xref: /linux/drivers/misc/cxl/pci.c (revision b85d45947951d23cb22d90caecf4c1eb81342c96)
1 /*
2  * Copyright 2014 IBM Corp.
3  *
4  * This program is free software; you can redistribute it and/or
5  * modify it under the terms of the GNU General Public License
6  * as published by the Free Software Foundation; either version
7  * 2 of the License, or (at your option) any later version.
8  */
9 
10 #include <linux/pci_regs.h>
11 #include <linux/pci_ids.h>
12 #include <linux/device.h>
13 #include <linux/module.h>
14 #include <linux/kernel.h>
15 #include <linux/slab.h>
16 #include <linux/sort.h>
17 #include <linux/pci.h>
18 #include <linux/of.h>
19 #include <linux/delay.h>
20 #include <asm/opal.h>
21 #include <asm/msi_bitmap.h>
22 #include <asm/pci-bridge.h> /* for struct pci_controller */
23 #include <asm/pnv-pci.h>
24 #include <asm/io.h>
25 
26 #include "cxl.h"
27 #include <misc/cxl.h>
28 
29 
30 #define CXL_PCI_VSEC_ID	0x1280
31 #define CXL_VSEC_MIN_SIZE 0x80
32 
33 #define CXL_READ_VSEC_LENGTH(dev, vsec, dest)			\
34 	{							\
35 		pci_read_config_word(dev, vsec + 0x6, dest);	\
36 		*dest >>= 4;					\
37 	}
38 #define CXL_READ_VSEC_NAFUS(dev, vsec, dest) \
39 	pci_read_config_byte(dev, vsec + 0x8, dest)
40 
41 #define CXL_READ_VSEC_STATUS(dev, vsec, dest) \
42 	pci_read_config_byte(dev, vsec + 0x9, dest)
43 #define CXL_STATUS_SECOND_PORT  0x80
44 #define CXL_STATUS_MSI_X_FULL   0x40
45 #define CXL_STATUS_MSI_X_SINGLE 0x20
46 #define CXL_STATUS_FLASH_RW     0x08
47 #define CXL_STATUS_FLASH_RO     0x04
48 #define CXL_STATUS_LOADABLE_AFU 0x02
49 #define CXL_STATUS_LOADABLE_PSL 0x01
50 /* If we see these features we won't try to use the card */
51 #define CXL_UNSUPPORTED_FEATURES \
52 	(CXL_STATUS_MSI_X_FULL | CXL_STATUS_MSI_X_SINGLE)
53 
54 #define CXL_READ_VSEC_MODE_CONTROL(dev, vsec, dest) \
55 	pci_read_config_byte(dev, vsec + 0xa, dest)
56 #define CXL_WRITE_VSEC_MODE_CONTROL(dev, vsec, val) \
57 	pci_write_config_byte(dev, vsec + 0xa, val)
58 #define CXL_VSEC_PROTOCOL_MASK   0xe0
59 #define CXL_VSEC_PROTOCOL_1024TB 0x80
60 #define CXL_VSEC_PROTOCOL_512TB  0x40
61 #define CXL_VSEC_PROTOCOL_256TB  0x20 /* Power 8 uses this */
62 #define CXL_VSEC_PROTOCOL_ENABLE 0x01
63 
64 #define CXL_READ_VSEC_PSL_REVISION(dev, vsec, dest) \
65 	pci_read_config_word(dev, vsec + 0xc, dest)
66 #define CXL_READ_VSEC_CAIA_MINOR(dev, vsec, dest) \
67 	pci_read_config_byte(dev, vsec + 0xe, dest)
68 #define CXL_READ_VSEC_CAIA_MAJOR(dev, vsec, dest) \
69 	pci_read_config_byte(dev, vsec + 0xf, dest)
70 #define CXL_READ_VSEC_BASE_IMAGE(dev, vsec, dest) \
71 	pci_read_config_word(dev, vsec + 0x10, dest)
72 
73 #define CXL_READ_VSEC_IMAGE_STATE(dev, vsec, dest) \
74 	pci_read_config_byte(dev, vsec + 0x13, dest)
75 #define CXL_WRITE_VSEC_IMAGE_STATE(dev, vsec, val) \
76 	pci_write_config_byte(dev, vsec + 0x13, val)
77 #define CXL_VSEC_USER_IMAGE_LOADED 0x80 /* RO */
78 #define CXL_VSEC_PERST_LOADS_IMAGE 0x20 /* RW */
79 #define CXL_VSEC_PERST_SELECT_USER 0x10 /* RW */
80 
81 #define CXL_READ_VSEC_AFU_DESC_OFF(dev, vsec, dest) \
82 	pci_read_config_dword(dev, vsec + 0x20, dest)
83 #define CXL_READ_VSEC_AFU_DESC_SIZE(dev, vsec, dest) \
84 	pci_read_config_dword(dev, vsec + 0x24, dest)
85 #define CXL_READ_VSEC_PS_OFF(dev, vsec, dest) \
86 	pci_read_config_dword(dev, vsec + 0x28, dest)
87 #define CXL_READ_VSEC_PS_SIZE(dev, vsec, dest) \
88 	pci_read_config_dword(dev, vsec + 0x2c, dest)
89 
90 
91 /* This works a little different than the p1/p2 register accesses to make it
92  * easier to pull out individual fields */
93 #define AFUD_READ(afu, off)		in_be64(afu->afu_desc_mmio + off)
94 #define AFUD_READ_LE(afu, off)		in_le64(afu->afu_desc_mmio + off)
95 #define EXTRACT_PPC_BIT(val, bit)	(!!(val & PPC_BIT(bit)))
96 #define EXTRACT_PPC_BITS(val, bs, be)	((val & PPC_BITMASK(bs, be)) >> PPC_BITLSHIFT(be))
97 
98 #define AFUD_READ_INFO(afu)		AFUD_READ(afu, 0x0)
99 #define   AFUD_NUM_INTS_PER_PROC(val)	EXTRACT_PPC_BITS(val,  0, 15)
100 #define   AFUD_NUM_PROCS(val)		EXTRACT_PPC_BITS(val, 16, 31)
101 #define   AFUD_NUM_CRS(val)		EXTRACT_PPC_BITS(val, 32, 47)
102 #define   AFUD_MULTIMODE(val)		EXTRACT_PPC_BIT(val, 48)
103 #define   AFUD_PUSH_BLOCK_TRANSFER(val)	EXTRACT_PPC_BIT(val, 55)
104 #define   AFUD_DEDICATED_PROCESS(val)	EXTRACT_PPC_BIT(val, 59)
105 #define   AFUD_AFU_DIRECTED(val)	EXTRACT_PPC_BIT(val, 61)
106 #define   AFUD_TIME_SLICED(val)		EXTRACT_PPC_BIT(val, 63)
107 #define AFUD_READ_CR(afu)		AFUD_READ(afu, 0x20)
108 #define   AFUD_CR_LEN(val)		EXTRACT_PPC_BITS(val, 8, 63)
109 #define AFUD_READ_CR_OFF(afu)		AFUD_READ(afu, 0x28)
110 #define AFUD_READ_PPPSA(afu)		AFUD_READ(afu, 0x30)
111 #define   AFUD_PPPSA_PP(val)		EXTRACT_PPC_BIT(val, 6)
112 #define   AFUD_PPPSA_PSA(val)		EXTRACT_PPC_BIT(val, 7)
113 #define   AFUD_PPPSA_LEN(val)		EXTRACT_PPC_BITS(val, 8, 63)
114 #define AFUD_READ_PPPSA_OFF(afu)	AFUD_READ(afu, 0x38)
115 #define AFUD_READ_EB(afu)		AFUD_READ(afu, 0x40)
116 #define   AFUD_EB_LEN(val)		EXTRACT_PPC_BITS(val, 8, 63)
117 #define AFUD_READ_EB_OFF(afu)		AFUD_READ(afu, 0x48)
118 
119 u16 cxl_afu_cr_read16(struct cxl_afu *afu, int cr, u64 off)
120 {
121 	u64 aligned_off = off & ~0x3L;
122 	u32 val;
123 
124 	val = cxl_afu_cr_read32(afu, cr, aligned_off);
125 	return (val >> ((off & 0x2) * 8)) & 0xffff;
126 }
127 
128 u8 cxl_afu_cr_read8(struct cxl_afu *afu, int cr, u64 off)
129 {
130 	u64 aligned_off = off & ~0x3L;
131 	u32 val;
132 
133 	val = cxl_afu_cr_read32(afu, cr, aligned_off);
134 	return (val >> ((off & 0x3) * 8)) & 0xff;
135 }
136 
137 static const struct pci_device_id cxl_pci_tbl[] = {
138 	{ PCI_DEVICE(PCI_VENDOR_ID_IBM, 0x0477), },
139 	{ PCI_DEVICE(PCI_VENDOR_ID_IBM, 0x044b), },
140 	{ PCI_DEVICE(PCI_VENDOR_ID_IBM, 0x04cf), },
141 	{ PCI_DEVICE_CLASS(0x120000, ~0), },
142 
143 	{ }
144 };
145 MODULE_DEVICE_TABLE(pci, cxl_pci_tbl);
146 
147 
148 /*
149  * Mostly using these wrappers to avoid confusion:
150  * priv 1 is BAR2, while priv 2 is BAR0
151  */
152 static inline resource_size_t p1_base(struct pci_dev *dev)
