1 // SPDX-License-Identifier: GPL-2.0
2 /* Copyright(c) 2019 Intel Corporation. All rights rsvd. */
3 #include <linux/init.h>
4 #include <linux/kernel.h>
5 #include <linux/module.h>
6 #include <linux/pci.h>
7 #include <linux/io-64-nonatomic-lo-hi.h>
8 #include <linux/dmaengine.h>
9 #include <uapi/linux/idxd.h>
10 #include "../dmaengine.h"
11 #include "idxd.h"
12 #include "registers.h"
13 
14 enum irq_work_type {
15 	IRQ_WORK_NORMAL = 0,
16 	IRQ_WORK_PROCESS_FAULT,
17 };
18 
19 struct idxd_fault {
20 	struct work_struct work;
21 	u64 addr;
22 	struct idxd_device *idxd;
23 };
24 
idxd_device_reinit(struct work_struct * work)25 static void idxd_device_reinit(struct work_struct *work)
26 {
27 	struct idxd_device *idxd = container_of(work, struct idxd_device, work);
28 	struct device *dev = &idxd->pdev->dev;
29 	int rc, i;
30 
31 	idxd_device_reset(idxd);
32 	rc = idxd_device_config(idxd);
33 	if (rc < 0)
34 		goto out;
35 
36 	rc = idxd_device_enable(idxd);
37 	if (rc < 0)
38 		goto out;
39 
40 	for (i = 0; i < idxd->max_wqs; i++) {
41 		struct idxd_wq *wq = idxd->wqs[i];
42 
43 		if (wq->state == IDXD_WQ_ENABLED) {
44 			rc = idxd_wq_enable(wq);
45 			if (rc < 0) {
46 				dev_warn(dev, "Unable to re-enable wq %s\n",
47 					 dev_name(wq_confdev(wq)));
48 			}
49 		}
50 	}
51 
52 	return;
53 
54  out:
55 	idxd_device_clear_state(idxd);
56 }
57 
process_misc_interrupts(struct idxd_device * idxd,u32 cause)58 static int process_misc_interrupts(struct idxd_device *idxd, u32 cause)
59 {
60 	struct device *dev = &idxd->pdev->dev;
61 	union gensts_reg gensts;
62 	u32 val = 0;
63 	int i;
64 	bool err = false;
65 
66 	if (cause & IDXD_INTC_HALT_STATE)
67 		goto halt;
68 
69 	if (cause & IDXD_INTC_ERR) {
70 		spin_lock(&idxd->dev_lock);
71 		for (i = 0; i < 4; i++)
72 			idxd->sw_err.bits[i] = ioread64(idxd->reg_base +
73 					IDXD_SWERR_OFFSET + i * sizeof(u64));
74 
75 		iowrite64(idxd->sw_err.bits[0] & IDXD_SWERR_ACK,
76 			  idxd->reg_base + IDXD_SWERR_OFFSET);
77 
78 		if (idxd->sw_err.valid && idxd->sw_err.wq_idx_valid) {
79 			int id = idxd->sw_err.wq_idx;
80 			struct idxd_wq *wq = idxd->wqs[id];
81 
82 			if (wq->type == IDXD_WQT_USER)
83 				wake_up_interruptible(&wq->err_queue);
84 		} else {
85 			int i;
86 
87 			for (i = 0; i < idxd->max_wqs; i++) {
88 				struct idxd_wq *wq = idxd->wqs[i];
89 
90 				if (wq->type == IDXD_WQT_USER)
91 					wake_up_interruptible(&wq->err_queue);
92 			}
93 		}
94 
95 		spin_unlock(&idxd->dev_lock);
96 		val |= IDXD_INTC_ERR;
97 
98 		for (i = 0; i < 4; i++)
99 			dev_warn(dev, "err[%d]: %#16.16llx\n",
100 				 i, idxd->sw_err.bits[i]);
101 		err = true;
102 	}
103 
104 	if (cause & IDXD_INTC_CMD) {
105 		val |= IDXD_INTC_CMD;
106 		complete(idxd->cmd_done);
107 	}
108 
109 	if (cause & IDXD_INTC_OCCUPY) {
110 		/* Driver does not utilize occupancy interrupt */
111 		val |= IDXD_INTC_OCCUPY;
112 	}
113 
114 	if (cause & IDXD_INTC_PERFMON_OVFL) {
115 		val |= IDXD_INTC_PERFMON_OVFL;
116 		perfmon_counter_overflow(idxd);
117 	}
118 
119 	val ^= cause;
120 	if (val)
121 		dev_warn_once(dev, "Unexpected interrupt cause bits set: %#x\n",
122 			      val);
123 
124 	if (!err)
125 		return 0;
126 
127 halt:
128 	gensts.bits = ioread32(idxd->reg_base + IDXD_GENSTATS_OFFSET);
129 	if (gensts.state == IDXD_DEVICE_STATE_HALT) {
130 		idxd->state = IDXD_DEV_HALTED;
131 		if (gensts.reset_type == IDXD_DEVICE_RESET_SOFTWARE) {
132 			/*
133 			 * If we need a software reset, we will throw the work
134 			 * on a system workqueue in order to allow interrupts
135 			 * for the device command completions.
