xref: /linux/drivers/dma/ioat/dca.c (revision e814f3fd16acfb7f9966773953de8f740a1e3202)
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 #include <asm/cpuid.h>
14 
15 /* either a kernel change is needed, or we need something like this in kernel */
16 #ifndef CONFIG_SMP
17 #include <asm/smp.h>
18 #undef cpu_physical_id
19 #define cpu_physical_id(cpu) (cpuid_ebx(1) >> 24)
20 #endif
21 
22 #include "dma.h"
23 #include "registers.h"
24 
25 /*
26  * Bit 7 of a tag map entry is the "valid" bit, if it is set then bits 0:6
27  * contain the bit number of the APIC ID to map into the DCA tag.  If the valid
28  * bit is not set, then the value must be 0 or 1 and defines the bit in the tag.
29  */
30 #define DCA_TAG_MAP_VALID 0x80
31 
32 #define DCA3_TAG_MAP_BIT_TO_INV 0x80
33 #define DCA3_TAG_MAP_BIT_TO_SEL 0x40
34 #define DCA3_TAG_MAP_LITERAL_VAL 0x1
35 
36 #define DCA_TAG_MAP_MASK 0xDF
37 
38 /* expected tag map bytes for I/OAT ver.2 */
39 #define DCA2_TAG_MAP_BYTE0 0x80
40 #define DCA2_TAG_MAP_BYTE1 0x0
41 #define DCA2_TAG_MAP_BYTE2 0x81
42 #define DCA2_TAG_MAP_BYTE3 0x82
43 #define DCA2_TAG_MAP_BYTE4 0x82
44 
45 /*
46  * "Legacy" DCA systems do not implement the DCA register set in the
47  * I/OAT device.  Software needs direct support for their tag mappings.
48  */
49 
50 #define APICID_BIT(x)		(DCA_TAG_MAP_VALID | (x))
51 #define IOAT_TAG_MAP_LEN	8
52 
53 /* pack PCI B/D/F into a u16 */
54 static inline u16 dcaid_from_pcidev(struct pci_dev *pci)
55 {
56 	return pci_dev_id(pci);
57 }
58 
59 static int dca_enabled_in_bios(struct pci_dev *pdev)
60 {
61 	/* CPUID level 9 returns DCA configuration */
62 	/* Bit 0 indicates DCA enabled by the BIOS */
63 	u32 eax;
64 	int res;
65 
66 	eax = cpuid_eax(CPUID_LEAF_DCA);
67 	res = eax & BIT(0);
68 	if (!res)
69 		dev_dbg(&pdev->dev, "DCA is disabled in BIOS\n");
70 
71 	return res;
72 }
73 
74 int system_has_dca_enabled(struct pci_dev *pdev)
75 {
76 	if (boot_cpu_has(X86_FEATURE_DCA))
77 		return dca_enabled_in_bios(pdev);
78 
79 	dev_dbg(&pdev->dev, "boot cpu doesn't have X86_FEATURE_DCA\n");
80 	return 0;
81 }
82 
83 struct ioat_dca_slot {
84 	struct pci_dev *pdev;	/* requester device */
85 	u16 rid;		/* requester id, as used by IOAT */
86 };
87 
88 #define IOAT_DCA_MAX_REQ 6
89 #define IOAT3_DCA_MAX_REQ 2
90 
91 struct ioat_dca_priv {
92 	void __iomem		*iobase;
93 	void __iomem		*dca_base;
94 	int			 max_requesters;
95 	int			 requester_count;
96 	u8			 tag_map[IOAT_TAG_MAP_LEN];
97 	struct ioat_dca_slot	 req_slots[];
98 };
99 
100 static int ioat_dca_dev_managed(struct dca_provider *dca,
101 				struct device *dev)
102 {
103 	struct ioat_dca_priv *ioatdca = dca_priv(dca);
104 	struct pci_dev *pdev;
105 	int i;
106 
107 	pdev = to_pci_dev(dev);
108 	for (i = 0; i < ioatdca->max_requesters; i++) {
109 		if (ioatdca->req_slots[i].pdev == pdev)
110 			return 1;
111 	}
112 	return 0;
113 }
114 
115 static int ioat_dca_add_requester(struct dca_provider *dca, struct device *dev)
116 {
117 	struct ioat_dca_priv *ioatdca = dca_priv(dca);
118 	struct pci_dev *pdev;
119 	int i;
120 	u16 id;
121 	u16 global_req_table;
122 
123 	/* This implementation only supports PCI-Express */
124 	if (!dev_is_pci(dev))
125 		return -ENODEV;
126 	pdev = to_pci_dev(dev);
127 	id = dcaid_from_pcidev(pdev);
128 
129 	if (ioatdca->requester_count == ioatdca->max_requesters)
