On Fri 17 Sep 07:59 CDT 2021, Alexandre Bailon wrote: > Some Mediatek SoC provides hardware accelerator for AI / ML. > This driver use the DRM driver to manage the shared memory, > and use rpmsg to execute jobs on the APU. > > Signed-off-by: Alexandre Bailon <abailon@xxxxxxxxxxxx> > --- > drivers/rpmsg/Kconfig | 10 +++ > drivers/rpmsg/Makefile | 1 + > drivers/rpmsg/apu_rpmsg.c | 184 ++++++++++++++++++++++++++++++++++++++ > 3 files changed, 195 insertions(+) > create mode 100644 drivers/rpmsg/apu_rpmsg.c > > diff --git a/drivers/rpmsg/Kconfig b/drivers/rpmsg/Kconfig > index 0b4407abdf138..fc1668f795004 100644 > --- a/drivers/rpmsg/Kconfig > +++ b/drivers/rpmsg/Kconfig > @@ -73,4 +73,14 @@ config RPMSG_VIRTIO > select RPMSG_NS > select VIRTIO > > +config RPMSG_APU > + tristate "APU RPMSG driver" > + select REMOTEPROC > + select RPMSG_VIRTIO > + select DRM_APU > + help > + This provides a RPMSG driver that provides some facilities to > + communicate with an accelerated processing unit (APU). > + This Uses the APU DRM driver to manage memory and job scheduling. Similar to how a driver for e.g. an I2C device doesn't live in drivers/i2c, this doesn't belong in drivers/rpmsg. Probably rather directly in the DRM driver. > + > endmenu > diff --git a/drivers/rpmsg/Makefile b/drivers/rpmsg/Makefile > index 8d452656f0ee3..8b336b9a817c1 100644 > --- a/drivers/rpmsg/Makefile > +++ b/drivers/rpmsg/Makefile > @@ -9,3 +9,4 @@ obj-$(CONFIG_RPMSG_QCOM_GLINK_RPM) += qcom_glink_rpm.o > obj-$(CONFIG_RPMSG_QCOM_GLINK_SMEM) += qcom_glink_smem.o > obj-$(CONFIG_RPMSG_QCOM_SMD) += qcom_smd.o > obj-$(CONFIG_RPMSG_VIRTIO) += virtio_rpmsg_bus.o > +obj-$(CONFIG_RPMSG_APU) += apu_rpmsg.o > diff --git a/drivers/rpmsg/apu_rpmsg.c b/drivers/rpmsg/apu_rpmsg.c > new file mode 100644 > index 0000000000000..7e504bd176a4d > --- /dev/null > +++ b/drivers/rpmsg/apu_rpmsg.c > @@ -0,0 +1,184 @@ > +// SPDX-License-Identifier: GPL-2.0 > +// > +// Copyright 2020 BayLibre SAS > + > +#include <asm/cacheflush.h> > + > +#include <linux/cdev.h> > +#include <linux/dma-buf.h> > +#include <linux/dma-map-ops.h> > +#include <linux/dma-mapping.h> > +#include <linux/iommu.h> > +#include <linux/iova.h> > +#include <linux/mm.h> > +#include <linux/module.h> > +#include <linux/of.h> > +#include <linux/platform_device.h> > +#include <linux/remoteproc.h> > +#include <linux/rpmsg.h> > +#include <linux/slab.h> > +#include <linux/types.h> > + > +#include <drm/apu_drm.h> > + > +#include "rpmsg_internal.h" > + > +#define APU_RPMSG_SERVICE_MT8183 "rpmsg-mt8183-apu0" > + > +struct rpmsg_apu { > + struct apu_core *core; > + struct rpmsg_device *rpdev; > +}; > + > +static int apu_rpmsg_callback(struct rpmsg_device *rpdev, void *data, int count, > + void *priv, u32 addr) > +{ > + struct rpmsg_apu *apu = dev_get_drvdata(&rpdev->dev); > + struct apu_core *apu_core = apu->core; > + > + return apu_drm_callback(apu_core, data, count); > +} > + > +static int apu_rpmsg_send(struct apu_core *apu_core, void *data, int len) > +{ > + struct rpmsg_apu *apu = apu_drm_priv(apu_core); > + struct rpmsg_device *rpdev = apu->rpdev; > + > + return rpmsg_send(rpdev->ept, data, len); The rpmsg API is exposed outside drivers/rpmsg, so as I said above, just implement this directly in your driver, no need to lug around a dummy wrapper for things like this. > +} > + > +static struct apu_drm_ops apu_rpmsg_ops = { > + .send = apu_rpmsg_send, > +}; > + > +static int apu_init_iovad(struct rproc *rproc, struct rpmsg_apu *apu) > +{ > + struct resource_table *table; > + struct fw_rsc_carveout *rsc; > + int i; > + > + if (!rproc->table_ptr) { > + dev_err(&apu->rpdev->dev, > + "No resource_table: has the firmware been loaded ?