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>
7 #include <linux/io-64-nonatomic-lo-hi.h>
8 #include <linux/dmaengine.h>
10 #include <uapi/linux/idxd.h>
11 #include "../dmaengine.h"
13 #include "registers.h"
15 static void idxd_cmd_exec(struct idxd_device *idxd, int cmd_code, u32 operand,
17 static void idxd_device_wqs_clear_state(struct idxd_device *idxd);
18 static void idxd_wq_disable_cleanup(struct idxd_wq *wq);
20 /* Interrupt control bits */
21 void idxd_unmask_error_interrupts(struct idxd_device *idxd)
23 union genctrl_reg genctrl;
25 genctrl.bits = ioread32(idxd->reg_base + IDXD_GENCTRL_OFFSET);
26 genctrl.softerr_int_en = 1;
27 genctrl.halt_int_en = 1;
28 iowrite32(genctrl.bits, idxd->reg_base + IDXD_GENCTRL_OFFSET);
31 void idxd_mask_error_interrupts(struct idxd_device *idxd)
33 union genctrl_reg genctrl;
35 genctrl.bits = ioread32(idxd->reg_base + IDXD_GENCTRL_OFFSET);
36 genctrl.softerr_int_en = 0;
37 genctrl.halt_int_en = 0;
38 iowrite32(genctrl.bits, idxd->reg_base + IDXD_GENCTRL_OFFSET);
41 static void free_hw_descs(struct idxd_wq *wq)
45 for (i = 0; i < wq->num_descs; i++)
46 kfree(wq->hw_descs[i]);
51 static int alloc_hw_descs(struct idxd_wq *wq, int num)
53 struct device *dev = &wq->idxd->pdev->dev;
55 int node = dev_to_node(dev);
57 wq->hw_descs = kcalloc_node(num, sizeof(struct dsa_hw_desc *),
62 for (i = 0; i < num; i++) {
63 wq->hw_descs[i] = kzalloc_node(sizeof(*wq->hw_descs[i]),
65 if (!wq->hw_descs[i]) {
74 static void free_descs(struct idxd_wq *wq)
78 for (i = 0; i < wq->num_descs; i++)
84 static int alloc_descs(struct idxd_wq *wq, int num)
86 struct device *dev = &wq->idxd->pdev->dev;
88 int node = dev_to_node(dev);
90 wq->descs = kcalloc_node(num, sizeof(struct idxd_desc *),
95 for (i = 0; i < num; i++) {
96 wq->descs[i] = kzalloc_node(sizeof(*wq->descs[i]),
107 /* WQ control bits */
108 int idxd_wq_alloc_resources(struct idxd_wq *wq)
110 struct idxd_device *idxd = wq->idxd;
111 struct device *dev = &idxd->pdev->dev;
112 int rc, num_descs, i;
114 if (wq->type != IDXD_WQT_KERNEL)
117 num_descs = wq_dedicated(wq) ? wq->size : wq->threshold;
118 wq->num_descs = num_descs;
120 rc = alloc_hw_descs(wq, num_descs);
124 wq->compls_size = num_descs * idxd->data->compl_size;
125 wq->compls = dma_alloc_coherent(dev, wq->compls_size, &wq->compls_addr, GFP_KERNEL);
128 goto fail_alloc_compls;
131 rc = alloc_descs(wq, num_descs);
133 goto fail_alloc_descs;
135 rc = sbitmap_queue_init_node(&wq->sbq, num_descs, -1, false, GFP_KERNEL,
138 goto fail_sbitmap_init;
140 for (i = 0; i < num_descs; i++) {
141 struct idxd_desc *desc = wq->descs[i];
143 desc->hw = wq->hw_descs[i];
144 if (idxd->data->type == IDXD_TYPE_DSA)
145 desc->completion = &wq->compls[i];
146 else if (idxd->data->type == IDXD_TYPE_IAX)
147 desc->iax_completion = &wq->iax_compls[i];
148 desc->compl_dma = wq->compls_addr + idxd->data->compl_size * i;
159 dma_free_coherent(dev, wq->compls_size, wq->compls, wq->compls_addr);
165 void idxd_wq_free_resources(struct idxd_wq *wq)
167 struct device *dev = &wq->idxd->pdev->dev;
169 if (wq->type != IDXD_WQT_KERNEL)
174 dma_free_coherent(dev, wq->compls_size, wq->compls, wq->compls_addr);
175 sbitmap_queue_free(&wq->sbq);
178 int idxd_wq_enable(struct idxd_wq *wq)
180 struct idxd_device *idxd = wq->idxd;
181 struct device *dev = &idxd->pdev->dev;
184 if (wq->state == IDXD_WQ_ENABLED) {
185 dev_dbg(dev, "WQ %d already enabled\n", wq->id);
189 idxd_cmd_exec(idxd, IDXD_CMD_ENABLE_WQ, wq->id, &status);
191 if (status != IDXD_CMDSTS_SUCCESS &&
192 status != IDXD_CMDSTS_ERR_WQ_ENABLED) {
193 dev_dbg(dev, "WQ enable failed: %#x\n", status);
197 wq->state = IDXD_WQ_ENABLED;
198 set_bit(wq->id, idxd->wq_enable_map);
199 dev_dbg(dev, "WQ %d enabled\n", wq->id);
203 int idxd_wq_disable(struct idxd_wq *wq, bool reset_config)
205 struct idxd_device *idxd = wq->idxd;
206 struct device *dev = &idxd->pdev->dev;
209 dev_dbg(dev, "Disabling WQ %d\n", wq->id);
211 if (wq->state != IDXD_WQ_ENABLED) {
212 dev_dbg(dev, "WQ %d in wrong state: %d\n", wq->id, wq->state);
