drm/v3d: Introduce a new DRM driver for Broadcom V3D V3.x+
authorEric Anholt <eric@anholt.net>
Mon, 30 Apr 2018 18:10:58 +0000 (11:10 -0700)
committerEric Anholt <eric@anholt.net>
Thu, 3 May 2018 23:26:30 +0000 (16:26 -0700)
This driver will be used to support Mesa on the Broadcom 7268 and 7278
platforms.

V3D 3.3 introduces an MMU, which means we no longer need CMA or vc4's
complicated CL/shader validation scheme.  This massively changes the
GEM behavior, so I've forked off to a new driver.

v2: Mark SUBMIT_CL as needing DRM_AUTH.  coccinelle fixes from kbuild
    test robot. Drop personal git link from MAINTAINERS.  Don't
    double-map dma-buf imported BOs.  Add kerneldoc about needing MMU
    eviction.  Drop prime vmap/unmap stubs.  Delay mmap offset setup
    to mmap time.  Use drm_dev_init instead of _alloc.  Use
    ktime_get() for wait_bo timeouts.  Drop drm_can_sleep() usage,
    since we don't modeset.  Switch page tables back to WC (debug
    change to coherent had slipped in).  Switch
    drm_gem_object_unreference_unlocked() to
    drm_gem_object_put_unlocked().  Simplify overflow mem handling by
    not sharing overflow mem between jobs.
v3: no changes
v4: align submit_cl to 64 bits (review by airlied), check zero flags in
    other ioctls.

Signed-off-by: Eric Anholt <eric@anholt.net>
Acked-by: Daniel Vetter <daniel.vetter@ffwll.ch> (v4)
Acked-by: Dave Airlie <airlied@linux.ie> (v3, requested submit_cl change)
Link: https://patchwork.freedesktop.org/patch/msgid/20180430181058.30181-3-eric@anholt.net
19 files changed:
Documentation/gpu/drivers.rst
MAINTAINERS
drivers/gpu/drm/Kconfig
drivers/gpu/drm/Makefile
drivers/gpu/drm/v3d/Kconfig [new file with mode: 0644]
drivers/gpu/drm/v3d/Makefile [new file with mode: 0644]
drivers/gpu/drm/v3d/v3d_bo.c [new file with mode: 0644]
drivers/gpu/drm/v3d/v3d_debugfs.c [new file with mode: 0644]
drivers/gpu/drm/v3d/v3d_drv.c [new file with mode: 0644]
drivers/gpu/drm/v3d/v3d_drv.h [new file with mode: 0644]
drivers/gpu/drm/v3d/v3d_fence.c [new file with mode: 0644]
drivers/gpu/drm/v3d/v3d_gem.c [new file with mode: 0644]
drivers/gpu/drm/v3d/v3d_irq.c [new file with mode: 0644]
drivers/gpu/drm/v3d/v3d_mmu.c [new file with mode: 0644]
drivers/gpu/drm/v3d/v3d_regs.h [new file with mode: 0644]
drivers/gpu/drm/v3d/v3d_sched.c [new file with mode: 0644]
drivers/gpu/drm/v3d/v3d_trace.h [new file with mode: 0644]
drivers/gpu/drm/v3d/v3d_trace_points.c [new file with mode: 0644]
include/uapi/drm/v3d_drm.h [new file with mode: 0644]

index d3ab6abae8385a531a7573a553cb557e5f3c0411..f982558fc25d4c2b9deeb9b7b529752f11e4916d 100644 (file)
@@ -10,6 +10,7 @@ GPU Driver Documentation
    tegra
    tinydrm
    tve200
+   v3d
    vc4
    bridge/dw-hdmi
    xen-front
index 4af7f6119530177f90bd5cb8ce8a45b720aadeb3..631a16f7fa1920e09f834f813e9bc37200337a97 100644 (file)
@@ -4795,6 +4795,14 @@ S:       Maintained
 F:     drivers/gpu/drm/omapdrm/
 F:     Documentation/devicetree/bindings/display/ti/
 
+DRM DRIVERS FOR V3D
+M:     Eric Anholt <eric@anholt.net>
+S:     Supported
+F:     drivers/gpu/drm/v3d/
+F:     include/uapi/drm/v3d_drm.h
+F:     Documentation/devicetree/bindings/display/brcm,bcm-v3d.txt
+T:     git git://anongit.freedesktop.org/drm/drm-misc
+
 DRM DRIVERS FOR VC4
 M:     Eric Anholt <eric@anholt.net>
 T:     git git://github.com/anholt/linux
index 757825ac60dfb7c09586a09724a0259c6405ef86..1c73a455fdb1346a5b3a16d224dc7de5131df7c9 100644 (file)
@@ -267,6 +267,8 @@ source "drivers/gpu/drm/amd/amdkfd/Kconfig"
 
 source "drivers/gpu/drm/imx/Kconfig"
 
+source "drivers/gpu/drm/v3d/Kconfig"
+
 source "drivers/gpu/drm/vc4/Kconfig"
 
 source "drivers/gpu/drm/etnaviv/Kconfig"
index 9d66657ea11743ef434ebfd4403d8e7b11020eae..7a401edd87613d94f67a49cc4c940ab8bc4f4a2f 100644 (file)
@@ -61,6 +61,7 @@ obj-$(CONFIG_DRM_MGA) += mga/
 obj-$(CONFIG_DRM_I810) += i810/
 obj-$(CONFIG_DRM_I915) += i915/
 obj-$(CONFIG_DRM_MGAG200) += mgag200/
+obj-$(CONFIG_DRM_V3D)  += v3d/
 obj-$(CONFIG_DRM_VC4)  += vc4/
 obj-$(CONFIG_DRM_CIRRUS_QEMU) += cirrus/
 obj-$(CONFIG_DRM_SIS)   += sis/
diff --git a/drivers/gpu/drm/v3d/Kconfig b/drivers/gpu/drm/v3d/Kconfig
new file mode 100644 (file)
index 0000000..a0c0259
--- /dev/null
@@ -0,0 +1,9 @@
+config DRM_V3D
+       tristate "Broadcom V3D 3.x and newer"
+       depends on ARCH_BCM || ARCH_BCMSTB || COMPILE_TEST
+       depends on DRM
+       depends on COMMON_CLK
+       select DRM_SCHED
+       help
+         Choose this option if you have a system that has a Broadcom
+         V3D 3.x or newer GPU, such as BCM7268.
diff --git a/drivers/gpu/drm/v3d/Makefile b/drivers/gpu/drm/v3d/Makefile
new file mode 100644 (file)
index 0000000..34446e1
--- /dev/null
@@ -0,0 +1,18 @@
+# Please keep these build lists sorted!
