674 		}
675 	}
676 
677 	return (0);
678 }
679 
680 static void
681 gicv3_its_conftable_init(struct gicv3_its_softc *sc)
682 {
683 	void *conf_table;
684 
685 	conf_table = atomic_load_ptr(&conf_base);
686 	if (conf_table == NULL) {
687 		conf_table = contigmalloc(LPI_CONFTAB_SIZE,
688 		    M_GICV3_ITS, M_WAITOK, 0, LPI_CONFTAB_MAX_ADDR,
689 		    LPI_CONFTAB_ALIGN, 0);
690 
691 		if (atomic_cmpset_ptr((uintptr_t *)&conf_base,
692 		    (uintptr_t)NULL, (uintptr_t)conf_table) == 0) {
693 			contigfree(conf_table, LPI_CONFTAB_SIZE, M_GICV3_ITS);
694 			conf_table = atomic_load_ptr(&conf_base);
695 		}
696 	}
697 	sc->sc_conf_base = conf_table;
698 
699 	/* Set the default configuration */
700 	memset(sc->sc_conf_base, GIC_PRIORITY_MAX | LPI_CONF_GROUP1,
701 	    LPI_CONFTAB_SIZE);
702 
703 	/* Flush the table to memory */
704 	cpu_dcache_wb_range((vm_offset_t)sc->sc_conf_base, LPI_CONFTAB_SIZE);
705 }
706 
707 static void
708 gicv3_its_pendtables_init(struct gicv3_its_softc *sc)
709 {
710 	int i;
711 
712 	for (i = 0; i <= mp_maxid; i++) {
713 		if (CPU_ISSET(i, &sc->sc_cpus) == 0)
714 			continue;
715 
716 		sc->sc_pend_base[i] = (vm_offset_t)contigmalloc(
717 		    LPI_PENDTAB_SIZE, M_GICV3_ITS, M_WAITOK | M_ZERO,
718 		    0, LPI_PENDTAB_MAX_ADDR, LPI_PENDTAB_ALIGN, 0);
719 
720 		/* Flush so the ITS can see the memory */
721 		cpu_dcache_wb_range((vm_offset_t)sc->sc_pend_base[i],
722 		    LPI_PENDTAB_SIZE);
723 	}
724 }
725 
726 static void
727 its_init_cpu_lpi(device_t dev, struct gicv3_its_softc *sc)
728 {
729 	device_t gicv3;
730 	uint64_t xbaser, tmp;
731 	uint32_t ctlr;
732 	u_int cpuid;
733 
734 	gicv3 = device_get_parent(dev);
735 	cpuid = PCPU_GET(cpuid);
736 
737 	/* Disable LPIs */
738 	ctlr = gic_r_read_4(gicv3, GICR_CTLR);
739 	ctlr &= ~GICR_CTLR_LPI_ENABLE;
740 	gic_r_write_4(gicv3, GICR_CTLR, ctlr);
741 
742 	/* Make sure changes are observable my the GIC */
743 	dsb(sy);
744 
745 	/*
746 	 * Set the redistributor base
747 	 */
748 	xbaser = vtophys(sc->sc_conf_base) |
749 	    (GICR_PROPBASER_SHARE_IS << GICR_PROPBASER_SHARE_SHIFT) |
750 	    (GICR_PROPBASER_CACHE_NIWAWB << GICR_PROPBASER_CACHE_SHIFT) |
751 	    (flsl(LPI_CONFTAB_SIZE | GIC_FIRST_LPI) - 1);
752 	gic_r_write_8(gicv3, GICR_PROPBASER, xbaser);
753 
754 	/* Check the cache attributes we set */
755 	tmp = gic_r_read_8(gicv3, GICR_PROPBASER);
756 
757 	if ((tmp & GICR_PROPBASER_SHARE_MASK) !=
758 	    (xbaser & GICR_PROPBASER_SHARE_MASK)) {
759 		if ((tmp & GICR_PROPBASER_SHARE_MASK) ==
760 		    (GICR_PROPBASER_SHARE_NS << GICR_PROPBASER_SHARE_SHIFT)) {
761 			/* We need to mark as non-cacheable */
762 			xbaser &= ~(GICR_PROPBASER_SHARE_MASK |
763 			    GICR_PROPBASER_CACHE_MASK);
764 			/* Non-cacheable */
765 			xbaser |= GICR_PROPBASER_CACHE_NIN <<
766 			    GICR_PROPBASER_CACHE_SHIFT;
767 			/* Non-shareable */
768 			xbaser |= GICR_PROPBASER_SHARE_NS <<
769 			    GICR_PROPBASER_SHARE_SHIFT;
770 			gic_r_write_8(gicv3, GICR_PROPBASER, xbaser);
771 		}
772 		sc->sc_its_flags |= ITS_FLAGS_LPI_CONF_FLUSH;
773 	}
774 
775 	/*
776 	 * Set the LPI pending table base
777 	 */
778 	xbaser = vtophys(sc->sc_pend_base[cpuid]) |
779 	    (GICR_PENDBASER_CACHE_NIWAWB << GICR_PENDBASER_CACHE_SHIFT) |
780 	    (GICR_PENDBASER_SHARE_IS << GICR_PENDBASER_SHARE_SHIFT);
781 
782 	gic_r_write_8(gicv3, GICR_PENDBASER, xbaser);
783 
784 	tmp = gic_r_read_8(gicv3, GICR_PENDBASER);
785 
786 	if ((tmp & GICR_PENDBASER_SHARE_MASK) ==
787 	    (GICR_PENDBASER_SHARE_NS << GICR_PENDBASER_SHARE_SHIFT)) {
788 		/* Clear the cahce and shareability bits */
789 		xbaser &= ~(GICR_PENDBASER_CACHE_MASK |
790 		    GICR_PENDBASER_SHARE_MASK);
791 		/* Mark as non-shareable */
792 		xbaser |= GICR_PENDBASER_SHARE_NS << GICR_PENDBASER_SHARE_SHIFT;
793 		/* And non-cacheable */
794 		xbaser |= GICR_PENDBASER_CACHE_NIN <<
795 		    GICR_PENDBASER_CACHE_SHIFT;
796 	}
797 
798 	/* Enable LPIs */
799 	ctlr = gic_r_read_4(gicv3, GICR_CTLR);
800 	ctlr |= GICR_CTLR_LPI_ENABLE;
801 	gic_r_write_4(gicv3, GICR_CTLR, ctlr);
802 
803 	/* Make sure the GIC has seen everything */
804 	dsb(sy);
805 }
806 
807 static int
808 its_init_cpu(device_t dev, struct gicv3_its_softc *sc)
809 {
810 	device_t gicv3;
811 	vm_paddr_t target;
812 	u_int cpuid;
813 	struct redist_pcpu *rpcpu;
814 
815 	gicv3 = device_get_parent(dev);
816 	cpuid = PCPU_GET(cpuid);
817 	if (!CPU_ISSET(cpuid, &sc->sc_cpus))
818 		return (0);
819 
820 	/* Check if the ITS is enabled on this CPU */
821 	if ((gic_r_read_8(gicv3, GICR_TYPER) & GICR_TYPER_PLPIS) == 0)
822 		return (ENXIO);
823 
824 	rpcpu = gicv3_get_redist(dev);
825 
826 	/* Do per-cpu LPI init once */
827 	if (!rpcpu->lpi_enabled) {
828 		its_init_cpu_lpi(dev, sc);
829 		rpcpu->lpi_enabled = true;
830 	}
831 
832 	if ((gic_its_read_8(sc, GITS_TYPER) & GITS_TYPER_PTA) != 0) {
833 		/* This ITS wants the redistributor physical address */
834 		target = vtophys((vm_offset_t)rman_get_virtual(rpcpu->res) +
835 		    rpcpu->offset);
836 	} else {
837 		/* This ITS wants the unique processor number */
838 		target = GICR_TYPER_CPUNUM(gic_r_read_8(gicv3, GICR_TYPER)) <<
839 		    CMD_TARGET_SHIFT;
840 	}
841 
842 	sc->sc_its_cols[cpuid]->col_target = target;
843 	sc->sc_its_cols[cpuid]->col_id = cpuid;
844 
845 	its_cmd_mapc(dev, sc->sc_its_cols[cpuid], 1);
846 	its_cmd_invall(dev, sc->sc_its_cols[cpuid]);
847 
848 	return (0);
849 }
850 
851 static int
852 gicv3_its_sysctl_trace_enable(SYSCTL_HANDLER_ARGS)
853 {
854 	struct gicv3_its_softc *sc;
855 	int rv;
856 
857 	sc = arg1;
858 
859 	rv = sysctl_handle_bool(oidp, &sc->trace_enable, 0, req);
