xref: /linux/drivers/misc/cxl/pci.c (revision 9938b04472d5c59f8bd8152a548533a8599596a2)
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/pnv-pci.h>
23 #include <asm/io.h>
24 
25 #include "cxl.h"
26 #include <misc/cxl.h>
27 
28 
29 #define CXL_PCI_VSEC_ID	0x1280
30 #define CXL_VSEC_MIN_SIZE 0x80
31 
32 #define CXL_READ_VSEC_LENGTH(dev, vsec, dest)			\
33 	{							\
34 		pci_read_config_word(dev, vsec + 0x6, dest);	\
35 		*dest >>= 4;					\
36 	}
37 #define CXL_READ_VSEC_NAFUS(dev, vsec, dest) \
38 	pci_read_config_byte(dev, vsec + 0x8, dest)
39 
40 #define CXL_READ_VSEC_STATUS(dev, vsec, dest) \
41 	pci_read_config_byte(dev, vsec + 0x9, dest)
42 #define CXL_STATUS_SECOND_PORT  0x80
43 #define CXL_STATUS_MSI_X_FULL   0x40
44 #define CXL_STATUS_MSI_X_SINGLE 0x20
45 #define CXL_STATUS_FLASH_RW     0x08
46 #define CXL_STATUS_FLASH_RO     0x04
47 #define CXL_STATUS_LOADABLE_AFU 0x02
48 #define CXL_STATUS_LOADABLE_PSL 0x01
49 /* If we see these features we won't try to use the card */
50 #define CXL_UNSUPPORTED_FEATURES \
51 	(CXL_STATUS_MSI_X_FULL | CXL_STATUS_MSI_X_SINGLE)
52 
53 #define CXL_READ_VSEC_MODE_CONTROL(dev, vsec, dest) \
54 	pci_read_config_byte(dev, vsec + 0xa, dest)
55 #define CXL_WRITE_VSEC_MODE_CONTROL(dev, vsec, val) \
56 	pci_write_config_byte(dev, vsec + 0xa, val)
57 #define CXL_VSEC_PROTOCOL_MASK   0xe0
58 #define CXL_VSEC_PROTOCOL_1024TB 0x80
59 #define CXL_VSEC_PROTOCOL_512TB  0x40
60 #define CXL_VSEC_PROTOCOL_256TB  0x20 /* Power 8 uses this */
61 #define CXL_VSEC_PROTOCOL_ENABLE 0x01
62 
63 #define CXL_READ_VSEC_PSL_REVISION(dev, vsec, dest) \
64 	pci_read_config_word(dev, vsec + 0xc, dest)
65 #define CXL_READ_VSEC_CAIA_MINOR(dev, vsec, dest) \
66 	pci_read_config_byte(dev, vsec + 0xe, dest)
67 #define CXL_READ_VSEC_CAIA_MAJOR(dev, vsec, dest) \
68 	pci_read_config_byte(dev, vsec + 0xf, dest)
69 #define CXL_READ_VSEC_BASE_IMAGE(dev, vsec, dest) \
70 	pci_read_config_word(dev, vsec + 0x10, dest)
71 
72 #define CXL_READ_VSEC_IMAGE_STATE(dev, vsec, dest) \
73 	pci_read_config_byte(dev, vsec + 0x13, dest)
74 #define CXL_WRITE_VSEC_IMAGE_STATE(dev, vsec, val) \
75 	pci_write_config_byte(dev, vsec + 0x13, val)
76 #define CXL_VSEC_USER_IMAGE_LOADED 0x80 /* RO */
77 #define CXL_VSEC_PERST_LOADS_IMAGE 0x20 /* RW */
78 #define CXL_VSEC_PERST_SELECT_USER 0x10 /* RW */
79 
80 #define CXL_READ_VSEC_AFU_DESC_OFF(dev, vsec, dest) \
81 	pci_read_config_dword(dev, vsec + 0x20, dest)
82 #define CXL_READ_VSEC_AFU_DESC_SIZE(dev, vsec, dest) \
83 	pci_read_config_dword(dev, vsec + 0x24, dest)
84 #define CXL_READ_VSEC_PS_OFF(dev, vsec, dest) \
85 	pci_read_config_dword(dev, vsec + 0x28, dest)
86 #define CXL_READ_VSEC_PS_SIZE(dev, vsec, dest) \
87 	pci_read_config_dword(dev, vsec + 0x2c, dest)
88 
89 
90 /* This works a little different than the p1/p2 register accesses to make it
91  * easier to pull out individual fields */
92 #define AFUD_READ(afu, off)		in_be64(afu->native->afu_desc_mmio + off)
93 #define AFUD_READ_LE(afu, off)		in_le64(afu->native->afu_desc_mmio + off)
94 #define EXTRACT_PPC_BIT(val, bit)	(!!(val & PPC_BIT(bit)))
95 #define EXTRACT_PPC_BITS(val, bs, be)	((val & PPC_BITMASK(bs, be)) >> PPC_BITLSHIFT(be))
96 
97 #define AFUD_READ_INFO(afu)		AFUD_READ(afu, 0x0)
98 #define   AFUD_NUM_INTS_PER_PROC(val)	EXTRACT_PPC_BITS(val,  0, 15)
99 #define   AFUD_NUM_PROCS(val)		EXTRACT_PPC_BITS(val, 16, 31)
100 #define   AFUD_NUM_CRS(val)		EXTRACT_PPC_BITS(val, 32, 47)
101 #define   AFUD_MULTIMODE(val)		EXTRACT_PPC_BIT(val, 48)
102 #define   AFUD_PUSH_BLOCK_TRANSFER(val)	EXTRACT_PPC_BIT(val, 55)
103 #define   AFUD_DEDICATED_PROCESS(val)	EXTRACT_PPC_BIT(val, 59)
104 #define   AFUD_AFU_DIRECTED(val)	EXTRACT_PPC_BIT(val, 61)
105 #define   AFUD_TIME_SLICED(val)		EXTRACT_PPC_BIT(val, 63)
106 #define AFUD_READ_CR(afu)		AFUD_READ(afu, 0x20)
107 #define   AFUD_CR_LEN(val)		EXTRACT_PPC_BITS(val, 8, 63)
108 #define AFUD_READ_CR_OFF(afu)		AFUD_READ(afu, 0x28)
109 #define AFUD_READ_PPPSA(afu)		AFUD_READ(afu, 0x30)
110 #define   AFUD_PPPSA_PP(val)		EXTRACT_PPC_BIT(val, 6)
111 #define   AFUD_PPPSA_PSA(val)		EXTRACT_PPC_BIT(val, 7)
112 #define   AFUD_PPPSA_LEN(val)		EXTRACT_PPC_BITS(val, 8, 63)
113 #define AFUD_READ_PPPSA_OFF(afu)	AFUD_READ(afu, 0x38)
114 #define AFUD_READ_EB(afu)		AFUD_READ(afu, 0x40)
115 #define   AFUD_EB_LEN(val)		EXTRACT_PPC_BITS(val, 8, 63)
116 #define AFUD_READ_EB_OFF(afu)		AFUD_READ(afu, 0x48)
117 
118 static const struct pci_device_id cxl_pci_tbl[] = {
119 	{ PCI_DEVICE(PCI_VENDOR_ID_IBM, 0x0477), },
120 	{ PCI_DEVICE(PCI_VENDOR_ID_IBM, 0x044b), },
121 	{ PCI_DEVICE(PCI_VENDOR_ID_IBM, 0x04cf), },
122 	{ PCI_DEVICE(PCI_VENDOR_ID_IBM, 0x0601), },
123 	{ PCI_DEVICE_CLASS(0x120000, ~0), },
124 
125 	{ }
126 };
127 MODULE_DEVICE_TABLE(pci, cxl_pci_tbl);
128 
129 
130 /*
131  * Mostly using these wrappers to avoid confusion:
132  * priv 1 is BAR2, while priv 2 is BAR0
133  */
134 static inline resource_size_t p1_base(struct pci_dev *dev)
135 {
136 	return pci_resource_start(dev, 2);
137 }
138 
139 static inline resource_size_t p1_size(struct pci_dev *dev)
140 {
141 	return pci_resource_len(dev, 2);
142 }
143 
144 static inline resource_size_t p2_base(struct pci_dev *dev)
145 {
146 	return pci_resource_start(dev, 0);
147 }
148 
149 static inline resource_size_t p2_size(struct pci_dev *dev)
150 {
151 	return pci_resource_len(dev, 0);
152 }
153 
154 static int find_cxl_vsec(struct pci_dev *dev)
155 {
156 	int vsec = 0;
157 	u16 val;
158 
159 	while ((vsec = pci_find_next_ext_capability(dev, vsec, PCI_EXT_CAP_ID_VNDR))) {
160 		pci_read_config_word(dev, vsec + 0x4, &val);
161 		if (val == CXL_PCI_VSEC_ID)
162 			return vsec;
163 	}
164 	return 0;
165 
166 }
167 
168 static void dump_cxl_config_space(struct pci_dev *dev)
169 {
170 	int vsec;
171 	u32 val;
172 
173 	dev_info(&dev->dev, "dump_cxl_config_space\n");
174 
175 	pci_read_config_dword(dev, PCI_BASE_ADDRESS_0, &val);
176 	dev_info(&dev->dev, "BAR0: %#.8x\n", val);
177 	pci_read_config_dword(dev, PCI_BASE_ADDRESS_1, &val);
178 	dev_info(&dev->dev, "BAR1: %#.8x\n", val);