153 {
154 	return pci_resource_start(dev, 2);
155 }
156 
157 static inline resource_size_t p1_size(struct pci_dev *dev)
158 {
159 	return pci_resource_len(dev, 2);
160 }
161 
162 static inline resource_size_t p2_base(struct pci_dev *dev)
163 {
164 	return pci_resource_start(dev, 0);
165 }
166 
167 static inline resource_size_t p2_size(struct pci_dev *dev)
168 {
169 	return pci_resource_len(dev, 0);
170 }
171 
172 static int find_cxl_vsec(struct pci_dev *dev)
173 {
174 	int vsec = 0;
175 	u16 val;
176 
177 	while ((vsec = pci_find_next_ext_capability(dev, vsec, PCI_EXT_CAP_ID_VNDR))) {
178 		pci_read_config_word(dev, vsec + 0x4, &val);
179 		if (val == CXL_PCI_VSEC_ID)
180 			return vsec;
181 	}
182 	return 0;
183 
184 }
185 
186 static void dump_cxl_config_space(struct pci_dev *dev)
187 {
188 	int vsec;
189 	u32 val;
190 
191 	dev_info(&dev->dev, "dump_cxl_config_space\n");
192 
193 	pci_read_config_dword(dev, PCI_BASE_ADDRESS_0, &val);
194 	dev_info(&dev->dev, "BAR0: %#.8x\n", val);
195 	pci_read_config_dword(dev, PCI_BASE_ADDRESS_1, &val);
196 	dev_info(&dev->dev, "BAR1: %#.8x\n", val);
197 	pci_read_config_dword(dev, PCI_BASE_ADDRESS_2, &val);
198 	dev_info(&dev->dev, "BAR2: %#.8x\n", val);
199 	pci_read_config_dword(dev, PCI_BASE_ADDRESS_3, &val);
200 	dev_info(&dev->dev, "BAR3: %#.8x\n", val);
201 	pci_read_config_dword(dev, PCI_BASE_ADDRESS_4, &val);
202 	dev_info(&dev->dev, "BAR4: %#.8x\n", val);
203 	pci_read_config_dword(dev, PCI_BASE_ADDRESS_5, &val);
204 	dev_info(&dev->dev, "BAR5: %#.8x\n", val);
205 
206 	dev_info(&dev->dev, "p1 regs: %#llx, len: %#llx\n",
207 		p1_base(dev), p1_size(dev));
208 	dev_info(&dev->dev, "p2 regs: %#llx, len: %#llx\n",
209 		p2_base(dev), p2_size(dev));
210 	dev_info(&dev->dev, "BAR 4/5: %#llx, len: %#llx\n",
211 		pci_resource_start(dev, 4), pci_resource_len(dev, 4));
212 
213 	if (!(vsec = find_cxl_vsec(dev)))
214 		return;
215 
216 #define show_reg(name, what) \
217 	dev_info(&dev->dev, "cxl vsec: %30s: %#x\n", name, what)
218 
219 	pci_read_config_dword(dev, vsec + 0x0, &val);
220 	show_reg("Cap ID", (val >> 0) & 0xffff);
221 	show_reg("Cap Ver", (val >> 16) & 0xf);
222 	show_reg("Next Cap Ptr", (val >> 20) & 0xfff);
223 	pci_read_config_dword(dev, vsec + 0x4, &val);
224 	show_reg("VSEC ID", (val >> 0) & 0xffff);
225 	show_reg("VSEC Rev", (val >> 16) & 0xf);
226 	show_reg("VSEC Length",	(val >> 20) & 0xfff);
227 	pci_read_config_dword(dev, vsec + 0x8, &val);
228 	show_reg("Num AFUs", (val >> 0) & 0xff);
229 	show_reg("Status", (val >> 8) & 0xff);
230 	show_reg("Mode Control", (val >> 16) & 0xff);
231 	show_reg("Reserved", (val >> 24) & 0xff);
232 	pci_read_config_dword(dev, vsec + 0xc, &val);
233 	show_reg("PSL Rev", (val >> 0) & 0xffff);
234 	show_reg("CAIA Ver", (val >> 16) & 0xffff);
235 	pci_read_config_dword(dev, vsec + 0x10, &val);
236 	show_reg("Base Image Rev", (val >> 0) & 0xffff);
237 	show_reg("Reserved", (val >> 16) & 0x0fff);
238 	show_reg("Image Control", (val >> 28) & 0x3);
239 	show_reg("Reserved", (val >> 30) & 0x1);
240 	show_reg("Image Loaded", (val >> 31) & 0x1);
241 
242 	pci_read_config_dword(dev, vsec + 0x14, &val);
243 	show_reg("Reserved", val);
244 	pci_read_config_dword(dev, vsec + 0x18, &val);
245 	show_reg("Reserved", val);
246 	pci_read_config_dword(dev, vsec + 0x1c, &val);
247 	show_reg("Reserved", val);
248 
249 	pci_read_config_dword(dev, vsec + 0x20, &val);
250 	show_reg("AFU Descriptor Offset", val);
251 	pci_read_config_dword(dev, vsec + 0x24, &val);
252 	show_reg("AFU Descriptor Size", val);
253 	pci_read_config_dword(dev, vsec + 0x28, &val);
254 	show_reg("Problem State Offset", val);
255 	pci_read_config_dword(dev, vsec + 0x2c, &val);
256 	show_reg("Problem State Size", val);
257 
258 	pci_read_config_dword(dev, vsec + 0x30, &val);
259 	show_reg("Reserved", val);
260 	pci_read_config_dword(dev, vsec + 0x34, &val);
261 	show_reg("Reserved", val);
262 	pci_read_config_dword(dev, vsec + 0x38, &val);
263 	show_reg("Reserved", val);
264 	pci_read_config_dword(dev, vsec + 0x3c, &val);
265 	show_reg("Reserved", val);
266 
267 	pci_read_config_dword(dev, vsec + 0x40, &val);
268 	show_reg("PSL Programming Port", val);
269 	pci_read_config_dword(dev, vsec + 0x44, &val);
270 	show_reg("PSL Programming Control", val);
271 
272 	pci_read_config_dword(dev, vsec + 0x48, &val);
273 	show_reg("Reserved", val);
274 	pci_read_config_dword(dev, vsec + 0x4c, &val);
275 	show_reg("Reserved", val);
276 
277 	pci_read_config_dword(dev, vsec + 0x50, &val);
278 	show_reg("Flash Address Register", val);
279 	pci_read_config_dword(dev, vsec + 0x54, &val);
280 	show_reg("Flash Size Register", val);
281 	pci_read_config_dword(dev, vsec + 0x58, &val);
282 	show_reg("Flash Status/Control Register", val);
283 	pci_read_config_dword(dev, vsec + 0x58, &val);
284 	show_reg("Flash Data Port", val);
285 
286 #undef show_reg
287 }
288 
289 static void dump_afu_descriptor(struct cxl_afu *afu)
290 {
291 	u64 val, afu_cr_num, afu_cr_off, afu_cr_len;
292 	int i;
293 
294 #define show_reg(name, what) \
295 	dev_info(&afu->dev, "afu desc: %30s: %#llx\n", name, what)
296 
297 	val = AFUD_READ_INFO(afu);
298 	show_reg("num_ints_per_process", AFUD_NUM_INTS_PER_PROC(val));
299 	show_reg("num_of_processes", AFUD_NUM_PROCS(val));
300 	show_reg("num_of_afu_CRs", AFUD_NUM_CRS(val));
301 	show_reg("req_prog_mode", val & 0xffffULL);
302 	afu_cr_num = AFUD_NUM_CRS(val);
303 
304 	val = AFUD_READ(afu, 0x8);
305 	show_reg("Reserved", val);
306 	val = AFUD_READ(afu, 0x10);
307 	show_reg("Reserved", val);
308 	val = AFUD_READ(afu, 0x18);
309 	show_reg("Reserved", val);
310 
311 	val = AFUD_READ_CR(afu);
312 	show_reg("Reserved", (val >> (63-7)) & 0xff);
313 	show_reg("AFU_CR_len", AFUD_CR_LEN(val));
314 	afu_cr_len = AFUD_CR_LEN(val) * 256;
315 
316 	val = AFUD_READ_CR_OFF(afu);
317 	afu_cr_off = val;
318 	show_reg("AFU_CR_offset", val);
319 
320 	val = AFUD_READ_PPPSA(afu);
321 	show_reg("PerProcessPSA_control", (val >> (63-7)) & 0xff);
322 	show_reg("PerProcessPSA Length", AFUD_PPPSA_LEN(val));