216 operand = BIT(wq->id % 16) | ((wq->id / 16) << 16);
217 idxd_cmd_exec(idxd, IDXD_CMD_DISABLE_WQ, operand, &status);
219 if (status != IDXD_CMDSTS_SUCCESS) {
220 dev_dbg(dev, "WQ disable failed: %#x\n", status);
225 idxd_wq_disable_cleanup(wq);
226 clear_bit(wq->id, idxd->wq_enable_map);
227 wq->state = IDXD_WQ_DISABLED;
228 dev_dbg(dev, "WQ %d disabled\n", wq->id);
232 void idxd_wq_drain(struct idxd_wq *wq)
234 struct idxd_device *idxd = wq->idxd;
235 struct device *dev = &idxd->pdev->dev;
238 if (wq->state != IDXD_WQ_ENABLED) {
239 dev_dbg(dev, "WQ %d in wrong state: %d\n", wq->id, wq->state);
243 dev_dbg(dev, "Draining WQ %d\n", wq->id);
244 operand = BIT(wq->id % 16) | ((wq->id / 16) << 16);
245 idxd_cmd_exec(idxd, IDXD_CMD_DRAIN_WQ, operand, NULL);
248 void idxd_wq_reset(struct idxd_wq *wq)
250 struct idxd_device *idxd = wq->idxd;
251 struct device *dev = &idxd->pdev->dev;
254 if (wq->state != IDXD_WQ_ENABLED) {
255 dev_dbg(dev, "WQ %d in wrong state: %d\n", wq->id, wq->state);
259 operand = BIT(wq->id % 16) | ((wq->id / 16) << 16);
260 idxd_cmd_exec(idxd, IDXD_CMD_RESET_WQ, operand, NULL);
261 idxd_wq_disable_cleanup(wq);
264 int idxd_wq_map_portal(struct idxd_wq *wq)
266 struct idxd_device *idxd = wq->idxd;
267 struct pci_dev *pdev = idxd->pdev;
268 struct device *dev = &pdev->dev;
269 resource_size_t start;
271 start = pci_resource_start(pdev, IDXD_WQ_BAR);
272 start += idxd_get_wq_portal_full_offset(wq->id, IDXD_PORTAL_LIMITED);
274 wq->portal = devm_ioremap(dev, start, IDXD_PORTAL_SIZE);
281 void idxd_wq_unmap_portal(struct idxd_wq *wq)
283 struct device *dev = &wq->idxd->pdev->dev;
285 devm_iounmap(dev, wq->portal);
287 wq->portal_offset = 0;
290 void idxd_wqs_unmap_portal(struct idxd_device *idxd)
294 for (i = 0; i < idxd->max_wqs; i++) {
295 struct idxd_wq *wq = idxd->wqs[i];
298 idxd_wq_unmap_portal(wq);
302 static void __idxd_wq_set_priv_locked(struct idxd_wq *wq, int priv)
304 struct idxd_device *idxd = wq->idxd;
308 offset = WQCFG_OFFSET(idxd, wq->id, WQCFG_PRIVL_IDX);
309 spin_lock(&idxd->dev_lock);
310 wqcfg.bits[WQCFG_PRIVL_IDX] = ioread32(idxd->reg_base + offset);
312 wq->wqcfg->bits[WQCFG_PRIVL_IDX] = wqcfg.bits[WQCFG_PRIVL_IDX];
313 iowrite32(wqcfg.bits[WQCFG_PRIVL_IDX], idxd->reg_base + offset);
314 spin_unlock(&idxd->dev_lock);
317 static void __idxd_wq_set_pasid_locked(struct idxd_wq *wq, int pasid)
319 struct idxd_device *idxd = wq->idxd;
323 offset = WQCFG_OFFSET(idxd, wq->id, WQCFG_PASID_IDX);
324 spin_lock(&idxd->dev_lock);
325 wqcfg.bits[WQCFG_PASID_IDX] = ioread32(idxd->reg_base + offset);
328 wq->wqcfg->bits[WQCFG_PASID_IDX] = wqcfg.bits[WQCFG_PASID_IDX];
329 iowrite32(wqcfg.bits[WQCFG_PASID_IDX], idxd->reg_base + offset);
330 spin_unlock(&idxd->dev_lock);
333 int idxd_wq_set_pasid(struct idxd_wq *wq, int pasid)
337 rc = idxd_wq_disable(wq, false);
341 __idxd_wq_set_pasid_locked(wq, pasid);
343 rc = idxd_wq_enable(wq);
350 int idxd_wq_disable_pasid(struct idxd_wq *wq)
352 struct idxd_device *idxd = wq->idxd;
357 rc = idxd_wq_disable(wq, false);
361 offset = WQCFG_OFFSET(idxd, wq->id, WQCFG_PASID_IDX);
362 spin_lock(&idxd->dev_lock);
363 wqcfg.bits[WQCFG_PASID_IDX] = ioread32(idxd->reg_base + offset);
366 iowrite32(wqcfg.bits[WQCFG_PASID_IDX], idxd->reg_base + offset);
367 spin_unlock(&idxd->dev_lock);
369 rc = idxd_wq_enable(wq);
376 static void idxd_wq_disable_cleanup(struct idxd_wq *wq)
378 struct idxd_device *idxd = wq->idxd;
380 lockdep_assert_held(&wq->wq_lock);
381 wq->state = IDXD_WQ_DISABLED;
382 memset(wq->wqcfg, 0, idxd->wqcfg_size);
383 wq->type = IDXD_WQT_NONE;
386 wq->enqcmds_retries = IDXD_ENQCMDS_RETRIES;
387 clear_bit(WQ_FLAG_DEDICATED, &wq->flags);
388 clear_bit(WQ_FLAG_BLOCK_ON_FAULT, &wq->flags);
389 clear_bit(WQ_FLAG_ATS_DISABLE, &wq->flags);
390 memset(wq->name, 0, WQ_NAME_SIZE);
391 wq->max_xfer_bytes = WQ_DEFAULT_MAX_XFER;
392 idxd_wq_set_max_batch_size(idxd->data->type, wq, WQ_DEFAULT_MAX_BATCH);
394 bitmap_copy(wq->opcap_bmap, idxd->opcap_bmap, IDXD_MAX_OPCAP_BITS);
397 static void idxd_wq_device_reset_cleanup(struct idxd_wq *wq)