+
+# core driver code
+v3d-y := \
+       v3d_bo.o \
+       v3d_drv.o \
+       v3d_fence.o \
+       v3d_gem.o \
+       v3d_irq.o \
+       v3d_mmu.o \
+       v3d_trace_points.o \
+       v3d_sched.o
+
+v3d-$(CONFIG_DEBUG_FS) += v3d_debugfs.o
+
+obj-$(CONFIG_DRM_V3D)  += v3d.o
+
+CFLAGS_v3d_trace_points.o := -I$(src)
diff --git a/drivers/gpu/drm/v3d/v3d_bo.c b/drivers/gpu/drm/v3d/v3d_bo.c
new file mode 100644 (file)
index 0000000..7b1e2a5
--- /dev/null
@@ -0,0 +1,389 @@
+// SPDX-License-Identifier: GPL-2.0+
+/* Copyright (C) 2015-2018 Broadcom */
+
+/**
+ * DOC: V3D GEM BO management support
+ *
+ * Compared to VC4 (V3D 2.x), V3D 3.3 introduces an MMU between the
+ * GPU and the bus, allowing us to use shmem objects for our storage
+ * instead of CMA.
+ *
+ * Physically contiguous objects may still be imported to V3D, but the
+ * driver doesn't allocate physically contiguous objects on its own.
+ * Display engines requiring physically contiguous allocations should
+ * look into Mesa's "renderonly" support (as used by the Mesa pl111
+ * driver) for an example of how to integrate with V3D.
+ *
+ * Long term, we should support evicting pages from the MMU when under
+ * memory pressure (thus the v3d_bo_get_pages() refcounting), but
+ * that's not a high priority since our systems tend to not have swap.
+ */
+
+#include <linux/dma-buf.h>
+#include <linux/pfn_t.h>
+
+#include "v3d_drv.h"
+#include "uapi/drm/v3d_drm.h"
+
+/* Pins the shmem pages, fills in the .pages and .sgt fields of the BO, and maps
+ * it for DMA.
+ */
+static int
+v3d_bo_get_pages(struct v3d_bo *bo)
+{
+       struct drm_gem_object *obj = &bo->base;
+       struct drm_device *dev = obj->dev;
+       int npages = obj->size >> PAGE_SHIFT;
+       int ret = 0;
+
+       mutex_lock(&bo->lock);
+       if (bo->pages_refcount++ != 0)
+               goto unlock;
+
+       if (!obj->import_attach) {
+               bo->pages = drm_gem_get_pages(obj);
+               if (IS_ERR(bo->pages)) {
+                       ret = PTR_ERR(bo->pages);
+                       goto unlock;
+               }
+
+               bo->sgt = drm_prime_pages_to_sg(bo->pages, npages);
+               if (IS_ERR(bo->sgt)) {
+                       ret = PTR_ERR(bo->sgt);
+                       goto put_pages;
+               }
+
+               /* Map the pages for use by the GPU. */
+               dma_map_sg(dev->dev, bo->sgt->sgl,
+                          bo->sgt->nents, DMA_BIDIRECTIONAL);
+       } else {
+               bo->pages = kcalloc(npages, sizeof(*bo->pages), GFP_KERNEL);
+               if (!bo->pages)
+                       goto put_pages;
+
+               drm_prime_sg_to_page_addr_arrays(bo->sgt, bo->pages,
+                                                NULL, npages);
+
+               /* Note that dma-bufs come in mapped. */
+       }
+
+       mutex_unlock(&bo->lock);
+
+       return 0;
+
+put_pages:
+       drm_gem_put_pages(obj, bo->pages, true, true);
+       bo->pages = NULL;
+unlock:
+       bo->pages_refcount--;
+       mutex_unlock(&bo->lock);
+       return ret;
+}
+
+static void
+v3d_bo_put_pages(struct v3d_bo *bo)
+{
+       struct drm_gem_object *obj = &bo->base;
+
+       mutex_lock(&bo->lock);
+       if (--bo->pages_refcount == 0) {
+               if (!obj->import_attach) {
+                       dma_unmap_sg(obj->dev->dev, bo->sgt->sgl,
+                                    bo->sgt->nents, DMA_BIDIRECTIONAL);
+                       sg_free_table(bo->sgt);
+                       kfree(bo->sgt);
+                       drm_gem_put_pages(obj, bo->pages, true, true);
+               } else {
+                       kfree(bo->pages);
+               }
+       }
+       mutex_unlock(&bo->lock);
+}
+
+static struct v3d_bo *v3d_bo_create_struct(struct drm_device *dev,
+                                          size_t unaligned_size)
+{
+       struct v3d_dev *v3d = to_v3d_dev(dev);
+       struct drm_gem_object *obj;
+       struct v3d_bo *bo;
+       size_t size = roundup(unaligned_size, PAGE_SIZE);
+       int ret;
+
+       if (size == 0)
+               return ERR_PTR(-EINVAL);
+
+       bo = kzalloc(sizeof(*bo), GFP_KERNEL);
+       if (!bo)
+               return ERR_PTR(-ENOMEM);
+       obj = &bo->base;
+
+       INIT_LIST_HEAD(&bo->vmas);
+       INIT_LIST_HEAD(&bo->unref_head);
+       mutex_init(&bo->lock);
+
+       ret = drm_gem_object_init(dev, obj, size);
+       if (ret)
+               goto free_bo;
+
+       spin_lock(&v3d->mm_lock);
+       ret = drm_mm_insert_node_generic(&v3d->mm, &bo->node,
+                                        obj->size >> PAGE_SHIFT,
+                                        GMP_GRANULARITY >> PAGE_SHIFT, 0, 0);
+       spin_unlock(&v3d->mm_lock);
+       if (ret)
+               goto free_obj;
+
+       return bo;
+
+free_obj:
+       drm_gem_object_release(obj);
+free_bo:
+       kfree(bo);
+       return ERR_PTR(ret);
+}
+
+struct v3d_bo *v3d_bo_create(struct drm_device *dev, struct drm_file *file_priv,
+                            size_t unaligned_size)
+{
+       struct v3d_dev *v3d = to_v3d_dev(dev);
+       struct drm_gem_object *obj;
+       struct v3d_bo *bo;
+       int ret;
+
+       bo = v3d_bo_create_struct(dev, unaligned_size);
+       if (IS_ERR(bo))
+               return bo;
+       obj = &bo->base;
+
+       bo->resv = &bo->_resv;
+       reservation_object_init(bo->resv);
+
+       ret = v3d_bo_get_pages(bo);
+       if (ret)
+               goto free_mm;
+
+       v3d_mmu_insert_ptes(bo);
+
+       mutex_lock(&v3d->bo_lock);
+       v3d->bo_stats.num_allocated++;
+       v3d->bo_stats.pages_allocated += obj->size >> PAGE_SHIFT;
+       mutex_unlock(&v3d->bo_lock);
+
+       return bo;
+
+free_mm:
+       spin_lock(&v3d->mm_lock);
+       drm_mm_remove_node(&bo->node);
+       spin_unlock(&v3d->mm_lock);
+
+       drm_gem_object_release(obj);
+       kfree(bo);
+       return ERR_PTR(ret);
+}
+
+/* Called DRM core on the last userspace/kernel unreference of the
+ * BO.