323 
324 	val = AFUD_READ_PPPSA_OFF(afu);
325 	show_reg("PerProcessPSA_offset", val);
326 
327 	val = AFUD_READ_EB(afu);
328 	show_reg("Reserved", (val >> (63-7)) & 0xff);
329 	show_reg("AFU_EB_len", AFUD_EB_LEN(val));
330 
331 	val = AFUD_READ_EB_OFF(afu);
332 	show_reg("AFU_EB_offset", val);
333 
334 	for (i = 0; i < afu_cr_num; i++) {
335 		val = AFUD_READ_LE(afu, afu_cr_off + i * afu_cr_len);
336 		show_reg("CR Vendor", val & 0xffff);
337 		show_reg("CR Device", (val >> 16) & 0xffff);
338 	}
339 #undef show_reg
340 }
341 
342 static int init_implementation_adapter_regs(struct cxl *adapter, struct pci_dev *dev)
343 {
344 	struct device_node *np;
345 	const __be32 *prop;
346 	u64 psl_dsnctl;
347 	u64 chipid;
348 
349 	if (!(np = pnv_pci_get_phb_node(dev)))
350 		return -ENODEV;
351 
352 	while (np && !(prop = of_get_property(np, "ibm,chip-id", NULL)))
353 		np = of_get_next_parent(np);
354 	if (!np)
355 		return -ENODEV;
356 	chipid = be32_to_cpup(prop);
357 	of_node_put(np);
358 
359 	/* Tell PSL where to route data to */
360 	psl_dsnctl = 0x02E8900002000000ULL | (chipid << (63-5));
361 	cxl_p1_write(adapter, CXL_PSL_DSNDCTL, psl_dsnctl);
362 	cxl_p1_write(adapter, CXL_PSL_RESLCKTO, 0x20000000200ULL);
363 	/* snoop write mask */
364 	cxl_p1_write(adapter, CXL_PSL_SNWRALLOC, 0x00000000FFFFFFFFULL);
365 	/* set fir_accum */
366 	cxl_p1_write(adapter, CXL_PSL_FIR_CNTL, 0x0800000000000000ULL);
367 	/* for debugging with trace arrays */
368 	cxl_p1_write(adapter, CXL_PSL_TRACE, 0x0000FF7C00000000ULL);
369 
370 	return 0;
371 }
372 
373 #define TBSYNC_CNT(n) (((u64)n & 0x7) << (63-6))
374 #define _2048_250MHZ_CYCLES 1
375 
376 static int cxl_setup_psl_timebase(struct cxl *adapter, struct pci_dev *dev)
377 {
378 	u64 psl_tb;
379 	int delta;
380 	unsigned int retry = 0;
381 	struct device_node *np;
382 
383 	if (!(np = pnv_pci_get_phb_node(dev)))
384 		return -ENODEV;
385 
386 	/* Do not fail when CAPP timebase sync is not supported by OPAL */
387 	of_node_get(np);
388 	if (! of_get_property(np, "ibm,capp-timebase-sync", NULL)) {
389 		of_node_put(np);
390 		pr_err("PSL: Timebase sync: OPAL support missing\n");
391 		return 0;
392 	}
393 	of_node_put(np);
394 
395 	/*
396 	 * Setup PSL Timebase Control and Status register
397 	 * with the recommended Timebase Sync Count value
398 	 */
399 	cxl_p1_write(adapter, CXL_PSL_TB_CTLSTAT,
400 		     TBSYNC_CNT(2 * _2048_250MHZ_CYCLES));
401 
402 	/* Enable PSL Timebase */
403 	cxl_p1_write(adapter, CXL_PSL_Control, 0x0000000000000000);
404 	cxl_p1_write(adapter, CXL_PSL_Control, CXL_PSL_Control_tb);
405 
406 	/* Wait until CORE TB and PSL TB difference <= 16usecs */
407 	do {
408 		msleep(1);
409 		if (retry++ > 5) {
410 			pr_err("PSL: Timebase sync: giving up!\n");
411 			return -EIO;
412 		}
413 		psl_tb = cxl_p1_read(adapter, CXL_PSL_Timebase);
414 		delta = mftb() - psl_tb;
415 		if (delta < 0)
416 			delta = -delta;
417 	} while (cputime_to_usecs(delta) > 16);
418 
419 	return 0;
420 }
421 
422 static int init_implementation_afu_regs(struct cxl_afu *afu)
423 {
424 	/* read/write masks for this slice */
425 	cxl_p1n_write(afu, CXL_PSL_APCALLOC_A, 0xFFFFFFFEFEFEFEFEULL);
426 	/* APC read/write masks for this slice */
427 	cxl_p1n_write(afu, CXL_PSL_COALLOC_A, 0xFF000000FEFEFEFEULL);
428 	/* for debugging with trace arrays */
429 	cxl_p1n_write(afu, CXL_PSL_SLICE_TRACE, 0x0000FFFF00000000ULL);
430 	cxl_p1n_write(afu, CXL_PSL_RXCTL_A, CXL_PSL_RXCTL_AFUHP_4S);
431 
432 	return 0;
433 }
434 
435 int cxl_setup_irq(struct cxl *adapter, unsigned int hwirq,
436 			 unsigned int virq)
437 {
438 	struct pci_dev *dev = to_pci_dev(adapter->dev.parent);
439 
440 	return pnv_cxl_ioda_msi_setup(dev, hwirq, virq);
441 }
442 
443 int cxl_update_image_control(struct cxl *adapter)
444 {
445 	struct pci_dev *dev = to_pci_dev(adapter->dev.parent);
446 	int rc;
447 	int vsec;
448 	u8 image_state;
449 
450 	if (!(vsec = find_cxl_vsec(dev))) {
451 		dev_err(&dev->dev, "ABORTING: CXL VSEC not found!\n");
452 		return -ENODEV;
453 	}
454 
455 	if ((rc = CXL_READ_VSEC_IMAGE_STATE(dev, vsec, &image_state))) {
456 		dev_err(&dev->dev, "failed to read image state: %i\n", rc);
457 		return rc;
458 	}
459 
460 	if (adapter->perst_loads_image)
461 		image_state |= CXL_VSEC_PERST_LOADS_IMAGE;
462 	else
463 		image_state &= ~CXL_VSEC_PERST_LOADS_IMAGE;
464 
465 	if (adapter->perst_select_user)
466 		image_state |= CXL_VSEC_PERST_SELECT_USER;
467 	else
468 		image_state &= ~CXL_VSEC_PERST_SELECT_USER;
469 
470 	if ((rc = CXL_WRITE_VSEC_IMAGE_STATE(dev, vsec, image_state))) {
471 		dev_err(&dev->dev, "failed to update image control: %i\n", rc);
472 		return rc;
473 	}
474 
475 	return 0;
476 }
477 
478 int cxl_alloc_one_irq(struct cxl *adapter)
479 {
480 	struct pci_dev *dev = to_pci_dev(adapter->dev.parent);
481 
482 	return pnv_cxl_alloc_hwirqs(dev, 1);
483 }
484 
485 void cxl_release_one_irq(struct cxl *adapter, int hwirq)
486 {
487 	struct pci_dev *dev = to_pci_dev(adapter->dev.parent);
488 
489 	return pnv_cxl_release_hwirqs(dev, hwirq, 1);
490 }
491 
492 int cxl_alloc_irq_ranges(struct cxl_irq_ranges *irqs, struct cxl *adapter, unsigned int num)
493 {
494 	struct pci_dev *dev = to_pci_dev(adapter->dev.parent);
495 
496 	return pnv_cxl_alloc_hwirq_ranges(irqs, dev, num);
497 }
498 
499 void cxl_release_irq_ranges(struct cxl_irq_ranges *irqs, struct cxl *adapter)
500 {
501 	struct pci_dev *dev = to_pci_dev(adapter->dev.parent);
502 
503 	pnv_cxl_release_hwirq_ranges(irqs, dev);
504 }
505 
506 static int setup_cxl_bars(struct pci_dev *dev)
507 {
508 	/* Safety check in case we get backported to < 3.17 without M64 */
509 	if ((p1_base(dev) < 0x100000000ULL) ||
510 	    (p2_base(dev) < 0x100000000ULL)) {
511 		dev_err(&dev->dev, "ABORTING: M32 BAR assignment incompatible with CXL\n");
512 		return -ENODEV;
513 	}
514 
515 	/*
516 	 * BAR 4/5 has a special meaning for CXL and must be programmed with a
517 	 * special value corresponding to the CXL protocol address range.
