1 // SPDX-License-Identifier: GPL-2.0+
2 // Copyright 2017 IBM Corp.
3 #include <asm/pnv-ocxl.h>
4 #include <asm/opal.h>
5 #include <misc/ocxl-config.h>
6 #include "pci.h"
7 
8 #define PNV_OCXL_TL_P9_RECV_CAP		0x000000000000000Full
9 #define PNV_OCXL_ACTAG_MAX		64
10 /* PASIDs are 20-bit, but on P9, NPU can only handle 15 bits */
11 #define PNV_OCXL_PASID_BITS		15
12 #define PNV_OCXL_PASID_MAX		((1 << PNV_OCXL_PASID_BITS) - 1)
13 
14 #define AFU_PRESENT (1 << 31)
15 #define AFU_INDEX_MASK 0x3F000000
16 #define AFU_INDEX_SHIFT 24
17 #define ACTAG_MASK 0xFFF
18 
19 
20 struct actag_range {
21 	u16 start;
22 	u16 count;
23 };
24 
25 struct npu_link {
26 	struct list_head list;
27 	int domain;
28 	int bus;
29 	int dev;
30 	u16 fn_desired_actags[8];
31 	struct actag_range fn_actags[8];
32 	bool assignment_done;
33 };
34 static struct list_head links_list = LIST_HEAD_INIT(links_list);
35 static DEFINE_MUTEX(links_list_lock);
36 
37 
38 /*
39  * opencapi actags handling:
40  *
41  * When sending commands, the opencapi device references the memory
42  * context it's targeting with an 'actag', which is really an alias
43  * for a (BDF, pasid) combination. When it receives a command, the NPU
44  * must do a lookup of the actag to identify the memory context. The
45  * hardware supports a finite number of actags per link (64 for
46  * POWER9).
47  *
48  * The device can carry multiple functions, and each function can have
49  * multiple AFUs. Each AFU advertises in its config space the number
50  * of desired actags. The host must configure in the config space of
51  * the AFU how many actags the AFU is really allowed to use (which can
52  * be less than what the AFU desires).
53  *
54  * When a PCI function is probed by the driver, it has no visibility
55  * about the other PCI functions and how many actags they'd like,
56  * which makes it impossible to distribute actags fairly among AFUs.
57  *
58  * Unfortunately, the only way to know how many actags a function
59  * desires is by looking at the data for each AFU in the config space
60  * and add them up. Similarly, the only way to know how many actags
61  * all the functions of the physical device desire is by adding the
62  * previously computed function counts. Then we can match that against
63  * what the hardware supports.
64  *
65  * To get a comprehensive view, we use a 'pci fixup': at the end of
66  * PCI enumeration, each function counts how many actags its AFUs
67  * desire and we save it in a 'npu_link' structure, shared between all
68  * the PCI functions of a same device. Therefore, when the first
69  * function is probed by the driver, we can get an idea of the total
70  * count of desired actags for the device, and assign the actags to
71  * the AFUs, by pro-rating if needed.