860 	if (rv != 0 || req->newptr == NULL)
861 		return (rv);
862 	if (sc->trace_enable)
863 		gic_its_write_8(sc, GITS_TRKCTLR, 3);
864 	else
865 		gic_its_write_8(sc, GITS_TRKCTLR, 0);
866 
867 	return (0);
868 }
869 
870 static int
871 gicv3_its_sysctl_trace_regs(SYSCTL_HANDLER_ARGS)
872 {
873 	struct gicv3_its_softc *sc;
874 	struct sbuf *sb;
875 	int err;
876 
877 	sc = arg1;
878 	sb = sbuf_new_for_sysctl(NULL, NULL, 128, req);
879 	if (sb == NULL) {
880 		device_printf(sc->dev, "Could not allocate sbuf for output.\n");
881 		return (ENOMEM);
882 	}
883 	sbuf_cat(sb, "\n");
884 	sbuf_printf(sb, "GITS_TRKCTLR: 0x%08X\n",
885 	    gic_its_read_4(sc, GITS_TRKCTLR));
886 	sbuf_printf(sb, "GITS_TRKR:    0x%08X\n",
887 	    gic_its_read_4(sc, GITS_TRKR));
888 	sbuf_printf(sb, "GITS_TRKDIDR: 0x%08X\n",
889 	    gic_its_read_4(sc, GITS_TRKDIDR));
890 	sbuf_printf(sb, "GITS_TRKPIDR: 0x%08X\n",
891 	    gic_its_read_4(sc, GITS_TRKPIDR));
892 	sbuf_printf(sb, "GITS_TRKVIDR: 0x%08X\n",
893 	    gic_its_read_4(sc, GITS_TRKVIDR));
894 	sbuf_printf(sb, "GITS_TRKTGTR: 0x%08X\n",
895 	   gic_its_read_4(sc, GITS_TRKTGTR));
896 
897 	err = sbuf_finish(sb);
898 	if (err)
899 		device_printf(sc->dev, "Error finishing sbuf: %d\n", err);
900 	sbuf_delete(sb);
901 	return(err);
902 }
903 
904 static int
905 gicv3_its_init_sysctl(struct gicv3_its_softc *sc)
906 {
907 	struct sysctl_oid *oid, *child;
908 	struct sysctl_ctx_list *ctx_list;
909 
910 	ctx_list = device_get_sysctl_ctx(sc->dev);
911 	child = device_get_sysctl_tree(sc->dev);
912 	oid = SYSCTL_ADD_NODE(ctx_list,
913 	    SYSCTL_CHILDREN(child), OID_AUTO, "tracing",
914 	    CTLFLAG_RD| CTLFLAG_MPSAFE, NULL, "Messages tracing");
915 	if (oid == NULL)
916 		return (ENXIO);
917 
918 	/* Add registers */
919 	SYSCTL_ADD_PROC(ctx_list,
920 	    SYSCTL_CHILDREN(oid), OID_AUTO, "enable",
921 	    CTLTYPE_U8 | CTLFLAG_RW | CTLFLAG_MPSAFE, sc, 0,
922 	    gicv3_its_sysctl_trace_enable, "CU", "Enable tracing");
923 	SYSCTL_ADD_PROC(ctx_list,
924 	    SYSCTL_CHILDREN(oid), OID_AUTO, "capture",
925 	    CTLTYPE_STRING | CTLFLAG_RW | CTLFLAG_MPSAFE, sc, 0,
926 	    gicv3_its_sysctl_trace_regs, "", "Captured tracing registers.");
927 
928 	return (0);
929 }
930 
931 static int
932 gicv3_its_attach(device_t dev)
933 {
934 	struct gicv3_its_softc *sc;
935 	int domain, err, i, rid;
936 	uint64_t phys;
937 	uint32_t ctlr, iidr;
938 
939 	sc = device_get_softc(dev);
940 
941 	sc->sc_dev_table_idx = -1;
942 	sc->sc_irq_length = gicv3_get_nirqs(dev);
943 	sc->sc_irq_base = GIC_FIRST_LPI;
944 	sc->sc_irq_base += device_get_unit(dev) * sc->sc_irq_length;
945 
946 	rid = 0;
947 	sc->sc_its_res = bus_alloc_resource_any(dev, SYS_RES_MEMORY, &rid,
948 	    RF_ACTIVE);
949 	if (sc->sc_its_res == NULL) {
950 		device_printf(dev, "Could not allocate memory\n");
951 		return (ENXIO);
952 	}
953 
954 	phys = rounddown2(vtophys(rman_get_virtual(sc->sc_its_res)) +
955 	    GITS_TRANSLATER, PAGE_SIZE);
956 	sc->ma = malloc(sizeof(struct vm_page), M_DEVBUF, M_WAITOK | M_ZERO);
957 	vm_page_initfake(sc->ma, phys, VM_MEMATTR_DEFAULT);
958 
959 	CPU_COPY(&all_cpus, &sc->sc_cpus);
960 	iidr = gic_its_read_4(sc, GITS_IIDR);
961 	for (i = 0; i < nitems(its_quirks); i++) {
962 		if ((iidr & its_quirks[i].iidr_mask) == its_quirks[i].iidr) {
963 			if (bootverbose) {
964 				device_printf(dev, "Applying %s\n",
965 				    its_quirks[i].desc);
966 			}
967 			its_quirks[i].func(dev);
968 			break;
969 		}
970 	}
971 
972 	if (bus_get_domain(dev, &domain) == 0 && domain < MAXMEMDOM) {
973 		sc->sc_ds = DOMAINSET_PREF(domain);
974 	} else {
975 		sc->sc_ds = DOMAINSET_RR();
976 	}
977 
978 	/*
979 	 * GIT_CTLR_EN is mandated to reset to 0 on a Warm reset, but we may be
980 	 * coming in via, for instance, a kexec/kboot style setup where a
981 	 * previous kernel has configured then relinquished control.  Clear it
982 	 * so that we can reconfigure GITS_BASER*.
983 	 */
984 	ctlr = gic_its_read_4(sc, GITS_CTLR);
985 	if ((ctlr & GITS_CTLR_EN) != 0) {
986 		ctlr &= ~GITS_CTLR_EN;
987 		gic_its_write_4(sc, GITS_CTLR, ctlr);
988 	}
989 
990 	/* Allocate the private tables */
991 	err = gicv3_its_table_init(dev, sc);
992 	if (err != 0)
993 		return (err);
994 
995 	/* Protects access to the device list */
996 	mtx_init(&sc->sc_its_dev_lock, "ITS device lock", NULL, MTX_SPIN);
997 
998 	/* Protects access to the ITS command circular buffer. */
999 	mtx_init(&sc->sc_its_cmd_lock, "ITS cmd lock", NULL, MTX_SPIN);
1000 
1001 	/* Allocate the command circular buffer */
1002 	gicv3_its_cmdq_init(sc);
1003 
1004 	/* Allocate the per-CPU collections */
1005 	for (int cpu = 0; cpu <= mp_maxid; cpu++)
1006 		if (CPU_ISSET(cpu, &sc->sc_cpus) != 0)
1007 			sc->sc_its_cols[cpu] = malloc_domainset(
1008 			    sizeof(*sc->sc_its_cols[0]), M_GICV3_ITS,
1009 			    DOMAINSET_PREF(pcpu_find(cpu)->pc_domain),
1010 			    M_WAITOK | M_ZERO);
1011 		else
1012 			sc->sc_its_cols[cpu] = NULL;
1013 
1014 	/* Enable the ITS */
1015 	gic_its_write_4(sc, GITS_CTLR, ctlr | GITS_CTLR_EN);
1016 
1017 	/* Create the LPI configuration table */
1018 	gicv3_its_conftable_init(sc);
1019 
1020 	/* And the pending tebles */
1021 	gicv3_its_pendtables_init(sc);
1022 
1023 	/* Enable LPIs on this CPU */
1024 	its_init_cpu(dev, sc);
1025 
1026 	TAILQ_INIT(&sc->sc_its_dev_list);
1027 	TAILQ_INIT(&sc->sc_free_irqs);
1028 
1029 	/*
1030 	 * Create the vmem object to allocate INTRNG IRQs from. We try to
1031 	 * use all IRQs not already used by the GICv3.
1032 	 * XXX: This assumes there are no other interrupt controllers in the
1033 	 * system.
