1 /*
2 * Coherency fabric (Aurora) support for Armada 370, 375, 38x and XP
3 * platforms.
4 *
5 * Copyright (C) 2012 Marvell
6 *
7 * Yehuda Yitschak <yehuday@marvell.com>
8 * Gregory Clement <gregory.clement@free-electrons.com>
9 * Thomas Petazzoni <thomas.petazzoni@free-electrons.com>
10 *
11 * This file is licensed under the terms of the GNU General Public
12 * License version 2. This program is licensed "as is" without any
13 * warranty of any kind, whether express or implied.
14 *
15 * The Armada 370, 375, 38x and XP SOCs have a coherency fabric which is
16 * responsible for ensuring hardware coherency between all CPUs and between
17 * CPUs and I/O masters. This file initializes the coherency fabric and
18 * supplies basic routines for configuring and controlling hardware coherency
19 */
20
21 #define pr_fmt(fmt) "mvebu-coherency: " fmt
22
23 #include <linux/kernel.h>
24 #include <linux/init.h>
25 #include <linux/of_address.h>
26 #include <linux/io.h>
27 #include <linux/smp.h>
28 #include <linux/dma-mapping.h>
29 #include <linux/platform_device.h>
30 #include <linux/slab.h>
31 #include <linux/mbus.h>
32 #include <linux/pci.h>
33 #include <asm/smp_plat.h>
34 #include <asm/cacheflush.h>
35 #include <asm/mach/map.h>
36 #include <asm/dma-mapping.h>
37 #include "coherency.h"
38 #include "mvebu-soc-id.h"
39
40 unsigned long coherency_phys_base;
41 void __iomem *coherency_base;
42 static void __iomem *coherency_cpu_base;
43 static void __iomem *cpu_config_base;
44
45 /* Coherency fabric registers */
46 #define IO_SYNC_BARRIER_CTL_OFFSET 0x0
47
48 enum {
49 COHERENCY_FABRIC_TYPE_NONE,
50 COHERENCY_FABRIC_TYPE_ARMADA_370_XP,
51 COHERENCY_FABRIC_TYPE_ARMADA_375,
52 COHERENCY_FABRIC_TYPE_ARMADA_380,
53 };
54
55 static const struct of_device_id of_coherency_table[] = {
56 {.compatible = "marvell,coherency-fabric",
57 .data = (void *) COHERENCY_FABRIC_TYPE_ARMADA_370_XP },
58 {.compatible = "marvell,armada-375-coherency-fabric",
59 .data = (void *) COHERENCY_FABRIC_TYPE_ARMADA_375 },
60 {.compatible = "marvell,armada-380-coherency-fabric",
61 .data = (void *) COHERENCY_FABRIC_TYPE_ARMADA_380 },
62 { /* end of list */ },
63 };
64
65 /* Functions defined in coherency_ll.S */
66 int ll_enable_coherency(void);
67 void ll_add_cpu_to_smp_group(void);
68
69 #define CPU_CONFIG_SHARED_L2 BIT(16)
70
71 /*
72 * Disable the "Shared L2 Present" bit in CPU Configuration register
73 * on Armada XP.
74 *
75 * The "Shared L2 Present" bit affects the "level of coherence" value
76 * in the clidr CP15 register. Cache operation functions such as
77 * "flush all" and "invalidate all" operate on all the cache levels
78 * that included in the defined level of coherence. When HW I/O
79 * coherency is used, this bit causes unnecessary flushes of the L2
80 * cache.
