xref: /linux/drivers/cxl/mem.c (revision 24b10e5f8e0d2bee1a10fc67011ea5d936c1a389)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /* Copyright(c) 2022 Intel Corporation. All rights reserved. */
3 #include <linux/debugfs.h>
4 #include <linux/device.h>
5 #include <linux/module.h>
6 #include <linux/pci.h>
7 
8 #include "cxlmem.h"
9 #include "cxlpci.h"
10 
11 /**
12  * DOC: cxl mem
13  *
14  * CXL memory endpoint devices and switches are CXL capable devices that are
15  * participating in CXL.mem protocol. Their functionality builds on top of the
16  * CXL.io protocol that allows enumerating and configuring components via
17  * standard PCI mechanisms.
18  *
19  * The cxl_mem driver owns kicking off the enumeration of this CXL.mem
20  * capability. With the detection of a CXL capable endpoint, the driver will
21  * walk up to find the platform specific port it is connected to, and determine
22  * if there are intervening switches in the path. If there are switches, a
23  * secondary action is to enumerate those (implemented in cxl_core). Finally the
24  * cxl_mem driver adds the device it is bound to as a CXL endpoint-port for use
25  * in higher level operations.
26  */
27 
28 static void enable_suspend(void *data)
29 {
30 	cxl_mem_active_dec();
31 }
32 
33 static void remove_debugfs(void *dentry)
34 {
35 	debugfs_remove_recursive(dentry);
36 }
37 
38 static int cxl_mem_dpa_show(struct seq_file *file, void *data)
39 {
40 	struct device *dev = file->private;
41 	struct cxl_memdev *cxlmd = to_cxl_memdev(dev);
42 
43 	cxl_dpa_debug(file, cxlmd->cxlds);
44 
45 	return 0;
46 }
47 
48 static int devm_cxl_add_endpoint(struct device *host, struct cxl_memdev *cxlmd,
49 				 struct cxl_dport *parent_dport)
50 {
51 	struct cxl_port *parent_port = parent_dport->port;
52 	struct cxl_port *endpoint, *iter, *down;
53 	int rc;
54 
55 	/*
56 	 * Now that the path to the root is established record all the
57 	 * intervening ports in the chain.
58 	 */
59 	for (iter = parent_port, down = NULL; !is_cxl_root(iter);
60 	     down = iter, iter = to_cxl_port(iter->dev.parent)) {
61 		struct cxl_ep *ep;
62 
63 		ep = cxl_ep_load(iter, cxlmd);
64 		ep->next = down;
65 	}
66 
67 	/* Note: endpoint port component registers are derived from @cxlds */
68 	endpoint = devm_cxl_add_port(host, &cxlmd->dev, CXL_RESOURCE_NONE,
69 				     parent_dport);
70 	if (IS_ERR(endpoint))
71 		return PTR_ERR(endpoint);
72 
73 	rc = cxl_endpoint_autoremove(cxlmd, endpoint);
74 	if (rc)
75 		return rc;
76 
77 	if (!endpoint->dev.driver) {
78 		dev_err(&cxlmd->dev, "%s failed probe\n",
79 			dev_name(&endpoint->dev));
80 		return -ENXIO;
81 	}
82 
83 	return 0;
84 }
85 
86 static int cxl_debugfs_poison_inject(void *data, u64 dpa)
87 {
88 	struct cxl_memdev *cxlmd = data;
89 
90 	return cxl_inject_poison(cxlmd, dpa);
91 }
92 
93 DEFINE_DEBUGFS_ATTRIBUTE(cxl_poison_inject_fops, NULL,
94 			 cxl_debugfs_poison_inject, "%llx\n");
95 
96 static int cxl_debugfs_poison_clear(void *data, u64 dpa)
97 {
98 	struct cxl_memdev *cxlmd = data;
99 
100 	return cxl_clear_poison(cxlmd, dpa);
101 }
102 
103 DEFINE_DEBUGFS_ATTRIBUTE(cxl_poison_clear_fops, NULL,
104 			 cxl_debugfs_poison_clear, "%llx\n");
105 
106 static int cxl_mem_probe(struct device *dev)
107 {
108 	struct cxl_memdev *cxlmd = to_cxl_memdev(dev);
109 	struct cxl_memdev_state *mds = to_cxl_memdev_state(cxlmd->cxlds);
110 	struct cxl_dev_state *cxlds = cxlmd->cxlds;
111 	struct device *endpoint_parent;
112 	struct cxl_port *parent_port;
113 	struct cxl_dport *dport;
114 	struct dentry *dentry;
115 	int rc;
116 
117 	if (!cxlds->media_ready)
118 		return -EBUSY;
119 
120 	/*
121 	 * Someone is trying to reattach this device after it lost its port
122 	 * connection (an endpoint port previously registered by this memdev was
123 	 * disabled). This racy check is ok because if the port is still gone,
124 	 * no harm done, and if the port hierarchy comes back it will re-trigger
125 	 * this probe. Port rescan and memdev detach work share the same
126 	 * single-threaded workqueue.
