1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright(c) 2020 Intel Corporation
5 #include <rte_bus_pci.h>
6 #include <rte_memzone.h>
8 #include "ioat_private.h"
11 #define IDXD_VENDOR_ID 0x8086
12 #define IDXD_DEVICE_ID_SPR 0x0B25
14 #define IDXD_PMD_RAWDEV_NAME_PCI rawdev_idxd_pci
16 const struct rte_pci_id pci_id_idxd_map[] = {
17 { RTE_PCI_DEVICE(IDXD_VENDOR_ID, IDXD_DEVICE_ID_SPR) },
18 { .vendor_id = 0, /* sentinel */ },
22 idxd_pci_dev_command(struct idxd_rawdev *idxd, enum rte_idxd_cmds command)
25 uint16_t qid = idxd->qid;
28 if (command >= idxd_disable_wq && command <= idxd_reset_wq)
30 rte_spinlock_lock(&idxd->u.pci->lk);
31 idxd->u.pci->regs->cmd = (command << IDXD_CMD_SHIFT) | qid;
35 err_code = idxd->u.pci->regs->cmdstatus;
37 IOAT_PMD_ERR("Timeout waiting for command response from HW");
38 rte_spinlock_unlock(&idxd->u.pci->lk);
41 } while (idxd->u.pci->regs->cmdstatus & CMDSTATUS_ACTIVE_MASK);
42 rte_spinlock_unlock(&idxd->u.pci->lk);
44 return err_code & CMDSTATUS_ERR_MASK;
48 idxd_is_wq_enabled(struct idxd_rawdev *idxd)
50 uint32_t state = idxd->u.pci->wq_regs[idxd->qid].wqcfg[WQ_STATE_IDX];
51 return ((state >> WQ_STATE_SHIFT) & WQ_STATE_MASK) == 0x1;
54 static const struct rte_rawdev_ops idxd_pci_ops = {
55 .dev_close = idxd_rawdev_close,
56 .dev_selftest = idxd_rawdev_test,
59 /* each portal uses 4 x 4k pages */
60 #define IDXD_PORTAL_SIZE (4096 * 4)
63 init_pci_device(struct rte_pci_device *dev, struct idxd_rawdev *idxd)
65 struct idxd_pci_common *pci;
66 uint8_t nb_groups, nb_engines, nb_wqs;
67 uint16_t grp_offset, wq_offset; /* how far into bar0 the regs are */
68 uint16_t wq_size, total_wq_size;
69 uint8_t lg2_max_batch, lg2_max_copy_size;
70 unsigned int i, err_code;
72 pci = malloc(sizeof(*pci));
74 IOAT_PMD_ERR("%s: Can't allocate memory", __func__);
77 rte_spinlock_init(&pci->lk);
79 /* assign the bar registers, and then configure device */
80 pci->regs = dev->mem_resource[0].addr;
81 grp_offset = (uint16_t)pci->regs->offsets[0];
82 pci->grp_regs = RTE_PTR_ADD(pci->regs, grp_offset * 0x100);
83 wq_offset = (uint16_t)(pci->regs->offsets[0] >> 16);
84 pci->wq_regs = RTE_PTR_ADD(pci->regs, wq_offset * 0x100);
85 pci->portals = dev->mem_resource[2].addr;
87 /* sanity check device status */
88 if (pci->regs->gensts & GENSTS_DEV_STATE_MASK) {
89 /* need function-level-reset (FLR) or is enabled */
90 IOAT_PMD_ERR("Device status is not disabled, cannot init");
93 if (pci->regs->cmdstatus & CMDSTATUS_ACTIVE_MASK) {
94 /* command in progress */
95 IOAT_PMD_ERR("Device has a command in progress, cannot init");
99 /* read basic info about the hardware for use when configuring */
100 nb_groups = (uint8_t)pci->regs->grpcap;
101 nb_engines = (uint8_t)pci->regs->engcap;
102 nb_wqs = (uint8_t)(pci->regs->wqcap >> 16);
103 total_wq_size = (uint16_t)pci->regs->wqcap;
104 lg2_max_copy_size = (uint8_t)(pci->regs->gencap >> 16) & 0x1F;
105 lg2_max_batch = (uint8_t)(pci->regs->gencap >> 21) & 0x0F;
107 IOAT_PMD_DEBUG("nb_groups = %u, nb_engines = %u, nb_wqs = %u",
108 nb_groups, nb_engines, nb_wqs);
110 /* zero out any old config */
111 for (i = 0; i < nb_groups; i++) {
112 pci->grp_regs[i].grpengcfg = 0;
113 pci->grp_regs[i].grpwqcfg[0] = 0;
115 for (i = 0; i < nb_wqs; i++)
116 pci->wq_regs[i].wqcfg[0] = 0;
118 /* put each engine into a separate group to avoid reordering */
119 if (nb_groups > nb_engines)
120 nb_groups = nb_engines;
121 if (nb_groups < nb_engines)
122 nb_engines = nb_groups;
124 /* assign engines to groups, round-robin style */
125 for (i = 0; i < nb_engines; i++) {
126 IOAT_PMD_DEBUG("Assigning engine %u to group %u",
128 pci->grp_regs[i % nb_groups].grpengcfg |= (1ULL << i);
131 /* now do the same for queues and give work slots to each queue */
132 wq_size = total_wq_size / nb_wqs;