179 	pci_read_config_dword(dev, PCI_BASE_ADDRESS_2, &val);
180 	dev_info(&dev->dev, "BAR2: %#.8x\n", val);
181 	pci_read_config_dword(dev, PCI_BASE_ADDRESS_3, &val);
182 	dev_info(&dev->dev, "BAR3: %#.8x\n", val);
183 	pci_read_config_dword(dev, PCI_BASE_ADDRESS_4, &val);
184 	dev_info(&dev->dev, "BAR4: %#.8x\n", val);
185 	pci_read_config_dword(dev, PCI_BASE_ADDRESS_5, &val);
186 	dev_info(&dev->dev, "BAR5: %#.8x\n", val);
187 
188 	dev_info(&dev->dev, "p1 regs: %#llx, len: %#llx\n",
189 		p1_base(dev), p1_size(dev));
190 	dev_info(&dev->dev, "p2 regs: %#llx, len: %#llx\n",
191 		p2_base(dev), p2_size(dev));
192 	dev_info(&dev->dev, "BAR 4/5: %#llx, len: %#llx\n",
193 		pci_resource_start(dev, 4), pci_resource_len(dev, 4));
194 
195 	if (!(vsec = find_cxl_vsec(dev)))
196 		return;
197 
198 #define show_reg(name, what) \
199 	dev_info(&dev->dev, "cxl vsec: %30s: %#x\n", name, what)
200 
201 	pci_read_config_dword(dev, vsec + 0x0, &val);
202 	show_reg("Cap ID", (val >> 0) & 0xffff);
203 	show_reg("Cap Ver", (val >> 16) & 0xf);
204 	show_reg("Next Cap Ptr", (val >> 20) & 0xfff);
205 	pci_read_config_dword(dev, vsec + 0x4, &val);
206 	show_reg("VSEC ID", (val >> 0) & 0xffff);
207 	show_reg("VSEC Rev", (val >> 16) & 0xf);
208 	show_reg("VSEC Length",	(val >> 20) & 0xfff);
209 	pci_read_config_dword(dev, vsec + 0x8, &val);
210 	show_reg("Num AFUs", (val >> 0) & 0xff);
211 	show_reg("Status", (val >> 8) & 0xff);
212 	show_reg("Mode Control", (val >> 16) & 0xff);
213 	show_reg("Reserved", (val >> 24) & 0xff);
214 	pci_read_config_dword(dev, vsec + 0xc, &val);
215 	show_reg("PSL Rev", (val >> 0) & 0xffff);
216 	show_reg("CAIA Ver", (val >> 16) & 0xffff);
217 	pci_read_config_dword(dev, vsec + 0x10, &val);
218 	show_reg("Base Image Rev", (val >> 0) & 0xffff);
219 	show_reg("Reserved", (val >> 16) & 0x0fff);
220 	show_reg("Image Control", (val >> 28) & 0x3);
221 	show_reg("Reserved", (val >> 30) & 0x1);
222 	show_reg("Image Loaded", (val >> 31) & 0x1);
223 
224 	pci_read_config_dword(dev, vsec + 0x14, &val);
225 	show_reg("Reserved", val);
226 	pci_read_config_dword(dev, vsec + 0x18, &val);
227 	show_reg("Reserved", val);
228 	pci_read_config_dword(dev, vsec + 0x1c, &val);
229 	show_reg("Reserved", val);
230 
231 	pci_read_config_dword(dev, vsec + 0x20, &val);
232 	show_reg("AFU Descriptor Offset", val);
233 	pci_read_config_dword(dev, vsec + 0x24, &val);
234 	show_reg("AFU Descriptor Size", val);
235 	pci_read_config_dword(dev, vsec + 0x28, &val);
236 	show_reg("Problem State Offset", val);
237 	pci_read_config_dword(dev, vsec + 0x2c, &val);
238 	show_reg("Problem State Size", val);
239 
240 	pci_read_config_dword(dev, vsec + 0x30, &val);
241 	show_reg("Reserved", val);
242 	pci_read_config_dword(dev, vsec + 0x34, &val);
243 	show_reg("Reserved", val);
244 	pci_read_config_dword(dev, vsec + 0x38, &val);
245 	show_reg("Reserved", val);
246 	pci_read_config_dword(dev, vsec + 0x3c, &val);
247 	show_reg("Reserved", val);
248 
249 	pci_read_config_dword(dev, vsec + 0x40, &val);
250 	show_reg("PSL Programming Port", val);
251 	pci_read_config_dword(dev, vsec + 0x44, &val);
252 	show_reg("PSL Programming Control", val);
253 
254 	pci_read_config_dword(dev, vsec + 0x48, &val);
255 	show_reg("Reserved", val);
256 	pci_read_config_dword(dev, vsec + 0x4c, &val);
257 	show_reg("Reserved", val);
258 
259 	pci_read_config_dword(dev, vsec + 0x50, &val);
260 	show_reg("Flash Address Register", val);
261 	pci_read_config_dword(dev, vsec + 0x54, &val);
262 	show_reg("Flash Size Register", val);
263 	pci_read_config_dword(dev, vsec + 0x58, &val);
264 	show_reg("Flash Status/Control Register", val);
265 	pci_read_config_dword(dev, vsec + 0x58, &val);
266 	show_reg("Flash Data Port", val);
267 
268 #undef show_reg
269 }
270 
271 static void dump_afu_descriptor(struct cxl_afu *afu)
272 {
273 	u64 val, afu_cr_num, afu_cr_off, afu_cr_len;
274 	int i;
275 
276 #define show_reg(name, what) \
277 	dev_info(&afu->dev, "afu desc: %30s: %#llx\n", name, what)
278 
279 	val = AFUD_READ_INFO(afu);
280 	show_reg("num_ints_per_process", AFUD_NUM_INTS_PER_PROC(val));
281 	show_reg("num_of_processes", AFUD_NUM_PROCS(val));
282 	show_reg("num_of_afu_CRs", AFUD_NUM_CRS(val));
283 	show_reg("req_prog_mode", val & 0xffffULL);
284 	afu_cr_num = AFUD_NUM_CRS(val);
285 
286 	val = AFUD_READ(afu, 0x8);
287 	show_reg("Reserved", val);
288 	val = AFUD_READ(afu, 0x10);
289 	show_reg("Reserved", val);
290 	val = AFUD_READ(afu, 0x18);
291 	show_reg("Reserved", val);
292 
293 	val = AFUD_READ_CR(afu);
294 	show_reg("Reserved", (val >> (63-7)) & 0xff);
295 	show_reg("AFU_CR_len", AFUD_CR_LEN(val));
296 	afu_cr_len = AFUD_CR_LEN(val) * 256;
297 
298 	val = AFUD_READ_CR_OFF(afu);
299 	afu_cr_off = val;
300 	show_reg("AFU_CR_offset", val);
301 
302 	val = AFUD_READ_PPPSA(afu);
303 	show_reg("PerProcessPSA_control", (val >> (63-7)) & 0xff);
304 	show_reg("PerProcessPSA Length", AFUD_PPPSA_LEN(val));
305 
306 	val = AFUD_READ_PPPSA_OFF(afu);
307 	show_reg("PerProcessPSA_offset", val);
308 
309 	val = AFUD_READ_EB(afu);
310 	show_reg("Reserved", (val >> (63-7)) & 0xff);
311 	show_reg("AFU_EB_len", AFUD_EB_LEN(val));
312 
313 	val = AFUD_READ_EB_OFF(afu);
314 	show_reg("AFU_EB_offset", val);
315 
316 	for (i = 0; i < afu_cr_num; i++) {
317 		val = AFUD_READ_LE(afu, afu_cr_off + i * afu_cr_len);
318 		show_reg("CR Vendor", val & 0xffff);
319 		show_reg("CR Device", (val >> 16) & 0xffff);
320 	}
321 #undef show_reg
322 }
323 
324 static int init_implementation_adapter_regs(struct cxl *adapter, struct pci_dev *dev)
325 {
326 	struct device_node *np;
327 	const __be32 *prop;
328 	u64 psl_dsnctl;
329 	u64 chipid;
330 
331 	if (!(np = pnv_pci_get_phb_node(dev)))
332 		return -ENODEV;
333 
334 	while (np && !(prop = of_get_property(np, "ibm,chip-id", NULL)))
335 		np = of_get_next_parent(np);
336 	if (!np)
337 		return -ENODEV;
338 	chipid = be32_to_cpup(prop);
339 	of_node_put(np);
340 
341 	/* Tell PSL where to route data to */
342 	psl_dsnctl = 0x02E8900002000000ULL | (chipid << (63-5));
343 	cxl_p1_write(adapter, CXL_PSL_DSNDCTL, psl_dsnctl);