1034 	 */
1035 	sc->sc_irq_alloc = vmem_create(device_get_nameunit(dev), 0,
1036 	    gicv3_get_nirqs(dev), 1, 0, M_FIRSTFIT | M_WAITOK);
1037 
1038 	sc->sc_irqs = malloc(sizeof(*sc->sc_irqs) * sc->sc_irq_length,
1039 	    M_GICV3_ITS, M_WAITOK | M_ZERO);
1040 
1041 	/* For GIC-500 install tracking sysctls. */
1042 	if ((iidr & (GITS_IIDR_PRODUCT_MASK | GITS_IIDR_IMPLEMENTOR_MASK)) ==
1043 	    GITS_IIDR_RAW(GITS_IIDR_IMPL_ARM, GITS_IIDR_PROD_GIC500, 0, 0))
1044 		gicv3_its_init_sysctl(sc);
1045 
1046 	return (0);
1047 }
1048 
1049 static int
1050 gicv3_its_detach(device_t dev)
1051 {
1052 
1053 	return (ENXIO);
1054 }
1055 
1056 static void
1057 its_quirk_cavium_22375(device_t dev)
1058 {
1059 	struct gicv3_its_softc *sc;
1060 	int domain;
1061 
1062 	sc = device_get_softc(dev);
1063 	sc->sc_its_flags |= ITS_FLAGS_ERRATA_CAVIUM_22375;
1064 
1065 	/*
1066 	 * We need to limit which CPUs we send these interrupts to on
1067 	 * the original dual socket ThunderX as it is unable to
1068 	 * forward them between the two sockets.
1069 	 */
1070 	if (bus_get_domain(dev, &domain) == 0) {
1071 		if (domain < MAXMEMDOM) {
1072 			CPU_COPY(&cpuset_domain[domain], &sc->sc_cpus);
1073 		} else {
1074 			CPU_ZERO(&sc->sc_cpus);
1075 		}
1076 	}
1077 }
1078 
1079 static void
1080 gicv3_its_disable_intr(device_t dev, struct intr_irqsrc *isrc)
1081 {
1082 	struct gicv3_its_softc *sc;
1083 	struct gicv3_its_irqsrc *girq;
1084 	uint8_t *conf;
1085 
1086 	sc = device_get_softc(dev);
1087 	girq = (struct gicv3_its_irqsrc *)isrc;
1088 	conf = sc->sc_conf_base;
1089 
1090 	conf[girq->gi_lpi] &= ~LPI_CONF_ENABLE;
1091 
1092 	if ((sc->sc_its_flags & ITS_FLAGS_LPI_CONF_FLUSH) != 0) {
1093 		/* Clean D-cache under command. */
1094 		cpu_dcache_wb_range((vm_offset_t)&conf[girq->gi_lpi], 1);
1095 	} else {
1096 		/* DSB inner shareable, store */
1097 		dsb(ishst);
1098 	}
1099 
1100 	its_cmd_inv(dev, girq->gi_its_dev, girq);
1101 }
1102 
1103 static void
1104 gicv3_its_enable_intr(device_t dev, struct intr_irqsrc *isrc)
1105 {
1106 	struct gicv3_its_softc *sc;
1107 	struct gicv3_its_irqsrc *girq;
1108 	uint8_t *conf;
1109 
1110 	sc = device_get_softc(dev);
1111 	girq = (struct gicv3_its_irqsrc *)isrc;
1112 	conf = sc->sc_conf_base;
1113 
1114 	conf[girq->gi_lpi] |= LPI_CONF_ENABLE;
1115 
1116 	if ((sc->sc_its_flags & ITS_FLAGS_LPI_CONF_FLUSH) != 0) {
1117 		/* Clean D-cache under command. */
1118 		cpu_dcache_wb_range((vm_offset_t)&conf[girq->gi_lpi], 1);
1119 	} else {
1120 		/* DSB inner shareable, store */
1121 		dsb(ishst);
1122 	}
1123 
1124 	its_cmd_inv(dev, girq->gi_its_dev, girq);
1125 }
1126 
1127 static int
1128 gicv3_its_intr(void *arg, uintptr_t irq)
1129 {
1130 	struct gicv3_its_softc *sc = arg;
1131 	struct gicv3_its_irqsrc *girq;
1132 	struct trapframe *tf;
1133 
1134 	irq -= sc->sc_irq_base;
1135 	girq = sc->sc_irqs[irq];
1136 	if (girq == NULL)
1137 		panic("gicv3_its_intr: Invalid interrupt %ld",
1138 		    irq + sc->sc_irq_base);
1139 
1140 	tf = curthread->td_intr_frame;
1141 	intr_isrc_dispatch(&girq->gi_isrc, tf);
1142 	return (FILTER_HANDLED);
1143 }
1144 
1145 static void
1146 gicv3_its_pre_ithread(device_t dev, struct intr_irqsrc *isrc)
1147 {
1148 	struct gicv3_its_irqsrc *girq;
1149 
1150 	girq = (struct gicv3_its_irqsrc *)isrc;
1151 	gic_icc_write(EOIR1, girq->gi_lpi + GIC_FIRST_LPI);
1152 }
1153 
1154 static void
1155 gicv3_its_post_ithread(device_t dev, struct intr_irqsrc *isrc)
1156 {
1157 
1158 }
1159 
1160 static void
1161 gicv3_its_post_filter(device_t dev, struct intr_irqsrc *isrc)
1162 {
1163 	struct gicv3_its_irqsrc *girq;
1164 
1165 	girq = (struct gicv3_its_irqsrc *)isrc;
1166 	gic_icc_write(EOIR1, girq->gi_lpi + GIC_FIRST_LPI);
1167 }
1168 
1169 static int
1170 gicv3_its_select_cpu(device_t dev, struct intr_irqsrc *isrc)
1171 {
1172 	struct gicv3_its_softc *sc;
1173 
1174 	sc = device_get_softc(dev);
1175 	if (CPU_EMPTY(&isrc->isrc_cpu)) {
1176 		sc->gic_irq_cpu = intr_irq_next_cpu(sc->gic_irq_cpu,
1177 		    &sc->sc_cpus);
1178 		CPU_SETOF(sc->gic_irq_cpu, &isrc->isrc_cpu);
1179 	}
1180 
1181 	return (0);
1182 }
1183 
1184 static int
1185 gicv3_its_bind_intr(device_t dev, struct intr_irqsrc *isrc)
1186 {
1187 	struct gicv3_its_irqsrc *girq;
1188 
1189 	gicv3_its_select_cpu(dev, isrc);
1190 
1191 	girq = (struct gicv3_its_irqsrc *)isrc;
1192 	its_cmd_movi(dev, girq);
1193 	return (0);
1194 }
1195 
1196 static int
1197 gicv3_its_map_intr(device_t dev, struct intr_map_data *data,
1198     struct intr_irqsrc **isrcp)
1199 {
1200 
1201 	/*
1202 	 * This should never happen, we only call this function to map
1203 	 * interrupts found before the controller driver is ready.
1204 	 */
1205 	panic("gicv3_its_map_intr: Unable to map a MSI interrupt");
1206 }
1207 
1208 static int
1209 gicv3_its_setup_intr(device_t dev, struct intr_irqsrc *isrc,
1210     struct resource *res, struct intr_map_data *data)
1211 {
1212 
1213 	/* Bind the interrupt to a CPU */
1214 	gicv3_its_bind_intr(dev, isrc);
1215 
1216 	return (0);
1217 }
1218 
1219 #ifdef SMP
1220 static void
1221 gicv3_its_init_secondary(device_t dev)
1222 {
1223 	struct gicv3_its_softc *sc;
1224 
1225 	sc = device_get_softc(dev);
1226 
1227 	/*
1228 	 * This is fatal as otherwise we may bind interrupts to this CPU.
1229 	 * We need a way to tell the interrupt framework to only bind to a
1230 	 * subset of given CPUs when it performs the shuffle.