518 	 * For POWER 8 that means bits 48:49 must be set to 10
519 	 */
520 	pci_write_config_dword(dev, PCI_BASE_ADDRESS_4, 0x00000000);
521 	pci_write_config_dword(dev, PCI_BASE_ADDRESS_5, 0x00020000);
522 
523 	return 0;
524 }
525 
526 /* pciex node: ibm,opal-m64-window = <0x3d058 0x0 0x3d058 0x0 0x8 0x0>; */
527 static int switch_card_to_cxl(struct pci_dev *dev)
528 {
529 	int vsec;
530 	u8 val;
531 	int rc;
532 
533 	dev_info(&dev->dev, "switch card to CXL\n");
534 
535 	if (!(vsec = find_cxl_vsec(dev))) {
536 		dev_err(&dev->dev, "ABORTING: CXL VSEC not found!\n");
537 		return -ENODEV;
538 	}
539 
540 	if ((rc = CXL_READ_VSEC_MODE_CONTROL(dev, vsec, &val))) {
541 		dev_err(&dev->dev, "failed to read current mode control: %i", rc);
542 		return rc;
543 	}
544 	val &= ~CXL_VSEC_PROTOCOL_MASK;
545 	val |= CXL_VSEC_PROTOCOL_256TB | CXL_VSEC_PROTOCOL_ENABLE;
546 	if ((rc = CXL_WRITE_VSEC_MODE_CONTROL(dev, vsec, val))) {
547 		dev_err(&dev->dev, "failed to enable CXL protocol: %i", rc);
548 		return rc;
549 	}
550 	/*
551 	 * The CAIA spec (v0.12 11.6 Bi-modal Device Support) states
552 	 * we must wait 100ms after this mode switch before touching
553 	 * PCIe config space.
554 	 */
555 	msleep(100);
556 
557 	return 0;
558 }
559 
560 static int cxl_map_slice_regs(struct cxl_afu *afu, struct cxl *adapter, struct pci_dev *dev)
561 {
562 	u64 p1n_base, p2n_base, afu_desc;
563 	const u64 p1n_size = 0x100;
564 	const u64 p2n_size = 0x1000;
565 
566 	p1n_base = p1_base(dev) + 0x10000 + (afu->slice * p1n_size);
567 	p2n_base = p2_base(dev) + (afu->slice * p2n_size);
568 	afu->psn_phys = p2_base(dev) + (adapter->ps_off + (afu->slice * adapter->ps_size));
569 	afu_desc = p2_base(dev) + adapter->afu_desc_off + (afu->slice * adapter->afu_desc_size);
570 
571 	if (!(afu->p1n_mmio = ioremap(p1n_base, p1n_size)))
572 		goto err;
573 	if (!(afu->p2n_mmio = ioremap(p2n_base, p2n_size)))
574 		goto err1;
575 	if (afu_desc) {
576 		if (!(afu->afu_desc_mmio = ioremap(afu_desc, adapter->afu_desc_size)))
577 			goto err2;
578 	}
579 
580 	return 0;
581 err2:
582 	iounmap(afu->p2n_mmio);
583 err1:
584 	iounmap(afu->p1n_mmio);
585 err:
586 	dev_err(&afu->dev, "Error mapping AFU MMIO regions\n");
587 	return -ENOMEM;
588 }
589 
590 static void cxl_unmap_slice_regs(struct cxl_afu *afu)
591 {
592 	if (afu->p2n_mmio) {
593 		iounmap(afu->p2n_mmio);
594 		afu->p2n_mmio = NULL;
595 	}
596 	if (afu->p1n_mmio) {
597 		iounmap(afu->p1n_mmio);
598 		afu->p1n_mmio = NULL;
599 	}
600 	if (afu->afu_desc_mmio) {
601 		iounmap(afu->afu_desc_mmio);
602 		afu->afu_desc_mmio = NULL;
603 	}
604 }
605 
606 static void cxl_release_afu(struct device *dev)
607 {
608 	struct cxl_afu *afu = to_cxl_afu(dev);
609 
610 	pr_devel("cxl_release_afu\n");
611 
612 	idr_destroy(&afu->contexts_idr);
613 	cxl_release_spa(afu);
614 
615 	kfree(afu);
616 }
617 
618 static struct cxl_afu *cxl_alloc_afu(struct cxl *adapter, int slice)
619 {
620 	struct cxl_afu *afu;
621 
622 	if (!(afu = kzalloc(sizeof(struct cxl_afu), GFP_KERNEL)))
623 		return NULL;
624 
625 	afu->adapter = adapter;
626 	afu->dev.parent = &adapter->dev;
627 	afu->dev.release = cxl_release_afu;
628 	afu->slice = slice;
629 	idr_init(&afu->contexts_idr);
630 	mutex_init(&afu->contexts_lock);
631 	spin_lock_init(&afu->afu_cntl_lock);
632 	mutex_init(&afu->spa_mutex);
633 
634 	afu->prefault_mode = CXL_PREFAULT_NONE;
635 	afu->irqs_max = afu->adapter->user_irqs;
636 
637 	return afu;
638 }
639 
640 /* Expects AFU struct to have recently been zeroed out */
641 static int cxl_read_afu_descriptor(struct cxl_afu *afu)
642 {
643 	u64 val;
644 
645 	val = AFUD_READ_INFO(afu);
646 	afu->pp_irqs = AFUD_NUM_INTS_PER_PROC(val);
647 	afu->max_procs_virtualised = AFUD_NUM_PROCS(val);
648 	afu->crs_num = AFUD_NUM_CRS(val);
649 
650 	if (AFUD_AFU_DIRECTED(val))
651 		afu->modes_supported |= CXL_MODE_DIRECTED;
652 	if (AFUD_DEDICATED_PROCESS(val))
653 		afu->modes_supported |= CXL_MODE_DEDICATED;
654 	if (AFUD_TIME_SLICED(val))
655 		afu->modes_supported |= CXL_MODE_TIME_SLICED;
656 
657 	val = AFUD_READ_PPPSA(afu);
658 	afu->pp_size = AFUD_PPPSA_LEN(val) * 4096;
659 	afu->psa = AFUD_PPPSA_PSA(val);
660 	if ((afu->pp_psa = AFUD_PPPSA_PP(val)))
661 		afu->pp_offset = AFUD_READ_PPPSA_OFF(afu);
662 
663 	val = AFUD_READ_CR(afu);
664 	afu->crs_len = AFUD_CR_LEN(val) * 256;
665 	afu->crs_offset = AFUD_READ_CR_OFF(afu);
666 
667 
668 	/* eb_len is in multiple of 4K */
669 	afu->eb_len = AFUD_EB_LEN(AFUD_READ_EB(afu)) * 4096;
670 	afu->eb_offset = AFUD_READ_EB_OFF(afu);
671 
672 	/* eb_off is 4K aligned so lower 12 bits are always zero */
673 	if (EXTRACT_PPC_BITS(afu->eb_offset, 0, 11) != 0) {
674 		dev_warn(&afu->dev,
675 			 "Invalid AFU error buffer offset %Lx\n",
676 			 afu->eb_offset);
677 		dev_info(&afu->dev,
678 			 "Ignoring AFU error buffer in the descriptor\n");
679 		/* indicate that no afu buffer exists */
680 		afu->eb_len = 0;
681 	}
682 
683 	return 0;
684 }
685 
686 static int cxl_afu_descriptor_looks_ok(struct cxl_afu *afu)
687 {
688 	int i;
689 
690 	if (afu->psa && afu->adapter->ps_size <
691 			(afu->pp_offset + afu->pp_size*afu->max_procs_virtualised)) {
692 		dev_err(&afu->dev, "per-process PSA can't fit inside the PSA!\n");
693 		return -ENODEV;
694 	}
695 
696 	if (afu->pp_psa && (afu->pp_size < PAGE_SIZE))
697 		dev_warn(&afu->dev, "AFU uses < PAGE_SIZE per-process PSA!");
698 
699 	for (i = 0; i < afu->crs_num; i++) {