136 			 */
137 			INIT_WORK(&idxd->work, idxd_device_reinit);
138 			queue_work(idxd->wq, &idxd->work);
139 		} else {
140 			idxd->state = IDXD_DEV_HALTED;
141 			idxd_wqs_quiesce(idxd);
142 			idxd_wqs_unmap_portal(idxd);
143 			spin_lock(&idxd->dev_lock);
144 			idxd_device_clear_state(idxd);
145 			dev_err(&idxd->pdev->dev,
146 				"idxd halted, need %s.\n",
147 				gensts.reset_type == IDXD_DEVICE_RESET_FLR ?
148 				"FLR" : "system reset");
149 			spin_unlock(&idxd->dev_lock);
150 			return -ENXIO;
151 		}
152 	}
153 
154 	return 0;
155 }
156 
idxd_misc_thread(int vec,void * data)157 irqreturn_t idxd_misc_thread(int vec, void *data)
158 {
159 	struct idxd_irq_entry *irq_entry = data;
160 	struct idxd_device *idxd = irq_entry->idxd;
161 	int rc;
162 	u32 cause;
163 
164 	cause = ioread32(idxd->reg_base + IDXD_INTCAUSE_OFFSET);
165 	if (cause)
166 		iowrite32(cause, idxd->reg_base + IDXD_INTCAUSE_OFFSET);
167 
168 	while (cause) {
169 		rc = process_misc_interrupts(idxd, cause);
170 		if (rc < 0)
171 			break;
172 		cause = ioread32(idxd->reg_base + IDXD_INTCAUSE_OFFSET);
173 		if (cause)
174 			iowrite32(cause, idxd->reg_base + IDXD_INTCAUSE_OFFSET);
175 	}
176 
177 	return IRQ_HANDLED;
178 }
179 
irq_process_pending_llist(struct idxd_irq_entry * irq_entry)180 static void irq_process_pending_llist(struct idxd_irq_entry *irq_entry)
181 {
182 	struct idxd_desc *desc, *t;
183 	struct llist_node *head;
184 
185 	head = llist_del_all(&irq_entry->pending_llist);
186 	if (!head)
187 		return;
188 
189 	llist_for_each_entry_safe(desc, t, head, llnode) {
190 		u8 status = desc->completion->status & DSA_COMP_STATUS_MASK;
191 
192 		if (status) {
193 			/*
194 			 * Check against the original status as ABORT is software defined
195 			 * and 0xff, which DSA_COMP_STATUS_MASK can mask out.
196 			 */
197 			if (unlikely(desc->completion->status == IDXD_COMP_DESC_ABORT)) {
198 				complete_desc(desc, IDXD_COMPLETE_ABORT);
199 				continue;
200 			}
201 
202 			complete_desc(desc, IDXD_COMPLETE_NORMAL);
203 		} else {
204 			spin_lock(&irq_entry->list_lock);
205 			list_add_tail(&desc->list,
206 				      &irq_entry->work_list);
207 			spin_unlock(&irq_entry->list_lock);
208 		}
209 	}
210 }
211 
irq_process_work_list(struct idxd_irq_entry * irq_entry)212 static void irq_process_work_list(struct idxd_irq_entry *irq_entry)
213 {
214 	LIST_HEAD(flist);
215 	struct idxd_desc *desc, *n;
216 
217 	/*
218 	 * This lock protects list corruption from access of list outside of the irq handler
219 	 * thread.
220 	 */
221 	spin_lock(&irq_entry->list_lock);
222 	if (list_empty(&irq_entry->work_list)) {
223 		spin_unlock(&irq_entry->list_lock);
224 		return;
225 	}
226 
227 	list_for_each_entry_safe(desc, n, &irq_entry->work_list, list) {
228 		if (desc->completion->status) {
229 			list_move_tail(&desc->list, &flist);
230 		}
231 	}
232 
233 	spin_unlock(&irq_entry->list_lock);
234 
235 	list_for_each_entry(desc, &flist, list) {
236 		/*
237 		 * Check against the original status as ABORT is software defined
238 		 * and 0xff, which DSA_COMP_STATUS_MASK can mask out.
239 		 */
240 		if (unlikely(desc->completion->status == IDXD_COMP_DESC_ABORT)) {
241 			complete_desc(desc, IDXD_COMPLETE_ABORT);
242 			continue;
243 		}
244 
245 		complete_desc(desc, IDXD_COMPLETE_NORMAL);
246 	}
247 }
248 
idxd_wq_thread(int irq,void * data)249 irqreturn_t idxd_wq_thread(int irq, void *data)
250 {
251 	struct idxd_irq_entry *irq_entry = data;
252 
253 	/*
254 	 * There are two lists we are processing. The pending_llist is where
255 	 * submmiter adds all the submitted descriptor after sending it to
256 	 * the workqueue. It's a lockless singly linked list. The work_list
257 	 * is the common linux double linked list. We are in a scenario of
258 	 * multiple producers and a single consumer. The producers are all
259 	 * the kernel submitters of descriptors, and the consumer is the
260 	 * kernel irq handler thread for the msix vector when using threaded
261 	 * irq. To work with the restrictions of llist to remain lockless,
262 	 * we are doing the following steps:
263 	 * 1. Iterate through the work_list and process any completed
264 	 *    descriptor. Delete the completed entries during iteration.
265 	 * 2. llist_del_all() from the pending list.
266 	 * 3. Iterate through the llist that was deleted from the pending list
267 	 *    and process the completed entries.
268 	 * 4. If the entry is still waiting on hardware, list_add_tail() to
269 	 *    the work_list.
270 	 */
271 	irq_process_work_list(irq_entry);
272 	irq_process_pending_llist(irq_entry);
273 
274 	return IRQ_HANDLED;
275 }
276