1231 	 */
1232 	if (its_init_cpu(dev, sc) != 0)
1233 		panic("gicv3_its_init_secondary: No usable ITS on CPU%d",
1234 		    PCPU_GET(cpuid));
1235 }
1236 #endif
1237 
1238 static uint32_t
1239 its_get_devid(device_t pci_dev)
1240 {
1241 	uintptr_t id;
1242 
1243 	if (pci_get_id(pci_dev, PCI_ID_MSI, &id) != 0)
1244 		panic("%s: %s: Unable to get the MSI DeviceID", __func__,
1245 		    device_get_nameunit(pci_dev));
1246 
1247 	return (id);
1248 }
1249 
1250 static struct its_dev *
1251 its_device_find(device_t dev, device_t child)
1252 {
1253 	struct gicv3_its_softc *sc;
1254 	struct its_dev *its_dev = NULL;
1255 
1256 	sc = device_get_softc(dev);
1257 
1258 	mtx_lock_spin(&sc->sc_its_dev_lock);
1259 	TAILQ_FOREACH(its_dev, &sc->sc_its_dev_list, entry) {
1260 		if (its_dev->pci_dev == child)
1261 			break;
1262 	}
1263 	mtx_unlock_spin(&sc->sc_its_dev_lock);
1264 
1265 	return (its_dev);
1266 }
1267 
1268 static bool
1269 its_device_alloc(struct gicv3_its_softc *sc, int devid)
1270 {
1271 	struct its_ptable *ptable;
1272 	vm_offset_t l2_table;
1273 	uint64_t *table;
1274 	uint32_t index;
1275 	bool shareable;
1276 
1277 	/* No device table */
1278 	if (sc->sc_dev_table_idx < 0) {
1279 		if (devid >= (1 << sc->sc_devbits)) {
1280 			if (bootverbose) {
1281 				device_printf(sc->dev,
1282 				    "%s: Device out of range for hardware "
1283 				    "(%x >= %x)\n", __func__, devid,
1284 				    1 << sc->sc_devbits);
1285 			}
1286 			return (false);
1287 		}
1288 		return (true);
1289 	}
1290 
1291 	ptable = &sc->sc_its_ptab[sc->sc_dev_table_idx];
1292 	/* Check the devid is within the table limit */
1293 	if (!ptable->ptab_indirect) {
1294 		if (devid >= ptable->ptab_l1_nidents) {
1295 			if (bootverbose) {
1296 				device_printf(sc->dev,
1297 				    "%s: Device out of range for table "
1298 				    "(%x >= %x)\n", __func__, devid,
1299 				    ptable->ptab_l1_nidents);
1300 			}
1301 			return (false);
1302 		}
1303 
1304 		return (true);
1305 	}
1306 
1307 	/* Check the devid is within the allocated range */
1308 	index = devid / ptable->ptab_l2_nidents;
1309 	if (index >= ptable->ptab_l1_nidents) {
1310 		if (bootverbose) {
1311 			device_printf(sc->dev,
1312 			    "%s: Index out of range for table (%x >= %x)\n",
1313 			    __func__, index, ptable->ptab_l1_nidents);
1314 		}
1315 		return (false);
1316 	}
1317 
1318 	table = (uint64_t *)ptable->ptab_vaddr;
1319 	/* We have an second level table */
1320 	if ((table[index] & GITS_BASER_VALID) != 0)
1321 		return (true);
1322 
1323 	shareable = true;
1324 	if ((ptable->ptab_share & GITS_BASER_SHARE_MASK) == GITS_BASER_SHARE_NS)
1325 		shareable = false;
1326 
1327 	l2_table = (vm_offset_t)contigmalloc_domainset(ptable->ptab_l2_size,
1328 	    M_GICV3_ITS, sc->sc_ds, M_WAITOK | M_ZERO, 0, (1ul << 48) - 1,
1329 	    ptable->ptab_page_size, 0);
1330 
1331 	if (!shareable)
1332 		cpu_dcache_wb_range((vm_offset_t)l2_table,
1333 		    ptable->ptab_l2_size);
1334 
1335 	table[index] = vtophys(l2_table) | GITS_BASER_VALID;
1336 	if (!shareable)
1337 		cpu_dcache_wb_range((vm_offset_t)&table[index],
1338 		    sizeof(table[index]));
1339 
1340 	dsb(sy);
1341 	return (true);
1342 }
1343 
1344 static struct its_dev *
1345 its_device_get(device_t dev, device_t child, u_int nvecs)
1346 {
1347 	struct gicv3_its_softc *sc;
1348 	struct its_dev *its_dev;
1349 	vmem_addr_t irq_base;
1350 	size_t esize;
1351 
1352 	sc = device_get_softc(dev);
1353 
1354 	its_dev = its_device_find(dev, child);
1355 	if (its_dev != NULL)
1356 		return (its_dev);
1357 
1358 	its_dev = malloc(sizeof(*its_dev), M_GICV3_ITS, M_NOWAIT | M_ZERO);
1359 	if (its_dev == NULL)
1360 		return (NULL);
1361 
1362 	its_dev->pci_dev = child;
1363 	its_dev->devid = its_get_devid(child);
1364 
1365 	its_dev->lpis.lpi_busy = 0;
1366 	its_dev->lpis.lpi_num = nvecs;
1367 	its_dev->lpis.lpi_free = nvecs;
1368 
1369 	if (!its_device_alloc(sc, its_dev->devid)) {
1370 		free(its_dev, M_GICV3_ITS);
1371 		return (NULL);
1372 	}
1373 
1374 	if (vmem_alloc(sc->sc_irq_alloc, nvecs, M_FIRSTFIT | M_NOWAIT,
1375 	    &irq_base) != 0) {
1376 		free(its_dev, M_GICV3_ITS);
1377 		return (NULL);
1378 	}
1379 	its_dev->lpis.lpi_base = irq_base;
1380 
1381 	/* Get ITT entry size */
1382 	esize = GITS_TYPER_ITTES(gic_its_read_8(sc, GITS_TYPER));
1383 
1384 	/*
1385 	 * Allocate ITT for this device.
1386 	 * PA has to be 256 B aligned. At least two entries for device.
1387 	 */
1388 	its_dev->itt_size = roundup2(MAX(nvecs, 2) * esize, 256);
1389 	its_dev->itt = (vm_offset_t)contigmalloc_domainset(its_dev->itt_size,
1390 	    M_GICV3_ITS, sc->sc_ds, M_NOWAIT | M_ZERO, 0,
1391 	    LPI_INT_TRANS_TAB_MAX_ADDR, LPI_INT_TRANS_TAB_ALIGN, 0);
1392 	if (its_dev->itt == 0) {
1393 		vmem_free(sc->sc_irq_alloc, its_dev->lpis.lpi_base, nvecs);
1394 		free(its_dev, M_GICV3_ITS);
1395 		return (NULL);
1396 	}
1397 
1398 	/* Make sure device sees zeroed ITT. */
1399 	if ((sc->sc_its_flags & ITS_FLAGS_CMDQ_FLUSH) != 0)
1400 		cpu_dcache_wb_range(its_dev->itt, its_dev->itt_size);
1401 
1402 	mtx_lock_spin(&sc->sc_its_dev_lock);
1403 	TAILQ_INSERT_TAIL(&sc->sc_its_dev_list, its_dev, entry);
1404 	mtx_unlock_spin(&sc->sc_its_dev_lock);
1405 
1406 	/* Map device to its ITT */
1407 	its_cmd_mapd(dev, its_dev, 1);
1408 
1409 	return (its_dev);
1410 }
1411 
1412 static void
1413 its_device_release(device_t dev, struct its_dev *its_dev)
1414 {
1415 	struct gicv3_its_softc *sc;
1416 
1417 	KASSERT(its_dev->lpis.lpi_busy == 0,
1418 	    ("its_device_release: Trying to release an inuse ITS device"));
1419 
1420 	/* Unmap device in ITS */
1421 	its_cmd_mapd(dev, its_dev, 0);
1422 
1423 	sc = device_get_softc(dev);
1424 
1425 	/* Remove the device from the list of devices */
1426 	mtx_lock_spin(&sc->sc_its_dev_lock);
1427 	TAILQ_REMOVE(&sc->sc_its_dev_list, its_dev, entry);
1428 	mtx_unlock_spin(&sc->sc_its_dev_lock);
1429 
1430 	/* Free ITT */
1431 	KASSERT(its_dev->itt != 0, ("Invalid ITT in valid ITS device"));
1432 	contigfree((void *)its_dev->itt, its_dev->itt_size, M_GICV3_ITS);
1433 
1434 	/* Free the IRQ allocation */
1435 	vmem_free(sc->sc_irq_alloc, its_dev->lpis.lpi_base,
1436 	    its_dev->lpis.lpi_num);
1437 
1438 	free(its_dev, M_GICV3_ITS);
1439 }
1440 
1441 static struct gicv3_its_irqsrc *
1442 gicv3_its_alloc_irqsrc(device_t dev, struct gicv3_its_softc *sc, u_int irq)
1443 {
1444 	struct gicv3_its_irqsrc *girq = NULL;
1445 
1446 	KASSERT(sc->sc_irqs[irq] == NULL,
1447 	    ("%s: Interrupt %u already allocated", __func__, irq));
1448 	mtx_lock_spin(&sc->sc_its_dev_lock);
1449 	if (!TAILQ_EMPTY(&sc->sc_free_irqs)) {
1450 		girq = TAILQ_FIRST(&sc->sc_free_irqs);
1451 		TAILQ_REMOVE(&sc->sc_free_irqs, girq, gi_link);
1452 	}
1453 	mtx_unlock_spin(&sc->sc_its_dev_lock);
1454 	if (girq == NULL) {
1455 		girq = malloc(sizeof(*girq), M_GICV3_ITS,
1456 		    M_NOWAIT | M_ZERO);
1457 		if (girq == NULL)
1458 			return (NULL);
1459 		girq->gi_id = -1;
