xref: /linux/drivers/dma/ioat/dca.c (revision e5a52fd2b8cdb700b3c07b030e050a49ef3156b9)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Intel I/OAT DMA Linux driver
4  * Copyright(c) 2007 - 2009 Intel Corporation.
5  */
6 
7 #include <linux/kernel.h>
8 #include <linux/pci.h>
9 #include <linux/smp.h>
10 #include <linux/interrupt.h>
11 #include <linux/dca.h>
12 
13 /* either a kernel change is needed, or we need something like this in kernel */
14 #ifndef CONFIG_SMP
15 #include <asm/smp.h>
16 #undef cpu_physical_id
17 #define cpu_physical_id(cpu) (cpuid_ebx(1) >> 24)
18 #endif
19 
20 #include "dma.h"
21 #include "registers.h"
22 
23 /*
24  * Bit 7 of a tag map entry is the "valid" bit, if it is set then bits 0:6
25  * contain the bit number of the APIC ID to map into the DCA tag.  If the valid
26  * bit is not set, then the value must be 0 or 1 and defines the bit in the tag.
27  */
28 #define DCA_TAG_MAP_VALID 0x80
29 
30 #define DCA3_TAG_MAP_BIT_TO_INV 0x80
31 #define DCA3_TAG_MAP_BIT_TO_SEL 0x40
32 #define DCA3_TAG_MAP_LITERAL_VAL 0x1
33 
34 #define DCA_TAG_MAP_MASK 0xDF
35 
36 /* expected tag map bytes for I/OAT ver.2 */
37 #define DCA2_TAG_MAP_BYTE0 0x80
38 #define DCA2_TAG_MAP_BYTE1 0x0
39 #define DCA2_TAG_MAP_BYTE2 0x81
40 #define DCA2_TAG_MAP_BYTE3 0x82
41 #define DCA2_TAG_MAP_BYTE4 0x82
42 
43 /* verify if tag map matches expected values */
44 static inline int dca2_tag_map_valid(u8 *tag_map)
45 {
46 	return ((tag_map[0] == DCA2_TAG_MAP_BYTE0) &&
47 		(tag_map[1] == DCA2_TAG_MAP_BYTE1) &&
48 		(tag_map[2] == DCA2_TAG_MAP_BYTE2) &&
49 		(tag_map[3] == DCA2_TAG_MAP_BYTE3) &&
50 		(tag_map[4] == DCA2_TAG_MAP_BYTE4));
51 }
52 
53 /*
54  * "Legacy" DCA systems do not implement the DCA register set in the
55  * I/OAT device.  Software needs direct support for their tag mappings.
56  */
57 
58 #define APICID_BIT(x)		(DCA_TAG_MAP_VALID | (x))
59 #define IOAT_TAG_MAP_LEN	8
60 
61 /* pack PCI B/D/F into a u16 */
62 static inline u16 dcaid_from_pcidev(struct pci_dev *pci)
63 {
64 	return (pci->bus->number << 8) | pci->devfn;
65 }
66 
67 static int dca_enabled_in_bios(struct pci_dev *pdev)
68 {
69 	/* CPUID level 9 returns DCA configuration */
70 	/* Bit 0 indicates DCA enabled by the BIOS */
71 	unsigned long cpuid_level_9;
72 	int res;
73 
74 	cpuid_level_9 = cpuid_eax(9);
75 	res = test_bit(0, &cpuid_level_9);
76 	if (!res)
77 		dev_dbg(&pdev->dev, "DCA is disabled in BIOS\n");
78 
79 	return res;
80 }
81 
82 int system_has_dca_enabled(struct pci_dev *pdev)
83 {
84 	if (boot_cpu_has(X86_FEATURE_DCA))
85 		return dca_enabled_in_bios(pdev);
86 
87 	dev_dbg(&pdev->dev, "boot cpu doesn't have X86_FEATURE_DCA\n");
88 	return 0;
89 }
90 
91 struct ioat_dca_slot {
92 	struct pci_dev *pdev;	/* requester device */
93 	u16 rid;		/* requester id, as used by IOAT */
94 };
95 
96 #define IOAT_DCA_MAX_REQ 6
97 #define IOAT3_DCA_MAX_REQ 2
98 
99 struct ioat_dca_priv {
100 	void __iomem		*iobase;
101 	void __iomem		*dca_base;
102 	int			 max_requesters;
103 	int			 requester_count;
104 	u8			 tag_map[IOAT_TAG_MAP_LEN];
105 	struct ioat_dca_slot	 req_slots[];
106 };
107 
108 static int ioat_dca_dev_managed(struct dca_provider *dca,
109 				struct device *dev)
110 {
111 	struct ioat_dca_priv *ioatdca = dca_priv(dca);
112 	struct pci_dev *pdev;
113 	int i;
114 
115 	pdev = to_pci_dev(dev);
116 	for (i = 0; i < ioatdca->max_requesters; i++) {
117 		if (ioatdca->req_slots[i].pdev == pdev)
118 			return 1;
119 	}
120 	return 0;
121 }
122 
123 static int ioat_dca_add_requester(struct dca_provider *dca, struct device *dev)
124 {
125 	struct ioat_dca_priv *ioatdca = dca_priv(dca);