399 lockdep_assert_held(&wq->wq_lock);
405 static void idxd_wq_ref_release(struct percpu_ref *ref)
407 struct idxd_wq *wq = container_of(ref, struct idxd_wq, wq_active);
409 complete(&wq->wq_dead);
412 int idxd_wq_init_percpu_ref(struct idxd_wq *wq)
416 memset(&wq->wq_active, 0, sizeof(wq->wq_active));
417 rc = percpu_ref_init(&wq->wq_active, idxd_wq_ref_release,
418 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL);
421 reinit_completion(&wq->wq_dead);
422 reinit_completion(&wq->wq_resurrect);
426 void __idxd_wq_quiesce(struct idxd_wq *wq)
428 lockdep_assert_held(&wq->wq_lock);
429 reinit_completion(&wq->wq_resurrect);
430 percpu_ref_kill(&wq->wq_active);
431 complete_all(&wq->wq_resurrect);
432 wait_for_completion(&wq->wq_dead);
435 void idxd_wq_quiesce(struct idxd_wq *wq)
437 mutex_lock(&wq->wq_lock);
438 __idxd_wq_quiesce(wq);
439 mutex_unlock(&wq->wq_lock);
442 /* Device control bits */
443 static inline bool idxd_is_enabled(struct idxd_device *idxd)
445 union gensts_reg gensts;
447 gensts.bits = ioread32(idxd->reg_base + IDXD_GENSTATS_OFFSET);
449 if (gensts.state == IDXD_DEVICE_STATE_ENABLED)
454 static inline bool idxd_device_is_halted(struct idxd_device *idxd)
456 union gensts_reg gensts;
458 gensts.bits = ioread32(idxd->reg_base + IDXD_GENSTATS_OFFSET);
460 return (gensts.state == IDXD_DEVICE_STATE_HALT);
464 * This is function is only used for reset during probe and will
465 * poll for completion. Once the device is setup with interrupts,
466 * all commands will be done via interrupt completion.
468 int idxd_device_init_reset(struct idxd_device *idxd)
470 struct device *dev = &idxd->pdev->dev;
471 union idxd_command_reg cmd;
473 if (idxd_device_is_halted(idxd)) {
474 dev_warn(&idxd->pdev->dev, "Device is HALTED!\n");
478 memset(&cmd, 0, sizeof(cmd));
479 cmd.cmd = IDXD_CMD_RESET_DEVICE;
480 dev_dbg(dev, "%s: sending reset for init.\n", __func__);
481 spin_lock(&idxd->cmd_lock);
482 iowrite32(cmd.bits, idxd->reg_base + IDXD_CMD_OFFSET);
484 while (ioread32(idxd->reg_base + IDXD_CMDSTS_OFFSET) &
487 spin_unlock(&idxd->cmd_lock);
491 static void idxd_cmd_exec(struct idxd_device *idxd, int cmd_code, u32 operand,
494 union idxd_command_reg cmd;
495 DECLARE_COMPLETION_ONSTACK(done);
498 if (idxd_device_is_halted(idxd)) {
499 dev_warn(&idxd->pdev->dev, "Device is HALTED!\n");
501 *status = IDXD_CMDSTS_HW_ERR;
505 memset(&cmd, 0, sizeof(cmd));
507 cmd.operand = operand;
510 spin_lock(&idxd->cmd_lock);
511 wait_event_lock_irq(idxd->cmd_waitq,
512 !test_bit(IDXD_FLAG_CMD_RUNNING, &idxd->flags),
515 dev_dbg(&idxd->pdev->dev, "%s: sending cmd: %#x op: %#x\n",
516 __func__, cmd_code, operand);
518 idxd->cmd_status = 0;
519 __set_bit(IDXD_FLAG_CMD_RUNNING, &idxd->flags);
520 idxd->cmd_done = &done;
521 iowrite32(cmd.bits, idxd->reg_base + IDXD_CMD_OFFSET);
524 * After command submitted, release lock and go to sleep until
525 * the command completes via interrupt.
527 spin_unlock(&idxd->cmd_lock);
528 wait_for_completion(&done);
529 stat = ioread32(idxd->reg_base + IDXD_CMDSTS_OFFSET);
530 spin_lock(&idxd->cmd_lock);
533 idxd->cmd_status = stat & GENMASK(7, 0);
535 __clear_bit(IDXD_FLAG_CMD_RUNNING, &idxd->flags);
536 /* Wake up other pending commands */
537 wake_up(&idxd->cmd_waitq);
538 spin_unlock(&idxd->cmd_lock);
541 int idxd_device_enable(struct idxd_device *idxd)
543 struct device *dev = &idxd->pdev->dev;
546 if (idxd_is_enabled(idxd)) {
547 dev_dbg(dev, "Device already enabled\n");
551 idxd_cmd_exec(idxd, IDXD_CMD_ENABLE_DEVICE, 0, &status);
553 /* If the command is successful or if the device was enabled */
554 if (status != IDXD_CMDSTS_SUCCESS &&
555 status != IDXD_CMDSTS_ERR_DEV_ENABLED) {
556 dev_dbg(dev, "%s: err_code: %#x\n", __func__, status);
560 idxd->state = IDXD_DEV_ENABLED;
564 int idxd_device_disable(struct idxd_device *idxd)
566 struct device *dev = &idxd->pdev->dev;
569 if (!idxd_is_enabled(idxd)) {
570 dev_dbg(dev, "Device is not enabled\n");
574 idxd_cmd_exec(idxd, IDXD_CMD_DISABLE_DEVICE, 0, &status);
576 /* If the command is successful or if the device was disabled */
