xref: /linux/drivers/vfio/pci/pds/pci_drv.c (revision c900529f3d9161bfde5cca0754f83b4d3c3e0220)
1 // SPDX-License-Identifier: GPL-2.0
2 /* Copyright(c) 2023 Advanced Micro Devices, Inc. */
3 
4 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
5 
6 #include <linux/module.h>
7 #include <linux/pci.h>
8 #include <linux/types.h>
9 #include <linux/vfio.h>
10 
11 #include <linux/pds/pds_common.h>
12 #include <linux/pds/pds_core_if.h>
13 #include <linux/pds/pds_adminq.h>
14 
15 #include "vfio_dev.h"
16 #include "pci_drv.h"
17 #include "cmds.h"
18 
19 #define PDS_VFIO_DRV_DESCRIPTION	"AMD/Pensando VFIO Device Driver"
20 #define PCI_VENDOR_ID_PENSANDO		0x1dd8
21 
22 static void pds_vfio_recovery(struct pds_vfio_pci_device *pds_vfio)
23 {
24 	bool deferred_reset_needed = false;
25 
26 	/*
27 	 * Documentation states that the kernel migration driver must not
28 	 * generate asynchronous device state transitions outside of
29 	 * manipulation by the user or the VFIO_DEVICE_RESET ioctl.
30 	 *
31 	 * Since recovery is an asynchronous event received from the device,
32 	 * initiate a deferred reset. Issue a deferred reset in the following
33 	 * situations:
34 	 *   1. Migration is in progress, which will cause the next step of
35 	 *	the migration to fail.
36 	 *   2. If the device is in a state that will be set to
37 	 *	VFIO_DEVICE_STATE_RUNNING on the next action (i.e. VM is
38 	 *	shutdown and device is in VFIO_DEVICE_STATE_STOP).
39 	 */
40 	mutex_lock(&pds_vfio->state_mutex);
41 	if ((pds_vfio->state != VFIO_DEVICE_STATE_RUNNING &&
42 	     pds_vfio->state != VFIO_DEVICE_STATE_ERROR) ||
43 	    (pds_vfio->state == VFIO_DEVICE_STATE_RUNNING &&
44 	     pds_vfio_dirty_is_enabled(pds_vfio)))
45 		deferred_reset_needed = true;
46 	mutex_unlock(&pds_vfio->state_mutex);
47 
48 	/*
49 	 * On the next user initiated state transition, the device will
50 	 * transition to the VFIO_DEVICE_STATE_ERROR. At this point it's the user's
51 	 * responsibility to reset the device.
52 	 *
53 	 * If a VFIO_DEVICE_RESET is requested post recovery and before the next
54 	 * state transition, then the deferred reset state will be set to
55 	 * VFIO_DEVICE_STATE_RUNNING.
56 	 */
57 	if (deferred_reset_needed) {
58 		spin_lock(&pds_vfio->reset_lock);
59 		pds_vfio->deferred_reset = true;
60 		pds_vfio->deferred_reset_state = VFIO_DEVICE_STATE_ERROR;
61 		spin_unlock(&pds_vfio->reset_lock);
62 	}
63 }
64 
65 static int pds_vfio_pci_notify_handler(struct notifier_block *nb,
66 				       unsigned long ecode, void *data)
67 {
68 	struct pds_vfio_pci_device *pds_vfio =
69 		container_of(nb, struct pds_vfio_pci_device, nb);
70 	struct device *dev = pds_vfio_to_dev(pds_vfio);
71 	union pds_core_notifyq_comp *event = data;
72 
73 	dev_dbg(dev, "%s: event code %lu\n", __func__, ecode);
74 
75 	/*
76 	 * We don't need to do anything for RESET state==0 as there is no notify
77 	 * or feedback mechanism available, and it is possible that we won't
78 	 * even see a state==0 event since the pds_core recovery is pending.
79 	 *
80 	 * Any requests from VFIO while state==0 will fail, which will return
81 	 * error and may cause migration to fail.