344 	cxl_p1_write(adapter, CXL_PSL_RESLCKTO, 0x20000000200ULL);
345 	/* snoop write mask */
346 	cxl_p1_write(adapter, CXL_PSL_SNWRALLOC, 0x00000000FFFFFFFFULL);
347 	/* set fir_accum */
348 	cxl_p1_write(adapter, CXL_PSL_FIR_CNTL, 0x0800000000000000ULL);
349 	/* for debugging with trace arrays */
350 	cxl_p1_write(adapter, CXL_PSL_TRACE, 0x0000FF7C00000000ULL);
351 
352 	return 0;
353 }
354 
355 #define TBSYNC_CNT(n) (((u64)n & 0x7) << (63-6))
356 #define _2048_250MHZ_CYCLES 1
357 
358 static int cxl_setup_psl_timebase(struct cxl *adapter, struct pci_dev *dev)
359 {
360 	u64 psl_tb;
361 	int delta;
362 	unsigned int retry = 0;
363 	struct device_node *np;
364 
365 	if (!(np = pnv_pci_get_phb_node(dev)))
366 		return -ENODEV;
367 
368 	/* Do not fail when CAPP timebase sync is not supported by OPAL */
369 	of_node_get(np);
370 	if (! of_get_property(np, "ibm,capp-timebase-sync", NULL)) {
371 		of_node_put(np);
372 		pr_err("PSL: Timebase sync: OPAL support missing\n");
373 		return 0;
374 	}
375 	of_node_put(np);
376 
377 	/*
378 	 * Setup PSL Timebase Control and Status register
379 	 * with the recommended Timebase Sync Count value
380 	 */
381 	cxl_p1_write(adapter, CXL_PSL_TB_CTLSTAT,
382 		     TBSYNC_CNT(2 * _2048_250MHZ_CYCLES));
383 
384 	/* Enable PSL Timebase */
385 	cxl_p1_write(adapter, CXL_PSL_Control, 0x0000000000000000);
386 	cxl_p1_write(adapter, CXL_PSL_Control, CXL_PSL_Control_tb);
387 
388 	/* Wait until CORE TB and PSL TB difference <= 16usecs */
389 	do {
390 		msleep(1);
391 		if (retry++ > 5) {
392 			pr_err("PSL: Timebase sync: giving up!\n");
393 			return -EIO;
394 		}
395 		psl_tb = cxl_p1_read(adapter, CXL_PSL_Timebase);
396 		delta = mftb() - psl_tb;
397 		if (delta < 0)
398 			delta = -delta;
399 	} while (tb_to_ns(delta) > 16000);
400 
401 	return 0;
402 }
403 
404 static int init_implementation_afu_regs(struct cxl_afu *afu)
405 {
406 	/* read/write masks for this slice */
407 	cxl_p1n_write(afu, CXL_PSL_APCALLOC_A, 0xFFFFFFFEFEFEFEFEULL);
408 	/* APC read/write masks for this slice */
409 	cxl_p1n_write(afu, CXL_PSL_COALLOC_A, 0xFF000000FEFEFEFEULL);
410 	/* for debugging with trace arrays */
411 	cxl_p1n_write(afu, CXL_PSL_SLICE_TRACE, 0x0000FFFF00000000ULL);
412 	cxl_p1n_write(afu, CXL_PSL_RXCTL_A, CXL_PSL_RXCTL_AFUHP_4S);
413 
414 	return 0;
415 }
416 
417 int cxl_pci_setup_irq(struct cxl *adapter, unsigned int hwirq,
418 		unsigned int virq)
419 {
420 	struct pci_dev *dev = to_pci_dev(adapter->dev.parent);
421 
422 	return pnv_cxl_ioda_msi_setup(dev, hwirq, virq);
423 }
424 
425 int cxl_update_image_control(struct cxl *adapter)
426 {
427 	struct pci_dev *dev = to_pci_dev(adapter->dev.parent);
428 	int rc;
429 	int vsec;
430 	u8 image_state;
431 
432 	if (!(vsec = find_cxl_vsec(dev))) {
433 		dev_err(&dev->dev, "ABORTING: CXL VSEC not found!\n");
434 		return -ENODEV;
435 	}
436 
437 	if ((rc = CXL_READ_VSEC_IMAGE_STATE(dev, vsec, &image_state))) {
438 		dev_err(&dev->dev, "failed to read image state: %i\n", rc);
439 		return rc;
440 	}
441 
442 	if (adapter->perst_loads_image)
443 		image_state |= CXL_VSEC_PERST_LOADS_IMAGE;
444 	else
445 		image_state &= ~CXL_VSEC_PERST_LOADS_IMAGE;
446 
447 	if (adapter->perst_select_user)
448 		image_state |= CXL_VSEC_PERST_SELECT_USER;
449 	else
450 		image_state &= ~CXL_VSEC_PERST_SELECT_USER;
451 
452 	if ((rc = CXL_WRITE_VSEC_IMAGE_STATE(dev, vsec, image_state))) {
453 		dev_err(&dev->dev, "failed to update image control: %i\n", rc);
454 		return rc;
455 	}
456 
457 	return 0;
458 }
459 
460 int cxl_pci_alloc_one_irq(struct cxl *adapter)
461 {
462 	struct pci_dev *dev = to_pci_dev(adapter->dev.parent);
463 
464 	return pnv_cxl_alloc_hwirqs(dev, 1);
465 }
466 
467 void cxl_pci_release_one_irq(struct cxl *adapter, int hwirq)
468 {
469 	struct pci_dev *dev = to_pci_dev(adapter->dev.parent);
470 
471 	return pnv_cxl_release_hwirqs(dev, hwirq, 1);
472 }
473 
474 int cxl_pci_alloc_irq_ranges(struct cxl_irq_ranges *irqs,
475 			struct cxl *adapter, unsigned int num)
476 {
477 	struct pci_dev *dev = to_pci_dev(adapter->dev.parent);
478 
479 	return pnv_cxl_alloc_hwirq_ranges(irqs, dev, num);
480 }
481 
482 void cxl_pci_release_irq_ranges(struct cxl_irq_ranges *irqs,
483 				struct cxl *adapter)
484 {
485 	struct pci_dev *dev = to_pci_dev(adapter->dev.parent);
486 
487 	pnv_cxl_release_hwirq_ranges(irqs, dev);
488 }
489 
490 static int setup_cxl_bars(struct pci_dev *dev)
491 {
492 	/* Safety check in case we get backported to < 3.17 without M64 */
493 	if ((p1_base(dev) < 0x100000000ULL) ||
494 	    (p2_base(dev) < 0x100000000ULL)) {
495 		dev_err(&dev->dev, "ABORTING: M32 BAR assignment incompatible with CXL\n");
496 		return -ENODEV;
497 	}
498 
499 	/*
500 	 * BAR 4/5 has a special meaning for CXL and must be programmed with a
501 	 * special value corresponding to the CXL protocol address range.
502 	 * For POWER 8 that means bits 48:49 must be set to 10
503 	 */
504 	pci_write_config_dword(dev, PCI_BASE_ADDRESS_4, 0x00000000);
505 	pci_write_config_dword(dev, PCI_BASE_ADDRESS_5, 0x00020000);
506 
507 	return 0;
508 }
509 
510 /* pciex node: ibm,opal-m64-window = <0x3d058 0x0 0x3d058 0x0 0x8 0x0>; */
511 static int switch_card_to_cxl(struct pci_dev *dev)
512 {
513 	int vsec;
514 	u8 val;
515 	int rc;
516 
517 	dev_info(&dev->dev, "switch card to CXL\n");
518 
519 	if (!(vsec = find_cxl_vsec(dev))) {
520 		dev_err(&dev->dev, "ABORTING: CXL VSEC not found!\n");
521 		return -ENODEV;
522 	}
523 
524 	if ((rc = CXL_READ_VSEC_MODE_CONTROL(dev, vsec, &val))) {
525 		dev_err(&dev->dev, "failed to read current mode control: %i", rc);
526 		return rc;
527 	}
528 	val &= ~CXL_VSEC_PROTOCOL_MASK;
529 	val |= CXL_VSEC_PROTOCOL_256TB | CXL_VSEC_PROTOCOL_ENABLE;
530 	if ((rc = CXL_WRITE_VSEC_MODE_CONTROL(dev, vsec, val))) {
531 		dev_err(&dev->dev, "failed to enable CXL protocol: %i", rc);
532 		return rc;
533 	}
534 	/*
535 	 * The CAIA spec (v0.12 11.6 Bi-modal Device Support) states
536 	 * we must wait 100ms after this mode switch before touching
537 	 * PCIe config space.
