1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright(c) 2021 Intel Corporation
5 #include <rte_bus_pci.h>
6 #include <rte_devargs.h>
7 #include <rte_dmadev_pmd.h>
8 #include <rte_malloc.h>
10 #include "idxd_internal.h"
12 #define IDXD_VENDOR_ID 0x8086
13 #define IDXD_DEVICE_ID_SPR 0x0B25
15 #define IDXD_PMD_DMADEV_NAME_PCI dmadev_idxd_pci
17 const struct rte_pci_id pci_id_idxd_map[] = {
18 { RTE_PCI_DEVICE(IDXD_VENDOR_ID, IDXD_DEVICE_ID_SPR) },
19 { .vendor_id = 0, /* sentinel */ },
23 idxd_pci_dev_command(struct idxd_dmadev *idxd, enum rte_idxd_cmds command)
26 uint16_t qid = idxd->qid;
29 if (command >= idxd_disable_wq && command <= idxd_reset_wq)
31 rte_spinlock_lock(&idxd->u.pci->lk);
32 idxd->u.pci->regs->cmd = (command << IDXD_CMD_SHIFT) | qid;
36 err_code = idxd->u.pci->regs->cmdstatus;
38 IDXD_PMD_ERR("Timeout waiting for command response from HW");
39 rte_spinlock_unlock(&idxd->u.pci->lk);
42 } while (err_code & CMDSTATUS_ACTIVE_MASK);
43 rte_spinlock_unlock(&idxd->u.pci->lk);
45 err_code &= CMDSTATUS_ERR_MASK;
50 idxd_get_wq_cfg(struct idxd_pci_common *pci, uint8_t wq_idx)
52 return RTE_PTR_ADD(pci->wq_regs_base,
53 (uintptr_t)wq_idx << (5 + pci->wq_cfg_sz));
57 idxd_is_wq_enabled(struct idxd_dmadev *idxd)
59 uint32_t state = idxd_get_wq_cfg(idxd->u.pci, idxd->qid)[wq_state_idx];
60 return ((state >> WQ_STATE_SHIFT) & WQ_STATE_MASK) == 0x1;
64 idxd_pci_dev_stop(struct rte_dma_dev *dev)
66 struct idxd_dmadev *idxd = dev->fp_obj->dev_private;
69 if (!idxd_is_wq_enabled(idxd)) {
70 IDXD_PMD_ERR("Work queue %d already disabled", idxd->qid);
74 err_code = idxd_pci_dev_command(idxd, idxd_disable_wq);
75 if (err_code || idxd_is_wq_enabled(idxd)) {
76 IDXD_PMD_ERR("Failed disabling work queue %d, error code: %#x",
78 return err_code == 0 ? -1 : -err_code;
80 IDXD_PMD_DEBUG("Work queue %d disabled OK", idxd->qid);
86 idxd_pci_dev_start(struct rte_dma_dev *dev)
88 struct idxd_dmadev *idxd = dev->fp_obj->dev_private;
91 if (idxd_is_wq_enabled(idxd)) {
92 IDXD_PMD_WARN("WQ %d already enabled", idxd->qid);
96 if (idxd->desc_ring == NULL) {
97 IDXD_PMD_ERR("WQ %d has not been fully configured", idxd->qid);
101 err_code = idxd_pci_dev_command(idxd, idxd_enable_wq);
102 if (err_code || !idxd_is_wq_enabled(idxd)) {
103 IDXD_PMD_ERR("Failed enabling work queue %d, error code: %#x",
104 idxd->qid, err_code);
105 return err_code == 0 ? -1 : -err_code;
107 IDXD_PMD_DEBUG("Work queue %d enabled OK", idxd->qid);
113 idxd_pci_dev_close(struct rte_dma_dev *dev)
115 struct idxd_dmadev *idxd = dev->fp_obj->dev_private;
118 /* disable the device */
119 err_code = idxd_pci_dev_command(idxd, idxd_disable_dev);
121 IDXD_PMD_ERR("Error disabling device: code %#x", err_code);
124 IDXD_PMD_DEBUG("IDXD Device disabled OK");
126 /* free device memory */
127 IDXD_PMD_DEBUG("Freeing device driver memory");
128 rte_free(idxd->batch_idx_ring);
129 rte_free(idxd->desc_ring);
134 static const struct rte_dma_dev_ops idxd_pci_ops = {
135 .dev_close = idxd_pci_dev_close,
136 .dev_dump = idxd_dump,
137 .dev_configure = idxd_configure,
138 .vchan_setup = idxd_vchan_setup,
139 .dev_info_get = idxd_info_get,
140 .stats_get = idxd_stats_get,
141 .stats_reset = idxd_stats_reset,
142 .dev_start = idxd_pci_dev_start,
143 .dev_stop = idxd_pci_dev_stop,