+ */
+void v3d_free_object(struct drm_gem_object *obj)
+{
+       struct v3d_dev *v3d = to_v3d_dev(obj->dev);
+       struct v3d_bo *bo = to_v3d_bo(obj);
+
+       mutex_lock(&v3d->bo_lock);
+       v3d->bo_stats.num_allocated--;
+       v3d->bo_stats.pages_allocated -= obj->size >> PAGE_SHIFT;
+       mutex_unlock(&v3d->bo_lock);
+
+       reservation_object_fini(&bo->_resv);
+
+       v3d_bo_put_pages(bo);
+
+       if (obj->import_attach)
+               drm_prime_gem_destroy(obj, bo->sgt);
+
+       v3d_mmu_remove_ptes(bo);
+       spin_lock(&v3d->mm_lock);
+       drm_mm_remove_node(&bo->node);
+       spin_unlock(&v3d->mm_lock);
+
+       mutex_destroy(&bo->lock);
+
+       drm_gem_object_release(obj);
+       kfree(bo);
+}
+
+struct reservation_object *v3d_prime_res_obj(struct drm_gem_object *obj)
+{
+       struct v3d_bo *bo = to_v3d_bo(obj);
+
+       return bo->resv;
+}
+
+static void
+v3d_set_mmap_vma_flags(struct vm_area_struct *vma)
+{
+       vma->vm_flags &= ~VM_PFNMAP;
+       vma->vm_flags |= VM_MIXEDMAP;
+       vma->vm_page_prot = pgprot_writecombine(vm_get_page_prot(vma->vm_flags));
+}
+
+int v3d_gem_fault(struct vm_fault *vmf)
+{
+       struct vm_area_struct *vma = vmf->vma;
+       struct drm_gem_object *obj = vma->vm_private_data;
+       struct v3d_bo *bo = to_v3d_bo(obj);
+       unsigned long pfn;
+       pgoff_t pgoff;
+       int ret;
+
+       /* We don't use vmf->pgoff since that has the fake offset: */
+       pgoff = (vmf->address - vma->vm_start) >> PAGE_SHIFT;
+       pfn = page_to_pfn(bo->pages[pgoff]);
+
+       ret = vm_insert_mixed(vma, vmf->address, __pfn_to_pfn_t(pfn, PFN_DEV));
+
+       switch (ret) {
+       case -EAGAIN:
+       case 0:
+       case -ERESTARTSYS:
+       case -EINTR:
+       case -EBUSY:
+               /*
+                * EBUSY is ok: this just means that another thread
+                * already did the job.
+                */
+               return VM_FAULT_NOPAGE;
+       case -ENOMEM:
+               return VM_FAULT_OOM;
+       default:
+               return VM_FAULT_SIGBUS;
+       }
+}
+
+int v3d_mmap(struct file *filp, struct vm_area_struct *vma)
+{
+       int ret;
+
+       ret = drm_gem_mmap(filp, vma);
+       if (ret)
+               return ret;
+
+       v3d_set_mmap_vma_flags(vma);
+
+       return ret;
+}
+
+int v3d_prime_mmap(struct drm_gem_object *obj, struct vm_area_struct *vma)
+{
+       int ret;
+
+       ret = drm_gem_mmap_obj(obj, obj->size, vma);
+       if (ret < 0)
+               return ret;
+
+       v3d_set_mmap_vma_flags(vma);
+
+       return 0;
+}
+
+struct sg_table *
+v3d_prime_get_sg_table(struct drm_gem_object *obj)
+{
+       struct v3d_bo *bo = to_v3d_bo(obj);
+       int npages = obj->size >> PAGE_SHIFT;
+
+       return drm_prime_pages_to_sg(bo->pages, npages);
+}
+
+struct drm_gem_object *
+v3d_prime_import_sg_table(struct drm_device *dev,
+                         struct dma_buf_attachment *attach,
+                         struct sg_table *sgt)
+{
+       struct drm_gem_object *obj;
+       struct v3d_bo *bo;
+
+       bo = v3d_bo_create_struct(dev, attach->dmabuf->size);
+       if (IS_ERR(bo))
+               return ERR_CAST(bo);
+       obj = &bo->base;
+
+       bo->resv = attach->dmabuf->resv;
+
+       bo->sgt = sgt;
+       v3d_bo_get_pages(bo);
+
+       v3d_mmu_insert_ptes(bo);
+
+       return obj;
+}
+
+int v3d_create_bo_ioctl(struct drm_device *dev, void *data,
+                       struct drm_file *file_priv)
+{
+       struct drm_v3d_create_bo *args = data;
+       struct v3d_bo *bo = NULL;
+       int ret;
+
+       if (args->flags != 0) {
+               DRM_INFO("unknown create_bo flags: %d\n", args->flags);
+               return -EINVAL;
+       }
+
+       bo = v3d_bo_create(dev, file_priv, PAGE_ALIGN(args->size));
+       if (IS_ERR(bo))
+               return PTR_ERR(bo);
+
+       args->offset = bo->node.start << PAGE_SHIFT;
+
+       ret = drm_gem_handle_create(file_priv, &bo->base, &args->handle);
+       drm_gem_object_put_unlocked(&bo->base);
+
+       return ret;
+}
+
+int v3d_mmap_bo_ioctl(struct drm_device *dev, void *data,
+                     struct drm_file *file_priv)
+{
+       struct drm_v3d_mmap_bo *args = data;
+       struct drm_gem_object *gem_obj;
+       int ret;
+
+       if (args->flags != 0) {
+               DRM_INFO("unknown mmap_bo flags: %d\n", args->flags);
+               return -EINVAL;