538 	 */
539 	msleep(100);
540 
541 	return 0;
542 }
543 
544 static int pci_map_slice_regs(struct cxl_afu *afu, struct cxl *adapter, struct pci_dev *dev)
545 {
546 	u64 p1n_base, p2n_base, afu_desc;
547 	const u64 p1n_size = 0x100;
548 	const u64 p2n_size = 0x1000;
549 
550 	p1n_base = p1_base(dev) + 0x10000 + (afu->slice * p1n_size);
551 	p2n_base = p2_base(dev) + (afu->slice * p2n_size);
552 	afu->psn_phys = p2_base(dev) + (adapter->native->ps_off + (afu->slice * adapter->ps_size));
553 	afu_desc = p2_base(dev) + adapter->native->afu_desc_off + (afu->slice * adapter->native->afu_desc_size);
554 
555 	if (!(afu->native->p1n_mmio = ioremap(p1n_base, p1n_size)))
556 		goto err;
557 	if (!(afu->p2n_mmio = ioremap(p2n_base, p2n_size)))
558 		goto err1;
559 	if (afu_desc) {
560 		if (!(afu->native->afu_desc_mmio = ioremap(afu_desc, adapter->native->afu_desc_size)))
561 			goto err2;
562 	}
563 
564 	return 0;
565 err2:
566 	iounmap(afu->p2n_mmio);
567 err1:
568 	iounmap(afu->native->p1n_mmio);
569 err:
570 	dev_err(&afu->dev, "Error mapping AFU MMIO regions\n");
571 	return -ENOMEM;
572 }
573 
574 static void pci_unmap_slice_regs(struct cxl_afu *afu)
575 {
576 	if (afu->p2n_mmio) {
577 		iounmap(afu->p2n_mmio);
578 		afu->p2n_mmio = NULL;
579 	}
580 	if (afu->native->p1n_mmio) {
581 		iounmap(afu->native->p1n_mmio);
582 		afu->native->p1n_mmio = NULL;
583 	}
584 	if (afu->native->afu_desc_mmio) {
585 		iounmap(afu->native->afu_desc_mmio);
586 		afu->native->afu_desc_mmio = NULL;
587 	}
588 }
589 
590 void cxl_pci_release_afu(struct device *dev)
591 {
592 	struct cxl_afu *afu = to_cxl_afu(dev);
593 
594 	pr_devel("%s\n", __func__);
595 
596 	idr_destroy(&afu->contexts_idr);
597 	cxl_release_spa(afu);
598 
599 	kfree(afu->native);
600 	kfree(afu);
601 }
602 
603 /* Expects AFU struct to have recently been zeroed out */
604 static int cxl_read_afu_descriptor(struct cxl_afu *afu)
605 {
606 	u64 val;
607 
608 	val = AFUD_READ_INFO(afu);
609 	afu->pp_irqs = AFUD_NUM_INTS_PER_PROC(val);
610 	afu->max_procs_virtualised = AFUD_NUM_PROCS(val);
611 	afu->crs_num = AFUD_NUM_CRS(val);
612 
613 	if (AFUD_AFU_DIRECTED(val))
614 		afu->modes_supported |= CXL_MODE_DIRECTED;
615 	if (AFUD_DEDICATED_PROCESS(val))
616 		afu->modes_supported |= CXL_MODE_DEDICATED;
617 	if (AFUD_TIME_SLICED(val))
618 		afu->modes_supported |= CXL_MODE_TIME_SLICED;
619 
620 	val = AFUD_READ_PPPSA(afu);
621 	afu->pp_size = AFUD_PPPSA_LEN(val) * 4096;
622 	afu->psa = AFUD_PPPSA_PSA(val);
623 	if ((afu->pp_psa = AFUD_PPPSA_PP(val)))
624 		afu->native->pp_offset = AFUD_READ_PPPSA_OFF(afu);
625 
626 	val = AFUD_READ_CR(afu);
627 	afu->crs_len = AFUD_CR_LEN(val) * 256;
628 	afu->crs_offset = AFUD_READ_CR_OFF(afu);
629 
630 
631 	/* eb_len is in multiple of 4K */
632 	afu->eb_len = AFUD_EB_LEN(AFUD_READ_EB(afu)) * 4096;
633 	afu->eb_offset = AFUD_READ_EB_OFF(afu);
634 
635 	/* eb_off is 4K aligned so lower 12 bits are always zero */
636 	if (EXTRACT_PPC_BITS(afu->eb_offset, 0, 11) != 0) {
637 		dev_warn(&afu->dev,
638 			 "Invalid AFU error buffer offset %Lx\n",
639 			 afu->eb_offset);
640 		dev_info(&afu->dev,
641 			 "Ignoring AFU error buffer in the descriptor\n");
642 		/* indicate that no afu buffer exists */
643 		afu->eb_len = 0;
644 	}
645 
646 	return 0;
647 }
648 
649 static int cxl_afu_descriptor_looks_ok(struct cxl_afu *afu)
650 {
651 	int i, rc;
652 	u32 val;
653 
654 	if (afu->psa && afu->adapter->ps_size <
655 			(afu->native->pp_offset + afu->pp_size*afu->max_procs_virtualised)) {
656 		dev_err(&afu->dev, "per-process PSA can't fit inside the PSA!\n");
657 		return -ENODEV;
658 	}
659 
660 	if (afu->pp_psa && (afu->pp_size < PAGE_SIZE))
661 		dev_warn(&afu->dev, "AFU uses < PAGE_SIZE per-process PSA!");
662 
663 	for (i = 0; i < afu->crs_num; i++) {
664 		rc = cxl_ops->afu_cr_read32(afu, i, 0, &val);
665 		if (rc || val == 0) {
666 			dev_err(&afu->dev, "ABORTING: AFU configuration record %i is invalid\n", i);
667 			return -EINVAL;
668 		}
669 	}
670 
671 	return 0;
672 }
673 
674 static int sanitise_afu_regs(struct cxl_afu *afu)
675 {
676 	u64 reg;
677 
678 	/*
679 	 * Clear out any regs that contain either an IVTE or address or may be
680 	 * waiting on an acknowledgement to try to be a bit safer as we bring
681 	 * it online
682 	 */
683 	reg = cxl_p2n_read(afu, CXL_AFU_Cntl_An);
684 	if ((reg & CXL_AFU_Cntl_An_ES_MASK) != CXL_AFU_Cntl_An_ES_Disabled) {
685 		dev_warn(&afu->dev, "WARNING: AFU was not disabled: %#016llx\n", reg);
686 		if (cxl_ops->afu_reset(afu))
687 			return -EIO;
688 		if (cxl_afu_disable(afu))
689 			return -EIO;
690 		if (cxl_psl_purge(afu))
691 			return -EIO;
692 	}
693 	cxl_p1n_write(afu, CXL_PSL_SPAP_An, 0x0000000000000000);
694 	cxl_p1n_write(afu, CXL_PSL_IVTE_Limit_An, 0x0000000000000000);
695 	cxl_p1n_write(afu, CXL_PSL_IVTE_Offset_An, 0x0000000000000000);
696 	cxl_p1n_write(afu, CXL_PSL_AMBAR_An, 0x0000000000000000);
697 	cxl_p1n_write(afu, CXL_PSL_SPOffset_An, 0x0000000000000000);
698 	cxl_p1n_write(afu, CXL_HAURP_An, 0x0000000000000000);
699 	cxl_p2n_write(afu, CXL_CSRP_An, 0x0000000000000000);
700 	cxl_p2n_write(afu, CXL_AURP1_An, 0x0000000000000000);
701 	cxl_p2n_write(afu, CXL_AURP0_An, 0x0000000000000000);
702 	cxl_p2n_write(afu, CXL_SSTP1_An, 0x0000000000000000);
703 	cxl_p2n_write(afu, CXL_SSTP0_An, 0x0000000000000000);
704 	reg = cxl_p2n_read(afu, CXL_PSL_DSISR_An);
705 	if (reg) {
706 		dev_warn(&afu->dev, "AFU had pending DSISR: %#016llx\n", reg);
707 		if (reg & CXL_PSL_DSISR_TRANS)
708 			cxl_p2n_write(afu, CXL_PSL_TFC_An, CXL_PSL_TFC_An_AE);
709 		else
710 			cxl_p2n_write(afu, CXL_PSL_TFC_An, CXL_PSL_TFC_An_A);
711 	}
712 	reg = cxl_p1n_read(afu, CXL_PSL_SERR_An);
713 	if (reg) {
714 		if (reg & ~0xffff)
715 			dev_warn(&afu->dev, "AFU had pending SERR: %#016llx\n", reg);
716 		cxl_p1n_write(afu, CXL_PSL_SERR_An, reg & ~0xffff);
717 	}
718 	reg = cxl_p2n_read(afu, CXL_PSL_ErrStat_An);
719 	if (reg) {
720 		dev_warn(&afu->dev, "AFU had pending error status: %#016llx\n", reg);
721 		cxl_p2n_write(afu, CXL_PSL_ErrStat_An, reg);
722 	}
723 
724 	return 0;
725 }
726 
727 #define ERR_BUFF_MAX_COPY_SIZE PAGE_SIZE
728 /*
729  * afu_eb_read:
730  * Called from sysfs and reads the afu error info buffer. The h/w only supports
731  * 4/8 bytes aligned access. So in case the requested offset/count arent 8 byte
732  * aligned the function uses a bounce buffer which can be max PAGE_SIZE.