127 	 */
128 	if (work_pending(&cxlmd->detach_work))
129 		return -EBUSY;
130 
131 	dentry = cxl_debugfs_create_dir(dev_name(dev));
132 	debugfs_create_devm_seqfile(dev, "dpamem", dentry, cxl_mem_dpa_show);
133 
134 	if (test_bit(CXL_POISON_ENABLED_INJECT, mds->poison.enabled_cmds))
135 		debugfs_create_file("inject_poison", 0200, dentry, cxlmd,
136 				    &cxl_poison_inject_fops);
137 	if (test_bit(CXL_POISON_ENABLED_CLEAR, mds->poison.enabled_cmds))
138 		debugfs_create_file("clear_poison", 0200, dentry, cxlmd,
139 				    &cxl_poison_clear_fops);
140 
141 	rc = devm_add_action_or_reset(dev, remove_debugfs, dentry);
142 	if (rc)
143 		return rc;
144 
145 	rc = devm_cxl_enumerate_ports(cxlmd);
146 	if (rc)
147 		return rc;
148 
149 	parent_port = cxl_mem_find_port(cxlmd, &dport);
150 	if (!parent_port) {
151 		dev_err(dev, "CXL port topology not found\n");
152 		return -ENXIO;
153 	}
154 
155 	if (dport->rch)
156 		endpoint_parent = parent_port->uport_dev;
157 	else
158 		endpoint_parent = &parent_port->dev;
159 
160 	cxl_setup_parent_dport(dev, dport);
161 
162 	device_lock(endpoint_parent);
163 	if (!endpoint_parent->driver) {
164 		dev_err(dev, "CXL port topology %s not enabled\n",
165 			dev_name(endpoint_parent));
166 		rc = -ENXIO;
167 		goto unlock;
168 	}
169 
170 	rc = devm_cxl_add_endpoint(endpoint_parent, cxlmd, dport);
171 unlock:
172 	device_unlock(endpoint_parent);
173 	put_device(&parent_port->dev);
174 	if (rc)
175 		return rc;
176 
177 	if (resource_size(&cxlds->pmem_res) && IS_ENABLED(CONFIG_CXL_PMEM)) {
178 		rc = devm_cxl_add_nvdimm(cxlmd);
179 		if (rc == -ENODEV)
180 			dev_info(dev, "PMEM disabled by platform\n");
181 		else
182 			return rc;
183 	}
184 
185 	/*
186 	 * The kernel may be operating out of CXL memory on this device,
187 	 * there is no spec defined way to determine whether this device
188 	 * preserves contents over suspend, and there is no simple way
189 	 * to arrange for the suspend image to avoid CXL memory which
190 	 * would setup a circular dependency between PCI resume and save
191 	 * state restoration.
192 	 *
193 	 * TODO: support suspend when all the regions this device is
194 	 * hosting are locked and covered by the system address map,
195 	 * i.e. platform firmware owns restoring the HDM configuration
196 	 * that it locked.