72  */
73 
74 static int find_dvsec_from_pos(struct pci_dev *dev, int dvsec_id, int pos)
75 {
76 	int vsec = pos;
77 	u16 vendor, id;
78 
79 	while ((vsec = pci_find_next_ext_capability(dev, vsec,
80 						    OCXL_EXT_CAP_ID_DVSEC))) {
81 		pci_read_config_word(dev, vsec + OCXL_DVSEC_VENDOR_OFFSET,
82 				&vendor);
83 		pci_read_config_word(dev, vsec + OCXL_DVSEC_ID_OFFSET, &id);
84 		if (vendor == PCI_VENDOR_ID_IBM && id == dvsec_id)
85 			return vsec;
86 	}
87 	return 0;
88 }
89 
90 static int find_dvsec_afu_ctrl(struct pci_dev *dev, u8 afu_idx)
91 {
92 	int vsec = 0;
93 	u8 idx;
94 
95 	while ((vsec = find_dvsec_from_pos(dev, OCXL_DVSEC_AFU_CTRL_ID,
96 					   vsec))) {
97 		pci_read_config_byte(dev, vsec + OCXL_DVSEC_AFU_CTRL_AFU_IDX,
98 				&idx);
99 		if (idx == afu_idx)
100 			return vsec;
101 	}
102 	return 0;
103 }
104 
105 static int get_max_afu_index(struct pci_dev *dev, int *afu_idx)
106 {
107 	int pos;
108 	u32 val;
109 
110 	pos = find_dvsec_from_pos(dev, OCXL_DVSEC_FUNC_ID, 0);
111 	if (!pos)
112 		return -ESRCH;
113 
114 	pci_read_config_dword(dev, pos + OCXL_DVSEC_FUNC_OFF_INDEX, &val);
115 	if (val & AFU_PRESENT)
116 		*afu_idx = (val & AFU_INDEX_MASK) >> AFU_INDEX_SHIFT;
117 	else
118 		*afu_idx = -1;
119 	return 0;
120 }
121 
122 static int get_actag_count(struct pci_dev *dev, int afu_idx, int *actag)
123 {
124 	int pos;
125 	u16 actag_sup;
126 
127 	pos = find_dvsec_afu_ctrl(dev, afu_idx);
128 	if (!pos)
129 		return -ESRCH;
130 
131 	pci_read_config_word(dev, pos + OCXL_DVSEC_AFU_CTRL_ACTAG_SUP,
132 			&actag_sup);
133 	*actag = actag_sup & ACTAG_MASK;
134 	return 0;
135 }
136 
137 static struct npu_link *find_link(struct pci_dev *dev)
138 {
139 	struct npu_link *link;
140 
141 	list_for_each_entry(link, &links_list, list) {
142 		/* The functions of a device all share the same link */
143 		if (link->domain == pci_domain_nr(dev->bus) &&
144 			link->bus == dev->bus->number &&
145 			link->dev == PCI_SLOT(dev->devfn)) {
146 			return link;
147 		}
148 	}
149 
150 	/* link doesn't exist yet. Allocate one */
151 	link = kzalloc(sizeof(struct npu_link), GFP_KERNEL);
152 	if (!link)
153 		return NULL;
154 	link->domain = pci_domain_nr(dev->bus);
155 	link->bus = dev->bus->number;
156 	link->dev = PCI_SLOT(dev->devfn);
157 	list_add(&link->list, &links_list);
158 	return link;
159 }
160 
161 static void pnv_ocxl_fixup_actag(struct pci_dev *dev)
162 {
163 	struct pci_controller *hose = pci_bus_to_host(dev->bus);
164 	struct pnv_phb *phb = hose->private_data;
165 	struct npu_link *link;
166 	int rc, afu_idx = -1, i, actag;
167 
168 	if (!machine_is(powernv))
169 		return;
170 
171 	if (phb->type != PNV_PHB_NPU_OCAPI)
172 		return;
173 
174 	mutex_lock(&links_list_lock);
175 
176 	link = find_link(dev);
177 	if (!link) {
178 		dev_warn(&dev->dev, "couldn't update actag information\n");
179 		mutex_unlock(&links_list_lock);
180 		return;
181 	}
182 
183 	/*
184 	 * Check how many actags are desired for the AFUs under that
185 	 * function and add it to the count for the link
186 	 */
187 	rc = get_max_afu_index(dev, &afu_idx);
188 	if (rc) {
189 		/* Most likely an invalid config space */
190 		dev_dbg(&dev->dev, "couldn't find AFU information\n");
191 		afu_idx = -1;