1460 		if (intr_isrc_register(&girq->gi_isrc, dev, 0,
1461 		    "%s,%u", device_get_nameunit(dev), irq) != 0) {
1462 			free(girq, M_GICV3_ITS);
1463 			return (NULL);
1464 		}
1465 	}
1466 	girq->gi_lpi = irq + sc->sc_irq_base - GIC_FIRST_LPI;
1467 	sc->sc_irqs[irq] = girq;
1468 
1469 	return (girq);
1470 }
1471 
1472 static void
1473 gicv3_its_release_irqsrc(struct gicv3_its_softc *sc,
1474     struct gicv3_its_irqsrc *girq)
1475 {
1476 	u_int irq;
1477 
1478 	mtx_assert(&sc->sc_its_dev_lock, MA_OWNED);
1479 
1480 	irq = girq->gi_lpi + GIC_FIRST_LPI - sc->sc_irq_base;
1481 	sc->sc_irqs[irq] = NULL;
1482 
1483 	girq->gi_id = -1;
1484 	girq->gi_its_dev = NULL;
1485 	TAILQ_INSERT_TAIL(&sc->sc_free_irqs, girq, gi_link);
1486 }
1487 
1488 static int
1489 gicv3_its_alloc_msi(device_t dev, device_t child, int count, int maxcount,
1490     device_t *pic, struct intr_irqsrc **srcs)
1491 {
1492 	struct gicv3_its_softc *sc;
1493 	struct gicv3_its_irqsrc *girq;
1494 	struct its_dev *its_dev;
1495 	u_int irq;
1496 	int i;
1497 
1498 	its_dev = its_device_get(dev, child, count);
1499 	if (its_dev == NULL)
1500 		return (ENXIO);
1501 
1502 	KASSERT(its_dev->lpis.lpi_free >= count,
1503 	    ("gicv3_its_alloc_msi: No free LPIs"));
1504 	sc = device_get_softc(dev);
1505 	irq = its_dev->lpis.lpi_base + its_dev->lpis.lpi_num -
1506 	    its_dev->lpis.lpi_free;
1507 
1508 	/* Allocate the irqsrc for each MSI */
1509 	for (i = 0; i < count; i++, irq++) {
1510 		its_dev->lpis.lpi_free--;
1511 		srcs[i] = (struct intr_irqsrc *)gicv3_its_alloc_irqsrc(dev,
1512 		    sc, irq);
1513 		if (srcs[i] == NULL)
1514 			break;
1515 	}
1516 
1517 	/* The allocation failed, release them */
1518 	if (i != count) {
1519 		mtx_lock_spin(&sc->sc_its_dev_lock);
1520 		for (i = 0; i < count; i++) {
1521 			girq = (struct gicv3_its_irqsrc *)srcs[i];
1522 			if (girq == NULL)
1523 				break;
1524 			gicv3_its_release_irqsrc(sc, girq);
1525 			srcs[i] = NULL;
1526 		}
1527 		mtx_unlock_spin(&sc->sc_its_dev_lock);
1528 		return (ENXIO);
1529 	}
1530 
1531 	/* Finish the allocation now we have all MSI irqsrcs */
1532 	for (i = 0; i < count; i++) {
1533 		girq = (struct gicv3_its_irqsrc *)srcs[i];
1534 		girq->gi_id = i;
1535 		girq->gi_its_dev = its_dev;
1536 
1537 		/* Map the message to the given IRQ */
1538 		gicv3_its_select_cpu(dev, (struct intr_irqsrc *)girq);
1539 		its_cmd_mapti(dev, girq);
1540 	}
1541 	its_dev->lpis.lpi_busy += count;
1542 	*pic = dev;
1543 
1544 	return (0);
1545 }
1546 
1547 static int
1548 gicv3_its_release_msi(device_t dev, device_t child, int count,
1549     struct intr_irqsrc **isrc)
1550 {
1551 	struct gicv3_its_softc *sc;
1552 	struct gicv3_its_irqsrc *girq;
1553 	struct its_dev *its_dev;
1554 	int i;
1555 
1556 	its_dev = its_device_find(dev, child);
1557 
1558 	KASSERT(its_dev != NULL,
1559 	    ("gicv3_its_release_msi: Releasing a MSI interrupt with "
1560 	     "no ITS device"));
1561 	KASSERT(its_dev->lpis.lpi_busy >= count,
1562 	    ("gicv3_its_release_msi: Releasing more interrupts than "
1563 	     "were allocated: releasing %d, allocated %d", count,
1564 	     its_dev->lpis.lpi_busy));
1565 
1566 	sc = device_get_softc(dev);
1567 	mtx_lock_spin(&sc->sc_its_dev_lock);
1568 	for (i = 0; i < count; i++) {
1569 		girq = (struct gicv3_its_irqsrc *)isrc[i];
1570 		gicv3_its_release_irqsrc(sc, girq);
1571 	}
1572 	mtx_unlock_spin(&sc->sc_its_dev_lock);
1573 	its_dev->lpis.lpi_busy -= count;
1574 
1575 	if (its_dev->lpis.lpi_busy == 0)
1576 		its_device_release(dev, its_dev);
1577 
1578 	return (0);
1579 }
1580 
1581 static int
1582 gicv3_its_alloc_msix(device_t dev, device_t child, device_t *pic,
1583     struct intr_irqsrc **isrcp)
1584 {
1585 	struct gicv3_its_softc *sc;
1586 	struct gicv3_its_irqsrc *girq;
1587 	struct its_dev *its_dev;
1588 	u_int nvecs, irq;
1589 
1590 	nvecs = pci_msix_count(child);
1591 	its_dev = its_device_get(dev, child, nvecs);
1592 	if (its_dev == NULL)
1593 		return (ENXIO);
1594 
1595 	KASSERT(its_dev->lpis.lpi_free > 0,
1596 	    ("gicv3_its_alloc_msix: No free LPIs"));
1597 	sc = device_get_softc(dev);
1598 	irq = its_dev->lpis.lpi_base + its_dev->lpis.lpi_num -
1599 	    its_dev->lpis.lpi_free;
1600 
1601 	girq = gicv3_its_alloc_irqsrc(dev, sc, irq);
1602 	if (girq == NULL)
1603 		return (ENXIO);
1604 	girq->gi_id = its_dev->lpis.lpi_busy;
1605 	girq->gi_its_dev = its_dev;
1606 
1607 	its_dev->lpis.lpi_free--;
1608 	its_dev->lpis.lpi_busy++;
1609 
1610 	/* Map the message to the given IRQ */
1611 	gicv3_its_select_cpu(dev, (struct intr_irqsrc *)girq);
1612 	its_cmd_mapti(dev, girq);
1613 
1614 	*pic = dev;
1615 	*isrcp = (struct intr_irqsrc *)girq;
1616 
1617 	return (0);
1618 }
1619 
1620 static int
1621 gicv3_its_release_msix(device_t dev, device_t child, struct intr_irqsrc *isrc)
1622 {
1623 	struct gicv3_its_softc *sc;
1624 	struct gicv3_its_irqsrc *girq;
1625 	struct its_dev *its_dev;
1626 
1627 	its_dev = its_device_find(dev, child);
1628 
1629 	KASSERT(its_dev != NULL,
1630 	    ("gicv3_its_release_msix: Releasing a MSI-X interrupt with "
1631 	     "no ITS device"));
1632 	KASSERT(its_dev->lpis.lpi_busy > 0,
1633 	    ("gicv3_its_release_msix: Releasing more interrupts than "
1634 	     "were allocated: allocated %d", its_dev->lpis.lpi_busy));
1635 
1636 	sc = device_get_softc(dev);
1637 	girq = (struct gicv3_its_irqsrc *)isrc;
1638 	mtx_lock_spin(&sc->sc_its_dev_lock);
1639 	gicv3_its_release_irqsrc(sc, girq);
1640 	mtx_unlock_spin(&sc->sc_its_dev_lock);
1641 	its_dev->lpis.lpi_busy--;
1642 
1643 	if (its_dev->lpis.lpi_busy == 0)
1644 		its_device_release(dev, its_dev);
1645 
1646 	return (0);
1647 }
1648 
1649 static int
1650 gicv3_its_map_msi(device_t dev, device_t child, struct intr_irqsrc *isrc,
1651     uint64_t *addr, uint32_t *data)
1652 {
1653 	struct gicv3_its_softc *sc;
1654 	struct gicv3_its_irqsrc *girq;
1655 
1656 	sc = device_get_softc(dev);
1657 	girq = (struct gicv3_its_irqsrc *)isrc;
1658 
1659 	*addr = vtophys(rman_get_virtual(sc->sc_its_res)) + GITS_TRANSLATER;
1660 	*data = girq->gi_id;
1661 
1662 	return (0);
1663 }
1664 
1665 #ifdef IOMMU
1666 static int
1667 gicv3_iommu_init(device_t dev, device_t child, struct iommu_domain **domain)
1668 {
1669 	struct gicv3_its_softc *sc;
1670 	struct iommu_ctx *ctx;
1671 	int error;
1672 
1673 	sc = device_get_softc(dev);
1674 	ctx = iommu_get_dev_ctx(child);
1675 	if (ctx == NULL)
1676 		return (ENXIO);
1677 	/* Map the page containing the GITS_TRANSLATER register. */
1678 	error = iommu_map_msi(ctx, PAGE_SIZE, 0,
1679 	    IOMMU_MAP_ENTRY_WRITE, IOMMU_MF_CANWAIT, &sc->ma);
1680 	*domain = iommu_get_ctx_domain(ctx);
1681 
1682 	return (error);
1683 }
1684 
1685 static void
1686 gicv3_iommu_deinit(device_t dev, device_t child)
1687 {
1688 	struct iommu_ctx *ctx;
1689 
1690 	ctx = iommu_get_dev_ctx(child);
1691 	if (ctx == NULL)
1692 		return;
1693 
1694 	iommu_unmap_msi(ctx);
1695 }
1696 #endif
1697 
1698 /*
1699  * Commands handling.