+       }
+
+       gem_obj = drm_gem_object_lookup(file_priv, args->handle);
+       if (!gem_obj) {
+               DRM_DEBUG("Failed to look up GEM BO %d\n", args->handle);
+               return -ENOENT;
+       }
+
+       ret = drm_gem_create_mmap_offset(gem_obj);
+       if (ret == 0)
+               args->offset = drm_vma_node_offset_addr(&gem_obj->vma_node);
+       drm_gem_object_put_unlocked(gem_obj);
+
+       return ret;
+}
+
+int v3d_get_bo_offset_ioctl(struct drm_device *dev, void *data,
+                           struct drm_file *file_priv)
+{
+       struct drm_v3d_get_bo_offset *args = data;
+       struct drm_gem_object *gem_obj;
+       struct v3d_bo *bo;
+
+       gem_obj = drm_gem_object_lookup(file_priv, args->handle);
+       if (!gem_obj) {
+               DRM_DEBUG("Failed to look up GEM BO %d\n", args->handle);
+               return -ENOENT;
+       }
+       bo = to_v3d_bo(gem_obj);
+
+       args->offset = bo->node.start << PAGE_SHIFT;
+
+       drm_gem_object_put_unlocked(gem_obj);
+       return 0;
+}
diff --git a/drivers/gpu/drm/v3d/v3d_debugfs.c b/drivers/gpu/drm/v3d/v3d_debugfs.c
new file mode 100644 (file)
index 0000000..4db62c5
--- /dev/null
@@ -0,0 +1,191 @@
+// SPDX-License-Identifier: GPL-2.0+
+/* Copyright (C) 2014-2018 Broadcom */
+
+#include <linux/circ_buf.h>
+#include <linux/ctype.h>
+#include <linux/debugfs.h>
+#include <linux/pm_runtime.h>
+#include <linux/seq_file.h>
+#include <drm/drmP.h>
+
+#include "v3d_drv.h"
+#include "v3d_regs.h"
+
+#define REGDEF(reg) { reg, #reg }
+struct v3d_reg_def {
+       u32 reg;
+       const char *name;
+};
+
+static const struct v3d_reg_def v3d_hub_reg_defs[] = {
+       REGDEF(V3D_HUB_AXICFG),
+       REGDEF(V3D_HUB_UIFCFG),
+       REGDEF(V3D_HUB_IDENT0),
+       REGDEF(V3D_HUB_IDENT1),
+       REGDEF(V3D_HUB_IDENT2),
+       REGDEF(V3D_HUB_IDENT3),
+       REGDEF(V3D_HUB_INT_STS),
+       REGDEF(V3D_HUB_INT_MSK_STS),
+};
+
+static const struct v3d_reg_def v3d_gca_reg_defs[] = {
+       REGDEF(V3D_GCA_SAFE_SHUTDOWN),
+       REGDEF(V3D_GCA_SAFE_SHUTDOWN_ACK),
+};
+
+static const struct v3d_reg_def v3d_core_reg_defs[] = {
+       REGDEF(V3D_CTL_IDENT0),
+       REGDEF(V3D_CTL_IDENT1),
+       REGDEF(V3D_CTL_IDENT2),
+       REGDEF(V3D_CTL_MISCCFG),
+       REGDEF(V3D_CTL_INT_STS),
+       REGDEF(V3D_CTL_INT_MSK_STS),
+       REGDEF(V3D_CLE_CT0CS),
+       REGDEF(V3D_CLE_CT0CA),
+       REGDEF(V3D_CLE_CT0EA),
+       REGDEF(V3D_CLE_CT1CS),
+       REGDEF(V3D_CLE_CT1CA),
+       REGDEF(V3D_CLE_CT1EA),
+
+       REGDEF(V3D_PTB_BPCA),
+       REGDEF(V3D_PTB_BPCS),
+
+       REGDEF(V3D_MMU_CTL),
+       REGDEF(V3D_MMU_VIO_ADDR),
+
+       REGDEF(V3D_GMP_STATUS),
+       REGDEF(V3D_GMP_CFG),
+       REGDEF(V3D_GMP_VIO_ADDR),
+};
+
+static int v3d_v3d_debugfs_regs(struct seq_file *m, void *unused)
+{
+       struct drm_info_node *node = (struct drm_info_node *)m->private;
+       struct drm_device *dev = node->minor->dev;
+       struct v3d_dev *v3d = to_v3d_dev(dev);
+       int i, core;
+
+       for (i = 0; i < ARRAY_SIZE(v3d_hub_reg_defs); i++) {
+               seq_printf(m, "%s (0x%04x): 0x%08x\n",
+                          v3d_hub_reg_defs[i].name, v3d_hub_reg_defs[i].reg,
+                          V3D_READ(v3d_hub_reg_defs[i].reg));
+       }
+
+       for (i = 0; i < ARRAY_SIZE(v3d_gca_reg_defs); i++) {
+               seq_printf(m, "%s (0x%04x): 0x%08x\n",
+                          v3d_gca_reg_defs[i].name, v3d_gca_reg_defs[i].reg,
+                          V3D_GCA_READ(v3d_gca_reg_defs[i].reg));
+       }
+
+       for (core = 0; core < v3d->cores; core++) {
+               for (i = 0; i < ARRAY_SIZE(v3d_core_reg_defs); i++) {
+                       seq_printf(m, "core %d %s (0x%04x): 0x%08x\n",
+                                  core,
+                                  v3d_core_reg_defs[i].name,
+                                  v3d_core_reg_defs[i].reg,
+                                  V3D_CORE_READ(core,
+                                                v3d_core_reg_defs[i].reg));
+               }
+       }
+
+       return 0;
+}
+
+static int v3d_v3d_debugfs_ident(struct seq_file *m, void *unused)
+{
+       struct drm_info_node *node = (struct drm_info_node *)m->private;
+       struct drm_device *dev = node->minor->dev;
+       struct v3d_dev *v3d = to_v3d_dev(dev);
+       u32 ident0, ident1, ident2, ident3, cores;