733  */
734 ssize_t cxl_pci_afu_read_err_buffer(struct cxl_afu *afu, char *buf,
735 				loff_t off, size_t count)
736 {
737 	loff_t aligned_start, aligned_end;
738 	size_t aligned_length;
739 	void *tbuf;
740 	const void __iomem *ebuf = afu->native->afu_desc_mmio + afu->eb_offset;
741 
742 	if (count == 0 || off < 0 || (size_t)off >= afu->eb_len)
743 		return 0;
744 
745 	/* calculate aligned read window */
746 	count = min((size_t)(afu->eb_len - off), count);
747 	aligned_start = round_down(off, 8);
748 	aligned_end = round_up(off + count, 8);
749 	aligned_length = aligned_end - aligned_start;
750 
751 	/* max we can copy in one read is PAGE_SIZE */
752 	if (aligned_length > ERR_BUFF_MAX_COPY_SIZE) {
753 		aligned_length = ERR_BUFF_MAX_COPY_SIZE;
754 		count = ERR_BUFF_MAX_COPY_SIZE - (off & 0x7);
755 	}
756 
757 	/* use bounce buffer for copy */
758 	tbuf = (void *)__get_free_page(GFP_TEMPORARY);
759 	if (!tbuf)
760 		return -ENOMEM;
761 
762 	/* perform aligned read from the mmio region */
763 	memcpy_fromio(tbuf, ebuf + aligned_start, aligned_length);
764 	memcpy(buf, tbuf + (off & 0x7), count);
765 
766 	free_page((unsigned long)tbuf);
767 
768 	return count;
769 }
770 
771 static int pci_configure_afu(struct cxl_afu *afu, struct cxl *adapter, struct pci_dev *dev)
772 {
773 	int rc;
774 
775 	if ((rc = pci_map_slice_regs(afu, adapter, dev)))
776 		return rc;
777 
778 	if ((rc = sanitise_afu_regs(afu)))
779 		goto err1;
780 
781 	/* We need to reset the AFU before we can read the AFU descriptor */
782 	if ((rc = cxl_ops->afu_reset(afu)))
783 		goto err1;
784 
785 	if (cxl_verbose)
786 		dump_afu_descriptor(afu);
787 
788 	if ((rc = cxl_read_afu_descriptor(afu)))
789 		goto err1;
790 
791 	if ((rc = cxl_afu_descriptor_looks_ok(afu)))
792 		goto err1;
793 
794 	if ((rc = init_implementation_afu_regs(afu)))
795 		goto err1;
796 
797 	if ((rc = cxl_native_register_serr_irq(afu)))
798 		goto err1;
799 
800 	if ((rc = cxl_native_register_psl_irq(afu)))
801 		goto err2;
802 
803 	return 0;
804 
805 err2:
806 	cxl_native_release_serr_irq(afu);
807 err1:
808 	pci_unmap_slice_regs(afu);
809 	return rc;
810 }
811 
812 static void pci_deconfigure_afu(struct cxl_afu *afu)
813 {
814 	cxl_native_release_psl_irq(afu);
815 	cxl_native_release_serr_irq(afu);
816 	pci_unmap_slice_regs(afu);
817 }
818 
819 static int pci_init_afu(struct cxl *adapter, int slice, struct pci_dev *dev)
820 {
821 	struct cxl_afu *afu;
822 	int rc = -ENOMEM;
823 
824 	afu = cxl_alloc_afu(adapter, slice);
825 	if (!afu)
826 		return -ENOMEM;
827 
828 	afu->native = kzalloc(sizeof(struct cxl_afu_native), GFP_KERNEL);
829 	if (!afu->native)
830 		goto err_free_afu;
831 
832 	mutex_init(&afu->native->spa_mutex);
833 
834 	rc = dev_set_name(&afu->dev, "afu%i.%i", adapter->adapter_num, slice);
835 	if (rc)
836 		goto err_free_native;
837 
838 	rc = pci_configure_afu(afu, adapter, dev);
839 	if (rc)
840 		goto err_free_native;
841 
842 	/* Don't care if this fails */
843 	cxl_debugfs_afu_add(afu);
844 
845 	/*
846 	 * After we call this function we must not free the afu directly, even
847 	 * if it returns an error!
848 	 */
849 	if ((rc = cxl_register_afu(afu)))
850 		goto err_put1;
851 
852 	if ((rc = cxl_sysfs_afu_add(afu)))
853 		goto err_put1;
854 
855 	adapter->afu[afu->slice] = afu;
856 
857 	if ((rc = cxl_pci_vphb_add(afu)))
858 		dev_info(&afu->dev, "Can't register vPHB\n");
859 
860 	return 0;
861 
862 err_put1:
863 	pci_deconfigure_afu(afu);
864 	cxl_debugfs_afu_remove(afu);
865 	device_unregister(&afu->dev);
866 	return rc;
867 
868 err_free_native:
869 	kfree(afu->native);
870 err_free_afu:
871 	kfree(afu);
872 	return rc;
873 
874 }
875 
876 static void cxl_pci_remove_afu(struct cxl_afu *afu)
877 {
878 	pr_devel("%s\n", __func__);
879 
880 	if (!afu)
881 		return;
882 
883 	cxl_pci_vphb_remove(afu);
884 	cxl_sysfs_afu_remove(afu);
885 	cxl_debugfs_afu_remove(afu);
886 
887 	spin_lock(&afu->adapter->afu_list_lock);
888 	afu->adapter->afu[afu->slice] = NULL;
889 	spin_unlock(&afu->adapter->afu_list_lock);
890 
891 	cxl_context_detach_all(afu);
892 	cxl_ops->afu_deactivate_mode(afu, afu->current_mode);
893 
894 	pci_deconfigure_afu(afu);
895 	device_unregister(&afu->dev);
896 }
897 
898 int cxl_pci_reset(struct cxl *adapter)
899 {
900 	struct pci_dev *dev = to_pci_dev(adapter->dev.parent);
901 	int rc;
902 
903 	if (adapter->perst_same_image) {
904 		dev_warn(&dev->dev,
905 			 "cxl: refusing to reset/reflash when perst_reloads_same_image is set.\n");
906 		return -EINVAL;
907 	}
908 
909 	dev_info(&dev->dev, "CXL reset\n");
910 
911 	/* pcie_warm_reset requests a fundamental pci reset which includes a
912 	 * PERST assert/deassert.  PERST triggers a loading of the image
913 	 * if "user" or "factory" is selected in sysfs */
914 	if ((rc = pci_set_pcie_reset_state(dev, pcie_warm_reset))) {
915 		dev_err(&dev->dev, "cxl: pcie_warm_reset failed\n");
916 		return rc;
917 	}
918 
919 	return rc;
920 }
921 
922 static int cxl_map_adapter_regs(struct cxl *adapter, struct pci_dev *dev)
923 {
924 	if (pci_request_region(dev, 2, "priv 2 regs"))
925 		goto err1;
926 	if (pci_request_region(dev, 0, "priv 1 regs"))
927 		goto err2;
928 
929 	pr_devel("cxl_map_adapter_regs: p1: %#016llx %#llx, p2: %#016llx %#llx",
930 			p1_base(dev), p1_size(dev), p2_base(dev), p2_size(dev));
931 
932 	if (!(adapter->native->p1_mmio = ioremap(p1_base(dev), p1_size(dev))))
933 		goto err3;
934 
935 	if (!(adapter->native->p2_mmio = ioremap(p2_base(dev), p2_size(dev))))