1700  */
1701 
1702 static __inline void
1703 cmd_format_command(struct its_cmd *cmd, uint8_t cmd_type)
1704 {
1705 	/* Command field: DW0 [7:0] */
1706 	cmd->cmd_dword[0] &= htole64(~CMD_COMMAND_MASK);
1707 	cmd->cmd_dword[0] |= htole64(cmd_type);
1708 }
1709 
1710 static __inline void
1711 cmd_format_devid(struct its_cmd *cmd, uint32_t devid)
1712 {
1713 	/* Device ID field: DW0 [63:32] */
1714 	cmd->cmd_dword[0] &= htole64(~CMD_DEVID_MASK);
1715 	cmd->cmd_dword[0] |= htole64((uint64_t)devid << CMD_DEVID_SHIFT);
1716 }
1717 
1718 static __inline void
1719 cmd_format_size(struct its_cmd *cmd, uint16_t size)
1720 {
1721 	/* Size field: DW1 [4:0] */
1722 	cmd->cmd_dword[1] &= htole64(~CMD_SIZE_MASK);
1723 	cmd->cmd_dword[1] |= htole64((size & CMD_SIZE_MASK));
1724 }
1725 
1726 static __inline void
1727 cmd_format_id(struct its_cmd *cmd, uint32_t id)
1728 {
1729 	/* ID field: DW1 [31:0] */
1730 	cmd->cmd_dword[1] &= htole64(~CMD_ID_MASK);
1731 	cmd->cmd_dword[1] |= htole64(id);
1732 }
1733 
1734 static __inline void
1735 cmd_format_pid(struct its_cmd *cmd, uint32_t pid)
1736 {
1737 	/* Physical ID field: DW1 [63:32] */
1738 	cmd->cmd_dword[1] &= htole64(~CMD_PID_MASK);
1739 	cmd->cmd_dword[1] |= htole64((uint64_t)pid << CMD_PID_SHIFT);
1740 }
1741 
1742 static __inline void
1743 cmd_format_col(struct its_cmd *cmd, uint16_t col_id)
1744 {
1745 	/* Collection field: DW2 [16:0] */
1746 	cmd->cmd_dword[2] &= htole64(~CMD_COL_MASK);
1747 	cmd->cmd_dword[2] |= htole64(col_id);
1748 }
1749 
1750 static __inline void
1751 cmd_format_target(struct its_cmd *cmd, uint64_t target)
1752 {
1753 	/* Target Address field: DW2 [47:16] */
1754 	cmd->cmd_dword[2] &= htole64(~CMD_TARGET_MASK);
1755 	cmd->cmd_dword[2] |= htole64(target & CMD_TARGET_MASK);
1756 }
1757 
1758 static __inline void
1759 cmd_format_itt(struct its_cmd *cmd, uint64_t itt)
1760 {
1761 	/* ITT Address field: DW2 [47:8] */
1762 	cmd->cmd_dword[2] &= htole64(~CMD_ITT_MASK);
1763 	cmd->cmd_dword[2] |= htole64(itt & CMD_ITT_MASK);
1764 }
1765 
1766 static __inline void
1767 cmd_format_valid(struct its_cmd *cmd, uint8_t valid)
1768 {
1769 	/* Valid field: DW2 [63] */
1770 	cmd->cmd_dword[2] &= htole64(~CMD_VALID_MASK);
1771 	cmd->cmd_dword[2] |= htole64((uint64_t)valid << CMD_VALID_SHIFT);
1772 }
1773 
1774 static inline bool
1775 its_cmd_queue_full(struct gicv3_its_softc *sc)
1776 {
1777 	size_t read_idx, next_write_idx;
1778 
1779 	/* Get the index of the next command */
1780 	next_write_idx = (sc->sc_its_cmd_next_idx + 1) %
1781 	    (ITS_CMDQ_SIZE / sizeof(struct its_cmd));
1782 	/* And the index of the current command being read */
1783 	read_idx = gic_its_read_4(sc, GITS_CREADR) / sizeof(struct its_cmd);
1784 
1785 	/*
1786 	 * The queue is full when the write offset points
1787 	 * at the command before the current read offset.
1788 	 */
1789 	return (next_write_idx == read_idx);
1790 }
1791 
1792 static inline void
1793 its_cmd_sync(struct gicv3_its_softc *sc, struct its_cmd *cmd)
1794 {
1795 
1796 	if ((sc->sc_its_flags & ITS_FLAGS_CMDQ_FLUSH) != 0) {
1797 		/* Clean D-cache under command. */
1798 		cpu_dcache_wb_range((vm_offset_t)cmd, sizeof(*cmd));
1799 	} else {
1800 		/* DSB inner shareable, store */
1801 		dsb(ishst);
1802 	}
1803 
1804 }
1805 
1806 static inline uint64_t
1807 its_cmd_cwriter_offset(struct gicv3_its_softc *sc, struct its_cmd *cmd)
1808 {
1809 	uint64_t off;
1810 
1811 	off = (cmd - sc->sc_its_cmd_base) * sizeof(*cmd);
1812 
1813 	return (off);
1814 }
1815 
1816 static void
1817 its_cmd_wait_completion(device_t dev, struct its_cmd *cmd_first,
1818     struct its_cmd *cmd_last)
1819 {
1820 	struct gicv3_its_softc *sc;
1821 	uint64_t first, last, read;
1822 	size_t us_left;
1823 
1824 	sc = device_get_softc(dev);
1825 
1826 	/*
1827 	 * XXX ARM64TODO: This is obviously a significant delay.
1828 	 * The reason for that is that currently the time frames for
1829 	 * the command to complete are not known.
1830 	 */
1831 	us_left = 1000000;
1832 
1833 	first = its_cmd_cwriter_offset(sc, cmd_first);
1834 	last = its_cmd_cwriter_offset(sc, cmd_last);
1835 
1836 	for (;;) {
1837 		read = gic_its_read_8(sc, GITS_CREADR);
1838 		if (first < last) {
1839 			if (read < first || read >= last)
1840 				break;
1841 		} else if (read < first && read >= last)
1842 			break;
1843 
1844 		if (us_left-- == 0) {
1845 			/* This means timeout */
1846 			device_printf(dev,
1847 			    "Timeout while waiting for CMD completion.\n");
1848 			return;
1849 		}
1850 		DELAY(1);
1851 	}
1852 }
1853 
1854 static struct its_cmd *
1855 its_cmd_alloc_locked(device_t dev)
1856 {
1857 	struct gicv3_its_softc *sc;
1858 	struct its_cmd *cmd;
1859 	size_t us_left;
1860 
1861 	sc = device_get_softc(dev);
1862 
1863 	/*
1864 	 * XXX ARM64TODO: This is obviously a significant delay.
