| Message ID | 20210816105934.28265-1-irui.wang@mediatek.com |
|---|---|
| Headers | show |
| Series | Enable two H264 encoder cores on MT8195 | expand |
On Mon, Aug 16, 2021 at 06:59:27PM +0800, Irui Wang wrote: > +static struct component_match *mtk_venc_match_add(struct mtk_vcodec_dev *dev) > +{ > + struct platform_device *pdev = dev->plat_dev; > + struct component_match *match = NULL; > + int i; > + > + for (i = 0; i < ARRAY_SIZE(mtk_venc_comp_ids); i++) { > + enum mtk_venc_hw_id comp_idx; > + struct device_node *comp_node; > + const struct of_device_id *of_id; To be neat, prefer to define the variables outside of the loop (i.e. at the beginning of the function). > + > + comp_node = of_find_compatible_node(NULL, NULL, > + mtk_venc_comp_ids[i].compatible); > + if (!comp_node) > + continue; > + > + of_id = of_match_node(mtk_venc_comp_ids, comp_node); > + if (!of_id) { > + dev_err(&pdev->dev, "Failed to get match node\n"); Need to call of_node_put() actually, but see comment below. > + return ERR_PTR(-EINVAL); > + } > + > + comp_idx = (enum mtk_venc_hw_id)of_id->data; For getting the comp_idx, mtk_venc_comp_ids[i].data should be sufficient. If so, of_match_node() can be removed so that the error handling path won't need to call of_node_put(). > @@ -239,6 +314,7 @@ static int mtk_vcodec_probe(struct platform_device *pdev) > phandle rproc_phandle; > enum mtk_vcodec_fw_type fw_type; > int ret; > + struct component_match *match = NULL; It doesn't need to be initialized. > - res = platform_get_resource(pdev, IORESOURCE_IRQ, 0); > - if (res == NULL) { > - dev_err(&pdev->dev, "failed to get irq resource"); > - ret = -ENOENT; > - goto err_res; > - } > + res = platform_get_resource(pdev, IORESOURCE_IRQ, 0); > + if (!res) { > + dev_err(&pdev->dev, "failed to get irq resource"); > + ret = -ENOENT; > + goto err_res; > + } res is not used. Can be removed in next version or in another patch. > diff --git a/drivers/media/platform/mtk-vcodec/mtk_vcodec_enc_hw.c b/drivers/media/platform/mtk-vcodec/mtk_vcodec_enc_hw.c > new file mode 100644 > index 000000000000..4e6a8a81ff67 > --- /dev/null > +++ b/drivers/media/platform/mtk-vcodec/mtk_vcodec_enc_hw.c > @@ -0,0 +1,179 @@ > +// SPDX-License-Identifier: GPL-2.0 > +/* > + * Copyright (c) 2021 MediaTek Inc. > + */ > + > +#include <linux/pm_runtime.h> > +#include <linux/interrupt.h> > +#include <linux/irq.h> > +#include <linux/of_platform.h> > +#include <linux/module.h> Would be better to maintain an order. > +#include "mtk_vcodec_enc_hw.h" > +#include "mtk_vcodec_enc.h" Would be better to maintain an order. > +static irqreturn_t mtk_enc_comp_irq_handler(int irq, void *priv) > +{ > + struct mtk_venc_comp_dev *dev = priv; > + struct mtk_vcodec_ctx *ctx; > + unsigned long flags; > + void __iomem *addr; > + > + spin_lock_irqsave(&dev->master_dev->irqlock, flags); > + ctx = dev->curr_ctx; > + spin_unlock_irqrestore(&dev->master_dev->irqlock, flags); > + if (!ctx) > + return IRQ_HANDLED; Here is a read lock for the curr_ctx. The patch doesn't contain the write lock part. I am not sure if the following situation would be happened: 1. curr_ctx is not NULL. 2. mtk_enc_comp_irq_handler() gets the curr_ctx. 3. The curr_ctx has been destroyed somewhere. 