+       int ret, core;
+
+       ret = pm_runtime_get_sync(v3d->dev);
+       if (ret < 0)
+               return ret;
+
+       ident0 = V3D_READ(V3D_HUB_IDENT0);
+       ident1 = V3D_READ(V3D_HUB_IDENT1);
+       ident2 = V3D_READ(V3D_HUB_IDENT2);
+       ident3 = V3D_READ(V3D_HUB_IDENT3);
+       cores = V3D_GET_FIELD(ident1, V3D_HUB_IDENT1_NCORES);
+
+       seq_printf(m, "Revision:   %d.%d.%d.%d\n",
+                  V3D_GET_FIELD(ident1, V3D_HUB_IDENT1_TVER),
+                  V3D_GET_FIELD(ident1, V3D_HUB_IDENT1_REV),
+                  V3D_GET_FIELD(ident3, V3D_HUB_IDENT3_IPREV),
+                  V3D_GET_FIELD(ident3, V3D_HUB_IDENT3_IPIDX));
+       seq_printf(m, "MMU:        %s\n",
+                  (ident2 & V3D_HUB_IDENT2_WITH_MMU) ? "yes" : "no");
+       seq_printf(m, "TFU:        %s\n",
+                  (ident1 & V3D_HUB_IDENT1_WITH_TFU) ? "yes" : "no");
+       seq_printf(m, "TSY:        %s\n",
+                  (ident1 & V3D_HUB_IDENT1_WITH_TSY) ? "yes" : "no");
+       seq_printf(m, "MSO:        %s\n",
+                  (ident1 & V3D_HUB_IDENT1_WITH_MSO) ? "yes" : "no");
+       seq_printf(m, "L3C:        %s (%dkb)\n",
+                  (ident1 & V3D_HUB_IDENT1_WITH_L3C) ? "yes" : "no",
+                  V3D_GET_FIELD(ident2, V3D_HUB_IDENT2_L3C_NKB));
+
+       for (core = 0; core < cores; core++) {
+               u32 misccfg;
+               u32 nslc, ntmu, qups;
+
+               ident0 = V3D_CORE_READ(core, V3D_CTL_IDENT0);
+               ident1 = V3D_CORE_READ(core, V3D_CTL_IDENT1);
+               ident2 = V3D_CORE_READ(core, V3D_CTL_IDENT2);
+               misccfg = V3D_CORE_READ(core, V3D_CTL_MISCCFG);
+
+               nslc = V3D_GET_FIELD(ident1, V3D_IDENT1_NSLC);
+               ntmu = V3D_GET_FIELD(ident1, V3D_IDENT1_NTMU);
+               qups = V3D_GET_FIELD(ident1, V3D_IDENT1_QUPS);
+
+               seq_printf(m, "Core %d:\n", core);
+               seq_printf(m, "  Revision:     %d.%d\n",
+                          V3D_GET_FIELD(ident0, V3D_IDENT0_VER),
+                          V3D_GET_FIELD(ident1, V3D_IDENT1_REV));
+               seq_printf(m, "  Slices:       %d\n", nslc);
+               seq_printf(m, "  TMUs:         %d\n", nslc * ntmu);
+               seq_printf(m, "  QPUs:         %d\n", nslc * qups);
+               seq_printf(m, "  Semaphores:   %d\n",
+                          V3D_GET_FIELD(ident1, V3D_IDENT1_NSEM));
+               seq_printf(m, "  BCG int:      %d\n",
+                          (ident2 & V3D_IDENT2_BCG_INT) != 0);
+               seq_printf(m, "  Override TMU: %d\n",
+                          (misccfg & V3D_MISCCFG_OVRTMUOUT) != 0);
+       }
+
+       pm_runtime_mark_last_busy(v3d->dev);
+       pm_runtime_put_autosuspend(v3d->dev);
+
+       return 0;
+}
+
+static int v3d_debugfs_bo_stats(struct seq_file *m, void *unused)
+{
+       struct drm_info_node *node = (struct drm_info_node *)m->private;
+       struct drm_device *dev = node->minor->dev;
+       struct v3d_dev *v3d = to_v3d_dev(dev);
+
+       mutex_lock(&v3d->bo_lock);
+       seq_printf(m, "allocated bos:          %d\n",
+                  v3d->bo_stats.num_allocated);
+       seq_printf(m, "allocated bo size (kb): %ld\n",
+                  (long)v3d->bo_stats.pages_allocated << (PAGE_SHIFT - 10));
+       mutex_unlock(&v3d->bo_lock);
+
+       return 0;
+}
+
+static const struct drm_info_list v3d_debugfs_list[] = {
+       {"v3d_ident", v3d_v3d_debugfs_ident, 0},
+       {"v3d_regs", v3d_v3d_debugfs_regs, 0},
+       {"bo_stats", v3d_debugfs_bo_stats, 0},
+};
+
+int
+v3d_debugfs_init(struct drm_minor *minor)
+{
+       return drm_debugfs_create_files(v3d_debugfs_list,
+                                       ARRAY_SIZE(v3d_debugfs_list),
+                                       minor->debugfs_root, minor);
+}
diff --git a/drivers/gpu/drm/v3d/v3d_drv.c b/drivers/gpu/drm/v3d/v3d_drv.c
new file mode 100644 (file)
index 0000000..38e8041
--- /dev/null
@@ -0,0 +1,371 @@
+// SPDX-License-Identifier: GPL-2.0+
+/* Copyright (C) 2014-2018 Broadcom */
+
+/**
+ * DOC: Broadcom V3D Graphics Driver
+ *
+ * This driver supports the Broadcom V3D 3.3 and 4.1 OpenGL ES GPUs.
+ * For V3D 2.x support, see the VC4 driver.
+ *
+ * Currently only single-core rendering using the binner and renderer
+ * is supported.  The TFU (texture formatting unit) and V3D 4.x's CSD
+ * (compute shader dispatch) are not yet supported.