192 	}
193 
194 	link->fn_desired_actags[PCI_FUNC(dev->devfn)] = 0;
195 	for (i = 0; i <= afu_idx; i++) {
196 		/*
197 		 * AFU index 'holes' are allowed. So don't fail if we
198 		 * can't read the actag info for an index
199 		 */
200 		rc = get_actag_count(dev, i, &actag);
201 		if (rc)
202 			continue;
203 		link->fn_desired_actags[PCI_FUNC(dev->devfn)] += actag;
204 	}
205 	dev_dbg(&dev->dev, "total actags for function: %d\n",
206 		link->fn_desired_actags[PCI_FUNC(dev->devfn)]);
207 
208 	mutex_unlock(&links_list_lock);
209 }
210 DECLARE_PCI_FIXUP_HEADER(PCI_ANY_ID, PCI_ANY_ID, pnv_ocxl_fixup_actag);
211 
212 static u16 assign_fn_actags(u16 desired, u16 total)
213 {
214 	u16 count;
215 
216 	if (total <= PNV_OCXL_ACTAG_MAX)
217 		count = desired;
218 	else
219 		count = PNV_OCXL_ACTAG_MAX * desired / total;
220 
221 	return count;
222 }
223 
224 static void assign_actags(struct npu_link *link)
225 {
226 	u16 actag_count, range_start = 0, total_desired = 0;
227 	int i;
228 
229 	for (i = 0; i < 8; i++)
230 		total_desired += link->fn_desired_actags[i];
231 
232 	for (i = 0; i < 8; i++) {
233 		if (link->fn_desired_actags[i]) {
234 			actag_count = assign_fn_actags(
235 				link->fn_desired_actags[i],
236 				total_desired);
237 			link->fn_actags[i].start = range_start;
238 			link->fn_actags[i].count = actag_count;
239 			range_start += actag_count;
240 			WARN_ON(range_start >= PNV_OCXL_ACTAG_MAX);
241 		}
242 		pr_debug("link %x:%x:%x fct %d actags: start=%d count=%d (desired=%d)\n",
243 			link->domain, link->bus, link->dev, i,
244 			link->fn_actags[i].start, link->fn_actags[i].count,
245 			link->fn_desired_actags[i]);
246 	}
247 	link->assignment_done = true;
248 }
249 
250 int pnv_ocxl_get_actag(struct pci_dev *dev, u16 *base, u16 *enabled,
251 		u16 *supported)
252 {
253 	struct npu_link *link;
254 
255 	mutex_lock(&links_list_lock);
256 
257 	link = find_link(dev);
258 	if (!link) {
259 		dev_err(&dev->dev, "actag information not found\n");
260 		mutex_unlock(&links_list_lock);
261 		return -ENODEV;
262 	}
263 	/*
264 	 * On p9, we only have 64 actags per link, so they must be
265 	 * shared by all the functions of the same adapter. We counted
266 	 * the desired actag counts during PCI enumeration, so that we
267 	 * can allocate a pro-rated number of actags to each function.
268 	 */
269 	if (!link->assignment_done)
270 		assign_actags(link);
271 
272 	*base      = link->fn_actags[PCI_FUNC(dev->devfn)].start;
273 	*enabled   = link->fn_actags[PCI_FUNC(dev->devfn)].count;
274 	*supported = link->fn_desired_actags[PCI_FUNC(dev->devfn)];
275 
276 	mutex_unlock(&links_list_lock);
277 	return 0;
278 }
279 EXPORT_SYMBOL_GPL(pnv_ocxl_get_actag);
280 
281 int pnv_ocxl_get_pasid_count(struct pci_dev *dev, int *count)
282 {
283 	struct npu_link *link;
284 	int i, rc = -EINVAL;
285 
286 	/*
287 	 * The number of PASIDs (process address space ID) which can
288 	 * be used by a function depends on how many functions exist
289 	 * on the device. The NPU needs to be configured to know how
290 	 * many bits are available to PASIDs and how many are to be
291 	 * used by the function BDF indentifier.
292 	 *
293 	 * We only support one AFU-carrying function for now.