126 	struct pci_dev *pdev;
127 	int i;
128 	u16 id;
129 	u16 global_req_table;
130 
131 	/* This implementation only supports PCI-Express */
132 	if (!dev_is_pci(dev))
133 		return -ENODEV;
134 	pdev = to_pci_dev(dev);
135 	id = dcaid_from_pcidev(pdev);
136 
137 	if (ioatdca->requester_count == ioatdca->max_requesters)
138 		return -ENODEV;
139 
140 	for (i = 0; i < ioatdca->max_requesters; i++) {
141 		if (ioatdca->req_slots[i].pdev == NULL) {
142 			/* found an empty slot */
143 			ioatdca->requester_count++;
144 			ioatdca->req_slots[i].pdev = pdev;
145 			ioatdca->req_slots[i].rid = id;
146 			global_req_table =
147 			      readw(ioatdca->dca_base + IOAT3_DCA_GREQID_OFFSET);
148 			writel(id | IOAT_DCA_GREQID_VALID,
149 			       ioatdca->iobase + global_req_table + (i * 4));
150 			return i;
151 		}
152 	}
153 	/* Error, ioatdma->requester_count is out of whack */
154 	return -EFAULT;
155 }
156 
157 static int ioat_dca_remove_requester(struct dca_provider *dca,
158 				      struct device *dev)
159 {
160 	struct ioat_dca_priv *ioatdca = dca_priv(dca);
161 	struct pci_dev *pdev;
162 	int i;
163 	u16 global_req_table;
164 
165 	/* This implementation only supports PCI-Express */
166 	if (!dev_is_pci(dev))
167 		return -ENODEV;
168 	pdev = to_pci_dev(dev);
169 
170 	for (i = 0; i < ioatdca->max_requesters; i++) {
171 		if (ioatdca->req_slots[i].pdev == pdev) {
172 			global_req_table =
173 			      readw(ioatdca->dca_base + IOAT3_DCA_GREQID_OFFSET);
174 			writel(0, ioatdca->iobase + global_req_table + (i * 4));
175 			ioatdca->req_slots[i].pdev = NULL;
176 			ioatdca->req_slots[i].rid = 0;
177 			ioatdca->requester_count--;
178 			return i;
179 		}
180 	}
181 	return -ENODEV;
182 }
183 
184 static u8 ioat_dca_get_tag(struct dca_provider *dca,
185 			    struct device *dev,
186 			    int cpu)
187 {
188 	u8 tag;
189 
190 	struct ioat_dca_priv *ioatdca = dca_priv(dca);
191 	int i, apic_id, bit, value;
192 	u8 entry;
193 
194 	tag = 0;
195 	apic_id = cpu_physical_id(cpu);
196 
197 	for (i = 0; i < IOAT_TAG_MAP_LEN; i++) {
198 		entry = ioatdca->tag_map[i];
199 		if (entry & DCA3_TAG_MAP_BIT_TO_SEL) {
200 			bit = entry &
201 				~(DCA3_TAG_MAP_BIT_TO_SEL | DCA3_TAG_MAP_BIT_TO_INV);
202 			value = (apic_id & (1 << bit)) ? 1 : 0;
203 		} else if (entry & DCA3_TAG_MAP_BIT_TO_INV) {
204 			bit = entry & ~DCA3_TAG_MAP_BIT_TO_INV;
205 			value = (apic_id & (1 << bit)) ? 0 : 1;
206 		} else {
207 			value = (entry & DCA3_TAG_MAP_LITERAL_VAL) ? 1 : 0;
208 		}
209 		tag |= (value << i);
210 	}
211 
212 	return tag;
213 }
214 
215 static const struct dca_ops ioat_dca_ops = {
216 	.add_requester		= ioat_dca_add_requester,
217 	.remove_requester	= ioat_dca_remove_requester,
218 	.get_tag		= ioat_dca_get_tag,
219 	.dev_managed		= ioat_dca_dev_managed,
220 };
221 
222 static int ioat_dca_count_dca_slots(void *iobase, u16 dca_offset)
223 {
224 	int slots = 0;
225 	u32 req;
226 	u16 global_req_table;
227 
228 	global_req_table = readw(iobase + dca_offset + IOAT3_DCA_GREQID_OFFSET);
229 	if (global_req_table == 0)
230 		return 0;
231 
232 	do {
233 		req = readl(iobase + global_req_table + (slots * sizeof(u32)));
234 		slots++;
235 	} while ((req & IOAT_DCA_GREQID_LASTID) == 0);
236 
237 	return slots;
238 }
239 
240 static inline int dca3_tag_map_invalid(u8 *tag_map)
241 {
242 	/*
243 	 * If the tag map is not programmed by the BIOS the default is:
244 	 * 0x80 0x80 0x80 0x80 0x80 0x00 0x00 0x00
245 	 *
246 	 * This an invalid map and will result in only 2 possible tags
247 	 * 0x1F and 0x00.  0x00 is an invalid DCA tag so we know that
248 	 * this entire definition is invalid.