+ */
+
+#include <linux/clk.h>
+#include <linux/device.h>
+#include <linux/io.h>
+#include <linux/module.h>
+#include <linux/of_platform.h>
+#include <linux/platform_device.h>
+#include <linux/pm_runtime.h>
+#include <drm/drm_fb_cma_helper.h>
+#include <drm/drm_fb_helper.h>
+
+#include "uapi/drm/v3d_drm.h"
+#include "v3d_drv.h"
+#include "v3d_regs.h"
+
+#define DRIVER_NAME "v3d"
+#define DRIVER_DESC "Broadcom V3D graphics"
+#define DRIVER_DATE "20180419"
+#define DRIVER_MAJOR 1
+#define DRIVER_MINOR 0
+#define DRIVER_PATCHLEVEL 0
+
+#ifdef CONFIG_PM
+static int v3d_runtime_suspend(struct device *dev)
+{
+       struct drm_device *drm = dev_get_drvdata(dev);
+       struct v3d_dev *v3d = to_v3d_dev(drm);
+
+       v3d_irq_disable(v3d);
+
+       clk_disable_unprepare(v3d->clk);
+
+       return 0;
+}
+
+static int v3d_runtime_resume(struct device *dev)
+{
+       struct drm_device *drm = dev_get_drvdata(dev);
+       struct v3d_dev *v3d = to_v3d_dev(drm);
+       int ret;
+
+       ret = clk_prepare_enable(v3d->clk);
+       if (ret != 0)
+               return ret;
+
+       /* XXX: VPM base */
+
+       v3d_mmu_set_page_table(v3d);
+       v3d_irq_enable(v3d);
+
+       return 0;
+}
+#endif
+
+static const struct dev_pm_ops v3d_v3d_pm_ops = {
+       SET_RUNTIME_PM_OPS(v3d_runtime_suspend, v3d_runtime_resume, NULL)
+};
+
+static int v3d_get_param_ioctl(struct drm_device *dev, void *data,
+                              struct drm_file *file_priv)
+{
+       struct v3d_dev *v3d = to_v3d_dev(dev);
+       struct drm_v3d_get_param *args = data;
+       int ret;
+       static const u32 reg_map[] = {
+               [DRM_V3D_PARAM_V3D_UIFCFG] = V3D_HUB_UIFCFG,
+               [DRM_V3D_PARAM_V3D_HUB_IDENT1] = V3D_HUB_IDENT1,
+               [DRM_V3D_PARAM_V3D_HUB_IDENT2] = V3D_HUB_IDENT2,
+               [DRM_V3D_PARAM_V3D_HUB_IDENT3] = V3D_HUB_IDENT3,
+               [DRM_V3D_PARAM_V3D_CORE0_IDENT0] = V3D_CTL_IDENT0,
+               [DRM_V3D_PARAM_V3D_CORE0_IDENT1] = V3D_CTL_IDENT1,
+               [DRM_V3D_PARAM_V3D_CORE0_IDENT2] = V3D_CTL_IDENT2,
+       };
+
+       if (args->pad != 0)
+               return -EINVAL;
+
+       /* Note that DRM_V3D_PARAM_V3D_CORE0_IDENT0 is 0, so we need
+        * to explicitly allow it in the "the register in our
+        * parameter map" check.
+        */
+       if (args->param < ARRAY_SIZE(reg_map) &&
+           (reg_map[args->param] ||
+            args->param == DRM_V3D_PARAM_V3D_CORE0_IDENT0)) {
+               u32 offset = reg_map[args->param];
+
+               if (args->value != 0)
+                       return -EINVAL;
+
+               ret = pm_runtime_get_sync(v3d->dev);
+               if (args->param >= DRM_V3D_PARAM_V3D_CORE0_IDENT0 &&
+                   args->param <= DRM_V3D_PARAM_V3D_CORE0_IDENT2) {
+                       args->value = V3D_CORE_READ(0, offset);
+               } else {
+                       args->value = V3D_READ(offset);
+               }
+               pm_runtime_mark_last_busy(v3d->dev);
+               pm_runtime_put_autosuspend(v3d->dev);
+               return 0;
+       }
+
+       /* Any params that aren't just register reads would go here. */
+
+       DRM_DEBUG("Unknown parameter %d\n", args->param);
+       return -EINVAL;
+}
+
+static int
+v3d_open(struct drm_device *dev, struct drm_file *file)
+{
+       struct v3d_dev *v3d = to_v3d_dev(dev);
+       struct v3d_file_priv *v3d_priv;
+       int i;
+
+       v3d_priv = kzalloc(sizeof(*v3d_priv), GFP_KERNEL);
+       if (!v3d_priv)
+               return -ENOMEM;
+
+       v3d_priv->v3d = v3d;
+
+       for (i = 0; i < V3D_MAX_QUEUES; i++) {
+               drm_sched_entity_init(&v3d->queue[i].sched,
+                                     &v3d_priv->sched_entity[i],
+                                     &v3d->queue[i].sched.sched_rq[DRM_SCHED_PRIORITY_NORMAL],
+                                     32, NULL);
+       }
+
+       file->driver_priv = v3d_priv;
+
+       return 0;
+}
+
+static void
+v3d_postclose(struct drm_device *dev, struct drm_file *file)
+{
+       struct v3d_dev *v3d = to_v3d_dev(dev);
+       struct v3d_file_priv *v3d_priv = file->driver_priv;
+       enum v3d_queue q;
+
+       for (q = 0; q < V3D_MAX_QUEUES; q++) {
+               drm_sched_entity_fini(&v3d->queue[q].sched,
+                                     &v3d_priv->sched_entity[q]);
+       }
+
+       kfree(v3d_priv);
+}
+
+static const struct file_operations v3d_drm_fops = {
+       .owner = THIS_MODULE,
+       .open = drm_open,
+       .release = drm_release,
+       .unlocked_ioctl = drm_ioctl,
+       .mmap = v3d_mmap,
+       .poll = drm_poll,
+       .read = drm_read,
+       .compat_ioctl = drm_compat_ioctl,
+       .llseek = noop_llseek,
+};
+
+/* DRM_AUTH is required on SUBMIT_CL for now, while we don't have GMP
+ * protection between clients.  Note that render nodes would be be
+ * able to submit CLs that could access BOs from clients authenticated
+ * with the master node.