197 	 */
198 	cxl_mem_active_inc();
199 	return devm_add_action_or_reset(dev, enable_suspend, NULL);
200 }
201 
202 static ssize_t trigger_poison_list_store(struct device *dev,
203 					 struct device_attribute *attr,
204 					 const char *buf, size_t len)
205 {
206 	bool trigger;
207 	int rc;
208 
209 	if (kstrtobool(buf, &trigger) || !trigger)
210 		return -EINVAL;
211 
212 	rc = cxl_trigger_poison_list(to_cxl_memdev(dev));
213 
214 	return rc ? rc : len;
215 }
216 static DEVICE_ATTR_WO(trigger_poison_list);
217 
218 static ssize_t ram_qos_class_show(struct device *dev,
219 				  struct device_attribute *attr, char *buf)
220 {
221 	struct cxl_memdev *cxlmd = to_cxl_memdev(dev);
222 	struct cxl_dev_state *cxlds = cxlmd->cxlds;
223 	struct cxl_memdev_state *mds = to_cxl_memdev_state(cxlds);
224 	struct cxl_dpa_perf *dpa_perf;
225 
226 	if (!dev->driver)
227 		return -ENOENT;
228 
229 	if (list_empty(&mds->ram_perf_list))
230 		return -ENOENT;
231 
232 	dpa_perf = list_first_entry(&mds->ram_perf_list, struct cxl_dpa_perf,
233 				    list);
234 
235 	return sysfs_emit(buf, "%d\n", dpa_perf->qos_class);
236 }
237 
238 static struct device_attribute dev_attr_ram_qos_class =
239 	__ATTR(qos_class, 0444, ram_qos_class_show, NULL);
240 
241 static ssize_t pmem_qos_class_show(struct device *dev,
242 				   struct device_attribute *attr, char *buf)
243 {
244 	struct cxl_memdev *cxlmd = to_cxl_memdev(dev);
245 	struct cxl_dev_state *cxlds = cxlmd->cxlds;
246 	struct cxl_memdev_state *mds = to_cxl_memdev_state(cxlds);
247 	struct cxl_dpa_perf *dpa_perf;
248 
249 	if (!dev->driver)
250 		return -ENOENT;
251 
252 	if (list_empty(&mds->pmem_perf_list))
253 		return -ENOENT;
254 
255 	dpa_perf = list_first_entry(&mds->pmem_perf_list, struct cxl_dpa_perf,
256 				    list);
257 
258 	return sysfs_emit(buf, "%d\n", dpa_perf->qos_class);
259 }
260 
261 static struct device_attribute dev_attr_pmem_qos_class =
262 	__ATTR(qos_class, 0444, pmem_qos_class_show, NULL);
263 
264 static umode_t cxl_mem_visible(struct kobject *kobj, struct attribute *a, int n)
265 {
266 	struct device *dev = kobj_to_dev(kobj);
267 	struct cxl_memdev *cxlmd = to_cxl_memdev(dev);
268 	struct cxl_memdev_state *mds = to_cxl_memdev_state(cxlmd->cxlds);
269 
270 	if (a == &dev_attr_trigger_poison_list.attr)
271 		if (!test_bit(CXL_POISON_ENABLED_LIST,
272 			      mds->poison.enabled_cmds))
273 			return 0;
274 
275 	if (a == &dev_attr_pmem_qos_class.attr)
276 		if (list_empty(&mds->pmem_perf_list))
277 			return 0;
278 
279 	if (a == &dev_attr_ram_qos_class.attr)
280 		if (list_empty(&mds->ram_perf_list))
281 			return 0;
282 
283 	return a->mode;
284 }
285 
286 static struct attribute *cxl_mem_attrs[] = {
287 	&dev_attr_trigger_poison_list.attr,
288 	&dev_attr_ram_qos_class.attr,
289 	&dev_attr_pmem_qos_class.attr,
290 	NULL
291 };
292 
293 static struct attribute_group cxl_mem_group = {
294 	.attrs = cxl_mem_attrs,
295 	.is_visible = cxl_mem_visible,
296 };
297 
298 __ATTRIBUTE_GROUPS(cxl_mem);
299 
300 static struct cxl_driver cxl_mem_driver = {
301 	.name = "cxl_mem",
302 	.probe = cxl_mem_probe,
303 	.id = CXL_DEVICE_MEMORY_EXPANDER,
304 	.drv = {
305 		.dev_groups = cxl_mem_groups,
306 	},
307 };
308 
309 module_cxl_driver(cxl_mem_driver);
310 
311 MODULE_LICENSE("GPL v2");
312 MODULE_IMPORT_NS(CXL);
313 MODULE_ALIAS_CXL(CXL_DEVICE_MEMORY_EXPANDER);
314 /*
315  * create_endpoint() wants to validate port driver attach immediately after
316  * endpoint registration.
317  */
318 MODULE_SOFTDEP("pre: cxl_port");
319