700 		if ((cxl_afu_cr_read32(afu, i, 0) == 0)) {
701 			dev_err(&afu->dev, "ABORTING: AFU configuration record %i is invalid\n", i);
702 			return -EINVAL;
703 		}
704 	}
705 
706 	return 0;
707 }
708 
709 static int sanitise_afu_regs(struct cxl_afu *afu)
710 {
711 	u64 reg;
712 
713 	/*
714 	 * Clear out any regs that contain either an IVTE or address or may be
715 	 * waiting on an acknowledgement to try to be a bit safer as we bring
716 	 * it online
717 	 */
718 	reg = cxl_p2n_read(afu, CXL_AFU_Cntl_An);
719 	if ((reg & CXL_AFU_Cntl_An_ES_MASK) != CXL_AFU_Cntl_An_ES_Disabled) {
720 		dev_warn(&afu->dev, "WARNING: AFU was not disabled: %#016llx\n", reg);
721 		if (__cxl_afu_reset(afu))
722 			return -EIO;
723 		if (cxl_afu_disable(afu))
724 			return -EIO;
725 		if (cxl_psl_purge(afu))
726 			return -EIO;
727 	}
728 	cxl_p1n_write(afu, CXL_PSL_SPAP_An, 0x0000000000000000);
729 	cxl_p1n_write(afu, CXL_PSL_IVTE_Limit_An, 0x0000000000000000);
730 	cxl_p1n_write(afu, CXL_PSL_IVTE_Offset_An, 0x0000000000000000);
731 	cxl_p1n_write(afu, CXL_PSL_AMBAR_An, 0x0000000000000000);
732 	cxl_p1n_write(afu, CXL_PSL_SPOffset_An, 0x0000000000000000);
733 	cxl_p1n_write(afu, CXL_HAURP_An, 0x0000000000000000);
734 	cxl_p2n_write(afu, CXL_CSRP_An, 0x0000000000000000);
735 	cxl_p2n_write(afu, CXL_AURP1_An, 0x0000000000000000);
736 	cxl_p2n_write(afu, CXL_AURP0_An, 0x0000000000000000);
737 	cxl_p2n_write(afu, CXL_SSTP1_An, 0x0000000000000000);
738 	cxl_p2n_write(afu, CXL_SSTP0_An, 0x0000000000000000);
739 	reg = cxl_p2n_read(afu, CXL_PSL_DSISR_An);
740 	if (reg) {
741 		dev_warn(&afu->dev, "AFU had pending DSISR: %#016llx\n", reg);
742 		if (reg & CXL_PSL_DSISR_TRANS)
743 			cxl_p2n_write(afu, CXL_PSL_TFC_An, CXL_PSL_TFC_An_AE);
744 		else
745 			cxl_p2n_write(afu, CXL_PSL_TFC_An, CXL_PSL_TFC_An_A);
746 	}
747 	reg = cxl_p1n_read(afu, CXL_PSL_SERR_An);
748 	if (reg) {
749 		if (reg & ~0xffff)
750 			dev_warn(&afu->dev, "AFU had pending SERR: %#016llx\n", reg);
751 		cxl_p1n_write(afu, CXL_PSL_SERR_An, reg & ~0xffff);
752 	}
753 	reg = cxl_p2n_read(afu, CXL_PSL_ErrStat_An);
754 	if (reg) {
755 		dev_warn(&afu->dev, "AFU had pending error status: %#016llx\n", reg);
756 		cxl_p2n_write(afu, CXL_PSL_ErrStat_An, reg);
757 	}
758 
759 	return 0;
760 }
761 
762 #define ERR_BUFF_MAX_COPY_SIZE PAGE_SIZE
763 /*
764  * afu_eb_read:
765  * Called from sysfs and reads the afu error info buffer. The h/w only supports
766  * 4/8 bytes aligned access. So in case the requested offset/count arent 8 byte
767  * aligned the function uses a bounce buffer which can be max PAGE_SIZE.
768  */
769 ssize_t cxl_afu_read_err_buffer(struct cxl_afu *afu, char *buf,
770 				loff_t off, size_t count)
771 {
772 	loff_t aligned_start, aligned_end;
773 	size_t aligned_length;
774 	void *tbuf;
775 	const void __iomem *ebuf = afu->afu_desc_mmio + afu->eb_offset;
776 
777 	if (count == 0 || off < 0 || (size_t)off >= afu->eb_len)
778 		return 0;
779 
780 	/* calculate aligned read window */
781 	count = min((size_t)(afu->eb_len - off), count);
782 	aligned_start = round_down(off, 8);
783 	aligned_end = round_up(off + count, 8);
784 	aligned_length = aligned_end - aligned_start;
785 
786 	/* max we can copy in one read is PAGE_SIZE */
787 	if (aligned_length > ERR_BUFF_MAX_COPY_SIZE) {
788 		aligned_length = ERR_BUFF_MAX_COPY_SIZE;
789 		count = ERR_BUFF_MAX_COPY_SIZE - (off & 0x7);
790 	}
791 
792 	/* use bounce buffer for copy */
793 	tbuf = (void *)__get_free_page(GFP_TEMPORARY);
794 	if (!tbuf)
795 		return -ENOMEM;
796 
797 	/* perform aligned read from the mmio region */
798 	memcpy_fromio(tbuf, ebuf + aligned_start, aligned_length);
799 	memcpy(buf, tbuf + (off & 0x7), count);
800 
801 	free_page((unsigned long)tbuf);
802 
803 	return count;
804 }
805 
806 static int cxl_configure_afu(struct cxl_afu *afu, struct cxl *adapter, struct pci_dev *dev)
807 {
808 	int rc;
809 
810 	if ((rc = cxl_map_slice_regs(afu, adapter, dev)))
811 		return rc;
812 
813 	if ((rc = sanitise_afu_regs(afu)))
814 		goto err1;
815 
816 	/* We need to reset the AFU before we can read the AFU descriptor */
817 	if ((rc = __cxl_afu_reset(afu)))
818 		goto err1;
819 
820 	if (cxl_verbose)
821 		dump_afu_descriptor(afu);
822 
823 	if ((rc = cxl_read_afu_descriptor(afu)))
824 		goto err1;
825 
826 	if ((rc = cxl_afu_descriptor_looks_ok(afu)))
827 		goto err1;
828 
829 	if ((rc = init_implementation_afu_regs(afu)))
830 		goto err1;
831 
832 	if ((rc = cxl_register_serr_irq(afu)))
833 		goto err1;
834 
835 	if ((rc = cxl_register_psl_irq(afu)))
836 		goto err2;
837 
838 	return 0;
839 
840 err2:
841 	cxl_release_serr_irq(afu);
842 err1:
843 	cxl_unmap_slice_regs(afu);
844 	return rc;
845 }
846 
847 static void cxl_deconfigure_afu(struct cxl_afu *afu)
848 {
849 	cxl_release_psl_irq(afu);
850 	cxl_release_serr_irq(afu);
851 	cxl_unmap_slice_regs(afu);
852 }
853 
854 static int cxl_init_afu(struct cxl *adapter, int slice, struct pci_dev *dev)
855 {
856 	struct cxl_afu *afu;
857 	int rc;
858 
859 	afu = cxl_alloc_afu(adapter, slice);
860 	if (!afu)
861 		return -ENOMEM;
862 
863 	rc = dev_set_name(&afu->dev, "afu%i.%i", adapter->adapter_num, slice);
864 	if (rc)
865 		goto err_free;
866 
867 	rc = cxl_configure_afu(afu, adapter, dev);
868 	if (rc)
869 		goto err_free;
870 
871 	/* Don't care if this fails */
872 	cxl_debugfs_afu_add(afu);
873 
874 	/*
875 	 * After we call this function we must not free the afu directly, even
876 	 * if it returns an error!