1865 	 * The reason for that is that currently the time frames for
1866 	 * the command to complete (and therefore free the descriptor)
1867 	 * are not known.
1868 	 */
1869 	us_left = 1000000;
1870 
1871 	mtx_assert(&sc->sc_its_cmd_lock, MA_OWNED);
1872 	while (its_cmd_queue_full(sc)) {
1873 		if (us_left-- == 0) {
1874 			/* Timeout while waiting for free command */
1875 			device_printf(dev,
1876 			    "Timeout while waiting for free command\n");
1877 			return (NULL);
1878 		}
1879 		DELAY(1);
1880 	}
1881 
1882 	cmd = &sc->sc_its_cmd_base[sc->sc_its_cmd_next_idx];
1883 	sc->sc_its_cmd_next_idx++;
1884 	sc->sc_its_cmd_next_idx %= ITS_CMDQ_SIZE / sizeof(struct its_cmd);
1885 
1886 	return (cmd);
1887 }
1888 
1889 static uint64_t
1890 its_cmd_prepare(struct its_cmd *cmd, struct its_cmd_desc *desc)
1891 {
1892 	uint64_t target;
1893 	uint8_t cmd_type;
1894 	u_int size;
1895 
1896 	cmd_type = desc->cmd_type;
1897 	target = ITS_TARGET_NONE;
1898 
1899 	switch (cmd_type) {
1900 	case ITS_CMD_MOVI:	/* Move interrupt ID to another collection */
1901 		target = desc->cmd_desc_movi.col->col_target;
1902 		cmd_format_command(cmd, ITS_CMD_MOVI);
1903 		cmd_format_id(cmd, desc->cmd_desc_movi.id);
1904 		cmd_format_col(cmd, desc->cmd_desc_movi.col->col_id);
1905 		cmd_format_devid(cmd, desc->cmd_desc_movi.its_dev->devid);
1906 		break;
1907 	case ITS_CMD_SYNC:	/* Wait for previous commands completion */
1908 		target = desc->cmd_desc_sync.col->col_target;
1909 		cmd_format_command(cmd, ITS_CMD_SYNC);
1910 		cmd_format_target(cmd, target);
1911 		break;
1912 	case ITS_CMD_MAPD:	/* Assign ITT to device */
1913 		cmd_format_command(cmd, ITS_CMD_MAPD);
1914 		cmd_format_itt(cmd, vtophys(desc->cmd_desc_mapd.its_dev->itt));
1915 		/*
1916 		 * Size describes number of bits to encode interrupt IDs
1917 		 * supported by the device minus one.
1918 		 * When V (valid) bit is zero, this field should be written
1919 		 * as zero.
1920 		 */
1921 		if (desc->cmd_desc_mapd.valid != 0) {
1922 			size = fls(desc->cmd_desc_mapd.its_dev->lpis.lpi_num);
1923 			size = MAX(1, size) - 1;
1924 		} else
1925 			size = 0;
1926 
1927 		cmd_format_size(cmd, size);
1928 		cmd_format_devid(cmd, desc->cmd_desc_mapd.its_dev->devid);
1929 		cmd_format_valid(cmd, desc->cmd_desc_mapd.valid);
1930 		break;
1931 	case ITS_CMD_MAPC:	/* Map collection to Re-Distributor */
1932 		target = desc->cmd_desc_mapc.col->col_target;
1933 		cmd_format_command(cmd, ITS_CMD_MAPC);
1934 		cmd_format_col(cmd, desc->cmd_desc_mapc.col->col_id);
1935 		cmd_format_valid(cmd, desc->cmd_desc_mapc.valid);
1936 		cmd_format_target(cmd, target);
1937 		break;
1938 	case ITS_CMD_MAPTI:
1939 		target = desc->cmd_desc_mapvi.col->col_target;
1940 		cmd_format_command(cmd, ITS_CMD_MAPTI);
1941 		cmd_format_devid(cmd, desc->cmd_desc_mapvi.its_dev->devid);
1942 		cmd_format_id(cmd, desc->cmd_desc_mapvi.id);
1943 		cmd_format_pid(cmd, desc->cmd_desc_mapvi.pid);
1944 		cmd_format_col(cmd, desc->cmd_desc_mapvi.col->col_id);
1945 		break;
1946 	case ITS_CMD_MAPI:
1947 		target = desc->cmd_desc_mapi.col->col_target;
1948 		cmd_format_command(cmd, ITS_CMD_MAPI);
1949 		cmd_format_devid(cmd, desc->cmd_desc_mapi.its_dev->devid);
1950 		cmd_format_id(cmd, desc->cmd_desc_mapi.pid);
1951 		cmd_format_col(cmd, desc->cmd_desc_mapi.col->col_id);
1952 		break;
1953 	case ITS_CMD_INV:
1954 		target = desc->cmd_desc_inv.col->col_target;
1955 		cmd_format_command(cmd, ITS_CMD_INV);
1956 		cmd_format_devid(cmd, desc->cmd_desc_inv.its_dev->devid);
1957 		cmd_format_id(cmd, desc->cmd_desc_inv.pid);
1958 		break;
1959 	case ITS_CMD_INVALL:
1960 		cmd_format_command(cmd, ITS_CMD_INVALL);
1961 		cmd_format_col(cmd, desc->cmd_desc_invall.col->col_id);
1962 		break;
1963 	default:
1964 		panic("its_cmd_prepare: Invalid command: %x", cmd_type);
1965 	}
1966 
1967 	return (target);
1968 }
1969 
1970 static int
1971 its_cmd_send(device_t dev, struct its_cmd_desc *desc)
1972 {
1973 	struct gicv3_its_softc *sc;
1974 	struct its_cmd *cmd, *cmd_sync, *cmd_write;
1975 	struct its_col col_sync;
1976 	struct its_cmd_desc desc_sync;
1977 	uint64_t target, cwriter;
1978 
1979 	sc = device_get_softc(dev);
1980 	mtx_lock_spin(&sc->sc_its_cmd_lock);
1981 	cmd = its_cmd_alloc_locked(dev);
1982 	if (cmd == NULL) {
1983 		device_printf(dev, "could not allocate ITS command\n");
1984 		mtx_unlock_spin(&sc->sc_its_cmd_lock);
1985 		return (EBUSY);
1986 	}
1987 
1988 	target = its_cmd_prepare(cmd, desc);
1989 	its_cmd_sync(sc, cmd);
1990 
1991 	if (target != ITS_TARGET_NONE) {
1992 		cmd_sync = its_cmd_alloc_locked(dev);
1993 		if (cmd_sync != NULL) {
1994 			desc_sync.cmd_type = ITS_CMD_SYNC;
1995 			col_sync.col_target = target;
1996 			desc_sync.cmd_desc_sync.col = &col_sync;
1997 			its_cmd_prepare(cmd_sync, &desc_sync);
1998 			its_cmd_sync(sc, cmd_sync);
1999 		}
2000 	}
2001 
2002 	/* Update GITS_CWRITER */
2003 	cwriter = sc->sc_its_cmd_next_idx * sizeof(struct its_cmd);
2004 	gic_its_write_8(sc, GITS_CWRITER, cwriter);
2005 	cmd_write = &sc->sc_its_cmd_base[sc->sc_its_cmd_next_idx];
2006 	mtx_unlock_spin(&sc->sc_its_cmd_lock);
2007 
2008 	its_cmd_wait_completion(dev, cmd, cmd_write);
2009 
2010 	return (0);
2011 }
2012 
2013 /* Handlers to send commands */
2014 static void
2015 its_cmd_movi(device_t dev, struct gicv3_its_irqsrc *girq)
2016 {
2017 	struct gicv3_its_softc *sc;
2018 	struct its_cmd_desc desc;
2019 	struct its_col *col;
2020 
2021 	sc = device_get_softc(dev);
2022 	col = sc->sc_its_cols[CPU_FFS(&girq->gi_isrc.isrc_cpu) - 1];
2023 
2024 	desc.cmd_type = ITS_CMD_MOVI;
2025 	desc.cmd_desc_movi.its_dev = girq->gi_its_dev;
2026 	desc.cmd_desc_movi.col = col;
2027 	desc.cmd_desc_movi.id = girq->gi_id;
2028 
2029 	its_cmd_send(dev, &desc);
2030 }
2031 
2032 static void
2033 its_cmd_mapc(device_t dev, struct its_col *col, uint8_t valid)
2034 {
2035 	struct its_cmd_desc desc;
2036 
2037 	desc.cmd_type = ITS_CMD_MAPC;
2038 	desc.cmd_desc_mapc.col = col;
2039 	/*
2040 	 * Valid bit set - map the collection.