4. mtk_enc_comp_irq_handler() finds the ctx is not NULL so that it continues to execute. 5. Something wrong in latter mtk_enc_comp_irq_handler() because the ctx has been destroyed. Does it make more sense to set curr_ctx to NULL to indicate the ownership has been transferred to mtk_enc_comp_irq_handler()? For example: spin_lock_irqsave(...); ctx = dev->curr_ctx; dev->curr_ctx = NULL; spin_unlock_irqrestore(...); > +static int mtk_venc_comp_bind(struct device *dev, > + struct device *master, void *data) > +{ > + struct mtk_venc_comp_dev *comp_dev = dev_get_drvdata(dev); > + struct mtk_vcodec_dev *master_dev = data; > + int i; > + > + for (i = 0; i < MTK_VENC_HW_MAX; i++) { > + if (dev->of_node != master_dev->enc_comp_node[i]) > + continue; > + > + /*add component device by order*/ > + if (comp_dev->core_id == MTK_VENC_CORE0) > + master_dev->enc_comp_dev[MTK_VENC_CORE0] = comp_dev; > + else if (comp_dev->core_id == MTK_VENC_CORE1) > + master_dev->enc_comp_dev[MTK_VENC_CORE1] = comp_dev; > + else > + return -EINVAL; if (comp_dev->core_id < 0 || comp_dev->core_id >= MTK_VENC_HW_MAX) return -EINVAL; master_dev->enc_comp_dev[comp_dev->core_id] = comp_dev;
On Mon, Aug 16, 2021 at 06:59:28PM +0800, Irui Wang wrote: > @@ -105,6 +106,14 @@ static int mtk_venc_comp_probe(struct platform_device *pdev) > > comp_dev->plat_dev = pdev; > > + ret = mtk_vcodec_init_enc_pm(pdev, &comp_dev->pm); > + if (ret < 0) { > + dev_err(&pdev->dev, "Failed to get venc component clock source!"); > + return ret; > + } > + > + pm_runtime_enable(&pdev->dev); mtk_vcodec_init_enc_pm() and mtk_vcodec_release_enc_pm() is more like a pair. Does it make more sense to call pm_runtime_enable() in mtk_vcodec_init_enc_pm()?
On Mon, Aug 16, 2021 at 06:59:30PM +0800, Irui Wang wrote: > -void mtk_vcodec_enc_clock_on(struct mtk_vcodec_pm *pm) > +void mtk_vcodec_enc_clock_on(struct mtk_vcodec_dev *dev, int core_id) > { > - struct mtk_vcodec_clk *enc_clk = &pm->venc_clk; > - int ret, i = 0; > + struct mtk_venc_comp_dev *venc; > + struct mtk_vcodec_pm *enc_pm; > + struct mtk_vcodec_clk *enc_clk; > + struct clk *clk; To be neat, remove the extra spaces. > - ret = mtk_smi_larb_get(pm->larbvenc); > - if (ret) { > - mtk_v4l2_err("mtk_smi_larb_get larb3 fail %d", ret); > - goto clkerr; I may miss the context but why does it remove mtk_smi_larb_get()? > -void mtk_vcodec_enc_clock_off(struct mtk_vcodec_pm *pm) > +void mtk_vcodec_enc_clock_off(struct mtk_vcodec_dev *dev, int core_id) > { > - struct mtk_vcodec_clk *enc_clk = &pm->venc_clk; > - int i = 0; > + struct mtk_venc_comp_dev *venc; > + struct mtk_vcodec_pm *enc_pm; > + struct mtk_vcodec_clk *enc_clk; > + int i; > > - mtk_smi_larb_put(pm->larbvenc); Same here. Why does it remove mtk_smi_larb_put()? > int mtk_venc_enable_comp_hw(struct mtk_vcodec_dev *dev) > { > int i, ret; > struct mtk_venc_comp_dev *venc_comp; > + struct mtk_vcodec_clk *enc_clk; > + int j = 0; It doesn't need to be initialized. Can inline to "int i, ret;".