\n"); > + return -ENODEV; > + } > + > + table = rproc->table_ptr; > + for (i = 0; i < table->num; i++) { > + int offset = table->offset[i]; > + struct fw_rsc_hdr *hdr = (void *)table + offset; > + > + if (hdr->type != RSC_CARVEOUT) > + continue; > + > + rsc = (void *)hdr + sizeof(*hdr); > + if (apu_drm_reserve_iova(apu->core, rsc->da, rsc->len)) { > + dev_err(&apu->rpdev->dev, > + "failed to reserve iova\n"); > + return -ENOMEM; > + } > + } > + > + return 0; > +} > + > +static struct rproc *apu_get_rproc(struct rpmsg_device *rpdev) > +{ > + /* > + * To work, the APU RPMsg driver need to get the rproc device. > + * Currently, we only use virtio so we could use that to find the > + * remoteproc parent. > + */ > + if (!rpdev->dev.parent && rpdev->dev.parent->bus) { > + dev_err(&rpdev->dev, "invalid rpmsg device\n"); > + return ERR_PTR(-EINVAL); > + } > + > + if (strcmp(rpdev->dev.parent->bus->name, "virtio")) { > + dev_err(&rpdev->dev, "unsupported bus\n"); > + return ERR_PTR(-EINVAL); > + } > + > + return vdev_to_rproc(dev_to_virtio(rpdev->dev.parent)); > +} > + > +static int apu_rpmsg_probe(struct rpmsg_device *rpdev) > +{ > + struct rpmsg_apu *apu; > + struct rproc *rproc; > + int ret; > + > + apu = devm_kzalloc(&rpdev->dev, sizeof(*apu), GFP_KERNEL); > + if (!apu) > + return -ENOMEM; > + apu->rpdev = rpdev; > + > + rproc = apu_get_rproc(rpdev); I believe that you can replace apu_get_rproc() with: rproc = rproc_get_by_child(&rpdev->dev); > + if (IS_ERR_OR_NULL(rproc)) > + return PTR_ERR(rproc); > + > + /* Make device dma capable by inheriting from parent's capabilities */ > + set_dma_ops(&rpdev->dev, get_dma_ops(rproc->dev.parent)); > + > + ret = dma_coerce_mask_and_coherent(&rpdev->dev, > + dma_get_mask(rproc->dev.parent)); > + if (ret) > + goto err_put_device; > + > + rpdev->dev.iommu_group = rproc->dev.parent->iommu_group; Would it be better or you if we have a device_node, so that you could specify the iommus property for this compute device? I'm asking because I've seen cases where multi-purpose remoteproc firmware operate using multiple different iommu streams... > + > + apu->core = apu_drm_register_core(rproc, &apu_rpmsg_ops, apu); > + if (!apu->core) { > + ret = -ENODEV; > + goto err_put_device; > + } > + > + ret = apu_init_iovad(rproc, apu); > + > + dev_set_drvdata(&rpdev->dev, apu); > + > + return ret; > + > +err_put_device: This label looks misplaced, and sure enough, if apu_init_iovad() fails you're not apu_drm_unregister_core(). But on that note, don't you want to apu_init_iovad() before you apu_drm_register_core()? > + devm_kfree(&rpdev->dev, apu); The reason for using devm_kzalloc() is that once you return unsuccessfully from probe, or from remove the memory is freed. So devm_kfree() should go in both cases. > + > + return ret; > +} > + > +static void apu_rpmsg_remove(struct rpmsg_device *rpdev) > +{ > + struct rpmsg_apu *apu = dev_get_drvdata(&rpdev->dev); > + > + apu_drm_unregister_core(apu); > + devm_kfree(&rpdev->dev, apu); No need to explicitly free devm resources. Regards, Bjorn > +} > + > +static const struct rpmsg_device_id apu_rpmsg_match[] = { > + { APU_RPMSG_SERVICE_MT8183 }, > + {} > +}; > + > +static struct rpmsg_driver apu_rpmsg_driver = { > + .probe = apu_rpmsg_probe, > + .remove = apu_rpmsg_remove, > + .callback = apu_rpmsg_callback, > + .id_table = apu_rpmsg_match, > + .drv = { > + .name = "apu_rpmsg", > + }, > +}; > + > +static int __init apu_rpmsg_init(void) > +{ > + return register_rpmsg_driver(&apu_rpmsg_driver); > +} > +arch_initcall(apu_rpmsg_init); > + > +static void __exit apu_rpmsg_exit(void) > +{ > + unregister_rpmsg_driver(&apu_rpmsg_driver); > +} > +module_exit(apu_rpmsg_exit); > + > +MODULE_LICENSE("GPL"); > +MODULE_DESCRIPTION("APU RPMSG driver"); > -- > 2.31.1 >