936 		goto err4;
937 
938 	return 0;
939 
940 err4:
941 	iounmap(adapter->native->p1_mmio);
942 	adapter->native->p1_mmio = NULL;
943 err3:
944 	pci_release_region(dev, 0);
945 err2:
946 	pci_release_region(dev, 2);
947 err1:
948 	return -ENOMEM;
949 }
950 
951 static void cxl_unmap_adapter_regs(struct cxl *adapter)
952 {
953 	if (adapter->native->p1_mmio) {
954 		iounmap(adapter->native->p1_mmio);
955 		adapter->native->p1_mmio = NULL;
956 		pci_release_region(to_pci_dev(adapter->dev.parent), 2);
957 	}
958 	if (adapter->native->p2_mmio) {
959 		iounmap(adapter->native->p2_mmio);
960 		adapter->native->p2_mmio = NULL;
961 		pci_release_region(to_pci_dev(adapter->dev.parent), 0);
962 	}
963 }
964 
965 static int cxl_read_vsec(struct cxl *adapter, struct pci_dev *dev)
966 {
967 	int vsec;
968 	u32 afu_desc_off, afu_desc_size;
969 	u32 ps_off, ps_size;
970 	u16 vseclen;
971 	u8 image_state;
972 
973 	if (!(vsec = find_cxl_vsec(dev))) {
974 		dev_err(&dev->dev, "ABORTING: CXL VSEC not found!\n");
975 		return -ENODEV;
976 	}
977 
978 	CXL_READ_VSEC_LENGTH(dev, vsec, &vseclen);
979 	if (vseclen < CXL_VSEC_MIN_SIZE) {
980 		dev_err(&dev->dev, "ABORTING: CXL VSEC too short\n");
981 		return -EINVAL;
982 	}
983 
984 	CXL_READ_VSEC_STATUS(dev, vsec, &adapter->vsec_status);
985 	CXL_READ_VSEC_PSL_REVISION(dev, vsec, &adapter->psl_rev);
986 	CXL_READ_VSEC_CAIA_MAJOR(dev, vsec, &adapter->caia_major);
987 	CXL_READ_VSEC_CAIA_MINOR(dev, vsec, &adapter->caia_minor);
988 	CXL_READ_VSEC_BASE_IMAGE(dev, vsec, &adapter->base_image);
989 	CXL_READ_VSEC_IMAGE_STATE(dev, vsec, &image_state);
990 	adapter->user_image_loaded = !!(image_state & CXL_VSEC_USER_IMAGE_LOADED);
991 	adapter->perst_select_user = !!(image_state & CXL_VSEC_USER_IMAGE_LOADED);
992 
993 	CXL_READ_VSEC_NAFUS(dev, vsec, &adapter->slices);
994 	CXL_READ_VSEC_AFU_DESC_OFF(dev, vsec, &afu_desc_off);
995 	CXL_READ_VSEC_AFU_DESC_SIZE(dev, vsec, &afu_desc_size);
996 	CXL_READ_VSEC_PS_OFF(dev, vsec, &ps_off);
997 	CXL_READ_VSEC_PS_SIZE(dev, vsec, &ps_size);
998 
999 	/* Convert everything to bytes, because there is NO WAY I'd look at the
1000 	 * code a month later and forget what units these are in ;-) */
1001 	adapter->native->ps_off = ps_off * 64 * 1024;
1002 	adapter->ps_size = ps_size * 64 * 1024;
1003 	adapter->native->afu_desc_off = afu_desc_off * 64 * 1024;
1004 	adapter->native->afu_desc_size = afu_desc_size * 64 * 1024;
1005 
1006 	/* Total IRQs - 1 PSL ERROR - #AFU*(1 slice error + 1 DSI) */
1007 	adapter->user_irqs = pnv_cxl_get_irq_count(dev) - 1 - 2*adapter->slices;
1008 
1009 	return 0;
1010 }
1011 
1012 /*
1013  * Workaround a PCIe Host Bridge defect on some cards, that can cause
1014  * malformed Transaction Layer Packet (TLP) errors to be erroneously
1015  * reported. Mask this error in the Uncorrectable Error Mask Register.
1016  *
1017  * The upper nibble of the PSL revision is used to distinguish between
1018  * different cards. The affected ones have it set to 0.
1019  */
1020 static void cxl_fixup_malformed_tlp(struct cxl *adapter, struct pci_dev *dev)
1021 {
1022 	int aer;
1023 	u32 data;
1024 
1025 	if (adapter->psl_rev & 0xf000)
1026 		return;
1027 	if (!(aer = pci_find_ext_capability(dev, PCI_EXT_CAP_ID_ERR)))
1028 		return;
1029 	pci_read_config_dword(dev, aer + PCI_ERR_UNCOR_MASK, &data);
1030 	if (data & PCI_ERR_UNC_MALF_TLP)
1031 		if (data & PCI_ERR_UNC_INTN)
1032 			return;
1033 	data |= PCI_ERR_UNC_MALF_TLP;
1034 	data |= PCI_ERR_UNC_INTN;
1035 	pci_write_config_dword(dev, aer + PCI_ERR_UNCOR_MASK, data);
1036 }
1037 
1038 static int cxl_vsec_looks_ok(struct cxl *adapter, struct pci_dev *dev)
1039 {
1040 	if (adapter->vsec_status & CXL_STATUS_SECOND_PORT)
1041 		return -EBUSY;
1042 
1043 	if (adapter->vsec_status & CXL_UNSUPPORTED_FEATURES) {
1044 		dev_err(&dev->dev, "ABORTING: CXL requires unsupported features\n");
1045 		return -EINVAL;
1046 	}
1047 
1048 	if (!adapter->slices) {
1049 		/* Once we support dynamic reprogramming we can use the card if
1050 		 * it supports loadable AFUs */
1051 		dev_err(&dev->dev, "ABORTING: Device has no AFUs\n");
1052 		return -EINVAL;
1053 	}
1054 
1055 	if (!adapter->native->afu_desc_off || !adapter->native->afu_desc_size) {
1056 		dev_err(&dev->dev, "ABORTING: VSEC shows no AFU descriptors\n");
1057 		return -EINVAL;
1058 	}
1059 
1060 	if (adapter->ps_size > p2_size(dev) - adapter->native->ps_off) {
1061 		dev_err(&dev->dev, "ABORTING: Problem state size larger than "
1062 				   "available in BAR2: 0x%llx > 0x%llx\n",
1063 			 adapter->ps_size, p2_size(dev) - adapter->native->ps_off);
1064 		return -EINVAL;
1065 	}
1066 
1067 	return 0;
1068 }
1069 
1070 ssize_t cxl_pci_read_adapter_vpd(struct cxl *adapter, void *buf, size_t len)
1071 {
1072 	return pci_read_vpd(to_pci_dev(adapter->dev.parent), 0, len, buf);
1073 }
1074 
1075 static void cxl_release_adapter(struct device *dev)
1076 {
1077 	struct cxl *adapter = to_cxl_adapter(dev);
1078 
1079 	pr_devel("cxl_release_adapter\n");
1080 
1081 	cxl_remove_adapter_nr(adapter);
1082 
1083 	kfree(adapter->native);
1084 	kfree(adapter);
1085 }
1086 
1087 #define CXL_PSL_ErrIVTE_tberror (0x1ull << (63-31))
1088 
1089 static int sanitise_adapter_regs(struct cxl *adapter)
1090 {
1091 	/* Clear PSL tberror bit by writing 1 to it */
1092 	cxl_p1_write(adapter, CXL_PSL_ErrIVTE, CXL_PSL_ErrIVTE_tberror);
1093 	return cxl_tlb_slb_invalidate(adapter);
1094 }
1095 
1096 /* This should contain *only* operations that can safely be done in
1097  * both creation and recovery.
