tjh.dev
Linux kernel mirror (for testing)
git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git
kernel
os
linux
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
11static 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
27struct 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
65void idxd_free_desc(struct idxd_wq *wq, struct idxd_desc *desc)
66{
67 int cpu = desc->cpu;
68
69 desc->cpu = -1;
70 sbitmap_queue_clear(&wq->sbq, desc->id, cpu);
71}
72
73static struct idxd_desc *list_abort_desc(struct idxd_wq *wq, struct idxd_irq_entry *ie,
74 struct idxd_desc *desc)
75{
76 struct idxd_desc *d, *n;
77
78 lockdep_assert_held(&ie->list_lock);
79 list_for_each_entry_safe(d, n, &ie->work_list, list) {
80 if (d == desc) {
81 list_del(&d->list);
82 return d;
83 }
84 }
85
86 /*
87 * At this point, the desc needs to be aborted is held by the completion
88 * handler where it has taken it off the pending list but has not added to the
89 * work list. It will be cleaned up by the interrupt handler when it sees the
90 * IDXD_COMP_DESC_ABORT for completion status.
91 */
92 return NULL;
93}
94
95static void llist_abort_desc(struct idxd_wq *wq, struct idxd_irq_entry *ie,
96 struct idxd_desc *desc)
97{
98 struct idxd_desc *d, *t, *found = NULL;
99 struct llist_node *head;
100 LIST_HEAD(flist);
101
102 desc->completion->status = IDXD_COMP_DESC_ABORT;
103 /*
104 * Grab the list lock so it will block the irq thread handler. This allows the
105 * abort code to locate the descriptor need to be aborted.
106 */
107 spin_lock(&ie->list_lock);
108 head = llist_del_all(&ie->pending_llist);
109 if (head) {
110 llist_for_each_entry_safe(d, t, head, llnode) {
111 if (d == desc) {
112 found = desc;
113 continue;
114 }
115
116 if (d->completion->status)
117 list_add_tail(&d->list, &flist);
118 else
119 list_add_tail(&d->list, &ie->work_list);
120 }
121 }
122
123 if (!found)
124 found = list_abort_desc(wq, ie, desc);
125 spin_unlock(&ie->list_lock);
126
127 if (found)
128 idxd_dma_complete_txd(found, IDXD_COMPLETE_ABORT, false);
129
130 /*
131 * completing the descriptor will return desc to allocator and
132 * the desc can be acquired by a different process and the
133 * desc->list can be modified. Delete desc from list so the
134 * list trasversing does not get corrupted by the other process.
135 */
136 list_for_each_entry_safe(d, t, &flist, list) {
137 list_del_init(&d->list);
138 idxd_dma_complete_txd(found, IDXD_COMPLETE_ABORT, true);
139 }
140}
141
142/*
143 * ENQCMDS typically fail when the WQ is inactive or busy. On host submission, the driver
144 * has better control of number of descriptors being submitted to a shared wq by limiting
145 * the number of driver allocated descriptors to the wq size. However, when the swq is
146 * exported to a guest kernel, it may be shared with multiple guest kernels. This means
147 * the likelihood of getting busy returned on the swq when submitting goes significantly up.
148 * Having a tunable retry mechanism allows the driver to keep trying for a bit before giving
149 * up. The sysfs knob can be tuned by the system administrator.
150 */
151int idxd_enqcmds(struct idxd_wq *wq, void __iomem *portal, const void *desc)
152{
153 int rc, retries = 0;
154
155 do {
156 rc = enqcmds(portal, desc);
157 if (rc == 0)
158 break;
159 cpu_relax();
160 } while (retries++ < wq->enqcmds_retries);
161
162 return rc;
163}
164
165int idxd_submit_desc(struct idxd_wq *wq, struct idxd_desc *desc)
166{
167 struct idxd_device *idxd = wq->idxd;
168 struct idxd_irq_entry *ie = NULL;
169 u32 desc_flags = desc->hw->flags;
170 void __iomem *portal;
171 int rc;
172
173 if (idxd->state != IDXD_DEV_ENABLED)
174 return -EIO;
175
176 if (!percpu_ref_tryget_live(&wq->wq_active)) {
177 wait_for_completion(&wq->wq_resurrect);
178 if (!percpu_ref_tryget_live(&wq->wq_active))
179 return -ENXIO;
180 }
181
182 portal = idxd_wq_portal_addr(wq);
183
184 /*
185 * The wmb() flushes writes to coherent DMA data before
186 * possibly triggering a DMA read. The wmb() is necessary
187 * even on UP because the recipient is a device.
188 */
189 wmb();
190
191 /*
192 * Pending the descriptor to the lockless list for the irq_entry
193 * that we designated the descriptor to.
194 */
195 if (desc_flags & IDXD_OP_FLAG_RCI) {
196 ie = &wq->ie;
197 desc->hw->int_handle = ie->int_handle;
198 llist_add(&desc->llnode, &ie->pending_llist);
199 }
200
201 if (wq_dedicated(wq)) {
202 iosubmit_cmds512(portal, desc->hw, 1);
203 } else {
204 rc = idxd_enqcmds(wq, portal, desc->hw);
205 if (rc < 0) {
206 percpu_ref_put(&wq->wq_active);
207 /* abort operation frees the descriptor */
208 if (ie)
209 llist_abort_desc(wq, ie, desc);
210 return rc;
211 }
212 }
213
214 percpu_ref_put(&wq->wq_active);
215 return 0;
216}