577 if (status != IDXD_CMDSTS_SUCCESS &&
578 !(status & IDXD_CMDSTS_ERR_DIS_DEV_EN)) {
579 dev_dbg(dev, "%s: err_code: %#x\n", __func__, status);
583 idxd_device_clear_state(idxd);
587 void idxd_device_reset(struct idxd_device *idxd)
589 idxd_cmd_exec(idxd, IDXD_CMD_RESET_DEVICE, 0, NULL);
590 idxd_device_clear_state(idxd);
591 spin_lock(&idxd->dev_lock);
592 idxd_unmask_error_interrupts(idxd);
593 spin_unlock(&idxd->dev_lock);
596 void idxd_device_drain_pasid(struct idxd_device *idxd, int pasid)
598 struct device *dev = &idxd->pdev->dev;
602 dev_dbg(dev, "cmd: %u operand: %#x\n", IDXD_CMD_DRAIN_PASID, operand);
603 idxd_cmd_exec(idxd, IDXD_CMD_DRAIN_PASID, operand, NULL);
604 dev_dbg(dev, "pasid %d drained\n", pasid);
607 int idxd_device_request_int_handle(struct idxd_device *idxd, int idx, int *handle,
608 enum idxd_interrupt_type irq_type)
610 struct device *dev = &idxd->pdev->dev;
613 if (!(idxd->hw.cmd_cap & BIT(IDXD_CMD_REQUEST_INT_HANDLE)))
616 dev_dbg(dev, "get int handle, idx %d\n", idx);
618 operand = idx & GENMASK(15, 0);
619 if (irq_type == IDXD_IRQ_IMS)
620 operand |= CMD_INT_HANDLE_IMS;
622 dev_dbg(dev, "cmd: %u operand: %#x\n", IDXD_CMD_REQUEST_INT_HANDLE, operand);
624 idxd_cmd_exec(idxd, IDXD_CMD_REQUEST_INT_HANDLE, operand, &status);
626 if ((status & IDXD_CMDSTS_ERR_MASK) != IDXD_CMDSTS_SUCCESS) {
627 dev_dbg(dev, "request int handle failed: %#x\n", status);
631 *handle = (status >> IDXD_CMDSTS_RES_SHIFT) & GENMASK(15, 0);
633 dev_dbg(dev, "int handle acquired: %u\n", *handle);
637 int idxd_device_release_int_handle(struct idxd_device *idxd, int handle,
638 enum idxd_interrupt_type irq_type)
640 struct device *dev = &idxd->pdev->dev;
642 union idxd_command_reg cmd;
644 if (!(idxd->hw.cmd_cap & BIT(IDXD_CMD_RELEASE_INT_HANDLE)))
647 dev_dbg(dev, "release int handle, handle %d\n", handle);
649 memset(&cmd, 0, sizeof(cmd));
650 operand = handle & GENMASK(15, 0);
652 if (irq_type == IDXD_IRQ_IMS)
653 operand |= CMD_INT_HANDLE_IMS;
655 cmd.cmd = IDXD_CMD_RELEASE_INT_HANDLE;
656 cmd.operand = operand;
658 dev_dbg(dev, "cmd: %u operand: %#x\n", IDXD_CMD_RELEASE_INT_HANDLE, operand);
660 spin_lock(&idxd->cmd_lock);
661 iowrite32(cmd.bits, idxd->reg_base + IDXD_CMD_OFFSET);
663 while (ioread32(idxd->reg_base + IDXD_CMDSTS_OFFSET) & IDXD_CMDSTS_ACTIVE)
665 status = ioread32(idxd->reg_base + IDXD_CMDSTS_OFFSET);
666 spin_unlock(&idxd->cmd_lock);
668 if ((status & IDXD_CMDSTS_ERR_MASK) != IDXD_CMDSTS_SUCCESS) {
669 dev_dbg(dev, "release int handle failed: %#x\n", status);
673 dev_dbg(dev, "int handle released.\n");
677 /* Device configuration bits */
678 static void idxd_engines_clear_state(struct idxd_device *idxd)
680 struct idxd_engine *engine;
683 lockdep_assert_held(&idxd->dev_lock);
684 for (i = 0; i < idxd->max_engines; i++) {
685 engine = idxd->engines[i];
686 engine->group = NULL;
690 static void idxd_groups_clear_state(struct idxd_device *idxd)
692 struct idxd_group *group;
695 lockdep_assert_held(&idxd->dev_lock);
696 for (i = 0; i < idxd->max_groups; i++) {
697 group = idxd->groups[i];
698 memset(&group->grpcfg, 0, sizeof(group->grpcfg));
699 group->num_engines = 0;
701 group->use_rdbuf_limit = false;
703 * The default value is the same as the value of
704 * total read buffers in GRPCAP.
706 group->rdbufs_allowed = idxd->max_rdbufs;
707 group->rdbufs_reserved = 0;
708 if (idxd->hw.version <= DEVICE_VERSION_2 && !tc_override) {
715 group->desc_progress_limit = 0;
716 group->batch_progress_limit = 0;
720 static void idxd_device_wqs_clear_state(struct idxd_device *idxd)
724 for (i = 0; i < idxd->max_wqs; i++) {
725 struct idxd_wq *wq = idxd->wqs[i];
727 mutex_lock(&wq->wq_lock);
728 idxd_wq_disable_cleanup(wq);
729 idxd_wq_device_reset_cleanup(wq);
730 mutex_unlock(&wq->wq_lock);
734 void idxd_device_clear_state(struct idxd_device *idxd)
736 /* IDXD is always disabled. Other states are cleared only when IDXD is configurable. */
737 if (test_bit(IDXD_FLAG_CONFIGURABLE, &idxd->flags)) {
739 * Clearing wq state is protected by wq lock.
740 * So no need to be protected by device lock.