1098  */
1099 static int cxl_configure_adapter(struct cxl *adapter, struct pci_dev *dev)
1100 {
1101 	int rc;
1102 
1103 	adapter->dev.parent = &dev->dev;
1104 	adapter->dev.release = cxl_release_adapter;
1105 	pci_set_drvdata(dev, adapter);
1106 
1107 	rc = pci_enable_device(dev);
1108 	if (rc) {
1109 		dev_err(&dev->dev, "pci_enable_device failed: %i\n", rc);
1110 		return rc;
1111 	}
1112 
1113 	if ((rc = cxl_read_vsec(adapter, dev)))
1114 		return rc;
1115 
1116 	if ((rc = cxl_vsec_looks_ok(adapter, dev)))
1117 	        return rc;
1118 
1119 	cxl_fixup_malformed_tlp(adapter, dev);
1120 
1121 	if ((rc = setup_cxl_bars(dev)))
1122 		return rc;
1123 
1124 	if ((rc = switch_card_to_cxl(dev)))
1125 		return rc;
1126 
1127 	if ((rc = cxl_update_image_control(adapter)))
1128 		return rc;
1129 
1130 	if ((rc = cxl_map_adapter_regs(adapter, dev)))
1131 		return rc;
1132 
1133 	if ((rc = sanitise_adapter_regs(adapter)))
1134 		goto err;
1135 
1136 	if ((rc = init_implementation_adapter_regs(adapter, dev)))
1137 		goto err;
1138 
1139 	if ((rc = pnv_phb_to_cxl_mode(dev, OPAL_PHB_CAPI_MODE_CAPI)))
1140 		goto err;
1141 
1142 	/* If recovery happened, the last step is to turn on snooping.
1143 	 * In the non-recovery case this has no effect */
1144 	if ((rc = pnv_phb_to_cxl_mode(dev, OPAL_PHB_CAPI_MODE_SNOOP_ON)))
1145 		goto err;
1146 
1147 	if ((rc = cxl_setup_psl_timebase(adapter, dev)))
1148 		goto err;
1149 
1150 	if ((rc = cxl_native_register_psl_err_irq(adapter)))
1151 		goto err;
1152 
1153 	return 0;
1154 
1155 err:
1156 	cxl_unmap_adapter_regs(adapter);
1157 	return rc;
1158 
1159 }
1160 
1161 static void cxl_deconfigure_adapter(struct cxl *adapter)
1162 {
1163 	struct pci_dev *pdev = to_pci_dev(adapter->dev.parent);
1164 
1165 	cxl_native_release_psl_err_irq(adapter);
1166 	cxl_unmap_adapter_regs(adapter);
1167 
1168 	pci_disable_device(pdev);
1169 }
1170 
1171 static struct cxl *cxl_pci_init_adapter(struct pci_dev *dev)
1172 {
1173 	struct cxl *adapter;
1174 	int rc;
1175 
1176 	adapter = cxl_alloc_adapter();
1177 	if (!adapter)
1178 		return ERR_PTR(-ENOMEM);
1179 
1180 	adapter->native = kzalloc(sizeof(struct cxl_native), GFP_KERNEL);
1181 	if (!adapter->native) {
1182 		rc = -ENOMEM;
1183 		goto err_release;
1184 	}
1185 
1186 	/* Set defaults for parameters which need to persist over
1187 	 * configure/reconfigure
1188 	 */
1189 	adapter->perst_loads_image = true;
1190 	adapter->perst_same_image = false;
1191 
1192 	rc = cxl_configure_adapter(adapter, dev);
1193 	if (rc) {
1194 		pci_disable_device(dev);
1195 		goto err_release;
1196 	}
1197 
1198 	/* Don't care if this one fails: */
1199 	cxl_debugfs_adapter_add(adapter);
1200 
1201 	/*
1202 	 * After we call this function we must not free the adapter directly,
1203 	 * even if it returns an error!
1204 	 */
1205 	if ((rc = cxl_register_adapter(adapter)))
1206 		goto err_put1;
1207 
1208 	if ((rc = cxl_sysfs_adapter_add(adapter)))
1209 		goto err_put1;
1210 
1211 	return adapter;
1212 
1213 err_put1:
1214 	/* This should mirror cxl_remove_adapter, except without the
1215 	 * sysfs parts
1216 	 */
1217 	cxl_debugfs_adapter_remove(adapter);
1218 	cxl_deconfigure_adapter(adapter);
1219 	device_unregister(&adapter->dev);
1220 	return ERR_PTR(rc);
1221 
1222 err_release:
1223 	cxl_release_adapter(&adapter->dev);
1224 	return ERR_PTR(rc);
1225 }
1226 
1227 static void cxl_pci_remove_adapter(struct cxl *adapter)
1228 {
1229 	pr_devel("cxl_remove_adapter\n");
1230 
1231 	cxl_sysfs_adapter_remove(adapter);
1232 	cxl_debugfs_adapter_remove(adapter);
1233 
1234 	cxl_deconfigure_adapter(adapter);
1235 
1236 	device_unregister(&adapter->dev);
1237 }
1238 
1239 static int cxl_probe(struct pci_dev *dev, const struct pci_device_id *id)
1240 {
1241 	struct cxl *adapter;
1242 	int slice;
1243 	int rc;
1244 
1245 	if (cxl_pci_is_vphb_device(dev)) {
1246 		dev_dbg(&dev->dev, "cxl_init_adapter: Ignoring cxl vphb device\n");
1247 		return -ENODEV;
1248 	}
1249 
1250 	if (cxl_verbose)
1251 		dump_cxl_config_space(dev);
1252 
1253 	adapter = cxl_pci_init_adapter(dev);
1254 	if (IS_ERR(adapter)) {
1255 		dev_err(&dev->dev, "cxl_init_adapter failed: %li\n", PTR_ERR(adapter));
1256 		return PTR_ERR(adapter);
1257 	}
1258 
1259 	for (slice = 0; slice < adapter->slices; slice++) {
1260 		if ((rc = pci_init_afu(adapter, slice, dev))) {
1261 			dev_err(&dev->dev, "AFU %i failed to initialise: %i\n", slice, rc);
1262 			continue;
1263 		}
1264 
1265 		rc = cxl_afu_select_best_mode(adapter->afu[slice]);
1266 		if (rc)
1267 			dev_err(&dev->dev, "AFU %i failed to start: %i\n", slice, rc);
1268 	}
1269 
1270 	return 0;
1271 }
1272 
1273 static void cxl_remove(struct pci_dev *dev)
1274 {
1275 	struct cxl *adapter = pci_get_drvdata(dev);
1276 	struct cxl_afu *afu;
1277 	int i;
1278 
1279 	/*
1280 	 * Lock to prevent someone grabbing a ref through the adapter list as
1281 	 * we are removing it
1282 	 */
1283 	for (i = 0; i < adapter->slices; i++) {
1284 		afu = adapter->afu[i];
1285 		cxl_pci_remove_afu(afu);
1286 	}
1287 	cxl_pci_remove_adapter(adapter);
1288 }
1289 
1290 static pci_ers_result_t cxl_vphb_error_detected(struct cxl_afu *afu,
1291 						pci_channel_state_t state)
1292 {
1293 	struct pci_dev *afu_dev;
1294 	pci_ers_result_t result = PCI_ERS_RESULT_NEED_RESET;
1295 	pci_ers_result_t afu_result = PCI_ERS_RESULT_NEED_RESET;
1296 
1297 	/* There should only be one entry, but go through the list
1298 	 * anyway
1299 	 */
1300 	list_for_each_entry(afu_dev, &afu->phb->bus->devices, bus_list) {
1301 		if (!afu_dev->driver)
1302 			continue;
1303 
1304 		afu_dev->error_state = state;
1305 
1306 		if (afu_dev->driver->err_handler)
1307 			afu_result = afu_dev->driver->err_handler->error_detected(afu_dev,
1308 										  state);
1309 		/* Disconnect trumps all, NONE trumps NEED_RESET */
1310 		if (afu_result == PCI_ERS_RESULT_DISCONNECT)
1311 			result = PCI_ERS_RESULT_DISCONNECT;
1312 		else if ((afu_result == PCI_ERS_RESULT_NONE) &&
1313 			 (result == PCI_ERS_RESULT_NEED_RESET))
1314 			result = PCI_ERS_RESULT_NONE;
1315 	}
1316 	return result;
1317 }
1318 
1319 static pci_ers_result_t cxl_pci_error_detected(struct pci_dev *pdev,
1320 					       pci_channel_state_t state)
1321 {
1322 	struct cxl *adapter = pci_get_drvdata(pdev);
1323 	struct cxl_afu *afu;
1324 	pci_ers_result_t result = PCI_ERS_RESULT_NEED_RESET;
1325 	int i;
1326 
1327 	/* At this point, we could still have an interrupt pending.
1328 	 * Let's try to get them out of the way before they do
1329 	 * anything we don't like.
1330 	 */
1331 	schedule();
1332 
1333 	/* If we're permanently dead, give up. */
1334 	if (state == pci_channel_io_perm_failure) {
1335 		/* Tell the AFU drivers; but we don't care what they
1336 		 * say, we're going away.
1337 		 */
1338 		for (i = 0; i < adapter->slices; i++) {
1339 			afu = adapter->afu[i];
1340 			cxl_vphb_error_detected(afu, state);
1341 		}
1342 		return PCI_ERS_RESULT_DISCONNECT;
1343 	}
1344 
1345 	/* Are we reflashing?