294 	 */
295 	mutex_lock(&links_list_lock);
296 
297 	link = find_link(dev);
298 	if (!link) {
299 		dev_err(&dev->dev, "actag information not found\n");
300 		mutex_unlock(&links_list_lock);
301 		return -ENODEV;
302 	}
303 
304 	for (i = 0; i < 8; i++)
305 		if (link->fn_desired_actags[i] && (i == PCI_FUNC(dev->devfn))) {
306 			*count = PNV_OCXL_PASID_MAX;
307 			rc = 0;
308 			break;
309 		}
310 
311 	mutex_unlock(&links_list_lock);
312 	dev_dbg(&dev->dev, "%d PASIDs available for function\n",
313 		rc ? 0 : *count);
314 	return rc;
315 }
316 EXPORT_SYMBOL_GPL(pnv_ocxl_get_pasid_count);
317 
318 static void set_templ_rate(unsigned int templ, unsigned int rate, char *buf)
319 {
320 	int shift, idx;
321 
322 	WARN_ON(templ > PNV_OCXL_TL_MAX_TEMPLATE);
323 	idx = (PNV_OCXL_TL_MAX_TEMPLATE - templ) / 2;
324 	shift = 4 * (1 - ((PNV_OCXL_TL_MAX_TEMPLATE - templ) % 2));
325 	buf[idx] |= rate << shift;
326 }
327 
328 int pnv_ocxl_get_tl_cap(struct pci_dev *dev, long *cap,
329 			char *rate_buf, int rate_buf_size)
330 {
331 	if (rate_buf_size != PNV_OCXL_TL_RATE_BUF_SIZE)
332 		return -EINVAL;
333 	/*
334 	 * The TL capabilities are a characteristic of the NPU, so
335 	 * we go with hard-coded values.
336 	 *
337 	 * The receiving rate of each template is encoded on 4 bits.
338 	 *
339 	 * On P9:
340 	 * - templates 0 -> 3 are supported
341 	 * - templates 0, 1 and 3 have a 0 receiving rate
342 	 * - template 2 has receiving rate of 1 (extra cycle)
343 	 */
344 	memset(rate_buf, 0, rate_buf_size);
345 	set_templ_rate(2, 1, rate_buf);
346 	*cap = PNV_OCXL_TL_P9_RECV_CAP;
347 	return 0;
348 }
349 EXPORT_SYMBOL_GPL(pnv_ocxl_get_tl_cap);
350 
351 int pnv_ocxl_set_tl_conf(struct pci_dev *dev, long cap,
352 			uint64_t rate_buf_phys, int rate_buf_size)
353 {
354 	struct pci_controller *hose = pci_bus_to_host(dev->bus);
355 	struct pnv_phb *phb = hose->private_data;
356 	int rc;
357 
358 	if (rate_buf_size != PNV_OCXL_TL_RATE_BUF_SIZE)
359 		return -EINVAL;
360 
361 	rc = opal_npu_tl_set(phb->opal_id, dev->devfn, cap,
362 			rate_buf_phys, rate_buf_size);
363 	if (rc) {
364 		dev_err(&dev->dev, "Can't configure host TL: %d\n", rc);
365 		return -EINVAL;
366 	}
367 	return 0;
368 }
369 EXPORT_SYMBOL_GPL(pnv_ocxl_set_tl_conf);
370 
371 int pnv_ocxl_get_xsl_irq(struct pci_dev *dev, int *hwirq)
372 {
373 	int rc;
374 
375 	rc = of_property_read_u32(dev->dev.of_node, "ibm,opal-xsl-irq", hwirq);
376 	if (rc) {
377 		dev_err(&dev->dev,
378 			"Can't get translation interrupt for device\n");
379 		return rc;
380 	}
381 	return 0;
382 }
383 EXPORT_SYMBOL_GPL(pnv_ocxl_get_xsl_irq);
384 
385 void pnv_ocxl_unmap_xsl_regs(void __iomem *dsisr, void __iomem *dar,
386 			void __iomem *tfc, void __iomem *pe_handle)
387 {
388 	iounmap(dsisr);
389 	iounmap(dar);
390 	iounmap(tfc);
391 	iounmap(pe_handle);
392 }
393 EXPORT_SYMBOL_GPL(pnv_ocxl_unmap_xsl_regs);
394 
395 int pnv_ocxl_map_xsl_regs(struct pci_dev *dev, void __iomem **dsisr,
396 			void __iomem **dar, void __iomem **tfc,
397 			void __iomem **pe_handle)
398 {
