diff options
| author | Jacek Lawrynowicz <jacek.lawrynowicz@linux.intel.com> | 2023-01-17 10:27:20 +0100 |
|---|---|---|
| committer | Daniel Vetter <daniel.vetter@ffwll.ch> | 2023-01-19 11:11:45 +0100 |
| commit | 5d7422cfb498bf25c4a9ea6b9d82253cb5236364 (patch) | |
| tree | 83f451dbd23a4a14de1304a9f856874f130920f9 /drivers/accel | |
| parent | 647371a6609ddf8700fe151af72e32daebb9baa7 (diff) | |
| download | linux-5d7422cfb498bf25c4a9ea6b9d82253cb5236364.tar.gz linux-5d7422cfb498bf25c4a9ea6b9d82253cb5236364.tar.bz2 linux-5d7422cfb498bf25c4a9ea6b9d82253cb5236364.zip | |
accel/ivpu: Add IPC driver and JSM messages
The IPC driver is used to send and receive messages to/from firmware
running on the VPU.
The only supported IPC message format is Job Submission Model (JSM)
defined in vpu_jsm_api.h header.
Co-developed-by: Andrzej Kacprowski <andrzej.kacprowski@linux.intel.com>
Signed-off-by: Andrzej Kacprowski <andrzej.kacprowski@linux.intel.com>
Co-developed-by: Krystian Pradzynski <krystian.pradzynski@linux.intel.com>
Signed-off-by: Krystian Pradzynski <krystian.pradzynski@linux.intel.com>
Signed-off-by: Jacek Lawrynowicz <jacek.lawrynowicz@linux.intel.com>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Reviewed-by: Jeffrey Hugo <quic_jhugo@quicinc.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Link: https://patchwork.freedesktop.org/patch/msgid/20230117092723.60441-5-jacek.lawrynowicz@linux.intel.com
Diffstat (limited to 'drivers/accel')
| -rw-r--r-- | drivers/accel/ivpu/Makefile | 2 | ||||
| -rw-r--r-- | drivers/accel/ivpu/ivpu_drv.c | 13 | ||||
| -rw-r--r-- | drivers/accel/ivpu/ivpu_drv.h | 2 | ||||
| -rw-r--r-- | drivers/accel/ivpu/ivpu_hw_mtl.c | 4 | ||||
| -rw-r--r-- | drivers/accel/ivpu/ivpu_ipc.c | 500 | ||||
| -rw-r--r-- | drivers/accel/ivpu/ivpu_ipc.h | 93 | ||||
| -rw-r--r-- | drivers/accel/ivpu/ivpu_jsm_msg.c | 169 | ||||
| -rw-r--r-- | drivers/accel/ivpu/ivpu_jsm_msg.h | 23 | ||||
| -rw-r--r-- | drivers/accel/ivpu/vpu_jsm_api.h | 999 |
9 files changed, 1805 insertions, 0 deletions
diff --git a/drivers/accel/ivpu/Makefile b/drivers/accel/ivpu/Makefile index 5d7c5862399c..46595f0112e3 100644 --- a/drivers/accel/ivpu/Makefile +++ b/drivers/accel/ivpu/Makefile @@ -5,6 +5,8 @@ intel_vpu-y := \ ivpu_drv.o \ ivpu_gem.o \ ivpu_hw_mtl.o \ + ivpu_ipc.o \ + ivpu_jsm_msg.o \ ivpu_mmu.o \ ivpu_mmu_context.o diff --git a/drivers/accel/ivpu/ivpu_drv.c b/drivers/accel/ivpu/ivpu_drv.c index 53d92d06814b..1eb12e157acd 100644 --- a/drivers/accel/ivpu/ivpu_drv.c +++ b/drivers/accel/ivpu/ivpu_drv.c @@ -17,6 +17,7 @@ #include "ivpu_drv.h" #include "ivpu_gem.h" #include "ivpu_hw.h" +#include "ivpu_ipc.h" #include "ivpu_mmu.h" #include "ivpu_mmu_context.h" @@ -229,6 +230,7 @@ int ivpu_shutdown(struct ivpu_device *vdev) int ret; ivpu_hw_irq_disable(vdev); + ivpu_ipc_disable(vdev); ivpu_mmu_disable(vdev); ret = ivpu_hw_power_down(vdev); @@ -339,6 +341,10 @@ static int ivpu_dev_init(struct ivpu_device *vdev) if (!vdev->mmu) return -ENOMEM; + vdev->ipc = drmm_kzalloc(&vdev->drm, sizeof(*vdev->ipc), GFP_KERNEL); + if (!vdev->ipc) + return -ENOMEM; + vdev->hw->ops = &ivpu_hw_mtl_ops; vdev->platform = IVPU_PLATFORM_INVALID; vdev->context_xa_limit.min = IVPU_GLOBAL_CONTEXT_MMU_SSID + 1; @@ -383,6 +389,12 @@ static int ivpu_dev_init(struct ivpu_device *vdev) goto err_mmu_gctx_fini; } + ret = ivpu_ipc_init(vdev); + if (ret) { + ivpu_err(vdev, "Failed to initialize IPC: %d\n", ret); + goto err_mmu_gctx_fini; + } + return 0; err_mmu_gctx_fini: @@ -397,6 +409,7 @@ err_xa_destroy: static void ivpu_dev_fini(struct ivpu_device *vdev) { ivpu_shutdown(vdev); + ivpu_ipc_fini(vdev); ivpu_mmu_global_context_fini(vdev); drm_WARN_ON(&vdev->drm, !xa_empty(&vdev->context_xa)); diff --git a/drivers/accel/ivpu/ivpu_drv.h b/drivers/accel/ivpu/ivpu_drv.h