81 */
armada_xp_clear_shared_l2(void)82 static void armada_xp_clear_shared_l2(void)
83 {
84 u32 reg;
85
86 if (!cpu_config_base)
87 return;
88
89 reg = readl(cpu_config_base);
90 reg &= ~CPU_CONFIG_SHARED_L2;
91 writel(reg, cpu_config_base);
92 }
93
mvebu_hwcc_notifier(struct notifier_block * nb,unsigned long event,void * __dev)94 static int mvebu_hwcc_notifier(struct notifier_block *nb,
95 unsigned long event, void *__dev)
96 {
97 struct device *dev = __dev;
98
99 if (event != BUS_NOTIFY_ADD_DEVICE)
100 return NOTIFY_DONE;
101 set_dma_ops(dev, &arm_coherent_dma_ops);
102
103 return NOTIFY_OK;
104 }
105
106 static struct notifier_block mvebu_hwcc_nb = {
107 .notifier_call = mvebu_hwcc_notifier,
108 };
109
110 static struct notifier_block mvebu_hwcc_pci_nb = {
111 .notifier_call = mvebu_hwcc_notifier,
112 };
113
armada_xp_clear_shared_l2_notifier_func(struct notifier_block * nfb,unsigned long action,void * hcpu)114 static int armada_xp_clear_shared_l2_notifier_func(struct notifier_block *nfb,
115 unsigned long action, void *hcpu)
116 {
117 if (action == CPU_STARTING || action == CPU_STARTING_FROZEN)
118 armada_xp_clear_shared_l2();
119
120 return NOTIFY_OK;
121 }
122
123 static struct notifier_block armada_xp_clear_shared_l2_notifier = {
124 .notifier_call = armada_xp_clear_shared_l2_notifier_func,
125 .priority = 100,
126 };
127
armada_370_coherency_init(struct device_node * np)128 static void __init armada_370_coherency_init(struct device_node *np)
129 {
130 struct resource res;
131 struct device_node *cpu_config_np;
132
133 of_address_to_resource(np, 0, &res);
134 coherency_phys_base = res.start;
135 /*
136 * Ensure secondary CPUs will see the updated value,
137 * which they read before they join the coherency
138 * fabric, and therefore before they are coherent with
139 * the boot CPU cache.
140 */
141 sync_cache_w(&coherency_phys_base);
142 coherency_base = of_iomap(np, 0);
143 coherency_cpu_base = of_iomap(np, 1);
144
145 cpu_config_np = of_find_compatible_node(NULL, NULL,
146 "marvell,armada-xp-cpu-config");
147 if (!cpu_config_np)
148 goto exit;
149
150 cpu_config_base = of_iomap(cpu_config_np, 0);
151 if (!cpu_config_base) {
152 of_node_put(cpu_config_np);
153 goto exit;
154 }
155
156 of_node_put(cpu_config_np);
157
158 register_cpu_notifier(&armada_xp_clear_shared_l2_notifier);
159
160 exit:
161 set_cpu_coherent();
162 }
163
164 /*
165 * This ioremap hook is used on Armada 375/38x to ensure that all MMIO
166 * areas are mapped as MT_UNCACHED instead of MT_DEVICE. This is
167 * needed for the HW I/O coherency mechanism to work properly without
168 * deadlock.
169 */
170 static void __iomem *
armada_wa_ioremap_caller(phys_addr_t phys_addr,size_t size,unsigned int mtype,void * caller)171 armada_wa_ioremap_caller(phys_addr_t phys_addr, size_t size,
172 unsigned int mtype, void *caller)
173 {
174 mtype = MT_UNCACHED;
175 return __arm_ioremap_caller(phys_addr, size, mtype, caller);
176 }
177
armada_375_380_coherency_init(struct device_node * np)178 static void __init armada_375_380_coherency_init(struct device_node *np)
179 {
180 struct device_node *cache_dn;
181
182 coherency_cpu_base = of_iomap(np, 0);
183 arch_ioremap_caller = armada_wa_ioremap_caller;
184
185 /*
186 * We should switch the PL310 to I/O coherency mode only if
187 * I/O coherency is actually enabled.
188 */
189 if (!coherency_available())
190 return;
191
192 /*
193 * Add the PL310 property "arm,io-coherent". This makes sure the
194 * outer sync operation is not used, which allows to
195 * workaround the system erratum that causes deadlocks when
196 * doing PCIe in an SMP situation on Armada 375 and Armada
197 * 38x.