742 idxd_device_wqs_clear_state(idxd);
744 spin_lock(&idxd->dev_lock);
745 idxd_groups_clear_state(idxd);
746 idxd_engines_clear_state(idxd);
748 spin_lock(&idxd->dev_lock);
751 idxd->state = IDXD_DEV_DISABLED;
752 spin_unlock(&idxd->dev_lock);
755 static void idxd_group_config_write(struct idxd_group *group)
757 struct idxd_device *idxd = group->idxd;
758 struct device *dev = &idxd->pdev->dev;
762 dev_dbg(dev, "Writing group %d cfg registers\n", group->id);
765 for (i = 0; i < GRPWQCFG_STRIDES; i++) {
766 grpcfg_offset = GRPWQCFG_OFFSET(idxd, group->id, i);
767 iowrite64(group->grpcfg.wqs[i], idxd->reg_base + grpcfg_offset);
768 dev_dbg(dev, "GRPCFG wq[%d:%d: %#x]: %#llx\n",
769 group->id, i, grpcfg_offset,
770 ioread64(idxd->reg_base + grpcfg_offset));
773 /* setup GRPENGCFG */
774 grpcfg_offset = GRPENGCFG_OFFSET(idxd, group->id);
775 iowrite64(group->grpcfg.engines, idxd->reg_base + grpcfg_offset);
776 dev_dbg(dev, "GRPCFG engs[%d: %#x]: %#llx\n", group->id,
777 grpcfg_offset, ioread64(idxd->reg_base + grpcfg_offset));
780 grpcfg_offset = GRPFLGCFG_OFFSET(idxd, group->id);
781 iowrite64(group->grpcfg.flags.bits, idxd->reg_base + grpcfg_offset);
782 dev_dbg(dev, "GRPFLAGS flags[%d: %#x]: %#llx\n",
783 group->id, grpcfg_offset,
784 ioread64(idxd->reg_base + grpcfg_offset));
787 static int idxd_groups_config_write(struct idxd_device *idxd)
790 union gencfg_reg reg;
792 struct device *dev = &idxd->pdev->dev;
794 /* Setup bandwidth rdbuf limit */
795 if (idxd->hw.gen_cap.config_en && idxd->rdbuf_limit) {
796 reg.bits = ioread32(idxd->reg_base + IDXD_GENCFG_OFFSET);
797 reg.rdbuf_limit = idxd->rdbuf_limit;
798 iowrite32(reg.bits, idxd->reg_base + IDXD_GENCFG_OFFSET);
801 dev_dbg(dev, "GENCFG(%#x): %#x\n", IDXD_GENCFG_OFFSET,
802 ioread32(idxd->reg_base + IDXD_GENCFG_OFFSET));
804 for (i = 0; i < idxd->max_groups; i++) {
805 struct idxd_group *group = idxd->groups[i];
807 idxd_group_config_write(group);
813 static bool idxd_device_pasid_priv_enabled(struct idxd_device *idxd)
815 struct pci_dev *pdev = idxd->pdev;
817 if (pdev->pasid_enabled && (pdev->pasid_features & PCI_PASID_CAP_PRIV))
822 static int idxd_wq_config_write(struct idxd_wq *wq)
824 struct idxd_device *idxd = wq->idxd;
825 struct device *dev = &idxd->pdev->dev;
833 * Instead of memset the entire shadow copy of WQCFG, copy from the hardware after
834 * wq reset. This will copy back the sticky values that are present on some devices.
836 for (i = 0; i < WQCFG_STRIDES(idxd); i++) {
837 wq_offset = WQCFG_OFFSET(idxd, wq->id, i);
838 wq->wqcfg->bits[i] |= ioread32(idxd->reg_base + wq_offset);
841 if (wq->size == 0 && wq->type != IDXD_WQT_NONE)
842 wq->size = WQ_DEFAULT_QUEUE_DEPTH;
845 wq->wqcfg->wq_size = wq->size;
848 wq->wqcfg->wq_thresh = wq->threshold;
851 if (wq_dedicated(wq))
855 * The WQ priv bit is set depending on the WQ type. priv = 1 if the
856 * WQ type is kernel to indicate privileged access. This setting only
857 * matters for dedicated WQ. According to the DSA spec:
858 * If the WQ is in dedicated mode, WQ PASID Enable is 1, and the
859 * Privileged Mode Enable field of the PCI Express PASID capability
860 * is 0, this field must be 0.
862 * In the case of a dedicated kernel WQ that is not able to support
863 * the PASID cap, then the configuration will be rejected.
865 if (wq_dedicated(wq) && wq->wqcfg->pasid_en &&
866 !idxd_device_pasid_priv_enabled(idxd) &&
867 wq->type == IDXD_WQT_KERNEL) {
868 idxd->cmd_status = IDXD_SCMD_WQ_NO_PRIV;
872 wq->wqcfg->priority = wq->priority;
874 if (idxd->hw.gen_cap.block_on_fault &&
875 test_bit(WQ_FLAG_BLOCK_ON_FAULT, &wq->flags))
878 if (idxd->hw.wq_cap.wq_ats_support)
879 wq->wqcfg->wq_ats_disable = test_bit(WQ_FLAG_ATS_DISABLE, &wq->flags);
882 wq->wqcfg->max_xfer_shift = ilog2(wq->max_xfer_bytes);
883 idxd_wqcfg_set_max_batch_shift(idxd->data->type, wq->wqcfg, ilog2(wq->max_batch_size));
886 if (idxd->hw.wq_cap.op_config && wq->opcap_bmap) {
887 memset(wq->wqcfg->op_config, 0, IDXD_MAX_OPCAP_BITS / 8);
888 for_each_set_bit(n, wq->opcap_bmap, IDXD_MAX_OPCAP_BITS) {
889 int pos = n % BITS_PER_LONG_LONG;
890 int idx = n / BITS_PER_LONG_LONG;
892 wq->wqcfg->op_config[idx] |= BIT(pos);
896 dev_dbg(dev, "WQ %d CFGs\n", wq->id);
897 for (i = 0; i < WQCFG_STRIDES(idxd); i++) {
898 wq_offset = WQCFG_OFFSET(idxd, wq->id, i);
899 iowrite32(wq->wqcfg->bits[i], idxd->reg_base + wq_offset);
900 dev_dbg(dev, "WQ[%d][%d][%#x]: %#x\n",
901 wq->id, i, wq_offset,
902 ioread32(idxd->reg_base + wq_offset));
908 static int idxd_wqs_config_write(struct idxd_device *idxd)