index 972f71cfaa12..39b706d82e68 100644 --- a/drivers/accel/ivpu/ivpu_drv.h +++ b/drivers/accel/ivpu/ivpu_drv.h @@ -74,6 +74,7 @@ struct ivpu_wa_table { struct ivpu_hw_info; struct ivpu_mmu_info; +struct ivpu_ipc_info; struct ivpu_device { struct drm_device drm; @@ -85,6 +86,7 @@ struct ivpu_device { struct ivpu_wa_table wa; struct ivpu_hw_info *hw; struct ivpu_mmu_info *mmu; + struct ivpu_ipc_info *ipc; struct ivpu_mmu_context gctx; struct xarray context_xa; diff --git a/drivers/accel/ivpu/ivpu_hw_mtl.c b/drivers/accel/ivpu/ivpu_hw_mtl.c index eaba2b8248b8..0e9ef4c40901 100644 --- a/drivers/accel/ivpu/ivpu_hw_mtl.c +++ b/drivers/accel/ivpu/ivpu_hw_mtl.c @@ -7,6 +7,7 @@ #include "ivpu_hw_mtl_reg.h" #include "ivpu_hw_reg_io.h" #include "ivpu_hw.h" +#include "ivpu_ipc.h" #include "ivpu_mmu.h" #define TILE_FUSE_ENABLE_BOTH 0x0 @@ -934,6 +935,9 @@ static u32 ivpu_hw_mtl_irqv_handler(struct ivpu_device *vdev, int irq) if (REG_TEST_FLD(MTL_VPU_HOST_SS_ICB_STATUS_0, MMU_IRQ_0_INT, status)) ivpu_mmu_irq_evtq_handler(vdev); + if (REG_TEST_FLD(MTL_VPU_HOST_SS_ICB_STATUS_0, HOST_IPC_FIFO_INT, status)) + ivpu_ipc_irq_handler(vdev); + if (REG_TEST_FLD(MTL_VPU_HOST_SS_ICB_STATUS_0, MMU_IRQ_1_INT, status)) ivpu_dbg(vdev, IRQ, "MMU sync complete\n"); diff --git a/drivers/accel/ivpu/ivpu_ipc.c b/drivers/accel/ivpu/ivpu_ipc.c new file mode 100644 index 000000000000..c756476bfc5b --- /dev/null +++ b/drivers/accel/ivpu/ivpu_ipc.c @@ -0,0 +1,500 @@ +// SPDX-License-Identifier: GPL-2.0-only +/* + * Copyright (C) 2020-2023 Intel Corporation + */ + +#include <linux/genalloc.h> +#include <linux/highmem.h> +#include <linux/kthread.h> +#include <linux/wait.h> + +#include "ivpu_drv.h" +#include "ivpu_gem.h" +#include "ivpu_hw.h" +#include "ivpu_hw_reg_io.h" +#include "ivpu_ipc.h" +#include "ivpu_jsm_msg.h" + +#define IPC_MAX_RX_MSG 128 +#define IS_KTHREAD() (get_current()->flags & PF_KTHREAD) + +struct ivpu_ipc_tx_buf { + struct ivpu_ipc_hdr ipc; + struct vpu_jsm_msg jsm; +}; + +struct ivpu_ipc_rx_msg { + struct list_head link; + struct ivpu_ipc_hdr *ipc_hdr; + struct vpu_jsm_msg *jsm_msg; +}; + +static void ivpu_ipc_msg_dump(struct ivpu_device *vdev, char *c, + struct ivpu_ipc_hdr *ipc_hdr, u32 vpu_addr) +{ + ivpu_dbg(vdev, IPC, + "%s: vpu:0x%x (data_addr:0x%08x, data_size:0x%x, channel:0x%x, src_node:0x%x, dst_node:0x%x, status:0x%x)", + c, vpu_addr, ipc_hdr->data_addr, ipc_hdr->data_size, ipc_hdr->channel, + ipc_hdr->src_node, ipc_hdr->dst_node, ipc_hdr->status); +} + +static void ivpu_jsm_msg_dump(struct ivpu_device *vdev, char *c, + struct vpu_jsm_msg *jsm_msg, u32 vpu_addr) +{ + u32 *payload = (u32 *)&jsm_msg->payload; + + ivpu_dbg(vdev, JSM, + "%s: vpu:0x%08x (type:0x%x, status:0x%x, id: 0x%x, result: 0x%x, payload:0x%x 0x%x 0x%x 0x%x 0x%x)\n", + c, vpu_addr, jsm_msg->type, jsm_msg->status, jsm_msg->request_id, jsm_msg->result, + payload[0], payload[1], payload[2], payload[3], payload[4]); +} + +static void +ivpu_ipc_rx_mark_free(struct ivpu_device *vdev, struct ivpu_ipc_hdr *ipc_hdr, + struct vpu_jsm_msg *jsm_msg) +{ + ipc_hdr->status = IVPU_IPC_HDR_FREE; + if (jsm_msg) + jsm_msg->status = VPU_JSM_MSG_FREE; + wmb(); /* Flush WC buffers for message statuses */ +} + +static void ivpu_ipc_mem_fini(struct ivpu_device *vdev) +{ + struct ivpu_ipc_info *ipc = vdev->ipc; + + ivpu_bo_free_internal(ipc->mem_rx); + ivpu_bo_free_internal(ipc->mem_tx); +} + +static int +ivpu_ipc_tx_prepare(struct ivpu_device *vdev, struct ivpu_ipc_consumer *cons, + struct vpu_jsm_msg *req) +{ + struct ivpu_ipc_info *ipc = vdev->ipc; + struct ivpu_ipc_tx_buf *tx_buf; + u32 tx_buf_vpu_addr; + u32 jsm_vpu_addr; + + tx_buf_vpu_addr = gen_pool_alloc(ipc->mm_tx, sizeof(*tx_buf)); + if (!tx_buf_vpu_addr) { + ivpu_err(vdev, "Failed to reserve IPC buffer, size %ld\n", + sizeof(*tx_buf)); + return -ENOMEM; + } + + tx_buf = ivpu_to_cpu_addr(ipc->mem_tx, tx_buf_vpu_addr); + if (drm_WARN_ON(&vdev->drm, !tx_buf)) { + gen_pool_free(ipc->mm_tx, tx_buf_vpu_addr, sizeof(*tx_buf)); + return -EIO; + } + + jsm_vpu_addr = tx_buf_vpu_addr + offsetof(struct ivpu_ipc_tx_buf, jsm); + + if (tx_buf->ipc.status != IVPU_IPC_HDR_FREE) + ivpu_warn(vdev, "IPC message vpu:0x%x not released by firmware\n", + tx_buf_vpu_addr); + + if (tx_buf->jsm.status != VPU_JSM_MSG_FREE) + ivpu_warn(vdev, "JSM message vpu:0x%x not released by firmware\n", + jsm_vpu_addr); + + memset(tx_buf, 0, sizeof(*tx_buf)); + tx_buf->ipc.data_addr = jsm_vpu_addr; + /* TODO: Set data_size to actual JSM message size, not union of all messages */ + tx_buf->ipc.data_size = sizeof(*req); + tx_buf->ipc.channel = cons->channel; + tx_buf->ipc.src_node = 0; + tx_buf->ipc.dst_node = 1; + tx_buf->ipc.status = IVPU_IPC_HDR_ALLOCATED; + tx_buf->jsm.type = req->type; + tx_buf->jsm.status = VPU_JSM_MSG_ALLOCATED; + tx_buf->jsm.payload = req->payload; + + req->request_id = atomic_inc_return(&ipc->request_id); + tx_buf->jsm.request_id = req->request_id; + cons->request_id = req->request_id; + wmb(); /* Flush WC buffers for IPC, JSM msgs */ + + cons->tx_vpu_addr = tx_buf_vpu_addr; + + ivpu_jsm_msg_dump(vdev, "TX", &tx_buf->jsm, jsm_vpu_addr); + ivpu_ipc_msg_dump(vdev, "TX", &tx_buf->ipc, tx_buf_vpu_addr); + + return 0; +} + +static void ivpu_ipc_tx_release(struct ivpu_device *vdev, u32 vpu_addr) +{ + struct ivpu_ipc_info *ipc = vdev->ipc; + + if (vpu_addr) + gen_pool_free(ipc->mm_tx, vpu_addr, sizeof(struct ivpu_ipc_tx_buf)); +} + +static void ivpu_ipc_tx(struct ivpu_device *vdev, u32 vpu_addr) +{ + ivpu_hw_reg_ipc_tx_set(vdev, vpu_addr); +} + +void +ivpu_ipc_consumer_add(struct ivpu_device *vdev, struct ivpu_ipc_consumer *cons, u32 channel) +{ + struct ivpu_ipc_info *ipc = vdev->ipc; + + INIT_LIST_HEAD(&cons->link); + cons->channel = channel; + cons->tx_vpu_addr = 0; + cons->request_id = 0; + spin_lock_init(&cons->rx_msg_lock); + INIT_LIST_HEAD(&cons->rx_msg_list); + init_waitqueue_head(&cons->rx_msg_wq); + + spin_lock_irq(&ipc->cons_list_lock); + list_add_tail(&cons->link, &ipc->cons_list); + spin_unlock_irq(&ipc->cons_list_lock); +} + +void ivpu_ipc_consumer_del(struct ivpu_device *vdev, struct ivpu_ipc_consumer *cons) +{ + struct ivpu_ipc_info *ipc = vdev->ipc; + struct ivpu_ipc_rx_msg *rx_msg, *r; + + spin_lock_irq(&ipc->cons_list_lock); + list_del(&cons->link); + spin_unlock_irq(&ipc->cons_list_lock); + + spin_lock_irq(&cons->rx_msg_lock); + list_for_each_entry_safe(rx_msg, r, &cons->rx_msg_list, link) { + list_del(&rx_msg->link); + ivpu_ipc_rx_mark_free(vdev, rx_msg->ipc_hdr, rx_msg->jsm_msg); + atomic_dec(&ipc->rx_msg_count); + kfree(rx_msg); + } + spin_unlock_irq(&cons->rx_msg_lock); + + ivpu_ipc_tx_release(vdev, cons->tx_vpu_addr); +} + +static int +ivpu_ipc_send(struct ivpu_device *vdev, struct ivpu_ipc_consumer *cons, struct vpu_jsm_msg *req) +{ + struct ivpu_ipc_info *ipc = vdev->ipc; + int ret; + + ret = mutex_lock_interruptible(&ipc->lock); + if (ret) + return ret; + + if (!ipc->on) { + ret = -EAGAIN; + goto unlock; + } + + ret = ivpu_ipc_tx_prepare(vdev, cons, req); + if (ret) + goto unlock; + + ivpu_ipc_tx(vdev, cons->tx_vpu_addr); + +unlock: + mutex_unlock(&ipc->lock); + return ret; +} + +int ivpu_ipc_receive(struct ivpu_device *vdev, struct ivpu_ipc_consumer *cons, + struct ivpu_ipc_hdr *ipc_buf, + struct vpu_jsm_msg *ipc_payload, unsigned long timeout_ms) +{ + struct ivpu_ipc_info *ipc = vdev->ipc; + struct ivpu_ipc_rx_msg *rx_msg; + int wait_ret, ret = 0; + + wait_ret = wait_event_interruptible_timeout(cons->rx_msg_wq, + (IS_KTHREAD() && kthread_should_stop()) || + !list_empty(&cons->rx_msg_list), + msecs_to_jiffies(timeout_ms)); + + if (IS_KTHREAD() && kthread_should_stop()) + return -EINTR; + + if (wait_ret == 0) + return -ETIMEDOUT; + + if (wait_ret < 0) + return -ERESTARTSYS; + + spin_lock_irq(&cons->rx_msg_lock); + rx_msg = list_first_entry_or_null(&cons->rx_msg_list, struct ivpu_ipc_rx_msg, link); + if (!rx_msg) { + spin_unlock_irq(&cons->rx_msg_lock); + return -EAGAIN; + } + list_del(&rx_msg->link); + spin_unlock_irq(&cons->rx_msg_lock); + + if (ipc_buf) + memcpy(ipc_buf, rx_msg->ipc_hdr, sizeof(*ipc_buf)); + if (rx_msg->jsm_msg) { + u32 size = min_t(int, rx_msg->ipc_hdr->data_size, sizeof(*ipc_payload)); + + if (rx_msg->jsm_msg->result != VPU_JSM_STATUS_SUCCESS) { + ivpu_dbg(vdev, IPC, "IPC resp result error: %d\n", rx_msg->jsm_msg->result); + ret = -EBADMSG; + } + + if (ipc_payload) + memcpy(ipc_payload, rx_msg->jsm_msg, size); + } + + ivpu_ipc_rx_mark_free(vdev, rx_msg->ipc_hdr, rx_msg->jsm_msg); + atomic_dec(&ipc->rx_msg_count); + kfree(rx_msg); + + return ret; +} + +static int +ivpu_ipc_send_receive_internal(struct ivpu_device *vdev, struct vpu_jsm_msg *req, + enum vpu_ipc_msg_type expected_resp_type, + struct vpu_jsm_msg *resp, u32 channel, + unsigned long timeout_ms) +{ + struct ivpu_ipc_consumer cons; + int ret; + + ivpu_ipc_consumer_add(vdev, &cons, channel); + + ret = ivpu_ipc_send(vdev, &cons, req); + if (ret) { + ivpu_warn(vdev, "IPC send failed: %d\n", ret); + goto consumer_del; + } + + ret = ivpu_ipc_receive(vdev, &cons, NULL, resp, timeout_ms); + if (ret) { + ivpu_warn(vdev, "IPC receive failed: type 0x%x, ret %d\n", req->type, ret); + goto consumer_del; + } + + if (resp->type != expected_resp_type) { + ivpu_warn(vdev, "Invalid JSM response type: 0x%x\n", resp->type); + ret = -EBADE; + } + +consumer_del: + ivpu_ipc_consumer_del(vdev, &cons); + return ret; +} + +int ivpu_ipc_send_receive(struct ivpu_device *vdev, struct vpu_jsm_msg *req, + enum vpu_ipc_msg_type expected_resp_type, + struct vpu_jsm_msg *resp, u32 channel, + unsigned long timeout_ms) +{ + struct vpu_jsm_msg hb_req = { .type = VPU_JSM_MSG_QUERY_ENGINE_HB }; + struct vpu_jsm_msg hb_resp; + int ret; + + ret = ivpu_ipc_send_receive_internal(vdev, req, expected_resp_type, resp, + channel, timeout_ms); + if (ret != -ETIMEDOUT) + return ret; + + ret = ivpu_ipc_send_receive_internal(vdev, &hb_req, VPU_JSM_MSG_QUERY_ENGINE_HB_DONE, + &hb_resp, VPU_IPC_CHAN_ASYNC_CMD, vdev->timeout.jsm); + if (ret == -ETIMEDOUT) + ivpu_hw_diagnose_failure(vdev); + + return ret; +} + +static bool +ivpu_ipc_match_consumer(struct ivpu_device *vdev, struct ivpu_ipc_consumer *cons, + struct ivpu_ipc_hdr *ipc_hdr, struct vpu_jsm_msg *jsm_msg) +{ + if (cons->channel != ipc_hdr->channel) + return false; + + if (!jsm_msg || jsm_msg->request_id == cons->request_id) + return true; + + return false; +} + +static void +ivpu_ipc_dispatch(struct ivpu_device *vdev, struct ivpu_ipc_consumer *cons, + struct ivpu_ipc_hdr *ipc_hdr, struct vpu_jsm_msg *jsm_msg) +{ + struct ivpu_ipc_info *ipc = vdev->ipc; + struct ivpu_ipc_rx_msg *rx_msg; + unsigned long flags; + + lockdep_assert_held(&ipc->cons_list_lock); + + rx_msg = kzalloc(sizeof(*rx_msg), GFP_ATOMIC); + if (!rx_msg) { + ivpu_ipc_rx_mark_free(vdev, ipc_hdr, jsm_msg); + return; + } + + atomic_inc(&ipc->rx_msg_count); + + rx_msg->ipc_hdr = ipc_hdr; + rx_msg->jsm_msg = jsm_msg; + + spin_lock_irqsave(&cons->rx_msg_lock, flags); + list_add_tail(&rx_msg->link, &cons->rx_msg_list); + spin_unlock_irqrestore(&cons->rx_msg_lock, flags); + + wake_up(&cons->rx_msg_wq); +} + +int ivpu_ipc_irq_handler(struct ivpu_device *vdev) +{ + struct ivpu_ipc_info *ipc = vdev->ipc; + struct ivpu_ipc_consumer *cons; + struct ivpu_ipc_hdr *ipc_hdr; + struct vpu_jsm_msg *jsm_msg; + unsigned long flags; + bool dispatched; + u32 vpu_addr; + + /* + * Driver