877 	 */
878 	if ((rc = cxl_register_afu(afu)))
879 		goto err_put1;
880 
881 	if ((rc = cxl_sysfs_afu_add(afu)))
882 		goto err_put1;
883 
884 	adapter->afu[afu->slice] = afu;
885 
886 	if ((rc = cxl_pci_vphb_add(afu)))
887 		dev_info(&afu->dev, "Can't register vPHB\n");
888 
889 	return 0;
890 
891 err_put1:
892 	cxl_deconfigure_afu(afu);
893 	cxl_debugfs_afu_remove(afu);
894 	device_unregister(&afu->dev);
895 	return rc;
896 
897 err_free:
898 	kfree(afu);
899 	return rc;
900 
901 }
902 
903 static void cxl_remove_afu(struct cxl_afu *afu)
904 {
905 	pr_devel("cxl_remove_afu\n");
906 
907 	if (!afu)
908 		return;
909 
910 	cxl_sysfs_afu_remove(afu);
911 	cxl_debugfs_afu_remove(afu);
912 
913 	spin_lock(&afu->adapter->afu_list_lock);
914 	afu->adapter->afu[afu->slice] = NULL;
915 	spin_unlock(&afu->adapter->afu_list_lock);
916 
917 	cxl_context_detach_all(afu);
918 	cxl_afu_deactivate_mode(afu);
919 
920 	cxl_deconfigure_afu(afu);
921 	device_unregister(&afu->dev);
922 }
923 
924 int cxl_reset(struct cxl *adapter)
925 {
926 	struct pci_dev *dev = to_pci_dev(adapter->dev.parent);
927 	int rc;
928 
929 	if (adapter->perst_same_image) {
930 		dev_warn(&dev->dev,
931 			 "cxl: refusing to reset/reflash when perst_reloads_same_image is set.\n");
932 		return -EINVAL;
933 	}
934 
935 	dev_info(&dev->dev, "CXL reset\n");
936 
937 	/* pcie_warm_reset requests a fundamental pci reset which includes a
938 	 * PERST assert/deassert.  PERST triggers a loading of the image
939 	 * if "user" or "factory" is selected in sysfs */
940 	if ((rc = pci_set_pcie_reset_state(dev, pcie_warm_reset))) {
941 		dev_err(&dev->dev, "cxl: pcie_warm_reset failed\n");
942 		return rc;
943 	}
944 
945 	return rc;
946 }
947 
948 static int cxl_map_adapter_regs(struct cxl *adapter, struct pci_dev *dev)
949 {
950 	if (pci_request_region(dev, 2, "priv 2 regs"))
951 		goto err1;
952 	if (pci_request_region(dev, 0, "priv 1 regs"))
953 		goto err2;
954 
955 	pr_devel("cxl_map_adapter_regs: p1: %#016llx %#llx, p2: %#016llx %#llx",
956 			p1_base(dev), p1_size(dev), p2_base(dev), p2_size(dev));
957 
958 	if (!(adapter->p1_mmio = ioremap(p1_base(dev), p1_size(dev))))
959 		goto err3;
960 
961 	if (!(adapter->p2_mmio = ioremap(p2_base(dev), p2_size(dev))))
962 		goto err4;
963 
964 	return 0;
965 
966 err4:
967 	iounmap(adapter->p1_mmio);
968 	adapter->p1_mmio = NULL;
969 err3:
970 	pci_release_region(dev, 0);
971 err2:
972 	pci_release_region(dev, 2);
973 err1:
974 	return -ENOMEM;
975 }
976 
977 static void cxl_unmap_adapter_regs(struct cxl *adapter)
978 {
979 	if (adapter->p1_mmio) {
980 		iounmap(adapter->p1_mmio);
981 		adapter->p1_mmio = NULL;
982 		pci_release_region(to_pci_dev(adapter->dev.parent), 2);
983 	}
984 	if (adapter->p2_mmio) {
985 		iounmap(adapter->p2_mmio);
986 		adapter->p2_mmio = NULL;
987 		pci_release_region(to_pci_dev(adapter->dev.parent), 0);
988 	}
989 }
990 
991 static int cxl_read_vsec(struct cxl *adapter, struct pci_dev *dev)
992 {
993 	int vsec;
994 	u32 afu_desc_off, afu_desc_size;
995 	u32 ps_off, ps_size;
996 	u16 vseclen;
997 	u8 image_state;
998 
999 	if (!(vsec = find_cxl_vsec(dev))) {
1000 		dev_err(&dev->dev, "ABORTING: CXL VSEC not found!\n");
1001 		return -ENODEV;
1002 	}
1003 
1004 	CXL_READ_VSEC_LENGTH(dev, vsec, &vseclen);
1005 	if (vseclen < CXL_VSEC_MIN_SIZE) {
1006 		dev_err(&dev->dev, "ABORTING: CXL VSEC too short\n");
1007 		return -EINVAL;
1008 	}
1009 
1010 	CXL_READ_VSEC_STATUS(dev, vsec, &adapter->vsec_status);
1011 	CXL_READ_VSEC_PSL_REVISION(dev, vsec, &adapter->psl_rev);
1012 	CXL_READ_VSEC_CAIA_MAJOR(dev, vsec, &adapter->caia_major);
1013 	CXL_READ_VSEC_CAIA_MINOR(dev, vsec, &adapter->caia_minor);
1014 	CXL_READ_VSEC_BASE_IMAGE(dev, vsec, &adapter->base_image);
1015 	CXL_READ_VSEC_IMAGE_STATE(dev, vsec, &image_state);
1016 	adapter->user_image_loaded = !!(image_state & CXL_VSEC_USER_IMAGE_LOADED);
1017 	adapter->perst_select_user = !!(image_state & CXL_VSEC_USER_IMAGE_LOADED);
1018 
1019 	CXL_READ_VSEC_NAFUS(dev, vsec, &adapter->slices);
1020 	CXL_READ_VSEC_AFU_DESC_OFF(dev, vsec, &afu_desc_off);
1021 	CXL_READ_VSEC_AFU_DESC_SIZE(dev, vsec, &afu_desc_size);
1022 	CXL_READ_VSEC_PS_OFF(dev, vsec, &ps_off);
1023 	CXL_READ_VSEC_PS_SIZE(dev, vsec, &ps_size);
1024 
1025 	/* Convert everything to bytes, because there is NO WAY I'd look at the
1026 	 * code a month later and forget what units these are in ;-) */
1027 	adapter->ps_off = ps_off * 64 * 1024;
1028 	adapter->ps_size = ps_size * 64 * 1024;
1029 	adapter->afu_desc_off = afu_desc_off * 64 * 1024;
1030 	adapter->afu_desc_size = afu_desc_size *64 * 1024;
1031 
1032 	/* Total IRQs - 1 PSL ERROR - #AFU*(1 slice error + 1 DSI) */
1033 	adapter->user_irqs = pnv_cxl_get_irq_count(dev) - 1 - 2*adapter->slices;
1034 
1035 	return 0;
1036 }
1037 
1038 /*
1039  * Workaround a PCIe Host Bridge defect on some cards, that can cause
1040  * malformed Transaction Layer Packet (TLP) errors to be erroneously
1041  * reported. Mask this error in the Uncorrectable Error Mask Register.
1042  *
1043  * The upper nibble of the PSL revision is used to distinguish between
1044  * different cards. The affected ones have it set to 0.
1045  */
1046 static void cxl_fixup_malformed_tlp(struct cxl *adapter, struct pci_dev *dev)
1047 {
1048 	int aer;
1049 	u32 data;
1050 
1051 	if (adapter->psl_rev & 0xf000)
1052 		return;
1053 	if (!(aer = pci_find_ext_capability(dev, PCI_EXT_CAP_ID_ERR)))
1054 		return;
1055 	pci_read_config_dword(dev, aer + PCI_ERR_UNCOR_MASK, &data);
1056 	if (data & PCI_ERR_UNC_MALF_TLP)
1057 		if (data & PCI_ERR_UNC_INTN)
1058 			return;
1059 	data |= PCI_ERR_UNC_MALF_TLP;
1060 	data |= PCI_ERR_UNC_INTN;
1061 	pci_write_config_dword(dev, aer + PCI_ERR_UNCOR_MASK, data);
1062 }
1063 
1064 static int cxl_vsec_looks_ok(struct cxl *adapter, struct pci_dev *dev)
1065 {
1066 	if (adapter->vsec_status & CXL_STATUS_SECOND_PORT)
1067 		return -EBUSY;
1068 
1069 	if (adapter->vsec_status & CXL_UNSUPPORTED_FEATURES) {
1070 		dev_err(&dev->dev, "ABORTING: CXL requires unsupported features\n");
1071 		return -EINVAL;
1072 	}
1073 
1074 	if (!adapter->slices) {
1075 		/* Once we support dynamic reprogramming we can use the card if
1076 		 * it supports loadable AFUs */
1077 		dev_err(&dev->dev, "ABORTING: Device has no AFUs\n");
1078 		return -EINVAL;
1079 	}
1080 
1081 	if (!adapter->afu_desc_off || !adapter->afu_desc_size) {
1082 		dev_err(&dev->dev, "ABORTING: VSEC shows no AFU descriptors\n");
1083 		return -EINVAL;
1084 	}
1085 
1086 	if (adapter->ps_size > p2_size(dev) - adapter->ps_off) {
1087 		dev_err(&dev->dev, "ABORTING: Problem state size larger than "
1088 				   "available in BAR2: 0x%llx > 0x%llx\n",
1089 			 adapter->ps_size, p2_size(dev) - adapter->ps_off);
1090 		return -EINVAL;
1091 	}
1092 
1093 	return 0;
1094 }
1095 
1096 static void cxl_release_adapter(struct device *dev)
1097 {
1098 	struct cxl *adapter = to_cxl_adapter(dev);
1099 
1100 	pr_devel("cxl_release_adapter\n");
1101 
1102 	cxl_remove_adapter_nr(adapter);
1103 
1104 	kfree(adapter);
1105 }
1106 
1107 static struct cxl *cxl_alloc_adapter(void)
1108 {
1109 	struct cxl *adapter;
1110 
1111 	if (!(adapter = kzalloc(sizeof(struct cxl), GFP_KERNEL)))
1112 		return NULL;
1113 
1114 	spin_lock_init(&adapter->afu_list_lock);
1115 
1116 	if (cxl_alloc_adapter_nr(adapter))
1117 		goto err1;
1118 
1119 	if (dev_set_name(&adapter->dev, "card%i", adapter->adapter_num))
1120 		goto err2;
1121 
1122 	return adapter;
1123 
1124 err2:
1125 	cxl_remove_adapter_nr(adapter);
1126 err1:
1127 	kfree(adapter);
1128 	return NULL;
1129 }
1130 
1131 #define CXL_PSL_ErrIVTE_tberror (0x1ull << (63-31))
1132 
1133 static int sanitise_adapter_regs(struct cxl *adapter)
1134 {
1135 	/* Clear PSL tberror bit by writing 1 to it */
1136 	cxl_p1_write(adapter, CXL_PSL_ErrIVTE, CXL_PSL_ErrIVTE_tberror);
1137 	return cxl_tlb_slb_invalidate(adapter);
1138 }
1139 
1140 /* This should contain *only* operations that can safely be done in
1141  * both creation and recovery.