249 	 */
250 	return ((tag_map[0] == DCA_TAG_MAP_VALID) &&
251 		(tag_map[1] == DCA_TAG_MAP_VALID) &&
252 		(tag_map[2] == DCA_TAG_MAP_VALID) &&
253 		(tag_map[3] == DCA_TAG_MAP_VALID) &&
254 		(tag_map[4] == DCA_TAG_MAP_VALID));
255 }
256 
257 struct dca_provider *ioat_dca_init(struct pci_dev *pdev, void __iomem *iobase)
258 {
259 	struct dca_provider *dca;
260 	struct ioat_dca_priv *ioatdca;
261 	int slots;
262 	int i;
263 	int err;
264 	u16 dca_offset;
265 	u16 csi_fsb_control;
266 	u16 pcie_control;
267 	u8 bit;
268 
269 	union {
270 		u64 full;
271 		struct {
272 			u32 low;
273 			u32 high;
274 		};
275 	} tag_map;
276 
277 	if (!system_has_dca_enabled(pdev))
278 		return NULL;
279 
280 	dca_offset = readw(iobase + IOAT_DCAOFFSET_OFFSET);
281 	if (dca_offset == 0)
282 		return NULL;
283 
284 	slots = ioat_dca_count_dca_slots(iobase, dca_offset);
285 	if (slots == 0)
286 		return NULL;
287 
288 	dca = alloc_dca_provider(&ioat_dca_ops,
289 				 struct_size(ioatdca, req_slots, slots));
290 	if (!dca)
291 		return NULL;
292 
293 	ioatdca = dca_priv(dca);
294 	ioatdca->iobase = iobase;
295 	ioatdca->dca_base = iobase + dca_offset;
296 	ioatdca->max_requesters = slots;
297 
298 	/* some bios might not know to turn these on */
299 	csi_fsb_control = readw(ioatdca->dca_base + IOAT3_CSI_CONTROL_OFFSET);
300 	if ((csi_fsb_control & IOAT3_CSI_CONTROL_PREFETCH) == 0) {
301 		csi_fsb_control |= IOAT3_CSI_CONTROL_PREFETCH;
302 		writew(csi_fsb_control,
303 		       ioatdca->dca_base + IOAT3_CSI_CONTROL_OFFSET);
304 	}
305 	pcie_control = readw(ioatdca->dca_base + IOAT3_PCI_CONTROL_OFFSET);
306 	if ((pcie_control & IOAT3_PCI_CONTROL_MEMWR) == 0) {
307 		pcie_control |= IOAT3_PCI_CONTROL_MEMWR;
308 		writew(pcie_control,
309 		       ioatdca->dca_base + IOAT3_PCI_CONTROL_OFFSET);
310 	}
311 
312 
313 	/* TODO version, compatibility and configuration checks */
314 
315 	/* copy out the APIC to DCA tag map */
316 	tag_map.low =
317 		readl(ioatdca->dca_base + IOAT3_APICID_TAG_MAP_OFFSET_LOW);
318 	tag_map.high =
319 		readl(ioatdca->dca_base + IOAT3_APICID_TAG_MAP_OFFSET_HIGH);
320 	for (i = 0; i < 8; i++) {
321 		bit = tag_map.full >> (8 * i);
322 		ioatdca->tag_map[i] = bit & DCA_TAG_MAP_MASK;
323 	}
324 
325 	if (dca3_tag_map_invalid(ioatdca->tag_map)) {
326 		add_taint(TAINT_FIRMWARE_WORKAROUND, LOCKDEP_STILL_OK);
327 		pr_warn_once("%s %s: APICID_TAG_MAP set incorrectly by BIOS, disabling DCA\n",
328 			     dev_driver_string(&pdev->dev),
329 			     dev_name(&pdev->dev));
330 		free_dca_provider(dca);
331 		return NULL;
332 	}
333 
334 	err = register_dca_provider(dca, &pdev->dev);
335 	if (err) {
336 		free_dca_provider(dca);
337 		return NULL;
338 	}
339 
340 	return dca;
341 }
342