82 	 */
83 	if (ecode == PDS_EVENT_RESET) {
84 		dev_info(dev, "%s: PDS_EVENT_RESET event received, state==%d\n",
85 			 __func__, event->reset.state);
86 		/*
87 		 * pds_core device finished recovery and sent us the
88 		 * notification (state == 1) to allow us to recover
89 		 */
90 		if (event->reset.state == 1)
91 			pds_vfio_recovery(pds_vfio);
92 	}
93 
94 	return 0;
95 }
96 
97 static int
98 pds_vfio_pci_register_event_handler(struct pds_vfio_pci_device *pds_vfio)
99 {
100 	struct device *dev = pds_vfio_to_dev(pds_vfio);
101 	struct notifier_block *nb = &pds_vfio->nb;
102 	int err;
103 
104 	if (!nb->notifier_call) {
105 		nb->notifier_call = pds_vfio_pci_notify_handler;
106 		err = pdsc_register_notify(nb);
107 		if (err) {
108 			nb->notifier_call = NULL;
109 			dev_err(dev,
110 				"failed to register pds event handler: %pe\n",
111 				ERR_PTR(err));
112 			return -EINVAL;
113 		}
114 		dev_dbg(dev, "pds event handler registered\n");
115 	}
116 
117 	return 0;
118 }
119 
120 static void
121 pds_vfio_pci_unregister_event_handler(struct pds_vfio_pci_device *pds_vfio)
122 {
123 	if (pds_vfio->nb.notifier_call) {
124 		pdsc_unregister_notify(&pds_vfio->nb);
125 		pds_vfio->nb.notifier_call = NULL;
126 	}
127 }
128 
129 static int pds_vfio_pci_probe(struct pci_dev *pdev,
130 			      const struct pci_device_id *id)
131 {
132 	struct pds_vfio_pci_device *pds_vfio;
133 	int err;
134 
135 	pds_vfio = vfio_alloc_device(pds_vfio_pci_device, vfio_coredev.vdev,
136 				     &pdev->dev, pds_vfio_ops_info());
137 	if (IS_ERR(pds_vfio))
138 		return PTR_ERR(pds_vfio);
139 
140 	dev_set_drvdata(&pdev->dev, &pds_vfio->vfio_coredev);
141 
142 	err = vfio_pci_core_register_device(&pds_vfio->vfio_coredev);
143 	if (err)
144 		goto out_put_vdev;
145 
146 	err = pds_vfio_register_client_cmd(pds_vfio);
147 	if (err) {
148 		dev_err(&pdev->dev, "failed to register as client: %pe\n",
149 			ERR_PTR(err));
150 		goto out_unregister_coredev;
151 	}
152 
153 	err = pds_vfio_pci_register_event_handler(pds_vfio);
154 	if (err)
155 		goto out_unregister_client;
156 
157 	return 0;
158 
159 out_unregister_client:
160 	pds_vfio_unregister_client_cmd(pds_vfio);
161 out_unregister_coredev:
162 	vfio_pci_core_unregister_device(&pds_vfio->vfio_coredev);
163 out_put_vdev:
164 	vfio_put_device(&pds_vfio->vfio_coredev.vdev);
165 	return err;
166 }
167 
168 static void pds_vfio_pci_remove(struct pci_dev *pdev)
169 {
170 	struct pds_vfio_pci_device *pds_vfio = pds_vfio_pci_drvdata(pdev);
171 
172 	pds_vfio_pci_unregister_event_handler(pds_vfio);
173 	pds_vfio_unregister_client_cmd(pds_vfio);
174 	vfio_pci_core_unregister_device(&pds_vfio->vfio_coredev);
175 	vfio_put_device(&pds_vfio->vfio_coredev.vdev);
176 }
177 
178 static const struct pci_device_id pds_vfio_pci_table[] = {
179 	{ PCI_DRIVER_OVERRIDE_DEVICE_VFIO(PCI_VENDOR_ID_PENSANDO, 0x1003) }, /* Ethernet VF */
180 	{ 0, }
181 };
182 MODULE_DEVICE_TABLE(pci, pds_vfio_pci_table);
183 
184 static void pds_vfio_pci_aer_reset_done(struct pci_dev *pdev)
185 {
186 	struct pds_vfio_pci_device *pds_vfio = pds_vfio_pci_drvdata(pdev);
187 
188 	pds_vfio_reset(pds_vfio);
189 }
190 
191 static const struct pci_error_handlers pds_vfio_pci_err_handlers = {
192 	.reset_done = pds_vfio_pci_aer_reset_done,
193 	.error_detected = vfio_pci_core_aer_err_detected,
194 };
195 
196 static struct pci_driver pds_vfio_pci_driver = {
197 	.name = KBUILD_MODNAME,
198 	.id_table = pds_vfio_pci_table,
199 	.probe = pds_vfio_pci_probe,
200 	.remove = pds_vfio_pci_remove,
201 	.err_handler = &pds_vfio_pci_err_handlers,
202 	.driver_managed_dma = true,
203 };
204 
205 module_pci_driver(pds_vfio_pci_driver);
206 
207 MODULE_DESCRIPTION(PDS_VFIO_DRV_DESCRIPTION);
208 MODULE_AUTHOR("Brett Creeley <brett.creeley@amd.com>");
209 MODULE_LICENSE("GPL");
210