146 /* each portal uses 4 x 4k pages */
147 #define IDXD_PORTAL_SIZE (4096 * 4)
150 init_pci_device(struct rte_pci_device *dev, struct idxd_dmadev *idxd,
151 unsigned int max_queues)
153 struct idxd_pci_common *pci;
154 uint8_t nb_groups, nb_engines, nb_wqs;
155 uint16_t grp_offset, wq_offset; /* how far into bar0 the regs are */
156 uint16_t wq_size, total_wq_size;
157 uint8_t lg2_max_batch, lg2_max_copy_size;
158 unsigned int i, err_code;
160 pci = malloc(sizeof(*pci));
162 IDXD_PMD_ERR("%s: Can't allocate memory", __func__);
166 rte_spinlock_init(&pci->lk);
168 /* assign the bar registers, and then configure device */
169 pci->regs = dev->mem_resource[0].addr;
170 grp_offset = (uint16_t)pci->regs->offsets[0];
171 pci->grp_regs = RTE_PTR_ADD(pci->regs, grp_offset * 0x100);
172 wq_offset = (uint16_t)(pci->regs->offsets[0] >> 16);
173 pci->wq_regs_base = RTE_PTR_ADD(pci->regs, wq_offset * 0x100);
174 pci->portals = dev->mem_resource[2].addr;
175 pci->wq_cfg_sz = (pci->regs->wqcap >> 24) & 0x0F;
177 /* sanity check device status */
178 if (pci->regs->gensts & GENSTS_DEV_STATE_MASK) {
179 /* need function-level-reset (FLR) or is enabled */
180 IDXD_PMD_ERR("Device status is not disabled, cannot init");
184 if (pci->regs->cmdstatus & CMDSTATUS_ACTIVE_MASK) {
185 /* command in progress */
186 IDXD_PMD_ERR("Device has a command in progress, cannot init");
191 /* read basic info about the hardware for use when configuring */
192 nb_groups = (uint8_t)pci->regs->grpcap;
193 nb_engines = (uint8_t)pci->regs->engcap;
194 nb_wqs = (uint8_t)(pci->regs->wqcap >> 16);
195 total_wq_size = (uint16_t)pci->regs->wqcap;
196 lg2_max_copy_size = (uint8_t)(pci->regs->gencap >> 16) & 0x1F;
197 lg2_max_batch = (uint8_t)(pci->regs->gencap >> 21) & 0x0F;
199 IDXD_PMD_DEBUG("nb_groups = %u, nb_engines = %u, nb_wqs = %u",
200 nb_groups, nb_engines, nb_wqs);
202 /* zero out any old config */
203 for (i = 0; i < nb_groups; i++) {
204 pci->grp_regs[i].grpengcfg = 0;
205 pci->grp_regs[i].grpwqcfg[0] = 0;
207 for (i = 0; i < nb_wqs; i++)
208 idxd_get_wq_cfg(pci, i)[0] = 0;
210 /* limit queues if necessary */
211 if (max_queues != 0 && nb_wqs > max_queues) {
213 if (nb_engines > max_queues)
214 nb_engines = max_queues;
215 if (nb_groups > max_queues)
216 nb_engines = max_queues;
217 IDXD_PMD_DEBUG("Limiting queues to %u", nb_wqs);
220 /* put each engine into a separate group to avoid reordering */
221 if (nb_groups > nb_engines)
222 nb_groups = nb_engines;
223 if (nb_groups < nb_engines)
224 nb_engines = nb_groups;
226 /* assign engines to groups, round-robin style */
227 for (i = 0; i < nb_engines; i++) {
228 IDXD_PMD_DEBUG("Assigning engine %u to group %u",
230 pci->grp_regs[i % nb_groups].grpengcfg |= (1ULL << i);
233 /* now do the same for queues and give work slots to each queue */
234 wq_size = total_wq_size / nb_wqs;
235 IDXD_PMD_DEBUG("Work queue size = %u, max batch = 2^%u, max copy = 2^%u",
236 wq_size, lg2_max_batch, lg2_max_copy_size);
237 for (i = 0; i < nb_wqs; i++) {
238 /* add engine "i" to a group */
239 IDXD_PMD_DEBUG("Assigning work queue %u to group %u",
241 pci->grp_regs[i % nb_groups].grpwqcfg[0] |= (1ULL << i);
242 /* now configure it, in terms of size, max batch, mode */
243 idxd_get_wq_cfg(pci, i)[wq_size_idx] = wq_size;