399 	u64 reg;
400 	int i, j, rc = 0;
401 	void __iomem *regs[4];
402 
403 	/*
404 	 * opal stores the mmio addresses of the DSISR, DAR, TFC and
405 	 * PE_HANDLE registers in a device tree property, in that
406 	 * order
407 	 */
408 	for (i = 0; i < 4; i++) {
409 		rc = of_property_read_u64_index(dev->dev.of_node,
410 						"ibm,opal-xsl-mmio", i, &reg);
411 		if (rc)
412 			break;
413 		regs[i] = ioremap(reg, 8);
414 		if (!regs[i]) {
415 			rc = -EINVAL;
416 			break;
417 		}
418 	}
419 	if (rc) {
420 		dev_err(&dev->dev, "Can't map translation mmio registers\n");
421 		for (j = i - 1; j >= 0; j--)
422 			iounmap(regs[j]);
423 	} else {
424 		*dsisr = regs[0];
425 		*dar = regs[1];
426 		*tfc = regs[2];
427 		*pe_handle = regs[3];
428 	}
429 	return rc;
430 }
431 EXPORT_SYMBOL_GPL(pnv_ocxl_map_xsl_regs);
432 
433 struct spa_data {
434 	u64 phb_opal_id;
435 	u32 bdfn;
436 };
437 
438 int pnv_ocxl_spa_setup(struct pci_dev *dev, void *spa_mem, int PE_mask,
439 		void **platform_data)
440 {
441 	struct pci_controller *hose = pci_bus_to_host(dev->bus);
442 	struct pnv_phb *phb = hose->private_data;
443 	struct spa_data *data;
444 	u32 bdfn;
445 	int rc;
446 
447 	data = kzalloc(sizeof(*data), GFP_KERNEL);
448 	if (!data)
449 		return -ENOMEM;
450 
451 	bdfn = (dev->bus->number << 8) | dev->devfn;
452 	rc = opal_npu_spa_setup(phb->opal_id, bdfn, virt_to_phys(spa_mem),
453 				PE_mask);
454 	if (rc) {
455 		dev_err(&dev->dev, "Can't setup Shared Process Area: %d\n", rc);
456 		kfree(data);
457 		return rc;
458 	}
459 	data->phb_opal_id = phb->opal_id;
460 	data->bdfn = bdfn;
461 	*platform_data = (void *) data;
462 	return 0;
463 }
464 EXPORT_SYMBOL_GPL(pnv_ocxl_spa_setup);
465 
466 void pnv_ocxl_spa_release(void *platform_data)
467 {
468 	struct spa_data *data = (struct spa_data *) platform_data;
469 	int rc;
470 
471 	rc = opal_npu_spa_setup(data->phb_opal_id, data->bdfn, 0, 0);
472 	WARN_ON(rc);
473 	kfree(data);
474 }
475 EXPORT_SYMBOL_GPL(pnv_ocxl_spa_release);
476 
477 int pnv_ocxl_spa_remove_pe_from_cache(void *platform_data, int pe_handle)
478 {
479 	struct spa_data *data = (struct spa_data *) platform_data;
480 	int rc;
481 
482 	rc = opal_npu_spa_clear_cache(data->phb_opal_id, data->bdfn, pe_handle);
483 	return rc;
484 }
485 EXPORT_SYMBOL_GPL(pnv_ocxl_spa_remove_pe_from_cache);
486 
487 int pnv_ocxl_map_lpar(struct pci_dev *dev, uint64_t lparid,
488 		      uint64_t lpcr, void __iomem **arva)
489 {
490 	struct pci_controller *hose = pci_bus_to_host(dev->bus);
491 	struct pnv_phb *phb = hose->private_data;
492 	u64 mmio_atsd;
493 	int rc;
494 
495 	/* ATSD physical address.
496 	 * ATSD LAUNCH register: write access initiates a shoot down to
497 	 * initiate the TLB Invalidate command.
498 	 */
499 	rc = of_property_read_u64_index(hose->dn, "ibm,mmio-atsd",
500 					0, &mmio_atsd);
501 	if (rc) {
502 		dev_info(&dev->dev, "No available ATSD found\n");
503 		return rc;
504 	}
505 
506 	/* Assign a register set to a Logical Partition and MMIO ATSD
507 	 * LPARID register to the required value.