2041 	 * Valid bit cleared - unmap the collection.
2042 	 */
2043 	desc.cmd_desc_mapc.valid = valid;
2044 
2045 	its_cmd_send(dev, &desc);
2046 }
2047 
2048 static void
2049 its_cmd_mapti(device_t dev, struct gicv3_its_irqsrc *girq)
2050 {
2051 	struct gicv3_its_softc *sc;
2052 	struct its_cmd_desc desc;
2053 	struct its_col *col;
2054 	u_int col_id;
2055 
2056 	sc = device_get_softc(dev);
2057 
2058 	col_id = CPU_FFS(&girq->gi_isrc.isrc_cpu) - 1;
2059 	col = sc->sc_its_cols[col_id];
2060 
2061 	desc.cmd_type = ITS_CMD_MAPTI;
2062 	desc.cmd_desc_mapvi.its_dev = girq->gi_its_dev;
2063 	desc.cmd_desc_mapvi.col = col;
2064 	/* The EventID sent to the device */
2065 	desc.cmd_desc_mapvi.id = girq->gi_id;
2066 	/* The physical interrupt presented to softeware */
2067 	desc.cmd_desc_mapvi.pid = girq->gi_lpi + GIC_FIRST_LPI;
2068 
2069 	its_cmd_send(dev, &desc);
2070 }
2071 
2072 static void
2073 its_cmd_mapd(device_t dev, struct its_dev *its_dev, uint8_t valid)
2074 {
2075 	struct its_cmd_desc desc;
2076 
2077 	desc.cmd_type = ITS_CMD_MAPD;
2078 	desc.cmd_desc_mapd.its_dev = its_dev;
2079 	desc.cmd_desc_mapd.valid = valid;
2080 
2081 	its_cmd_send(dev, &desc);
2082 }
2083 
2084 static void
2085 its_cmd_inv(device_t dev, struct its_dev *its_dev,
2086     struct gicv3_its_irqsrc *girq)
2087 {
2088 	struct gicv3_its_softc *sc;
2089 	struct its_cmd_desc desc;
2090 	struct its_col *col;
2091 
2092 	sc = device_get_softc(dev);
2093 	col = sc->sc_its_cols[CPU_FFS(&girq->gi_isrc.isrc_cpu) - 1];
2094 
2095 	desc.cmd_type = ITS_CMD_INV;
2096 	/* The EventID sent to the device */
2097 	desc.cmd_desc_inv.pid = girq->gi_id;
2098 	desc.cmd_desc_inv.its_dev = its_dev;
2099 	desc.cmd_desc_inv.col = col;
2100 
2101 	its_cmd_send(dev, &desc);
2102 }
2103 
2104 static void
2105 its_cmd_invall(device_t dev, struct its_col *col)
2106 {
2107 	struct its_cmd_desc desc;
2108 
2109 	desc.cmd_type = ITS_CMD_INVALL;
2110 	desc.cmd_desc_invall.col = col;
2111 
2112 	its_cmd_send(dev, &desc);
2113 }
2114 
2115 #ifdef FDT
2116 static device_probe_t gicv3_its_fdt_probe;
2117 static device_attach_t gicv3_its_fdt_attach;
2118 
2119 static device_method_t gicv3_its_fdt_methods[] = {
2120 	/* Device interface */
2121 	DEVMETHOD(device_probe,		gicv3_its_fdt_probe),
2122 	DEVMETHOD(device_attach,	gicv3_its_fdt_attach),
2123 
2124 	/* End */
2125 	DEVMETHOD_END
2126 };
2127 
2128 #define its_baseclasses its_fdt_baseclasses
2129 DEFINE_CLASS_1(its, gicv3_its_fdt_driver, gicv3_its_fdt_methods,
2130     sizeof(struct gicv3_its_softc), gicv3_its_driver);
2131 #undef its_baseclasses
2132 
2133 EARLY_DRIVER_MODULE(its_fdt, gic, gicv3_its_fdt_driver, 0, 0,
2134     BUS_PASS_INTERRUPT + BUS_PASS_ORDER_MIDDLE);
2135 
2136 static int
2137 gicv3_its_fdt_probe(device_t dev)
2138 {
2139 
2140 	if (!ofw_bus_status_okay(dev))
2141 		return (ENXIO);
2142 
2143 	if (!ofw_bus_is_compatible(dev, "arm,gic-v3-its"))
2144 		return (ENXIO);
2145 
2146 	device_set_desc(dev, "ARM GIC Interrupt Translation Service");
2147 	return (BUS_PROBE_DEFAULT);
2148 }
2149 
2150 static int
2151 gicv3_its_fdt_attach(device_t dev)
2152 {
2153 	struct gicv3_its_softc *sc;
2154 	phandle_t xref;
2155 	int err;
2156 
2157 	sc = device_get_softc(dev);
2158 	sc->dev = dev;
2159 	err = gicv3_its_attach(dev);
2160 	if (err != 0)
2161 		return (err);
2162 
2163 	/* Register this device as a interrupt controller */
2164 	xref = OF_xref_from_node(ofw_bus_get_node(dev));
2165 	sc->sc_pic = intr_pic_register(dev, xref);
2166 	err = intr_pic_add_handler(device_get_parent(dev), sc->sc_pic,
2167 	    gicv3_its_intr, sc, sc->sc_irq_base, sc->sc_irq_length);
2168 	if (err != 0) {
2169 		device_printf(dev, "Failed to add PIC handler: %d\n", err);
2170 		return (err);
2171 	}
2172 
2173 	/* Register this device to handle MSI interrupts */
2174 	err = intr_msi_register(dev, xref);
2175 	if (err != 0) {
2176 		device_printf(dev, "Failed to register for MSIs: %d\n", err);
2177 		return (err);
2178 	}
2179 
2180 	return (0);
2181 }
2182 #endif
2183 
2184 #ifdef DEV_ACPI
2185 static device_probe_t gicv3_its_acpi_probe;
2186 static device_attach_t gicv3_its_acpi_attach;
2187 
2188 static device_method_t gicv3_its_acpi_methods[] = {
2189 	/* Device interface */
2190 	DEVMETHOD(device_probe,		gicv3_its_acpi_probe),
2191 	DEVMETHOD(device_attach,	gicv3_its_acpi_attach),
2192 
2193 	/* End */
2194 	DEVMETHOD_END
2195 };
2196 
2197 #define its_baseclasses its_acpi_baseclasses
2198 DEFINE_CLASS_1(its, gicv3_its_acpi_driver, gicv3_its_acpi_methods,
2199     sizeof(struct gicv3_its_softc), gicv3_its_driver);
2200 #undef its_baseclasses
2201 
2202 EARLY_DRIVER_MODULE(its_acpi, gic, gicv3_its_acpi_driver, 0, 0,
2203     BUS_PASS_INTERRUPT + BUS_PASS_ORDER_MIDDLE);
2204 
2205 static int
2206 gicv3_its_acpi_probe(device_t dev)
2207 {
2208 
2209 	if (gic_get_bus(dev) != GIC_BUS_ACPI)
2210 		return (EINVAL);
2211 
2212 	if (gic_get_hw_rev(dev) < 3)
2213 		return (EINVAL);
2214 
2215 	device_set_desc(dev, "ARM GIC Interrupt Translation Service");
2216 	return (BUS_PROBE_DEFAULT);
2217 }
2218 
2219 static int
2220 gicv3_its_acpi_attach(device_t dev)
2221 {
2222 	struct gicv3_its_softc *sc;
2223 	struct gic_v3_devinfo *di;
2224 	int err;
2225 
2226 	sc = device_get_softc(dev);
2227 	sc->dev = dev;
2228 	err = gicv3_its_attach(dev);
2229 	if (err != 0)
2230 		return (err);
2231 
2232 	di = device_get_ivars(dev);
2233 	sc->sc_pic = intr_pic_register(dev, di->msi_xref);
2234 	err = intr_pic_add_handler(device_get_parent(dev), sc->sc_pic,
2235 	    gicv3_its_intr, sc, sc->sc_irq_base, sc->sc_irq_length);
2236 	if (err != 0) {
2237 		device_printf(dev, "Failed to add PIC handler: %d\n", err);
2238 		return (err);
2239 	}
2240 
2241 	/* Register this device to handle MSI interrupts */
2242 	err = intr_msi_register(dev, di->msi_xref);
2243 	if (err != 0) {
2244 		device_printf(dev, "Failed to register for MSIs: %d\n", err);
2245 		return (err);
2246 	}
2247 
2248 	return (0);
2249 }
2250 #endif
2251