912 for (i = 0; i < idxd->max_wqs; i++) {
913 struct idxd_wq *wq = idxd->wqs[i];
915 rc = idxd_wq_config_write(wq);
923 static void idxd_group_flags_setup(struct idxd_device *idxd)
927 /* TC-A 0 and TC-B 1 should be defaults */
928 for (i = 0; i < idxd->max_groups; i++) {
929 struct idxd_group *group = idxd->groups[i];
931 if (group->tc_a == -1)
932 group->tc_a = group->grpcfg.flags.tc_a = 0;
934 group->grpcfg.flags.tc_a = group->tc_a;
935 if (group->tc_b == -1)
936 group->tc_b = group->grpcfg.flags.tc_b = 1;
938 group->grpcfg.flags.tc_b = group->tc_b;
939 group->grpcfg.flags.use_rdbuf_limit = group->use_rdbuf_limit;
940 group->grpcfg.flags.rdbufs_reserved = group->rdbufs_reserved;
941 group->grpcfg.flags.rdbufs_allowed = group->rdbufs_allowed;
942 group->grpcfg.flags.desc_progress_limit = group->desc_progress_limit;
943 group->grpcfg.flags.batch_progress_limit = group->batch_progress_limit;
947 static int idxd_engines_setup(struct idxd_device *idxd)
950 struct idxd_engine *eng;
951 struct idxd_group *group;
953 for (i = 0; i < idxd->max_groups; i++) {
954 group = idxd->groups[i];
955 group->grpcfg.engines = 0;
958 for (i = 0; i < idxd->max_engines; i++) {
959 eng = idxd->engines[i];
965 group->grpcfg.engines |= BIT(eng->id);
975 static int idxd_wqs_setup(struct idxd_device *idxd)
978 struct idxd_group *group;
979 int i, j, configured = 0;
980 struct device *dev = &idxd->pdev->dev;
982 for (i = 0; i < idxd->max_groups; i++) {
983 group = idxd->groups[i];
984 for (j = 0; j < 4; j++)
985 group->grpcfg.wqs[j] = 0;
988 for (i = 0; i < idxd->max_wqs; i++) {
995 if (wq_shared(wq) && !wq_shared_supported(wq)) {
996 idxd->cmd_status = IDXD_SCMD_WQ_NO_SWQ_SUPPORT;
997 dev_warn(dev, "No shared wq support but configured.\n");
1001 group->grpcfg.wqs[wq->id / 64] |= BIT(wq->id % 64);
1005 if (configured == 0) {
1006 idxd->cmd_status = IDXD_SCMD_WQ_NONE_CONFIGURED;
1013 int idxd_device_config(struct idxd_device *idxd)
1017 lockdep_assert_held(&idxd->dev_lock);
1018 rc = idxd_wqs_setup(idxd);
1022 rc = idxd_engines_setup(idxd);
1026 idxd_group_flags_setup(idxd);
1028 rc = idxd_wqs_config_write(idxd);
1032 rc = idxd_groups_config_write(idxd);
1039 static int idxd_wq_load_config(struct idxd_wq *wq)
1041 struct idxd_device *idxd = wq->idxd;
1042 struct device *dev = &idxd->pdev->dev;
1046 wqcfg_offset = WQCFG_OFFSET(idxd, wq->id, 0);
1047 memcpy_fromio(wq->wqcfg, idxd->reg_base + wqcfg_offset, idxd->wqcfg_size);
1049 wq->size = wq->wqcfg->wq_size;
1050 wq->threshold = wq->wqcfg->wq_thresh;
1052 /* The driver does not support shared WQ mode in read-only config yet */
1053 if (wq->wqcfg->mode == 0 || wq->wqcfg->pasid_en)
1056 set_bit(WQ_FLAG_DEDICATED, &wq->flags);
1058 wq->priority = wq->wqcfg->priority;
1060 wq->max_xfer_bytes = 1ULL << wq->wqcfg->max_xfer_shift;
1061 idxd_wq_set_max_batch_size(idxd->data->type, wq, 1U << wq->wqcfg->max_batch_shift);
1063 for (i = 0; i < WQCFG_STRIDES(idxd); i++) {
1064 wqcfg_offset = WQCFG_OFFSET(idxd, wq->id, i);
1065 dev_dbg(dev, "WQ[%d][%d][%#x]: %#x\n", wq->id, i, wqcfg_offset, wq->wqcfg->bits[i]);
1071 static void idxd_group_load_config(struct idxd_group *group)
1073 struct idxd_device *idxd = group->idxd;
1074 struct device *dev = &idxd->pdev->dev;
1075 int i, j, grpcfg_offset;
1078 * Load WQS bit fields
1079 * Iterate through all 256 bits 64 bits at a time
1081 for (i = 0; i < GRPWQCFG_STRIDES; i++) {
1084 grpcfg_offset = GRPWQCFG_OFFSET(idxd, group->id, i);
1085 group->grpcfg.wqs[i] = ioread64(idxd->reg_base + grpcfg_offset);
1086 dev_dbg(dev, "GRPCFG wq[%d:%d: %#x]: %#llx\n",
1087 group->id, i, grpcfg_offset, group->grpcfg.wqs[i]);
1089 if (i * 64 >= idxd->max_wqs)
1092 /* Iterate through all 64 bits and check for wq set */
1093 for (j = 0; j < 64; j++) {
1094 int id = i * 64 + j;
1096 /* No need to check beyond max wqs */
1097 if (id >= idxd->max_wqs)
1100 /* Set group assignment for wq if wq bit is set */
1101 if (group->grpcfg.wqs[i] & BIT(j)) {
1108 grpcfg_offset = GRPENGCFG_OFFSET(idxd, group->id);
1109 group->grpcfg.engines = ioread64(idxd->reg_base + grpcfg_offset);
1110 dev_dbg(dev, "GRPCFG engs[%d: %#x]: %#llx\n", group->id,
1111 grpcfg_offset, group->grpcfg.engines);
1113 /* Iterate through all 64 bits to check engines set */
1114 for (i = 0; i < 64; i++) {
1115 if (i >= idxd->max_engines)
1118 if (group->grpcfg.engines & BIT(i)) {
1119 struct idxd_engine *engine = idxd->engines[i];
1121 engine->group = group;
1125 grpcfg_offset = GRPFLGCFG_OFFSET(idxd, group->id);
1126 group->grpcfg.flags.bits = ioread64(idxd->reg_base + grpcfg_offset);