+ */
+static const struct drm_ioctl_desc v3d_drm_ioctls[] = {
+       DRM_IOCTL_DEF_DRV(V3D_SUBMIT_CL, v3d_submit_cl_ioctl, DRM_RENDER_ALLOW | DRM_AUTH),
+       DRM_IOCTL_DEF_DRV(V3D_WAIT_BO, v3d_wait_bo_ioctl, DRM_RENDER_ALLOW),
+       DRM_IOCTL_DEF_DRV(V3D_CREATE_BO, v3d_create_bo_ioctl, DRM_RENDER_ALLOW),
+       DRM_IOCTL_DEF_DRV(V3D_MMAP_BO, v3d_mmap_bo_ioctl, DRM_RENDER_ALLOW),
+       DRM_IOCTL_DEF_DRV(V3D_GET_PARAM, v3d_get_param_ioctl, DRM_RENDER_ALLOW),
+       DRM_IOCTL_DEF_DRV(V3D_GET_BO_OFFSET, v3d_get_bo_offset_ioctl, DRM_RENDER_ALLOW),
+};
+
+static const struct vm_operations_struct v3d_vm_ops = {
+       .fault = v3d_gem_fault,
+       .open = drm_gem_vm_open,
+       .close = drm_gem_vm_close,
+};
+
+static struct drm_driver v3d_drm_driver = {
+       .driver_features = (DRIVER_GEM |
+                           DRIVER_RENDER |
+                           DRIVER_PRIME |
+                           DRIVER_SYNCOBJ),
+
+       .open = v3d_open,
+       .postclose = v3d_postclose,
+
+#if defined(CONFIG_DEBUG_FS)
+       .debugfs_init = v3d_debugfs_init,
+#endif
+
+       .gem_free_object_unlocked = v3d_free_object,
+       .gem_vm_ops = &v3d_vm_ops,
+
+       .prime_handle_to_fd = drm_gem_prime_handle_to_fd,
+       .prime_fd_to_handle = drm_gem_prime_fd_to_handle,
+       .gem_prime_import = drm_gem_prime_import,
+       .gem_prime_export = drm_gem_prime_export,
+       .gem_prime_res_obj = v3d_prime_res_obj,
+       .gem_prime_get_sg_table = v3d_prime_get_sg_table,
+       .gem_prime_import_sg_table = v3d_prime_import_sg_table,
+       .gem_prime_mmap = v3d_prime_mmap,
+
+       .ioctls = v3d_drm_ioctls,
+       .num_ioctls = ARRAY_SIZE(v3d_drm_ioctls),
+       .fops = &v3d_drm_fops,
+
+       .name = DRIVER_NAME,
+       .desc = DRIVER_DESC,
+       .date = DRIVER_DATE,
+       .major = DRIVER_MAJOR,
+       .minor = DRIVER_MINOR,
+       .patchlevel = DRIVER_PATCHLEVEL,
+};
+
+static const struct of_device_id v3d_of_match[] = {
+       { .compatible = "brcm,7268-v3d" },
+       { .compatible = "brcm,7278-v3d" },
+       {},
+};
+MODULE_DEVICE_TABLE(of, v3d_of_match);
+
+static int
+map_regs(struct v3d_dev *v3d, void __iomem **regs, const char *name)
+{
+       struct resource *res =
+               platform_get_resource_byname(v3d->pdev, IORESOURCE_MEM, name);
+
+       *regs = devm_ioremap_resource(v3d->dev, res);
+       return PTR_ERR_OR_ZERO(*regs);
+}
+
+static int v3d_platform_drm_probe(struct platform_device *pdev)
+{
+       struct device *dev = &pdev->dev;
+       struct drm_device *drm;
+       struct v3d_dev *v3d;
+       int ret;
+       u32 ident1;
+
+       dev->coherent_dma_mask = DMA_BIT_MASK(36);
+
+       v3d = kzalloc(sizeof(*v3d), GFP_KERNEL);
+       if (!v3d)
+               return -ENOMEM;
+       v3d->dev = dev;
+       v3d->pdev = pdev;
+       drm = &v3d->drm;
+
+       ret = map_regs(v3d, &v3d->bridge_regs, "bridge");
+       if (ret)
+               goto dev_free;
+
+       ret = map_regs(v3d, &v3d->hub_regs, "hub");
+       if (ret)
+               goto dev_free;
+
+       ret = map_regs(v3d, &v3d->core_regs[0], "core0");
+       if (ret)
+               goto dev_free;
+
+       ident1 = V3D_READ(V3D_HUB_IDENT1);
+       v3d->ver = (V3D_GET_FIELD(ident1, V3D_HUB_IDENT1_TVER) * 10 +
+                   V3D_GET_FIELD(ident1, V3D_HUB_IDENT1_REV));
+       v3d->cores = V3D_GET_FIELD(ident1, V3D_HUB_IDENT1_NCORES);
+       WARN_ON(v3d->cores > 1); /* multicore not yet implemented */
+
+       if (v3d->ver < 41) {
+               ret = map_regs(v3d, &v3d->gca_regs, "gca");
+               if (ret)
+                       goto dev_free;
+       }
+
+       v3d->mmu_scratch = dma_alloc_wc(dev, 4096, &v3d->mmu_scratch_paddr,
+                                       GFP_KERNEL | __GFP_NOWARN | __GFP_ZERO);
+       if (!v3d->mmu_scratch) {
+               dev_err(dev, "Failed to allocate MMU scratch page\n");
+               ret = -ENOMEM;
+               goto dev_free;
+       }
+
+       pm_runtime_use_autosuspend(dev);
+       pm_runtime_set_autosuspend_delay(dev, 50);
+       pm_runtime_enable(dev);
+
+       ret = drm_dev_init(&v3d->drm, &v3d_drm_driver, dev);
+       if (ret)
+               goto dma_free;
+
+       platform_set_drvdata(pdev, drm);
+       drm->dev_private = v3d;
+
+       ret = v3d_gem_init(drm);
+       if (ret)
+               goto dev_destroy;
+
+       v3d_irq_init(v3d);
+
+       ret = drm_dev_register(drm, 0);
+       if (ret)
+               goto gem_destroy;
+
+       return 0;
+
+gem_destroy:
+       v3d_gem_destroy(drm);
+dev_destroy:
+       drm_dev_put(drm);
+dma_free:
+       dma_free_wc(dev, 4096, v3d->mmu_scratch, v3d->mmu_scratch_paddr);
+dev_free:
+       kfree(v3d);
+       return ret;
+}
+
+static int v3d_platform_drm_remove(struct platform_device *pdev)
+{
+       struct drm_device *drm = platform_get_drvdata(pdev);
+       struct v3d_dev *v3d = to_v3d_dev(drm);
+
+       drm_dev_unregister(drm);
+
+       v3d_gem_destroy(drm);
+
+       drm_dev_put(drm);
+
+       dma_free_wc(v3d->dev, 4096, v3d->mmu_scratch, v3d->mmu_scratch_paddr);
+
+       return 0;
+}
+
+static struct platform_driver v3d_platform_driver = {
+       .probe          = v3d_platform_drm_probe,
+       .remove         = v3d_platform_drm_remove,
+       .driver         = {
+               .name   = "v3d",
+               .of_match_table = v3d_of_match,
+       },
+};
+
+static int __init v3d_drm_register(void)
+{
+       return platform_driver_register(&v3d_platform_driver);
+}
+
+static void __exit v3d_drm_unregister(void)
+{
+       platform_driver_unregister(&v3d_platform_driver);
+}
+
+module_init(v3d_drm_register);
+module_exit(v3d_drm_unregister);
+
+MODULE_ALIAS("platform:v3d-drm");
+MODULE_DESCRIPTION("Broadcom V3D DRM Driver");
+MODULE_AUTHOR("Eric Anholt <eric@anholt.net>");
+MODULE_LICENSE("GPL v2");
diff --git a/drivers/gpu/drm/v3d/v3d_drv.h b/drivers/gpu/drm/v3d/v3d_drv.h
new file mode 100644 (file)
index 0000000..a043ac3
--- /dev/null
@@ -0,0 +1,294 @@
+// SPDX-License-Identifier: GPL-2.0+
+/* Copyright (C) 2015-2018 Broadcom */
+
+#include <linux/reservation.h>
+#include <drm/drmP.h>
+#include <drm/drm_encoder.h>
+#include <drm/drm_gem.h>
+#include <drm/gpu_scheduler.h>
+
+#define GMP_GRANULARITY (128 * 1024)
+
+/* Enum for each of the V3D queues.  We maintain various queue
+ * tracking as an array because at some point we'll want to support
+ * the TFU (texture formatting unit) as another queue.