needs to purge all messages from IPC FIFO to clear IPC interrupt. + * Without purge IPC FIFO to 0 next IPC interrupts won't be generated. + */ + while (ivpu_hw_reg_ipc_rx_count_get(vdev)) { + vpu_addr = ivpu_hw_reg_ipc_rx_addr_get(vdev); + if (vpu_addr == REG_IO_ERROR) { + ivpu_err(vdev, "Failed to read IPC rx addr register\n"); + return -EIO; + } + + ipc_hdr = ivpu_to_cpu_addr(ipc->mem_rx, vpu_addr); + if (!ipc_hdr) { + ivpu_warn(vdev, "IPC msg 0x%x out of range\n", vpu_addr); + continue; + } + ivpu_ipc_msg_dump(vdev, "RX", ipc_hdr, vpu_addr); + + jsm_msg = NULL; + if (ipc_hdr->channel != IVPU_IPC_CHAN_BOOT_MSG) { + jsm_msg = ivpu_to_cpu_addr(ipc->mem_rx, ipc_hdr->data_addr); + if (!jsm_msg) { + ivpu_warn(vdev, "JSM msg 0x%x out of range\n", ipc_hdr->data_addr); + ivpu_ipc_rx_mark_free(vdev, ipc_hdr, NULL); + continue; + } + ivpu_jsm_msg_dump(vdev, "RX", jsm_msg, ipc_hdr->data_addr); + } + + if (atomic_read(&ipc->rx_msg_count) > IPC_MAX_RX_MSG) { + ivpu_warn(vdev, "IPC RX msg dropped, msg count %d\n", IPC_MAX_RX_MSG); + ivpu_ipc_rx_mark_free(vdev, ipc_hdr, jsm_msg); + continue; + } + + dispatched = false; + spin_lock_irqsave(&ipc->cons_list_lock, flags); + list_for_each_entry(cons, &ipc->cons_list, link) { + if (ivpu_ipc_match_consumer(vdev, cons, ipc_hdr, jsm_msg)) { + ivpu_ipc_dispatch(vdev, cons, ipc_hdr, jsm_msg); + dispatched = true; + break; + } + } + spin_unlock_irqrestore(&ipc->cons_list_lock, flags); + + if (!dispatched) { + ivpu_dbg(vdev, IPC, "IPC RX msg 0x%x dropped (no consumer)\n", vpu_addr); + ivpu_ipc_rx_mark_free(vdev, ipc_hdr, jsm_msg); + } + } + + return 0; +} + +int ivpu_ipc_init(struct ivpu_device *vdev) +{ + struct ivpu_ipc_info *ipc = vdev->ipc; + int ret = -ENOMEM; + + ipc->mem_tx = ivpu_bo_alloc_internal(vdev, 0, SZ_16K, DRM_IVPU_BO_WC); + if (!ipc->mem_tx) + return ret; + + ipc->mem_rx = ivpu_bo_alloc_internal(vdev, 0, SZ_16K, DRM_IVPU_BO_WC); + if (!ipc->mem_rx) + goto err_free_tx; + + ipc->mm_tx = devm_gen_pool_create(vdev->drm.dev, __ffs(IVPU_IPC_ALIGNMENT), + -1, "TX_IPC_JSM"); + if (IS_ERR(ipc->mm_tx)) { + ret = PTR_ERR(ipc->mm_tx); + ivpu_err(vdev, "Failed to create gen pool, %pe\n", ipc->mm_tx); + goto err_free_rx; + } + + ret = gen_pool_add(ipc->mm_tx, ipc->mem_tx->vpu_addr, ipc->mem_tx->base.size, -1); + if (ret) { + ivpu_err(vdev, "gen_pool_add failed, ret %d\n", ret); + goto err_free_rx; + } + + INIT_LIST_HEAD(&ipc->cons_list); + spin_lock_init(&ipc->cons_list_lock); + drmm_mutex_init(&vdev->drm, &ipc->lock); + + ivpu_ipc_reset(vdev); + return 0; + +err_free_rx: + ivpu_bo_free_internal(ipc->mem_rx); +err_free_tx: + ivpu_bo_free_internal(ipc->mem_tx); + return ret; +} + +void ivpu_ipc_fini(struct ivpu_device *vdev) +{ + ivpu_ipc_mem_fini(vdev); +} + +void ivpu_ipc_enable(struct ivpu_device *vdev) +{ + struct ivpu_ipc_info *ipc = vdev->ipc; + + mutex_lock(&ipc->lock); + ipc->on = true; + mutex_unlock(&ipc->lock); +} + +void ivpu_ipc_disable(struct ivpu_device *vdev) +{ + struct ivpu_ipc_info *ipc = vdev->ipc; + struct ivpu_ipc_consumer *cons, *c; + unsigned long flags; + + mutex_lock(&ipc->lock); + ipc->on = false; + mutex_unlock(&ipc->lock); + + spin_lock_irqsave(&ipc->cons_list_lock, flags); + list_for_each_entry_safe(cons, c, &ipc->cons_list, link) + wake_up(&cons->rx_msg_wq); + spin_unlock_irqrestore(&ipc->cons_list_lock, flags); +} + +void ivpu_ipc_reset(struct ivpu_device *vdev) +{ + struct ivpu_ipc_info *ipc = vdev->ipc; + + mutex_lock(&ipc->lock); + + memset(ipc->mem_tx->kvaddr, 0, ipc->mem_tx->base.size); + memset(ipc->mem_rx->kvaddr, 0, ipc->mem_rx->base.size); + wmb(); /* Flush WC buffers for TX and RX rings */ + + mutex_unlock(&ipc->lock); +} diff --git a/drivers/accel/ivpu/ivpu_ipc.h b/drivers/accel/ivpu/ivpu_ipc.h new file mode 100644 index 000000000000..9838202ecfad --- /dev/null +++ b/drivers/accel/ivpu/ivpu_ipc.h @@ -0,0 +1,93 @@ +/* SPDX-License-Identifier: GPL-2.0-only */ +/* + * Copyright (C) 2020-2023 Intel Corporation + */ + +#ifndef __IVPU_IPC_H__ +#define __IVPU_IPC_H__ + +#include <linux/interrupt.h> +#include <linux/spinlock.h> + +#include "vpu_jsm_api.h" + +struct ivpu_bo; + +/* VPU FW boot notification */ +#define IVPU_IPC_CHAN_BOOT_MSG 0x3ff +#define IVPU_IPC_BOOT_MSG_DATA_ADDR 0x424f4f54 + +/* The alignment to be used for IPC Buffers and IPC Data. */ +#define IVPU_IPC_ALIGNMENT 64 + +#define IVPU_IPC_HDR_FREE 0 +#define IVPU_IPC_HDR_ALLOCATED 0 + +/** + * struct ivpu_ipc_hdr - The IPC message header structure, exchanged + * with the VPU device firmware. + * @data_addr: The VPU address of the payload (JSM message) + * @data_size: The size of the payload. + * @channel: The channel used. + * @src_node: The Node ID of the sender. + * @dst_node: The Node ID of the intended receiver. + * @status: IPC buffer usage status + */ +struct ivpu_ipc_hdr { + u32 data_addr; + u32 data_size; + u16 channel; + u8 src_node; + u8 dst_node; + u8 status; +} __packed __aligned(IVPU_IPC_ALIGNMENT); + +struct ivpu_ipc_consumer { + struct list_head link; + u32 channel; + u32 tx_vpu_addr; + u32 request_id; + + spinlock_t rx_msg_lock; /* Protects rx_msg_list */ + struct list_head rx_msg_list; + wait_queue_head_t rx_msg_wq; +}; + +struct ivpu_ipc_info { + struct gen_pool *mm_tx; + struct ivpu_bo *mem_tx; + struct ivpu_bo *mem_rx; + + atomic_t rx_msg_count; + + spinlock_t cons_list_lock; /* Protects cons_list */ + struct list_head cons_list; + + atomic_t request_id; + struct mutex lock; /* Lock on status */ + bool on; +}; + +int ivpu_ipc_init(struct ivpu_device *vdev); +void ivpu_ipc_fini(struct ivpu_device *vdev); + +void ivpu_ipc_enable(struct ivpu_device *vdev); +void ivpu_ipc_disable(struct ivpu_device *vdev); +void ivpu_ipc_reset(struct ivpu_device *vdev); + +int ivpu_ipc_irq_handler(struct ivpu_device *vdev); + +void ivpu_ipc_consumer_add(struct ivpu_device *vdev, struct ivpu_ipc_consumer *cons, + u32 channel); +void ivpu_ipc_consumer_del(struct ivpu_device *vdev, struct ivpu_ipc_consumer *cons); + +int ivpu_ipc_receive(struct ivpu_device *vdev, struct ivpu_ipc_consumer *cons, + struct ivpu_ipc_hdr *ipc_buf, struct vpu_jsm_msg *ipc_payload, + unsigned long timeout_ms); + +int ivpu_ipc_send_receive(struct ivpu_device *vdev, struct vpu_jsm_msg *req, + enum vpu_ipc_msg_type expected_resp_type, + struct vpu_jsm_msg *resp, u32 channel, + unsigned long timeout_ms); + +#endif /* __IVPU_IPC_H__ */ diff --git a/drivers/accel/ivpu/ivpu_jsm_msg.c b/drivers/accel/ivpu/ivpu_jsm_msg.c new file mode 100644 index 000000000000..af77dafac97e --- /dev/null +++ b/drivers/accel/ivpu/ivpu_jsm_msg.c @@ -0,0 +1,169 @@ +// SPDX-License-Identifier: GPL-2.0-only +/* + * Copyright (C) 2020-2023 Intel Corporation + */ + +#include "ivpu_drv.h" +#include "ivpu_ipc.h" +#include "ivpu_jsm_msg.h" + +int ivpu_jsm_register_db(struct ivpu_device *vdev, u32 ctx_id, u32 db_id, + u64 jobq_base, u32 jobq_size) +{ + struct vpu_jsm_msg req = { .type = VPU_JSM_MSG_REGISTER_DB }; + struct vpu_jsm_msg resp; + int ret = 0; + + req.payload.register_db.db_idx = db_id; + req.payload.register_db.jobq_base = jobq_base; + req.payload.register_db.jobq_size = jobq_size; + req.payload.register_db.host_ssid = ctx_id; + + ret = ivpu_ipc_send_receive(vdev, &req, VPU_JSM_MSG_REGISTER_DB_DONE, &resp, + VPU_IPC_CHAN_ASYNC_CMD, vdev->timeout.jsm); + if (ret) { + ivpu_err(vdev, "Failed to register doorbell %d: %d\n", db_id, ret); + return ret; + } + + ivpu_dbg(vdev, JSM, "Doorbell %d registered to context %d\n", db_id, ctx_id); + + return 0; +} + +int ivpu_jsm_unregister_db(struct ivpu_device *vdev, u32 db_id) +{ + struct vpu_jsm_msg req = { .type = VPU_JSM_MSG_UNREGISTER_DB }; + struct vpu_jsm_msg resp; + int ret = 0; + + req.payload.unregister_db.db_idx = db_id; + + ret = ivpu_ipc_send_receive(vdev, &req, VPU_JSM_MSG_UNREGISTER_DB_DONE, &resp, + VPU_IPC_CHAN_ASYNC_CMD, vdev->timeout.jsm); + if (ret) { + ivpu_warn(vdev, "Failed to unregister doorbell %d: %d\n", db_id, ret); + return ret; + } + + ivpu_dbg(vdev, JSM, "Doorbell %d unregistered\n", db_id); + + return 0; +} + +int ivpu_jsm_get_heartbeat(struct ivpu_device *vdev, u32 engine, u64 *heartbeat) +{ + struct vpu_jsm_msg req = { .type = VPU_JSM_MSG_QUERY_ENGINE_HB }; + struct vpu_jsm_msg resp; + int ret; + + if (engine > VPU_ENGINE_COPY) + return -EINVAL; + + req.payload.query_engine_hb.engine_idx = engine; + + ret = ivpu_ipc_send_receive(vdev, &req, VPU_JSM_MSG_QUERY_ENGINE_HB_DONE, &resp, + VPU_IPC_CHAN_ASYNC_CMD, vdev->timeout.jsm); + if (ret) { + ivpu_err(vdev, "Failed to get heartbeat from engine %d: %d\n", engine, ret); + return ret; + } + + *heartbeat = resp.payload.query_engine_hb_done.heartbeat; + return ret; +} + +int ivpu_jsm_reset_engine(struct ivpu_device *vdev, u32 engine) +{ + struct vpu_jsm_msg req = { .type = VPU_JSM_MSG_ENGINE_RESET }; + struct vpu_jsm_msg resp; + int ret; + + if (engine > VPU_ENGINE_COPY) + return -EINVAL; + + req.payload.engine_reset.engine_idx = engine; + + ret = ivpu_ipc_send_receive(vdev, &req, VPU_JSM_MSG_ENGINE_RESET_DONE, &resp, + VPU_IPC_CHAN_ASYNC_CMD, vdev->timeout.jsm); + if (ret) + ivpu_err(vdev, "Failed to reset engine %d: %d\n", engine, ret); + + return ret; +} + +int ivpu_jsm_preempt_engine(struct ivpu_device *vdev, u32 engine, u32 preempt_id) +{ + struct vpu_jsm_msg req = { .type = VPU_JSM_MSG_ENGINE_PREEMPT }; + struct vpu_jsm_msg resp; + int ret; + + if (engine > VPU_ENGINE_COPY) + return -EINVAL; + + req.payload.engine_preempt.engine_idx = engine; + req.payload.engine_preempt.preempt_id = preempt_id; + + ret = ivpu_ipc_send_receive(vdev, &req, VPU_JSM_MSG_ENGINE_PREEMPT_DONE, &resp, + VPU_IPC_CHAN_ASYNC_CMD, vdev->timeout.jsm); + if (ret) + ivpu_err(vdev, "Failed to preempt engine %d: %d\n", engine, ret); + + return ret; +} + +int ivpu_jsm_dyndbg_control(struct ivpu_device *vdev, char *command, size_t size) +{ + struct vpu_jsm_msg req = { .type = VPU_JSM_MSG_DYNDBG_CONTROL }; + struct vpu_jsm_msg resp; + int ret; + + if (!strncpy(req.payload.dyndbg_control.dyndbg_cmd, command, VPU_DYNDBG_CMD_MAX_LEN - 1)) + return -ENOMEM; + + ret = ivpu_ipc_send_receive(vdev, &req, VPU_JSM_MSG_DYNDBG_CONTROL_RSP, &resp, + VPU_IPC_CHAN_ASYNC_CMD, vdev->timeout.jsm); + if (ret) + ivpu_warn(vdev, "Failed to send command \"%s\": ret %d\n", command, ret); + + return ret; +} + +int ivpu_jsm_trace_get_capability(struct ivpu_device *vdev, u32 *trace_destination_mask, + u64 *trace_hw_component_mask) +{ + struct vpu_jsm_msg req = { .type = VPU_JSM_MSG_TRACE_GET_CAPABILITY }; + struct vpu_jsm_msg resp; + int ret; + + ret = ivpu_ipc_send_receive(vdev, &req, VPU_JSM_MSG_TRACE_GET_CAPABILITY_RSP, &resp, + VPU_IPC_CHAN_ASYNC_CMD, vdev->timeout.jsm); + if (ret) { + ivpu_warn(vdev, "Failed to get trace capability: %d\n", ret); + return ret; + } + + *trace_destination_mask = resp.payload.trace_capability.trace_destination_mask; + *trace_hw_component_mask = resp.payload.trace_capability.trace_hw_component_mask; + + return ret; +} + +int ivpu_jsm_trace_set_config(struct ivpu_device *vdev, u32 trace_level, u32 trace_destination_mask, + u64 trace_hw_component_mask) +{ + struct vpu_jsm_msg req = { .type = VPU_JSM_MSG_TRACE_SET_CONFIG }; + struct vpu_jsm_msg resp; + int ret; + + req.payload.trace_config.trace_level = trace_level; + req.payload.trace_config.trace_destination_mask = trace_destination_mask; + req.payload.trace_config.trace_hw_component_mask = trace_hw_component_mask; + + ret = ivpu_ipc_send_receive(vdev, &req, VPU_JSM_MSG_TRACE_SET_CONFIG_RSP, &resp, + VPU_IPC_CHAN_ASYNC_CMD, vdev->timeout.jsm); + if (ret) + ivpu_warn(vdev, "Failed to set config: %d\n", ret); + + return ret; +} diff --git a/drivers/accel/ivpu/ivpu_jsm_msg.h b/drivers/accel/ivpu/ivpu_jsm_msg.h new file mode 100644 index 000000000000..1a3e2e2740bd --- /dev/null +++ b/drivers/accel/ivpu/ivpu_jsm_msg.h @@ -0,0 +1,23 @@ +/* SPDX-License-Identifier: GPL-2.0-only */ +/* + * Copyright (C) 2020-2023 Intel Corporation + */ + +#ifndef __IVPU_JSM_MSG_H__ +#define __IVPU_JSM_MSG_H__ + +#include "vpu_jsm_api.h" + +int ivpu_jsm_register_db(struct ivpu_device *vdev, u32 ctx_id, u32 db_id, + u64 jobq_base, u32 jobq_size); +int ivpu_jsm_unregister_db(struct ivpu_device *vdev, u32 db_id); +int ivpu_jsm_get_heartbeat(struct ivpu_device *vdev, u32 engine, u64 *heartbeat); +int ivpu_jsm_reset_engine(struct ivpu_device *vdev, u32 engine); +int ivpu_jsm_preempt_engine(struct ivpu_device *vdev, u32 engine, u32 preempt_id); +int ivpu_jsm_dyndbg_control(struct ivpu_device *vdev, char *command, size_t size); +int ivpu_jsm_trace_get_capability(struct ivpu_device *vdev, u32 *trace_destination_mask, + u64 *trace_hw_component_mask); +int ivpu_jsm_trace_set_config(struct ivpu_device *vdev, u32 trace_level, u32 trace_destination_mask, + u64 trace_hw_component_mask); + +#endif diff --git a/drivers/accel/ivpu/vpu_jsm_api.h b/drivers/accel/ivpu/vpu_jsm_api.h new file mode 100644 index 000000000000..1096cab0334e --- /dev/null +++ b/drivers/accel/ivpu/vpu_jsm_api.h @@ -0,0 +1,999 @@ +/* SPDX-License-Identifier: MIT */ +/* + * Copyright (C) 2020-2023 Intel Corporation + */ + +/** + * @file + * @brief JSM shared definitions + * + * @ingroup Jsm + * @brief JSM shared definitions + * @{ + */ +#ifndef VPU_JSM_API_H +#define VPU_JSM_API_H + +/* + * Major version changes that break backward compatibility + */ +#define VPU_JSM_API_VER_MAJOR 2 + +/* + * Minor version changes when API backward compatibility is preserved. + */ +#define VPU_JSM_API_VER_MINOR 10 + +/* + * API header changed (field names, documentation, formatting) but API itself has not been changed + */ +#define VPU_JSM_API_VER_PATCH 1 + +/* + * Index in the API version table + */ +#define VPU_JSM_API_VER_INDEX 4 + +/* + * Number of Priority Bands for Hardware Scheduling + * Bands: RealTime, Focus, Normal, Idle + */ +#define VPU_HWS_NUM_PRIORITY_BANDS 4 + +/* Max number of impacted contexts that can be dealt with the engine reset command */ +#define VPU_MAX_ENGINE_RESET_IMPACTED_CONTEXTS 3 + +/** Pack the API structures for now, once alignment issues are fixed this can be removed */ +#pragma pack(push, 1) + +/* + * Engine indexes. + */ +#define VPU_ENGINE_COMPUTE 0 +#define VPU_ENGINE_COPY 1 +#define VPU_ENGINE_NB 2 + +/* + * VPU status values. + */ +#define VPU_JSM_STATUS_SUCCESS 0x0U +#define VPU_JSM_STATUS_PARSING_ERR 0x1U +#define VPU_JSM_STATUS_PROCESSING_ERR 0x2U +#define VPU_JSM_STATUS_PREEMPTED 0x3U +#define VPU_JSM_STATUS_ABORTED 0x4U +#define VPU_JSM_STATUS_USER_CTX_VIOL_ERR 0x5U +#define VPU_JSM_STATUS_GLOBAL_CTX_VIOL_ERR 0x6U +#define VPU_JSM_STATUS_MVNCI_WRONG_INPUT_FORMAT 0x7U +#define VPU_JSM_STATUS_MVNCI_UNSUPPORTED_NETWORK_ELEMENT 0x8U +#define VPU_JSM_STATUS_MVNCI_INVALID_HANDLE 0x9U +#define VPU_JSM_STATUS_MVNCI_OUT_OF_RESOURCES 0xAU +#define VPU_JSM_STATUS_MVNCI_NOT_IMPLEMENTED 0xBU +#define VPU_JSM_STATUS_MVNCI_INTERNAL_ERROR 0xCU +/* Job status returned when the job was preempted mid-inference */ +#define VPU_JSM_STATUS_PREEMPTED_MID_INFERENCE 0xDU + +/* + * Host <-> VPU IPC channels. + * ASYNC commands use a high priority channel, other messages use low-priority ones. + */ +#define VPU_IPC_CHAN_ASYNC_CMD 0 +#define VPU_IPC_CHAN_GEN_CMD 10 +#define VPU_IPC_CHAN_ |