130 		return -ENODEV;
131 
132 	for (i = 0; i < ioatdca->max_requesters; i++) {
133 		if (ioatdca->req_slots[i].pdev == NULL) {
134 			/* found an empty slot */
135 			ioatdca->requester_count++;
136 			ioatdca->req_slots[i].pdev = pdev;
137 			ioatdca->req_slots[i].rid = id;
138 			global_req_table =
139 			      readw(ioatdca->dca_base + IOAT3_DCA_GREQID_OFFSET);
140 			writel(id | IOAT_DCA_GREQID_VALID,
141 			       ioatdca->iobase + global_req_table + (i * 4));
142 			return i;
143 		}
144 	}
145 	/* Error, ioatdma->requester_count is out of whack */
146 	return -EFAULT;
147 }
148 
149 static int ioat_dca_remove_requester(struct dca_provider *dca,
150 				      struct device *dev)
151 {
152 	struct ioat_dca_priv *ioatdca = dca_priv(dca);
153 	struct pci_dev *pdev;
154 	int i;
155 	u16 global_req_table;
156 
157 	/* This implementation only supports PCI-Express */
158 	if (!dev_is_pci(dev))
159 		return -ENODEV;
160 	pdev = to_pci_dev(dev);
161 
162 	for (i = 0; i < ioatdca->max_requesters; i++) {
163 		if (ioatdca->req_slots[i].pdev == pdev) {
164 			global_req_table =
165 			      readw(ioatdca->dca_base + IOAT3_DCA_GREQID_OFFSET);
166 			writel(0, ioatdca->iobase + global_req_table + (i * 4));
167 			ioatdca->req_slots[i].pdev = NULL;
168 			ioatdca->req_slots[i].rid = 0;
169 			ioatdca->requester_count--;
170 			return i;
171 		}
172 	}
173 	return -ENODEV;
174 }
175 
176 static u8 ioat_dca_get_tag(struct dca_provider *dca,
177 			    struct device *dev,
178 			    int cpu)
179 {
180 	u8 tag;
181 
182 	struct ioat_dca_priv *ioatdca = dca_priv(dca);
183 	int i, apic_id, bit, value;
184 	u8 entry;
185 
186 	tag = 0;
187 	apic_id = cpu_physical_id(cpu);
188 
189 	for (i = 0; i < IOAT_TAG_MAP_LEN; i++) {
190 		entry = ioatdca->tag_map[i];
191 		if (entry & DCA3_TAG_MAP_BIT_TO_SEL) {
192 			bit = entry &
193 				~(DCA3_TAG_MAP_BIT_TO_SEL | DCA3_TAG_MAP_BIT_TO_INV);
194 			value = (apic_id & (1 << bit)) ? 1 : 0;
195 		} else if (entry & DCA3_TAG_MAP_BIT_TO_INV) {
196 			bit = entry & ~DCA3_TAG_MAP_BIT_TO_INV;
197 			value = (apic_id & (1 << bit)) ? 0 : 1;
198 		} else {
199 			value = (entry & DCA3_TAG_MAP_LITERAL_VAL) ? 1 : 0;
200 		}
201 		tag |= (value << i);
202 	}
203 
204 	return tag;
205 }
206 
207 static const struct dca_ops ioat_dca_ops = {
208 	.add_requester		= ioat_dca_add_requester,
209 	.remove_requester	= ioat_dca_remove_requester,
210 	.get_tag		= ioat_dca_get_tag,
211 	.dev_managed		= ioat_dca_dev_managed,
212 };
213 
214 static int ioat_dca_count_dca_slots(void *iobase, u16 dca_offset)
215 {
216 	int slots = 0;
217 	u32 req;
218 	u16 global_req_table;
219 
220 	global_req_table = readw(iobase + dca_offset + IOAT3_DCA_GREQID_OFFSET);
221 	if (global_req_table == 0)
222 		return 0;
223 
224 	do {
225 		req = readl(iobase + global_req_table + (slots * sizeof(u32)));
226 		slots++;
227 	} while ((req & IOAT_DCA_GREQID_LASTID) == 0);
228 
229 	return slots;
230 }
231 
232 static inline int dca3_tag_map_invalid(u8 *tag_map)
233 {
234 	/*
235 	 * If the tag map is not programmed by the BIOS the default is:
236 	 * 0x80 0x80 0x80 0x80 0x80 0x00 0x00 0x00
237 	 *
238 	 * This an invalid map and will result in only 2 possible tags
239 	 * 0x1F and 0x00.  0x00 is an invalid DCA tag so we know that
240 	 * this entire definition is invalid.
