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 <uapi/linux/idxd.h>
8 #include "idxd.h"
9 #include "registers.h"
10
__get_desc(struct idxd_wq * wq,int idx,int cpu)11 static struct idxd_desc *__get_desc(struct idxd_wq *wq, int idx, int cpu)
12 {
13 struct idxd_desc *desc;
14 struct idxd_device *idxd = wq->idxd;
15
16 desc = wq->descs[idx];
17 memset(desc->hw, 0, sizeof(struct dsa_hw_desc));
18 memset(desc->completion, 0, idxd->data->compl_size);
19 desc->cpu = cpu;
20
21 if (device_pasid_enabled(idxd))
22 desc->hw->pasid = idxd->pasid;
23
24 /*
25 * On host, MSIX vecotr 0 is used for misc interrupt. Therefore when we match
26 * vector 1:1 to the WQ id, we need to add 1
27 */
28 if (!idxd->int_handles)
29 desc->hw->int_handle = wq->id + 1;
30 else
31 desc->hw->int_handle = idxd->int_handles[wq->id];
32
33 return desc;
34 }
35
idxd_alloc_desc(struct idxd_wq * wq,enum idxd_op_type optype)36 struct idxd_desc *idxd_alloc_desc(struct idxd_wq *wq, enum idxd_op_type optype)
37 {
38 int cpu, idx;
39 struct idxd_device *idxd = wq->idxd;
40 DEFINE_SBQ_WAIT(wait);
41 struct sbq_wait_state *ws;
42 struct sbitmap_queue *sbq;
43
44 if (idxd->state != IDXD_DEV_ENABLED)
45 return ERR_PTR(-EIO);
46
47 sbq = &wq->sbq;
48 idx = sbitmap_queue_get(sbq, &cpu);
49 if (idx < 0) {
50 if (optype == IDXD_OP_NONBLOCK)
51 return ERR_PTR(-EAGAIN);
52 } else {
53 return __get_desc(wq, idx, cpu);
54 }
55
56 ws = &sbq->ws[0];
57 for (;;) {
58 sbitmap_prepare_to_wait(sbq, ws, &wait, TASK_INTERRUPTIBLE);
59 if (signal_pending_state(TASK_INTERRUPTIBLE, current))
60 break;
61 idx = sbitmap_queue_get(sbq, &cpu);
62 if (idx >= 0)
63 break;
64 schedule();
65 }
66
67 sbitmap_finish_wait(sbq, ws, &wait);
68 if (idx < 0)
69 return ERR_PTR(-EAGAIN);
70
71 return __get_desc(wq, idx, cpu);
72 }
73
idxd_free_desc(struct idxd_wq * wq,struct idxd_desc * desc)74 void idxd_free_desc(struct idxd_wq *wq, struct idxd_desc *desc)
75 {
76 int cpu = desc->cpu;
77
78 desc->cpu = -1;
79 sbitmap_queue_clear(&wq->sbq, desc->id, cpu);
80 }
81
list_abort_desc(struct idxd_wq * wq,struct idxd_irq_entry * ie,struct idxd_desc * desc)82 static struct idxd_desc *list_abort_desc(struct idxd_wq *wq, struct idxd_irq_entry *ie,
83 struct idxd_desc *desc)
84 {
85 struct idxd_desc *d, *n;
86
87 lockdep_assert_held(&ie->list_lock);
88 list_for_each_entry_safe(d, n, &ie->work_list, list) {
89 if (d == desc) {
90 list_del(&d->list);
91 return d;
92 }
93 }
94
95 /*
96 * At this point, the desc needs to be aborted is held by the completion
97 * handler where it has taken it off the pending list but has not added to the
98 * work list. It will be cleaned up by the interrupt handler when it sees the
99 * IDXD_COMP_DESC_ABORT for completion status.
100 */
101 return NULL;
102 }
103
llist_abort_desc(struct idxd_wq * wq,struct idxd_irq_entry * ie,struct idxd_desc * desc)104 static void llist_abort_desc(struct idxd_wq *wq, struct idxd_irq_entry *ie,
105 struct idxd_desc *desc)
106 {
107 struct idxd_desc *d, *t, *found = NULL;
108 struct llist_node *head;
109 LIST_HEAD(flist);
110
111 desc->completion->status = IDXD_COMP_DESC_ABORT;
112 /*
113 * Grab the list lock so it will block the irq thread handler. This allows the
114 * abort code to locate the descriptor need to be aborted.
115 */
116 spin_lock(&ie->list_lock);
117 head = llist_del_all(&ie->pending_llist);
118 if (head) {
119 llist_for_each_entry_safe(d, t, head, llnode) {
120 if (d == desc) {
121 found = desc;
122 continue;
123 }
124
125 if (d->completion->status)
126 list_add_tail(&d->list, &flist);
127 else
128 list_add_tail(&d->list, &ie->work_list);
129 }
130 }
131
132 if (!found)
133 found = list_abort_desc(wq, ie, desc);
134 spin_unlock(&ie->list_lock);
135
136 if (found)
137 complete_desc(found, IDXD_COMPLETE_ABORT);
138
139 /*
140 * complete_desc() will return desc to allocator and the desc can be
141 * acquired by a different process and the desc->list can be modified.
142 * Delete desc from list so the list trasversing does not get corrupted
143 * by the other process.
144 */
145 list_for_each_entry_safe(d, t, &flist, list) {
146 list_del_init(&d->list);
147 complete_desc(d, IDXD_COMPLETE_NORMAL);
148 }
149 }
150
idxd_submit_desc(struct idxd_wq * wq,struct idxd_desc * desc)151 int idxd_submit_desc(struct idxd_wq *wq, struct idxd_desc *desc)
152 {
153 struct idxd_device *idxd = wq->idxd;
154 struct idxd_irq_entry *ie = NULL;
155 void __iomem *portal;
156 int rc;
157
158 if (idxd->state != IDXD_DEV_ENABLED) {
159 idxd_free_desc(wq, desc);
160 return -EIO;
161 }
162
163 if (!percpu_ref_tryget_live(&wq->wq_active)) {
164 idxd_free_desc(wq, desc);
165 return -ENXIO;
166 }
167
168 portal = idxd_wq_portal_addr(wq);
169
170 /*
171 * The wmb() flushes writes to coherent DMA data before
172 * possibly triggering a DMA read. The wmb() is necessary
173 * even on UP because the recipient is a device.
174 */
175 wmb();
176
177 /*
178 * Pending the descriptor to the lockless list for the irq_entry
179 * that we designated the descriptor to.
180 */
181 if (desc->hw->flags & IDXD_OP_FLAG_RCI) {
182 ie = &idxd->irq_entries[wq->id + 1];
183 llist_add(&desc->llnode, &ie->pending_llist);
184 }
185
186 if (wq_dedicated(wq)) {
187 iosubmit_cmds512(portal, desc->hw, 1);
188 } else {
189 /*
190 * It's not likely that we would receive queue full rejection
191 * since the descriptor allocation gates at wq size. If we
192 * receive a -EAGAIN, that means something went wrong such as the
193 * device is not accepting descriptor at all.
194 */
195 rc = enqcmds(portal, desc->hw);
196 if (rc < 0) {
197 percpu_ref_put(&wq->wq_active);
198 /* abort operation frees the descriptor */
199 if (ie)
200 llist_abort_desc(wq, ie, desc);
201 else
202 idxd_free_desc(wq, desc);
203 return rc;
204 }
205 }
206
207 percpu_ref_put(&wq->wq_active);
208 return 0;
209 }
210