508 	 */
509 	rc = opal_npu_map_lpar(phb->opal_id, pci_dev_id(dev),
510 			       lparid, lpcr);
511 	if (rc) {
512 		dev_err(&dev->dev, "Error mapping device to LPAR: %d\n", rc);
513 		return rc;
514 	}
515 
516 	*arva = ioremap(mmio_atsd, 24);
517 	if (!(*arva)) {
518 		dev_warn(&dev->dev, "ioremap failed - mmio_atsd: %#llx\n", mmio_atsd);
519 		rc = -ENOMEM;
520 	}
521 
522 	return rc;
523 }
524 EXPORT_SYMBOL_GPL(pnv_ocxl_map_lpar);
525 
526 void pnv_ocxl_unmap_lpar(void __iomem *arva)
527 {
528 	iounmap(arva);
529 }
530 EXPORT_SYMBOL_GPL(pnv_ocxl_unmap_lpar);
531 
532 void pnv_ocxl_tlb_invalidate(void __iomem *arva,
533 			     unsigned long pid,
534 			     unsigned long addr,
535 			     unsigned long page_size)
536 {
537 	unsigned long timeout = jiffies + (HZ * PNV_OCXL_ATSD_TIMEOUT);
538 	u64 val = 0ull;
539 	int pend;
540 	u8 size;
541 
542 	if (!(arva))
543 		return;
544 
545 	if (addr) {
546 		/* load Abbreviated Virtual Address register with
547 		 * the necessary value
548 		 */
549 		val |= FIELD_PREP(PNV_OCXL_ATSD_AVA_AVA, addr >> (63-51));
550 		out_be64(arva + PNV_OCXL_ATSD_AVA, val);
551 	}
552 
553 	/* Write access initiates a shoot down to initiate the
554 	 * TLB Invalidate command
555 	 */
556 	val = PNV_OCXL_ATSD_LNCH_R;
557 	val |= FIELD_PREP(PNV_OCXL_ATSD_LNCH_RIC, 0b10);
558 	if (addr)
559 		val |= FIELD_PREP(PNV_OCXL_ATSD_LNCH_IS, 0b00);
560 	else {
561 		val |= FIELD_PREP(PNV_OCXL_ATSD_LNCH_IS, 0b01);
562 		val |= PNV_OCXL_ATSD_LNCH_OCAPI_SINGLETON;
563 	}
564 	val |= PNV_OCXL_ATSD_LNCH_PRS;
565 	/* Actual Page Size to be invalidated
566 	 * 000 4KB
567 	 * 101 64KB
568 	 * 001 2MB
569 	 * 010 1GB
570 	 */
571 	size = 0b101;
572 	if (page_size == 0x1000)
573 		size = 0b000;
574 	if (page_size == 0x200000)
575 		size = 0b001;
576 	if (page_size == 0x40000000)
577 		size = 0b010;
578 	val |= FIELD_PREP(PNV_OCXL_ATSD_LNCH_AP, size);
579 	val |= FIELD_PREP(PNV_OCXL_ATSD_LNCH_PID, pid);
580 	out_be64(arva + PNV_OCXL_ATSD_LNCH, val);
581 
582 	/* Poll the ATSD status register to determine when the
583 	 * TLB Invalidate has been completed.
584 	 */
585 	val = in_be64(arva + PNV_OCXL_ATSD_STAT);
586 	pend = val >> 63;
587 
588 	while (pend) {
589 		if (time_after_eq(jiffies, timeout)) {
590 			pr_err("%s - Timeout while reading XTS MMIO ATSD status register (val=%#llx, pidr=0x%lx)\n",
591 			       __func__, val, pid);
592 			return;
593 		}
594 		cpu_relax();
595 		val = in_be64(arva + PNV_OCXL_ATSD_STAT);
596 		pend = val >> 63;
597 	}
598 }
599 EXPORT_SYMBOL_GPL(pnv_ocxl_tlb_invalidate);
600