1127 dev_dbg(dev, "GRPFLAGS flags[%d: %#x]: %#llx\n",
1128 group->id, grpcfg_offset, group->grpcfg.flags.bits);
1131 int idxd_device_load_config(struct idxd_device *idxd)
1133 union gencfg_reg reg;
1136 reg.bits = ioread32(idxd->reg_base + IDXD_GENCFG_OFFSET);
1137 idxd->rdbuf_limit = reg.rdbuf_limit;
1139 for (i = 0; i < idxd->max_groups; i++) {
1140 struct idxd_group *group = idxd->groups[i];
1142 idxd_group_load_config(group);
1145 for (i = 0; i < idxd->max_wqs; i++) {
1146 struct idxd_wq *wq = idxd->wqs[i];
1148 rc = idxd_wq_load_config(wq);
1156 static void idxd_flush_pending_descs(struct idxd_irq_entry *ie)
1158 struct idxd_desc *desc, *itr;
1159 struct llist_node *head;
1161 enum idxd_complete_type ctype;
1163 spin_lock(&ie->list_lock);
1164 head = llist_del_all(&ie->pending_llist);
1166 llist_for_each_entry_safe(desc, itr, head, llnode)
1167 list_add_tail(&desc->list, &ie->work_list);
1170 list_for_each_entry_safe(desc, itr, &ie->work_list, list)
1171 list_move_tail(&desc->list, &flist);
1172 spin_unlock(&ie->list_lock);
1174 list_for_each_entry_safe(desc, itr, &flist, list) {
1175 struct dma_async_tx_descriptor *tx;
1177 list_del(&desc->list);
1178 ctype = desc->completion->status ? IDXD_COMPLETE_NORMAL : IDXD_COMPLETE_ABORT;
1180 * wq is being disabled. Any remaining descriptors are
1181 * likely to be stuck and can be dropped. callback could
1182 * point to code that is no longer accessible, for example
1183 * if dmatest module has been unloaded.
1186 tx->callback = NULL;
1187 tx->callback_result = NULL;
1188 idxd_dma_complete_txd(desc, ctype, true);
1192 static void idxd_device_set_perm_entry(struct idxd_device *idxd,
1193 struct idxd_irq_entry *ie)
1195 union msix_perm mperm;
1197 if (ie->pasid == INVALID_IOASID)
1201 mperm.pasid = ie->pasid;
1203 iowrite32(mperm.bits, idxd->reg_base + idxd->msix_perm_offset + ie->id * 8);
1206 static void idxd_device_clear_perm_entry(struct idxd_device *idxd,
1207 struct idxd_irq_entry *ie)
1209 iowrite32(0, idxd->reg_base + idxd->msix_perm_offset + ie->id * 8);
1212 void idxd_wq_free_irq(struct idxd_wq *wq)
1214 struct idxd_device *idxd = wq->idxd;
1215 struct idxd_irq_entry *ie = &wq->ie;
1217 if (wq->type != IDXD_WQT_KERNEL)
1220 free_irq(ie->vector, ie);
1221 idxd_flush_pending_descs(ie);
1222 if (idxd->request_int_handles)
1223 idxd_device_release_int_handle(idxd, ie->int_handle, IDXD_IRQ_MSIX);
1224 idxd_device_clear_perm_entry(idxd, ie);
1226 ie->int_handle = INVALID_INT_HANDLE;
1227 ie->pasid = INVALID_IOASID;
1230 int idxd_wq_request_irq(struct idxd_wq *wq)
1232 struct idxd_device *idxd = wq->idxd;
1233 struct pci_dev *pdev = idxd->pdev;
1234 struct device *dev = &pdev->dev;
1235 struct idxd_irq_entry *ie;
1238 if (wq->type != IDXD_WQT_KERNEL)
1242 ie->vector = pci_irq_vector(pdev, ie->id);
1243 ie->pasid = device_pasid_enabled(idxd) ? idxd->pasid : INVALID_IOASID;
1244 idxd_device_set_perm_entry(idxd, ie);
1246 rc = request_threaded_irq(ie->vector, NULL, idxd_wq_thread, 0, "idxd-portal", ie);
1248 dev_err(dev, "Failed to request irq %d.\n", ie->vector);
1252 if (idxd->request_int_handles) {
1253 rc = idxd_device_request_int_handle(idxd, ie->id, &ie->int_handle,
1256 goto err_int_handle;
1258 ie->int_handle = ie->id;
1264 ie->int_handle = INVALID_INT_HANDLE;
1265 free_irq(ie->vector, ie);
1267 idxd_device_clear_perm_entry(idxd, ie);
1268 ie->pasid = INVALID_IOASID;
1272 int drv_enable_wq(struct idxd_wq *wq)
1274 struct idxd_device *idxd = wq->idxd;
1275 struct device *dev = &idxd->pdev->dev;
1278 lockdep_assert_held(&wq->wq_lock);
1280 if (idxd->state != IDXD_DEV_ENABLED) {
1281 idxd->cmd_status = IDXD_SCMD_DEV_NOT_ENABLED;
1285 if (wq->state != IDXD_WQ_DISABLED) {
1286 dev_dbg(dev, "wq %d already enabled.\n", wq->id);
1287 idxd->cmd_status = IDXD_SCMD_WQ_ENABLED;
1293 dev_dbg(dev, "wq %d not attached to group.\n", wq->id);
1294 idxd->cmd_status = IDXD_SCMD_WQ_NO_GRP;
1298 if (strlen(wq->name) == 0) {
1299 idxd->cmd_status = IDXD_SCMD_WQ_NO_NAME;
1300 dev_dbg(dev, "wq %d name not set.\n", wq->id);
1304 /* Shared WQ checks */
1305 if (wq_shared(wq)) {
1306 if (!wq_shared_supported(wq)) {
1307 idxd->cmd_status = IDXD_SCMD_WQ_NO_SVM;
1308 dev_dbg(dev, "PASID not enabled and shared wq.\n");
1312 * Shared wq with the threshold set to 0 means the user
1313 * did not set the threshold or transitioned from a
1314 * dedicated wq but did not set threshold. A value
1315 * of 0 would effectively disable the shared wq. The
1316 * driver does not allow a value of 0 to be set for
1317 * threshold via sysfs.