MT8195 has two H264 encoder cores, they have their own power-domains, clocks, interrupts, register base. The two H264 encoder cores can work together to achieve higher performance. This series of patches is to use enable two h264 encoder cores. path[1..2]: use linux component framework to manage encoder hardware, user call "mt8195-vcodec-enc" driver can get the encoder master device, the encoding work is done by the two encoder core device. The hw_mode variable is added to distinguish from old platform, two encoder cores called "FRAME_RACING_MODE". The hardware mode of two encoder cores work together(overlap, another word called) on MT8195 called "frame_racing_mode", the two encoder power-domains should be power on together while encoding, the encoding process look like this: VENC Core0 frm#0....frm#2....frm#4 VENC Core1 .frm#1....frm#3....frm#5 patch[3..5]: due to the component device, the master device has no power-domains/clocks properties in dtsi, so the power/clock init function can't use for "frame_racing_mode" device in master device probe process, it should be called in component device probe process. Power on the hardware power and clock on demand. patch[6]: "frame_racing_mode" encoding need a new set of memory buffer for two encoder cores. For compatibility, we should new a encoder driver interface. patch[7..9]: add "frame_racing_mode" encoding process: As-Is: Synchronous VIDIOC_QBUF#0 --> device_run(triger encoder) --> wait encoder IRQ --> encode done with result --> job_finish VIDIOC_QBUF#1 --> device_run(triger encoder) --> wait encoder IRQ --> encode done with result --> job_finish ... To-Be: Asynchronous VIDIOC_QBUF#0 --> device_run(triger encoder core0) --> job_finish ..VIDIOC_QBUF#1 --> device_run(triger encoder core1) --> job_finish (core0 may encode done here, return encode result to client) VIDIOC_QBUF#2 --> device_run(triger encoder core0) --> job_finish Thers is no "wait encoder IRQ" synchronous call during "frame_racing_mode" encoding process, which can full use the two encoder cores to achieve higher performance. Irui Wang (9): dt-bindings: media: mtk-vcodec: Add binding for MT8195 two venc cores media: mtk-vcodec: Use component framework to manage encoder hardware media: mtk-vcodec: Rewrite venc power manage interface media: mtk-vcodec: Add venc power on/off interface media: mtk-vcodec: Rewrite venc clock interface media: mtk-vcodec: Add new venc drv interface for frame_racing mode media: mtk-vcodec: Add frame racing mode encode process media: mtk-vcodec: Return encode result to client media: mtk-vcodec: Add delayed worker for encode timeout .../bindings/media/mediatek-vcodec.txt | 2 + drivers/media/platform/mtk-vcodec/Makefile | 2 + .../platform/mtk-vcodec/mtk_vcodec_drv.h | 34 +- .../platform/mtk-vcodec/mtk_vcodec_enc.c | 120 +++- .../platform/mtk-vcodec/mtk_vcodec_enc.h | 10 +- .../platform/mtk-vcodec/mtk_vcodec_enc_drv.c | 204 +++++- .../platform/mtk-vcodec/mtk_vcodec_enc_hw.c | 253 +++++++ .../platform/mtk-vcodec/mtk_vcodec_enc_hw.h | 38 + .../platform/mtk-vcodec/mtk_vcodec_enc_pm.c | 213 ++++-- .../platform/mtk-vcodec/mtk_vcodec_enc_pm.h | 13 +- .../platform/mtk-vcodec/mtk_vcodec_util.c | 19 + .../platform/mtk-vcodec/mtk_vcodec_util.h | 5 + .../platform/mtk-vcodec/venc/venc_common_if.c | 675 ++++++++++++++++++ .../platform/mtk-vcodec/venc/venc_h264_if.c | 6 +- .../platform/mtk-vcodec/venc/venc_vp8_if.c | 2 +- .../media/platform/mtk-vcodec/venc_drv_if.c | 96 ++- .../media/platform/mtk-vcodec/venc_drv_if.h | 7 + .../media/platform/mtk-vcodec/venc_vpu_if.c | 11 +- .../media/platform/mtk-vcodec/venc_vpu_if.h | 3 +- 19 files changed, 1564 insertions(+), 149 deletions(-) create mode 100644 drivers/media/platform/mtk-vcodec/mtk_vcodec_enc_hw.c create mode 100644 drivers/media/platform/mtk-vcodec/mtk_vcodec_enc_hw.h create mode 100644 drivers/media/platform/mtk-vcodec/venc/venc_common_if.c