198 */
199 for_each_compatible_node(cache_dn, NULL, "arm,pl310-cache") {
200 struct property *p;
201
202 p = kzalloc(sizeof(*p), GFP_KERNEL);
203 p->name = kstrdup("arm,io-coherent", GFP_KERNEL);
204 of_add_property(cache_dn, p);
205 }
206 }
207
coherency_type(void)208 static int coherency_type(void)
209 {
210 struct device_node *np;
211 const struct of_device_id *match;
212 int type;
213
214 /*
215 * The coherency fabric is needed:
216 * - For coherency between processors on Armada XP, so only
217 * when SMP is enabled.
218 * - For coherency between the processor and I/O devices, but
219 * this coherency requires many pre-requisites (write
220 * allocate cache policy, shareable pages, SMP bit set) that
221 * are only meant in SMP situations.
222 *
223 * Note that this means that on Armada 370, there is currently
224 * no way to use hardware I/O coherency, because even when
225 * CONFIG_SMP is enabled, is_smp() returns false due to the
226 * Armada 370 being a single-core processor. To lift this
227 * limitation, we would have to find a way to make the cache
228 * policy set to write-allocate (on all Armada SoCs), and to
229 * set the shareable attribute in page tables (on all Armada
230 * SoCs except the Armada 370). Unfortunately, such decisions
231 * are taken very early in the kernel boot process, at a point
232 * where we don't know yet on which SoC we are running.
233
234 */
235 if (!is_smp())
236 return COHERENCY_FABRIC_TYPE_NONE;
237
238 np = of_find_matching_node_and_match(NULL, of_coherency_table, &match);
239 if (!np)
240 return COHERENCY_FABRIC_TYPE_NONE;
241
242 type = (int) match->data;
243
244 of_node_put(np);
245
246 return type;
247 }
248
set_cpu_coherent(void)249 int set_cpu_coherent(void)
250 {
251 int type = coherency_type();
252
253 if (type == COHERENCY_FABRIC_TYPE_ARMADA_370_XP) {
254 if (!coherency_base) {
255 pr_warn("Can't make current CPU cache coherent.\n");
256 pr_warn("Coherency fabric is not initialized\n");
257 return 1;
258 }
259
260 armada_xp_clear_shared_l2();
261 ll_add_cpu_to_smp_group();
262 return ll_enable_coherency();
263 }
264
265 return 0;
266 }
267
coherency_available(void)268 int coherency_available(void)
269 {
270 return coherency_type() != COHERENCY_FABRIC_TYPE_NONE;
271 }
272
coherency_init(void)273 int __init coherency_init(void)
274 {
275 int type = coherency_type();
276 struct device_node *np;
277
278 np = of_find_matching_node(NULL, of_coherency_table);
279
280 if (type == COHERENCY_FABRIC_TYPE_ARMADA_370_XP)
281 armada_370_coherency_init(np);
282 else if (type == COHERENCY_FABRIC_TYPE_ARMADA_375 ||
283 type == COHERENCY_FABRIC_TYPE_ARMADA_380)
284 armada_375_380_coherency_init(np);
285
286 of_node_put(np);
287
288 return 0;
289 }
290
coherency_late_init(void)291 static int __init coherency_late_init(void)
292 {
293 if (coherency_available())
294 bus_register_notifier(&platform_bus_type,
295 &mvebu_hwcc_nb);
296 return 0;
297 }
298
299 postcore_initcall(coherency_late_init);
300
301 #if IS_ENABLED(CONFIG_PCI)
coherency_pci_init(void)302 static int __init coherency_pci_init(void)
303 {
304 if (coherency_available())
305 bus_register_notifier(&pci_bus_type,
306 &mvebu_hwcc_pci_nb);
307 return 0;
308 }
309
310 arch_initcall(coherency_pci_init);
311 #endif
312