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