241 	 */
242 	return ((tag_map[0] == DCA_TAG_MAP_VALID) &&
243 		(tag_map[1] == DCA_TAG_MAP_VALID) &&
244 		(tag_map[2] == DCA_TAG_MAP_VALID) &&
245 		(tag_map[3] == DCA_TAG_MAP_VALID) &&
246 		(tag_map[4] == DCA_TAG_MAP_VALID));
247 }
248 
249 struct dca_provider *ioat_dca_init(struct pci_dev *pdev, void __iomem *iobase)
250 {
251 	struct dca_provider *dca;
252 	struct ioat_dca_priv *ioatdca;
253 	int slots;
254 	int i;
255 	int err;
256 	u16 dca_offset;
257 	u16 csi_fsb_control;
258 	u16 pcie_control;
259 	u8 bit;
260 
261 	union {
262 		u64 full;
263 		struct {
264 			u32 low;
265 			u32 high;
266 		};
267 	} tag_map;
268 
269 	if (!system_has_dca_enabled(pdev))
270 		return NULL;
271 
272 	dca_offset = readw(iobase + IOAT_DCAOFFSET_OFFSET);
273 	if (dca_offset == 0)
274 		return NULL;
275 
276 	slots = ioat_dca_count_dca_slots(iobase, dca_offset);
277 	if (slots == 0)
278 		return NULL;
279 
280 	dca = alloc_dca_provider(&ioat_dca_ops,
281 				 struct_size(ioatdca, req_slots, slots));
282 	if (!dca)
283 		return NULL;
284 
285 	ioatdca = dca_priv(dca);
286 	ioatdca->iobase = iobase;
287 	ioatdca->dca_base = iobase + dca_offset;
288 	ioatdca->max_requesters = slots;
289 
290 	/* some bios might not know to turn these on */
291 	csi_fsb_control = readw(ioatdca->dca_base + IOAT3_CSI_CONTROL_OFFSET);
292 	if ((csi_fsb_control & IOAT3_CSI_CONTROL_PREFETCH) == 0) {
293 		csi_fsb_control |= IOAT3_CSI_CONTROL_PREFETCH;
294 		writew(csi_fsb_control,
295 		       ioatdca->dca_base + IOAT3_CSI_CONTROL_OFFSET);
296 	}
297 	pcie_control = readw(ioatdca->dca_base + IOAT3_PCI_CONTROL_OFFSET);
298 	if ((pcie_control & IOAT3_PCI_CONTROL_MEMWR) == 0) {
299 		pcie_control |= IOAT3_PCI_CONTROL_MEMWR;
300 		writew(pcie_control,
301 		       ioatdca->dca_base + IOAT3_PCI_CONTROL_OFFSET);
302 	}
303 
304 
305 	/* TODO version, compatibility and configuration checks */
306 
307 	/* copy out the APIC to DCA tag map */
308 	tag_map.low =
309 		readl(ioatdca->dca_base + IOAT3_APICID_TAG_MAP_OFFSET_LOW);
310 	tag_map.high =
311 		readl(ioatdca->dca_base + IOAT3_APICID_TAG_MAP_OFFSET_HIGH);
312 	for (i = 0; i < 8; i++) {
313 		bit = tag_map.full >> (8 * i);
314 		ioatdca->tag_map[i] = bit & DCA_TAG_MAP_MASK;
315 	}
316 
317 	if (dca3_tag_map_invalid(ioatdca->tag_map)) {
318 		add_taint(TAINT_FIRMWARE_WORKAROUND, LOCKDEP_STILL_OK);
319 		pr_warn_once("%s %s: APICID_TAG_MAP set incorrectly by BIOS, disabling DCA\n",
320 			     dev_driver_string(&pdev->dev),
321 			     dev_name(&pdev->dev));
322 		free_dca_provider(dca);
323 		return NULL;
324 	}
325 
326 	err = register_dca_provider(dca, &pdev->dev);
327 	if (err) {
328 		free_dca_provider(dca);
329 		return NULL;
330 	}
331 
332 	return dca;
333 }
334