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 return desc;
25 }
26
idxd_alloc_desc(struct idxd_wq * wq,enum idxd_op_type optype)27 struct idxd_desc *idxd_alloc_desc(struct idxd_wq *wq, enum idxd_op_type optype)
28 {
29 int cpu, idx;
30 struct idxd_device *idxd = wq->idxd;
31 DEFINE_SBQ_WAIT(wait);
32 struct sbq_wait_state *ws;
33 struct sbitmap_queue *sbq;
34
35 if (idxd->state != IDXD_DEV_ENABLED)
36 return ERR_PTR(-EIO);
37
38 sbq = &wq->sbq;
39 idx = sbitmap_queue_get(sbq, &cpu);
40 if (idx < 0) {
41 if (optype == IDXD_OP_NONBLOCK)
42 return ERR_PTR(-EAGAIN);
43 } else {
44 return __get_desc(wq, idx, cpu);
45 }
46
47 ws = &sbq->ws[0];
48 for (;;) {
49 sbitmap_prepare_to_wait(sbq, ws, &wait, TASK_INTERRUPTIBLE);
50 if (signal_pending_state(TASK_INTERRUPTIBLE, current))
51 break;
52 idx = sbitmap_queue_get(sbq, &cpu);
53 if (idx >= 0)
54 break;
55 schedule();
56 }
57
58 sbitmap_finish_wait(sbq, ws, &wait);
59 if (idx < 0)
60 return ERR_PTR(-EAGAIN);
61
62 return __get_desc(wq, idx, cpu);
63 }
64 EXPORT_SYMBOL_NS_GPL(idxd_alloc_desc, "IDXD");
65
idxd_free_desc(struct idxd_wq * wq,struct idxd_desc * desc)66 void idxd_free_desc(struct idxd_wq *wq, struct idxd_desc *desc)
67 {
68 int cpu = desc->cpu;
69
70 desc->cpu = -1;
71 sbitmap_queue_clear(&wq->sbq, desc->id, cpu);
72 }
73 EXPORT_SYMBOL_NS_GPL(idxd_free_desc, "IDXD");
74
list_abort_desc(struct idxd_wq * wq,struct idxd_irq_entry * ie,struct idxd_desc * desc)75 static struct idxd_desc *list_abort_desc(struct idxd_wq *wq, struct idxd_irq_entry *ie,
76 struct idxd_desc *desc)
77 {
78 struct idxd_desc *d, *n;
79
80 lockdep_assert_held(&ie->list_lock);
81 list_for_each_entry_safe(d, n, &ie->work_list, list) {
82 if (d == desc) {
83 list_del(&d->list);
84 return d;
85 }
86 }
87
88 /*
89 * At this point, the desc needs to be aborted is held by the completion
90 * handler where it has taken it off the pending list but has not added to the
91 * work list. It will be cleaned up by the interrupt handler when it sees the
92 * IDXD_COMP_DESC_ABORT for completion status.
93 */
94 return NULL;
95 }
96
llist_abort_desc(struct idxd_wq * wq,struct idxd_irq_entry * ie,struct idxd_desc * desc)97 static void llist_abort_desc(struct idxd_wq *wq, struct idxd_irq_entry *ie,
98 struct idxd_desc *desc)
99 {
100 struct idxd_desc *d, *t, *found = NULL;
101 struct llist_node *head;
102 LIST_HEAD(flist);
103
104 desc->completion->status = IDXD_COMP_DESC_ABORT;
105 /*
106 * Grab the list lock so it will block the irq thread handler. This allows the
107 * abort code to locate the descriptor need to be aborted.