1346 	 *
1347 	 * If we reflash, we could come back as something entirely
1348 	 * different, including a non-CAPI card. As such, by default
1349 	 * we don't participate in the process. We'll be unbound and
1350 	 * the slot re-probed. (TODO: check EEH doesn't blindly rebind
1351 	 * us!)
1352 	 *
1353 	 * However, this isn't the entire story: for reliablity
1354 	 * reasons, we usually want to reflash the FPGA on PERST in
1355 	 * order to get back to a more reliable known-good state.
1356 	 *
1357 	 * This causes us a bit of a problem: if we reflash we can't
1358 	 * trust that we'll come back the same - we could have a new
1359 	 * image and been PERSTed in order to load that
1360 	 * image. However, most of the time we actually *will* come
1361 	 * back the same - for example a regular EEH event.
1362 	 *
1363 	 * Therefore, we allow the user to assert that the image is
1364 	 * indeed the same and that we should continue on into EEH
1365 	 * anyway.
1366 	 */
1367 	if (adapter->perst_loads_image && !adapter->perst_same_image) {
1368 		/* TODO take the PHB out of CXL mode */
1369 		dev_info(&pdev->dev, "reflashing, so opting out of EEH!\n");
1370 		return PCI_ERS_RESULT_NONE;
1371 	}
1372 
1373 	/*
1374 	 * At this point, we want to try to recover.  We'll always
1375 	 * need a complete slot reset: we don't trust any other reset.
1376 	 *
1377 	 * Now, we go through each AFU:
1378 	 *  - We send the driver, if bound, an error_detected callback.
1379 	 *    We expect it to clean up, but it can also tell us to give
1380 	 *    up and permanently detach the card. To simplify things, if
1381 	 *    any bound AFU driver doesn't support EEH, we give up on EEH.
1382 	 *
1383 	 *  - We detach all contexts associated with the AFU. This
1384 	 *    does not free them, but puts them into a CLOSED state
1385 	 *    which causes any the associated files to return useful
1386 	 *    errors to userland. It also unmaps, but does not free,
1387 	 *    any IRQs.
1388 	 *
1389 	 *  - We clean up our side: releasing and unmapping resources we hold
1390 	 *    so we can wire them up again when the hardware comes back up.
1391 	 *
1392 	 * Driver authors should note:
1393 	 *
1394 	 *  - Any contexts you create in your kernel driver (except
1395 	 *    those associated with anonymous file descriptors) are
1396 	 *    your responsibility to free and recreate. Likewise with
1397 	 *    any attached resources.
1398 	 *
1399 	 *  - We will take responsibility for re-initialising the
1400 	 *    device context (the one set up for you in
1401 	 *    cxl_pci_enable_device_hook and accessed through
1402 	 *    cxl_get_context). If you've attached IRQs or other
1403 	 *    resources to it, they remains yours to free.
1404 	 *
1405 	 * You can call the same functions to release resources as you
1406 	 * normally would: we make sure that these functions continue
1407 	 * to work when the hardware is down.
1408 	 *
1409 	 * Two examples:
1410 	 *
1411 	 * 1) If you normally free all your resources at the end of
1412 	 *    each request, or if you use anonymous FDs, your
1413 	 *    error_detected callback can simply set a flag to tell
1414 	 *    your driver not to start any new calls. You can then
1415 	 *    clear the flag in the resume callback.
1416 	 *
1417 	 * 2) If you normally allocate your resources on startup:
1418 	 *     * Set a flag in error_detected as above.
1419 	 *     * Let CXL detach your contexts.
1420 	 *     * In slot_reset, free the old resources and allocate new ones.
1421 	 *     * In resume, clear the flag to allow things to start.
1422 	 */
1423 	for (i = 0; i < adapter->slices; i++) {
1424 		afu = adapter->afu[i];
1425 
1426 		result = cxl_vphb_error_detected(afu, state);
1427 
1428 		/* Only continue if everyone agrees on NEED_RESET */
1429 		if (result != PCI_ERS_RESULT_NEED_RESET)
1430 			return result;
1431 
1432 		cxl_context_detach_all(afu);
1433 		cxl_ops->afu_deactivate_mode(afu, afu->current_mode);
1434 		pci_deconfigure_afu(afu);
1435 	}
1436 	cxl_deconfigure_adapter(adapter);
1437 
1438 	return result;
1439 }
1440 
1441 static pci_ers_result_t cxl_pci_slot_reset(struct pci_dev *pdev)
1442 {
1443 	struct cxl *adapter = pci_get_drvdata(pdev);
1444 	struct cxl_afu *afu;
1445 	struct cxl_context *ctx;
1446 	struct pci_dev *afu_dev;
1447 	pci_ers_result_t afu_result = PCI_ERS_RESULT_RECOVERED;
1448 	pci_ers_result_t result = PCI_ERS_RESULT_RECOVERED;
1449 	int i;
1450 
1451 	if (cxl_configure_adapter(adapter, pdev))
1452 		goto err;
1453 
1454 	for (i = 0; i < adapter->slices; i++) {
1455 		afu = adapter->afu[i];
1456 
1457 		if (pci_configure_afu(afu, adapter, pdev))
1458 			goto err;
1459 
1460 		if (cxl_afu_select_best_mode(afu))
1461 			goto err;
1462 
1463 		list_for_each_entry(afu_dev, &afu->phb->bus->devices, bus_list) {
1464 			/* Reset the device context.
1465 			 * TODO: make this less disruptive
1466 			 */
1467 			ctx = cxl_get_context(afu_dev);
1468 
1469 			if (ctx && cxl_release_context(ctx))
1470 				goto err;
1471 
1472 			ctx = cxl_dev_context_init(afu_dev);
1473 			if (!ctx)
1474 				goto err;
1475 
1476 			afu_dev->dev.archdata.cxl_ctx = ctx;
1477 
1478 			if (cxl_ops->afu_check_and_enable(afu))
1479 				goto err;
1480 
1481 			afu_dev->error_state = pci_channel_io_normal;
1482 
1483 			/* If there's a driver attached, allow it to
1484 			 * chime in on recovery. Drivers should check
1485 			 * if everything has come back OK, but
1486 			 * shouldn't start new work until we call
1487 			 * their resume function.
1488 			 */
1489 			if (!afu_dev->driver)
1490 				continue;
1491 
1492 			if (afu_dev->driver->err_handler &&
1493 			    afu_dev->driver->err_handler->slot_reset)
1494 				afu_result = afu_dev->driver->err_handler->slot_reset(afu_dev);
1495 
1496 			if (afu_result == PCI_ERS_RESULT_DISCONNECT)
1497 				result = PCI_ERS_RESULT_DISCONNECT;
1498 		}
1499 	}
1500 	return result;
1501 
1502 err:
1503 	/* All the bits that happen in both error_detected and cxl_remove
1504 	 * should be idempotent, so we don't need to worry about leaving a mix
1505 	 * of unconfigured and reconfigured resources.
1506 	 */
1507 	dev_err(&pdev->dev, "EEH recovery failed. Asking to be disconnected.\n");
1508 	return PCI_ERS_RESULT_DISCONNECT;
1509 }
1510 
1511 static void cxl_pci_resume(struct pci_dev *pdev)
1512 {
1513 	struct cxl *adapter = pci_get_drvdata(pdev);
1514 	struct cxl_afu *afu;
1515 	struct pci_dev *afu_dev;
1516 	int i;
1517 
1518 	/* Everything is back now. Drivers should restart work now.
1519 	 * This is not the place to be checking if everything came back up
1520 	 * properly, because there's no return value: do that in slot_reset.
1521 	 */
1522 	for (i = 0; i < adapter->slices; i++) {
1523 		afu = adapter->afu[i];
1524 
1525 		list_for_each_entry(afu_dev, &afu->phb->bus->devices, bus_list) {
1526 			if (afu_dev->driver && afu_dev->driver->err_handler &&
1527 			    afu_dev->driver->err_handler->resume)
1528 				afu_dev->driver->err_handler->resume(afu_dev);
1529 		}
1530 	}
1531 }
1532 
1533 static const struct pci_error_handlers cxl_err_handler = {
1534 	.error_detected = cxl_pci_error_detected,
1535 	.slot_reset = cxl_pci_slot_reset,
1536 	.resume = cxl_pci_resume,
1537 };
1538 
1539 struct pci_driver cxl_pci_driver = {
1540 	.name = "cxl-pci",
1541 	.id_table = cxl_pci_tbl,
1542 	.probe = cxl_probe,
1543 	.remove = cxl_remove,
1544 	.shutdown = cxl_remove,
1545 	.err_handler = &cxl_err_handler,
1546 };
1547