133 IOAT_PMD_DEBUG("Work queue size = %u, max batch = 2^%u, max copy = 2^%u",
134 wq_size, lg2_max_batch, lg2_max_copy_size);
135 for (i = 0; i < nb_wqs; i++) {
136 /* add engine "i" to a group */
137 IOAT_PMD_DEBUG("Assigning work queue %u to group %u",
139 pci->grp_regs[i % nb_groups].grpwqcfg[0] |= (1ULL << i);
140 /* now configure it, in terms of size, max batch, mode */
141 pci->wq_regs[i].wqcfg[WQ_SIZE_IDX] = wq_size;
142 pci->wq_regs[i].wqcfg[WQ_MODE_IDX] = (1 << WQ_PRIORITY_SHIFT) |
144 pci->wq_regs[i].wqcfg[WQ_SIZES_IDX] = lg2_max_copy_size |
145 (lg2_max_batch << WQ_BATCH_SZ_SHIFT);
148 /* dump the group configuration to output */
149 for (i = 0; i < nb_groups; i++) {
150 IOAT_PMD_DEBUG("## Group %d", i);
151 IOAT_PMD_DEBUG(" GRPWQCFG: %"PRIx64, pci->grp_regs[i].grpwqcfg[0]);
152 IOAT_PMD_DEBUG(" GRPENGCFG: %"PRIx64, pci->grp_regs[i].grpengcfg);
153 IOAT_PMD_DEBUG(" GRPFLAGS: %"PRIx32, pci->grp_regs[i].grpflags);
157 idxd->max_batches = wq_size;
159 /* enable the device itself */
160 err_code = idxd_pci_dev_command(idxd, idxd_enable_dev);
162 IOAT_PMD_ERR("Error enabling device: code %#x", err_code);
165 IOAT_PMD_DEBUG("IDXD Device enabled OK");
175 idxd_rawdev_probe_pci(struct rte_pci_driver *drv, struct rte_pci_device *dev)
177 struct idxd_rawdev idxd = {{0}}; /* Double {} to avoid error on BSD12 */
180 char name[PCI_PRI_STR_SIZE];
182 rte_pci_device_name(&dev->addr, name, sizeof(name));
183 IOAT_PMD_INFO("Init %s on NUMA node %d", name, dev->device.numa_node);
184 dev->device.driver = &drv->driver;
186 ret = init_pci_device(dev, &idxd);
188 IOAT_PMD_ERR("Error initializing PCI hardware");
191 nb_wqs = (uint8_t)ret;
193 /* set up one device for each queue */
194 for (qid = 0; qid < nb_wqs; qid++) {
197 /* add the queue number to each device name */
198 snprintf(qname, sizeof(qname), "%s-q%d", name, qid);
200 idxd.public.portal = RTE_PTR_ADD(idxd.u.pci->portals,
201 qid * IDXD_PORTAL_SIZE);
202 if (idxd_is_wq_enabled(&idxd))
203 IOAT_PMD_ERR("Error, WQ %u seems enabled", qid);
204 ret = idxd_rawdev_create(qname, &dev->device,
205 &idxd, &idxd_pci_ops);
207 IOAT_PMD_ERR("Failed to create rawdev %s", name);
208 if (qid == 0) /* if no devices using this, free pci */
218 idxd_rawdev_destroy(const char *name)
222 struct rte_rawdev *rdev;
223 struct idxd_rawdev *idxd;
226 IOAT_PMD_ERR("Invalid device name");
230 rdev = rte_rawdev_pmd_get_named_dev(name);
232 IOAT_PMD_ERR("Invalid device name (%s)", name);
236 idxd = rdev->dev_private;
238 /* disable the device */
239 err_code = idxd_pci_dev_command(idxd, idxd_disable_dev);
241 IOAT_PMD_ERR("Error disabling device: code %#x", err_code);
244 IOAT_PMD_DEBUG("IDXD Device disabled OK");
246 /* free device memory */
247 if (rdev->dev_private != NULL) {
248 IOAT_PMD_DEBUG("Freeing device driver memory");
249 rdev->dev_private = NULL;
250 rte_free(idxd->public.batch_ring);
251 rte_free(idxd->public.hdl_ring);
252 rte_memzone_free(idxd->mz);
255 /* rte_rawdev_close is called by pmd_release */
256 ret = rte_rawdev_pmd_release(rdev);
258 IOAT_PMD_DEBUG("Device cleanup failed");
264 idxd_rawdev_remove_pci(struct rte_pci_device *dev)
266 char name[PCI_PRI_STR_SIZE];
269 rte_pci_device_name(&dev->addr, name, sizeof(name));
271 IOAT_PMD_INFO("Closing %s on NUMA node %d",
272 name, dev->device.numa_node);
274 ret = idxd_rawdev_destroy(name);
279 struct rte_pci_driver idxd_pmd_drv_pci = {
280 .id_table = pci_id_idxd_map,
281 .drv_flags = RTE_PCI_DRV_NEED_MAPPING,
282 .probe = idxd_rawdev_probe_pci,
283 .remove = idxd_rawdev_remove_pci,
286 RTE_PMD_REGISTER_PCI(IDXD_PMD_RAWDEV_NAME_PCI, idxd_pmd_drv_pci);
287 RTE_PMD_REGISTER_PCI_TABLE(IDXD_PMD_RAWDEV_NAME_PCI, pci_id_idxd_map);
288 RTE_PMD_REGISTER_KMOD_DEP(IDXD_PMD_RAWDEV_NAME_PCI,
289 "* igb_uio | uio_pci_generic | vfio-pci");