1142  */
1143 static int cxl_configure_adapter(struct cxl *adapter, struct pci_dev *dev)
1144 {
1145 	int rc;
1146 
1147 	adapter->dev.parent = &dev->dev;
1148 	adapter->dev.release = cxl_release_adapter;
1149 	pci_set_drvdata(dev, adapter);
1150 
1151 	rc = pci_enable_device(dev);
1152 	if (rc) {
1153 		dev_err(&dev->dev, "pci_enable_device failed: %i\n", rc);
1154 		return rc;
1155 	}
1156 
1157 	if ((rc = cxl_read_vsec(adapter, dev)))
1158 		return rc;
1159 
1160 	if ((rc = cxl_vsec_looks_ok(adapter, dev)))
1161 	        return rc;
1162 
1163 	cxl_fixup_malformed_tlp(adapter, dev);
1164 
1165 	if ((rc = setup_cxl_bars(dev)))
1166 		return rc;
1167 
1168 	if ((rc = switch_card_to_cxl(dev)))
1169 		return rc;
1170 
1171 	if ((rc = cxl_update_image_control(adapter)))
1172 		return rc;
1173 
1174 	if ((rc = cxl_map_adapter_regs(adapter, dev)))
1175 		return rc;
1176 
1177 	if ((rc = sanitise_adapter_regs(adapter)))
1178 		goto err;
1179 
1180 	if ((rc = init_implementation_adapter_regs(adapter, dev)))
1181 		goto err;
1182 
1183 	if ((rc = pnv_phb_to_cxl_mode(dev, OPAL_PHB_CAPI_MODE_CAPI)))
1184 		goto err;
1185 
1186 	/* If recovery happened, the last step is to turn on snooping.
1187 	 * In the non-recovery case this has no effect */
1188 	if ((rc = pnv_phb_to_cxl_mode(dev, OPAL_PHB_CAPI_MODE_SNOOP_ON)))
1189 		goto err;
1190 
1191 	if ((rc = cxl_setup_psl_timebase(adapter, dev)))
1192 		goto err;
1193 
1194 	if ((rc = cxl_register_psl_err_irq(adapter)))
1195 		goto err;
1196 
1197 	return 0;
1198 
1199 err:
1200 	cxl_unmap_adapter_regs(adapter);
1201 	return rc;
1202 
1203 }
1204 
1205 static void cxl_deconfigure_adapter(struct cxl *adapter)
1206 {
1207 	struct pci_dev *pdev = to_pci_dev(adapter->dev.parent);
1208 
1209 	cxl_release_psl_err_irq(adapter);
1210 	cxl_unmap_adapter_regs(adapter);
1211 
1212 	pci_disable_device(pdev);
1213 }
1214 
1215 static struct cxl *cxl_init_adapter(struct pci_dev *dev)
1216 {
1217 	struct cxl *adapter;
1218 	int rc;
1219 
1220 	adapter = cxl_alloc_adapter();
1221 	if (!adapter)
1222 		return ERR_PTR(-ENOMEM);
1223 
1224 	/* Set defaults for parameters which need to persist over
1225 	 * configure/reconfigure
1226 	 */
1227 	adapter->perst_loads_image = true;
1228 	adapter->perst_same_image = false;
1229 
1230 	rc = cxl_configure_adapter(adapter, dev);
1231 	if (rc) {
1232 		pci_disable_device(dev);
1233 		cxl_release_adapter(&adapter->dev);
1234 		return ERR_PTR(rc);
1235 	}
1236 
1237 	/* Don't care if this one fails: */
1238 	cxl_debugfs_adapter_add(adapter);
1239 
1240 	/*
1241 	 * After we call this function we must not free the adapter directly,
1242 	 * even if it returns an error!
1243 	 */
1244 	if ((rc = cxl_register_adapter(adapter)))
1245 		goto err_put1;
1246 
1247 	if ((rc = cxl_sysfs_adapter_add(adapter)))
1248 		goto err_put1;
1249 
1250 	return adapter;
1251 
1252 err_put1:
1253 	/* This should mirror cxl_remove_adapter, except without the
1254 	 * sysfs parts
1255 	 */
1256 	cxl_debugfs_adapter_remove(adapter);
1257 	cxl_deconfigure_adapter(adapter);
1258 	device_unregister(&adapter->dev);
1259 	return ERR_PTR(rc);
1260 }
1261 
1262 static void cxl_remove_adapter(struct cxl *adapter)
1263 {
1264 	pr_devel("cxl_remove_adapter\n");
1265 
1266 	cxl_sysfs_adapter_remove(adapter);
1267 	cxl_debugfs_adapter_remove(adapter);
1268 
1269 	cxl_deconfigure_adapter(adapter);
1270 
1271 	device_unregister(&adapter->dev);
1272 }
1273 
1274 static int cxl_probe(struct pci_dev *dev, const struct pci_device_id *id)
1275 {
1276 	struct cxl *adapter;
1277 	int slice;
1278 	int rc;
1279 
1280 	if (cxl_verbose)
1281 		dump_cxl_config_space(dev);
1282 
1283 	adapter = cxl_init_adapter(dev);
1284 	if (IS_ERR(adapter)) {
1285 		dev_err(&dev->dev, "cxl_init_adapter failed: %li\n", PTR_ERR(adapter));
1286 		return PTR_ERR(adapter);
1287 	}
1288 
1289 	for (slice = 0; slice < adapter->slices; slice++) {
1290 		if ((rc = cxl_init_afu(adapter, slice, dev))) {
1291 			dev_err(&dev->dev, "AFU %i failed to initialise: %i\n", slice, rc);
1292 			continue;
1293 		}
1294 
1295 		rc = cxl_afu_select_best_mode(adapter->afu[slice]);
1296 		if (rc)
1297 			dev_err(&dev->dev, "AFU %i failed to start: %i\n", slice, rc);
1298 	}
1299 
1300 	return 0;
1301 }
1302 
1303 static void cxl_remove(struct pci_dev *dev)
1304 {
1305 	struct cxl *adapter = pci_get_drvdata(dev);
1306 	struct cxl_afu *afu;
1307 	int i;
1308 
1309 	/*
1310 	 * Lock to prevent someone grabbing a ref through the adapter list as
1311 	 * we are removing it
1312 	 */
1313 	for (i = 0; i < adapter->slices; i++) {
1314 		afu = adapter->afu[i];
1315 		cxl_pci_vphb_remove(afu);
1316 		cxl_remove_afu(afu);
1317 	}
1318 	cxl_remove_adapter(adapter);
1319 }
1320 
1321 static pci_ers_result_t cxl_vphb_error_detected(struct cxl_afu *afu,
1322 						pci_channel_state_t state)
1323 {
1324 	struct pci_dev *afu_dev;
1325 	pci_ers_result_t result = PCI_ERS_RESULT_NEED_RESET;
1326 	pci_ers_result_t afu_result = PCI_ERS_RESULT_NEED_RESET;
1327 
1328 	/* There should only be one entry, but go through the list
1329 	 * anyway
1330 	 */
1331 	list_for_each_entry(afu_dev, &afu->phb->bus->devices, bus_list) {
1332 		if (!afu_dev->driver)
1333 			continue;
1334 
1335 		afu_dev->error_state = state;
1336 
1337 		if (afu_dev->driver->err_handler)
1338 			afu_result = afu_dev->driver->err_handler->error_detected(afu_dev,
1339 										  state);
1340 		/* Disconnect trumps all, NONE trumps NEED_RESET */
1341 		if (afu_result == PCI_ERS_RESULT_DISCONNECT)
1342 			result = PCI_ERS_RESULT_DISCONNECT;
1343 		else if ((afu_result == PCI_ERS_RESULT_NONE) &&
1344 			 (result == PCI_ERS_RESULT_NEED_RESET))
1345 			result = PCI_ERS_RESULT_NONE;
1346 	}
1347 	return result;
1348 }
1349 
1350 static pci_ers_result_t cxl_pci_error_detected(struct pci_dev *pdev,
1351 					       pci_channel_state_t state)
1352 {
1353 	struct cxl *adapter = pci_get_drvdata(pdev);
1354 	struct cxl_afu *afu;
1355 	pci_ers_result_t result = PCI_ERS_RESULT_NEED_RESET;
1356 	int i;
1357 
1358 	/* At this point, we could still have an interrupt pending.
1359 	 * Let's try to get them out of the way before they do
1360 	 * anything we don't like.
1361 	 */
1362 	schedule();
1363 
1364 	/* If we're permanently dead, give up. */
1365 	if (state == pci_channel_io_perm_failure) {
1366 		/* Tell the AFU drivers; but we don't care what they
1367 		 * say, we're going away.