108 */
109 spin_lock(&ie->list_lock);
110 head = llist_del_all(&ie->pending_llist);
111 if (head) {
112 llist_for_each_entry_safe(d, t, head, llnode) {
113 if (d == desc) {
114 found = desc;
115 continue;
116 }
117
118 if (d->completion->status)
119 list_add_tail(&d->list, &flist);
120 else
121 list_add_tail(&d->list, &ie->work_list);
122 }
123 }
124
125 if (!found)
126 found = list_abort_desc(wq, ie, desc);
127 spin_unlock(&ie->list_lock);
128
129 if (found)
130 idxd_dma_complete_txd(found, IDXD_COMPLETE_ABORT, false,
131 NULL, NULL);
132
133 /*
134 * completing the descriptor will return desc to allocator and
135 * the desc can be acquired by a different process and the
136 * desc->list can be modified. Delete desc from list so the
137 * list traversing does not get corrupted by the other process.
138 */
139 list_for_each_entry_safe(d, t, &flist, list) {
140 list_del_init(&d->list);
141 idxd_dma_complete_txd(found, IDXD_COMPLETE_ABORT, true,
142 NULL, NULL);
143 }
144 }
145
146 /*
147 * ENQCMDS typically fail when the WQ is inactive or busy. On host submission, the driver
148 * has better control of number of descriptors being submitted to a shared wq by limiting
149 * the number of driver allocated descriptors to the wq size. However, when the swq is
150 * exported to a guest kernel, it may be shared with multiple guest kernels. This means
151 * the likelihood of getting busy returned on the swq when submitting goes significantly up.
152 * Having a tunable retry mechanism allows the driver to keep trying for a bit before giving
153 * up. The sysfs knob can be tuned by the system administrator.
154 */
idxd_enqcmds(struct idxd_wq * wq,void __iomem * portal,const void * desc)155 int idxd_enqcmds(struct idxd_wq *wq, void __iomem *portal, const void *desc)
156 {
157 unsigned int retries = wq->enqcmds_retries;
158 int rc;
159
160 do {
161 rc = enqcmds(portal, desc);
162 if (rc == 0)
163 break;
164 cpu_relax();
165 } while (retries--);
166
167 return rc;
168 }
169
idxd_submit_desc(struct idxd_wq * wq,struct idxd_desc * desc)170 int idxd_submit_desc(struct idxd_wq *wq, struct idxd_desc *desc)
171 {
172 struct idxd_device *idxd = wq->idxd;
173 struct idxd_irq_entry *ie = NULL;
174 u32 desc_flags = desc->hw->flags;
175 void __iomem *portal;
176 int rc;
177
178 if (idxd->state != IDXD_DEV_ENABLED)
179 return -EIO;
180
181 if (!percpu_ref_tryget_live(&wq->wq_active)) {
182 wait_for_completion(&wq->wq_resurrect);
183 if (!percpu_ref_tryget_live(&wq->wq_active))
184 return -ENXIO;
185 }
186
187 portal = idxd_wq_portal_addr(wq);
188
189 /*
190 * Pending the descriptor to the lockless list for the irq_entry
191 * that we designated the descriptor to.
192 */
193 if (desc_flags & IDXD_OP_FLAG_RCI) {
194 ie = &wq->ie;
195 desc->hw->int_handle = ie->int_handle;
196 llist_add(&desc->llnode, &ie->pending_llist);
197 }
198
199 /*
200 * The wmb() flushes writes to coherent DMA data before
201 * possibly triggering a DMA read. The wmb() is necessary
202 * even on UP because the recipient is a device.
203 */
204 wmb();
205
206 if (wq_dedicated(wq)) {
207 iosubmit_cmds512(portal, desc->hw, 1);
208 } else {
209 rc = idxd_enqcmds(wq, portal, desc->hw);
210 if (rc < 0) {
211 percpu_ref_put(&wq->wq_active);
212 /* abort operation frees the descriptor */
213 if (ie)
214 llist_abort_desc(wq, ie, desc);
215 return rc;
216 }
217 }
218
219 percpu_ref_put(&wq->wq_active);
220 return 0;
221 }
222 EXPORT_SYMBOL_NS_GPL(idxd_submit_desc, "IDXD");
223