1319 if (wq->threshold == 0) {
1320 idxd->cmd_status = IDXD_SCMD_WQ_NO_THRESH;
1321 dev_dbg(dev, "Shared wq and threshold 0.\n");
1327 * In the event that the WQ is configurable for pasid and priv bits.
1328 * For kernel wq, the driver should setup the pasid, pasid_en, and priv bit.
1329 * However, for non-kernel wq, the driver should only set the pasid_en bit for
1330 * shared wq. A dedicated wq that is not 'kernel' type will configure pasid and
1331 * pasid_en later on so there is no need to setup.
1333 if (test_bit(IDXD_FLAG_CONFIGURABLE, &idxd->flags)) {
1336 if (wq_pasid_enabled(wq)) {
1337 if (is_idxd_wq_kernel(wq) || wq_shared(wq)) {
1338 u32 pasid = wq_dedicated(wq) ? idxd->pasid : 0;
1340 __idxd_wq_set_pasid_locked(wq, pasid);
1344 if (is_idxd_wq_kernel(wq))
1346 __idxd_wq_set_priv_locked(wq, priv);
1350 spin_lock(&idxd->dev_lock);
1351 if (test_bit(IDXD_FLAG_CONFIGURABLE, &idxd->flags))
1352 rc = idxd_device_config(idxd);
1353 spin_unlock(&idxd->dev_lock);
1355 dev_dbg(dev, "Writing wq %d config failed: %d\n", wq->id, rc);
1359 rc = idxd_wq_enable(wq);
1361 dev_dbg(dev, "wq %d enabling failed: %d\n", wq->id, rc);
1365 rc = idxd_wq_map_portal(wq);
1367 idxd->cmd_status = IDXD_SCMD_WQ_PORTAL_ERR;
1368 dev_dbg(dev, "wq %d portal mapping failed: %d\n", wq->id, rc);
1369 goto err_map_portal;
1372 wq->client_count = 0;
1374 rc = idxd_wq_request_irq(wq);
1376 idxd->cmd_status = IDXD_SCMD_WQ_IRQ_ERR;
1377 dev_dbg(dev, "WQ %d irq setup failed: %d\n", wq->id, rc);
1381 rc = idxd_wq_alloc_resources(wq);
1383 idxd->cmd_status = IDXD_SCMD_WQ_RES_ALLOC_ERR;
1384 dev_dbg(dev, "WQ resource alloc failed\n");
1388 rc = idxd_wq_init_percpu_ref(wq);
1390 idxd->cmd_status = IDXD_SCMD_PERCPU_ERR;
1391 dev_dbg(dev, "percpu_ref setup failed\n");
1398 idxd_wq_free_resources(wq);
1400 idxd_wq_free_irq(wq);
1402 idxd_wq_unmap_portal(wq);
1404 if (idxd_wq_disable(wq, false))
1405 dev_dbg(dev, "wq %s disable failed\n", dev_name(wq_confdev(wq)));
1410 void drv_disable_wq(struct idxd_wq *wq)
1412 struct idxd_device *idxd = wq->idxd;
1413 struct device *dev = &idxd->pdev->dev;
1415 lockdep_assert_held(&wq->wq_lock);
1417 if (idxd_wq_refcount(wq))
1418 dev_warn(dev, "Clients has claim on wq %d: %d\n",
1419 wq->id, idxd_wq_refcount(wq));
1421 idxd_wq_unmap_portal(wq);
1423 idxd_wq_free_irq(wq);
1425 idxd_wq_free_resources(wq);
1426 percpu_ref_exit(&wq->wq_active);
1427 wq->type = IDXD_WQT_NONE;
1428 wq->client_count = 0;
1431 int idxd_device_drv_probe(struct idxd_dev *idxd_dev)
1433 struct idxd_device *idxd = idxd_dev_to_idxd(idxd_dev);
1437 * Device should be in disabled state for the idxd_drv to load. If it's in
1438 * enabled state, then the device was altered outside of driver's control.
1439 * If the state is in halted state, then we don't want to proceed.
1441 if (idxd->state != IDXD_DEV_DISABLED) {
1442 idxd->cmd_status = IDXD_SCMD_DEV_ENABLED;
1446 /* Device configuration */
1447 spin_lock(&idxd->dev_lock);
1448 if (test_bit(IDXD_FLAG_CONFIGURABLE, &idxd->flags))
1449 rc = idxd_device_config(idxd);
1450 spin_unlock(&idxd->dev_lock);
1455 rc = idxd_device_enable(idxd);
1459 /* Setup DMA device without channels */
1460 rc = idxd_register_dma_device(idxd);
1462 idxd_device_disable(idxd);
1463 idxd->cmd_status = IDXD_SCMD_DEV_DMA_ERR;
1467 idxd->cmd_status = 0;
1471 void idxd_device_drv_remove(struct idxd_dev *idxd_dev)
1473 struct device *dev = &idxd_dev->conf_dev;
1474 struct idxd_device *idxd = idxd_dev_to_idxd(idxd_dev);
1477 for (i = 0; i < idxd->max_wqs; i++) {
1478 struct idxd_wq *wq = idxd->wqs[i];
1479 struct device *wq_dev = wq_confdev(wq);
1481 if (wq->state == IDXD_WQ_DISABLED)
1483 dev_warn(dev, "Active wq %d on disable %s.\n", i, dev_name(wq_dev));
1484 device_release_driver(wq_dev);
1487 idxd_unregister_dma_device(idxd);
1488 idxd_device_disable(idxd);
1489 if (test_bit(IDXD_FLAG_CONFIGURABLE, &idxd->flags))
1490 idxd_device_reset(idxd);
1493 static enum idxd_dev_type dev_types[] = {
1499 struct idxd_device_driver idxd_drv = {
1501 .probe = idxd_device_drv_probe,
1502 .remove = idxd_device_drv_remove,
1505 EXPORT_SYMBOL_GPL(idxd_drv);