1368 		 */
1369 		for (i = 0; i < adapter->slices; i++) {
1370 			afu = adapter->afu[i];
1371 			cxl_vphb_error_detected(afu, state);
1372 		}
1373 		return PCI_ERS_RESULT_DISCONNECT;
1374 	}
1375 
1376 	/* Are we reflashing?
1377 	 *
1378 	 * If we reflash, we could come back as something entirely
1379 	 * different, including a non-CAPI card. As such, by default
1380 	 * we don't participate in the process. We'll be unbound and
1381 	 * the slot re-probed. (TODO: check EEH doesn't blindly rebind
1382 	 * us!)
1383 	 *
1384 	 * However, this isn't the entire story: for reliablity
1385 	 * reasons, we usually want to reflash the FPGA on PERST in
1386 	 * order to get back to a more reliable known-good state.
1387 	 *
1388 	 * This causes us a bit of a problem: if we reflash we can't
1389 	 * trust that we'll come back the same - we could have a new
1390 	 * image and been PERSTed in order to load that
1391 	 * image. However, most of the time we actually *will* come
1392 	 * back the same - for example a regular EEH event.
1393 	 *
1394 	 * Therefore, we allow the user to assert that the image is
1395 	 * indeed the same and that we should continue on into EEH
1396 	 * anyway.
1397 	 */
1398 	if (adapter->perst_loads_image && !adapter->perst_same_image) {
1399 		/* TODO take the PHB out of CXL mode */
1400 		dev_info(&pdev->dev, "reflashing, so opting out of EEH!\n");
1401 		return PCI_ERS_RESULT_NONE;
1402 	}
1403 
1404 	/*
1405 	 * At this point, we want to try to recover.  We'll always
1406 	 * need a complete slot reset: we don't trust any other reset.
1407 	 *
1408 	 * Now, we go through each AFU:
1409 	 *  - We send the driver, if bound, an error_detected callback.
1410 	 *    We expect it to clean up, but it can also tell us to give
1411 	 *    up and permanently detach the card. To simplify things, if
1412 	 *    any bound AFU driver doesn't support EEH, we give up on EEH.
1413 	 *
1414 	 *  - We detach all contexts associated with the AFU. This
1415 	 *    does not free them, but puts them into a CLOSED state
1416 	 *    which causes any the associated files to return useful
1417 	 *    errors to userland. It also unmaps, but does not free,
1418 	 *    any IRQs.
1419 	 *
1420 	 *  - We clean up our side: releasing and unmapping resources we hold
1421 	 *    so we can wire them up again when the hardware comes back up.
1422 	 *
1423 	 * Driver authors should note:
1424 	 *
1425 	 *  - Any contexts you create in your kernel driver (except
1426 	 *    those associated with anonymous file descriptors) are
1427 	 *    your responsibility to free and recreate. Likewise with
1428 	 *    any attached resources.
1429 	 *
1430 	 *  - We will take responsibility for re-initialising the
1431 	 *    device context (the one set up for you in
1432 	 *    cxl_pci_enable_device_hook and accessed through
1433 	 *    cxl_get_context). If you've attached IRQs or other
1434 	 *    resources to it, they remains yours to free.
1435 	 *
1436 	 * You can call the same functions to release resources as you
1437 	 * normally would: we make sure that these functions continue
1438 	 * to work when the hardware is down.
1439 	 *
1440 	 * Two examples:
1441 	 *
1442 	 * 1) If you normally free all your resources at the end of
1443 	 *    each request, or if you use anonymous FDs, your
1444 	 *    error_detected callback can simply set a flag to tell
1445 	 *    your driver not to start any new calls. You can then
1446 	 *    clear the flag in the resume callback.
1447 	 *
1448 	 * 2) If you normally allocate your resources on startup:
1449 	 *     * Set a flag in error_detected as above.
1450 	 *     * Let CXL detach your contexts.
1451 	 *     * In slot_reset, free the old resources and allocate new ones.
1452 	 *     * In resume, clear the flag to allow things to start.
1453 	 */
1454 	for (i = 0; i < adapter->slices; i++) {
1455 		afu = adapter->afu[i];
1456 
1457 		result = cxl_vphb_error_detected(afu, state);
1458 
1459 		/* Only continue if everyone agrees on NEED_RESET */
1460 		if (result != PCI_ERS_RESULT_NEED_RESET)
1461 			return result;
1462 
1463 		cxl_context_detach_all(afu);
1464 		cxl_afu_deactivate_mode(afu);
1465 		cxl_deconfigure_afu(afu);
1466 	}
1467 	cxl_deconfigure_adapter(adapter);
1468 
1469 	return result;
1470 }
1471 
1472 static pci_ers_result_t cxl_pci_slot_reset(struct pci_dev *pdev)
1473 {
1474 	struct cxl *adapter = pci_get_drvdata(pdev);
1475 	struct cxl_afu *afu;
1476 	struct cxl_context *ctx;
1477 	struct pci_dev *afu_dev;
1478 	pci_ers_result_t afu_result = PCI_ERS_RESULT_RECOVERED;
1479 	pci_ers_result_t result = PCI_ERS_RESULT_RECOVERED;
1480 	int i;
1481 
1482 	if (cxl_configure_adapter(adapter, pdev))
1483 		goto err;
1484 
1485 	for (i = 0; i < adapter->slices; i++) {
1486 		afu = adapter->afu[i];
1487 
1488 		if (cxl_configure_afu(afu, adapter, pdev))
1489 			goto err;
1490 
1491 		if (cxl_afu_select_best_mode(afu))
1492 			goto err;
1493 
1494 		cxl_pci_vphb_reconfigure(afu);
1495 
1496 		list_for_each_entry(afu_dev, &afu->phb->bus->devices, bus_list) {
1497 			/* Reset the device context.
1498 			 * TODO: make this less disruptive
1499 			 */
1500 			ctx = cxl_get_context(afu_dev);
1501 
1502 			if (ctx && cxl_release_context(ctx))
1503 				goto err;
1504 
1505 			ctx = cxl_dev_context_init(afu_dev);
1506 			if (!ctx)
1507 				goto err;
1508 
1509 			afu_dev->dev.archdata.cxl_ctx = ctx;
1510 
1511 			if (cxl_afu_check_and_enable(afu))
1512 				goto err;
1513 
1514 			afu_dev->error_state = pci_channel_io_normal;
1515 
1516 			/* If there's a driver attached, allow it to
1517 			 * chime in on recovery. Drivers should check
1518 			 * if everything has come back OK, but
1519 			 * shouldn't start new work until we call
1520 			 * their resume function.
1521 			 */
1522 			if (!afu_dev->driver)
1523 				continue;
1524 
1525 			if (afu_dev->driver->err_handler &&
1526 			    afu_dev->driver->err_handler->slot_reset)
1527 				afu_result = afu_dev->driver->err_handler->slot_reset(afu_dev);
1528 
1529 			if (afu_result == PCI_ERS_RESULT_DISCONNECT)
1530 				result = PCI_ERS_RESULT_DISCONNECT;
1531 		}
1532 	}
1533 	return result;
1534 
1535 err:
1536 	/* All the bits that happen in both error_detected and cxl_remove
1537 	 * should be idempotent, so we don't need to worry about leaving a mix
1538 	 * of unconfigured and reconfigured resources.
1539 	 */
1540 	dev_err(&pdev->dev, "EEH recovery failed. Asking to be disconnected.\n");
1541 	return PCI_ERS_RESULT_DISCONNECT;
1542 }
1543 
1544 static void cxl_pci_resume(struct pci_dev *pdev)
1545 {
1546 	struct cxl *adapter = pci_get_drvdata(pdev);
1547 	struct cxl_afu *afu;
1548 	struct pci_dev *afu_dev;
1549 	int i;
1550 
1551 	/* Everything is back now. Drivers should restart work now.
1552 	 * This is not the place to be checking if everything came back up
1553 	 * properly, because there's no return value: do that in slot_reset.
1554 	 */
1555 	for (i = 0; i < adapter->slices; i++) {
1556 		afu = adapter->afu[i];
1557 
1558 		list_for_each_entry(afu_dev, &afu->phb->bus->devices, bus_list) {
1559 			if (afu_dev->driver && afu_dev->driver->err_handler &&
1560 			    afu_dev->driver->err_handler->resume)
1561 				afu_dev->driver->err_handler->resume(afu_dev);
1562 		}
1563 	}
1564 }
1565 
1566 static const struct pci_error_handlers cxl_err_handler = {
1567 	.error_detected = cxl_pci_error_detected,
1568 	.slot_reset = cxl_pci_slot_reset,
1569 	.resume = cxl_pci_resume,
1570 };
1571 
1572 struct pci_driver cxl_pci_driver = {
1573 	.name = "cxl-pci",
1574 	.id_table = cxl_pci_tbl,
1575 	.probe = cxl_probe,
1576 	.remove = cxl_remove,
1577 	.shutdown = cxl_remove,
1578 	.err_handler = &cxl_err_handler,
1579 };
1580