244 idxd_get_wq_cfg(pci, i)[wq_mode_idx] = (1 << WQ_PRIORITY_SHIFT) |
246 idxd_get_wq_cfg(pci, i)[wq_sizes_idx] = lg2_max_copy_size |
247 (lg2_max_batch << WQ_BATCH_SZ_SHIFT);
250 /* dump the group configuration to output */
251 for (i = 0; i < nb_groups; i++) {
252 IDXD_PMD_DEBUG("## Group %d", i);
253 IDXD_PMD_DEBUG(" GRPWQCFG: %"PRIx64, pci->grp_regs[i].grpwqcfg[0]);
254 IDXD_PMD_DEBUG(" GRPENGCFG: %"PRIx64, pci->grp_regs[i].grpengcfg);
255 IDXD_PMD_DEBUG(" GRPFLAGS: %"PRIx32, pci->grp_regs[i].grpflags);
259 idxd->max_batches = wq_size;
261 /* enable the device itself */
262 err_code = idxd_pci_dev_command(idxd, idxd_enable_dev);
264 IDXD_PMD_ERR("Error enabling device: code %#x", err_code);
267 IDXD_PMD_DEBUG("IDXD Device enabled OK");
277 idxd_dmadev_probe_pci(struct rte_pci_driver *drv, struct rte_pci_device *dev)
279 struct idxd_dmadev idxd = {0};
282 char name[PCI_PRI_STR_SIZE];
283 unsigned int max_queues = 0;
285 rte_pci_device_name(&dev->addr, name, sizeof(name));
286 IDXD_PMD_INFO("Init %s on NUMA node %d", name, dev->device.numa_node);
287 dev->device.driver = &drv->driver;
289 if (dev->device.devargs && dev->device.devargs->args[0] != '\0') {
290 /* if the number of devargs grows beyond just 1, use rte_kvargs */
291 if (sscanf(dev->device.devargs->args,
292 "max_queues=%u", &max_queues) != 1) {
293 IDXD_PMD_ERR("Invalid device parameter: '%s'",
294 dev->device.devargs->args);
299 ret = init_pci_device(dev, &idxd, max_queues);
301 IDXD_PMD_ERR("Error initializing PCI hardware");
304 if (idxd.u.pci->portals == NULL) {
305 IDXD_PMD_ERR("Error, invalid portal assigned during initialization\n");
309 nb_wqs = (uint8_t)ret;
311 /* set up one device for each queue */
312 for (qid = 0; qid < nb_wqs; qid++) {
315 /* add the queue number to each device name */
316 snprintf(qname, sizeof(qname), "%s-q%d", name, qid);
318 idxd.portal = RTE_PTR_ADD(idxd.u.pci->portals,
319 qid * IDXD_PORTAL_SIZE);
320 if (idxd_is_wq_enabled(&idxd))
321 IDXD_PMD_ERR("Error, WQ %u seems enabled", qid);
322 ret = idxd_dmadev_create(qname, &dev->device,
323 &idxd, &idxd_pci_ops);
325 IDXD_PMD_ERR("Failed to create dmadev %s", name);
326 if (qid == 0) /* if no devices using this, free pci */
336 idxd_dmadev_destroy(const char *name)
340 /* rte_dma_close is called by pmd_release */
341 ret = rte_dma_pmd_release(name);
343 IDXD_PMD_DEBUG("Device cleanup failed");
349 idxd_dmadev_remove_pci(struct rte_pci_device *dev)
352 char name[PCI_PRI_STR_SIZE];
354 rte_pci_device_name(&dev->addr, name, sizeof(name));
356 IDXD_PMD_INFO("Closing %s on NUMA node %d", name, dev->device.numa_node);
358 RTE_DMA_FOREACH_DEV(i) {
359 struct rte_dma_info *info = {0};
360 rte_dma_info_get(i, info);
361 if (strncmp(name, info->dev_name, strlen(name)) == 0)
362 idxd_dmadev_destroy(info->dev_name);
368 struct rte_pci_driver idxd_pmd_drv_pci = {
369 .id_table = pci_id_idxd_map,
370 .drv_flags = RTE_PCI_DRV_NEED_MAPPING,
371 .probe = idxd_dmadev_probe_pci,
372 .remove = idxd_dmadev_remove_pci,
375 RTE_PMD_REGISTER_PCI(IDXD_PMD_DMADEV_NAME_PCI, idxd_pmd_drv_pci);
376 RTE_PMD_REGISTER_PCI_TABLE(IDXD_PMD_DMADEV_NAME_PCI, pci_id_idxd_map);
377 RTE_PMD_REGISTER_KMOD_DEP(IDXD_PMD_DMADEV_NAME_PCI, "vfio-pci");
378 RTE_PMD_REGISTER_PARAM_STRING(dmadev_idxd_pci, "max_queues=0");