+ */
+enum v3d_queue {
+       V3D_BIN,
+       V3D_RENDER,
+};
+
+#define V3D_MAX_QUEUES (V3D_RENDER + 1)
+
+struct v3d_queue_state {
+       struct drm_gpu_scheduler sched;
+
+       u64 fence_context;
+       u64 emit_seqno;
+       u64 finished_seqno;
+};
+
+struct v3d_dev {
+       struct drm_device drm;
+
+       /* Short representation (e.g. 33, 41) of the V3D tech version
+        * and revision.
+        */
+       int ver;
+
+       struct device *dev;
+       struct platform_device *pdev;
+       void __iomem *hub_regs;
+       void __iomem *core_regs[3];
+       void __iomem *bridge_regs;
+       void __iomem *gca_regs;
+       struct clk *clk;
+
+       /* Virtual and DMA addresses of the single shared page table. */
+       volatile u32 *pt;
+       dma_addr_t pt_paddr;
+
+       /* Virtual and DMA addresses of the MMU's scratch page.  When
+        * a read or write is invalid in the MMU, it will be
+        * redirected here.
+        */
+       void *mmu_scratch;
+       dma_addr_t mmu_scratch_paddr;
+
+       /* Number of V3D cores. */
+       u32 cores;
+
+       /* Allocator managing the address space.  All units are in
+        * number of pages.
+        */
+       struct drm_mm mm;
+       spinlock_t mm_lock;
+
+       struct work_struct overflow_mem_work;
+
+       struct v3d_exec_info *bin_job;
+       struct v3d_exec_info *render_job;
+
+       struct v3d_queue_state queue[V3D_MAX_QUEUES];
+
+       /* Spinlock used to synchronize the overflow memory
+        * management against bin job submission.
+        */
+       spinlock_t job_lock;
+
+       /* Protects bo_stats */
+       struct mutex bo_lock;
+
+       /* Lock taken when resetting the GPU, to keep multiple
+        * processes from trying to park the scheduler threads and
+        * reset at once.
+        */
+       struct mutex reset_lock;
+
+       struct {
+               u32 num_allocated;
+               u32 pages_allocated;
+       } bo_stats;
+};
+
+static inline struct v3d_dev *
+to_v3d_dev(struct drm_device *dev)
+{
+       return (struct v3d_dev *)dev->dev_private;
+}
+
+/* The per-fd struct, which tracks the MMU mappings. */
+struct v3d_file_priv {
+       struct v3d_dev *v3d;
+
+       struct drm_sched_entity sched_entity[V3D_MAX_QUEUES];
+};
+
+/* Tracks a mapping of a BO into a per-fd address space */
+struct v3d_vma {
+       struct v3d_page_table *pt;
+       struct list_head list; /* entry in v3d_bo.vmas */
+};
+
+struct v3d_bo {
+       struct drm_gem_object base;
+
+       struct mutex lock;
+
+       struct drm_mm_node node;
+
+       u32 pages_refcount;
+       struct page **pages;
+       struct sg_table *sgt;
+       void *vaddr;
+
+       struct list_head vmas;    /* list of v3d_vma */
+
+       /* List entry for the BO's position in
+        * v3d_exec_info->unref_list
+        */
+       struct list_head unref_head;
+
+       /* normally (resv == &_resv) except for imported bo's */
+       struct reservation_object *resv;
+       struct reservation_object _resv;
+};
+
+static inline struct v3d_bo *
+to_v3d_bo(struct drm_gem_object *bo)
+{
+       return (struct v3d_bo *)bo;
+}
+
+struct v3d_fence {
+       struct dma_fence base;
+       struct drm_device *dev;
+       /* v3d seqno for signaled() test */
+       u64 seqno;
+       enum v3d_queue queue;
+};
+
+static inline struct v3d_fence *
+to_v3d_fence(struct dma_fence *fence)
+{
+       return (struct v3d_fence *)fence;
+}
+
+#define V3D_READ(offset) readl(v3d->hub_regs + offset)
+#define V3D_WRITE(offset, val) writel(val, v3d->hub_regs + offset)
+
+#define V3D_BRIDGE_READ(offset) readl(v3d->bridge_regs + offset)
+#define V3D_BRIDGE_WRITE(offset, val) writel(val, v3d->bridge_regs + offset)
+
+#define V3D_GCA_READ(offset) readl(v3d->gca_regs + offset)
+#define V3D_GCA_WRITE(offset, val) writel(val, v3d->gca_regs + offset)
+
+#define V3D_CORE_READ(core, offset) readl(v3d->core_regs[core] + offset)
+#define V3D_CORE_WRITE(core, offset, val) writel(val, v3d->core_regs[core] + offset)
+
+struct v3d_job {
+       struct drm_sched_job base;
+
+       struct v3d_exec_info *exec;
+
+       /* An optional fence userspace can pass in for the job to depend on. */
+       struct dma_fence *in_fence;
+
+       /* v3d fence to be signaled by IRQ handler when the job is complete. */
+       struct dma_fence *done_fence;
+
+       /* GPU virtual addresses of the start/end of the CL job. */
+       u32 start, end;
+};
+
+struct v3d_exec_info {
+       struct v3d_dev *v3d;
+
+       struct v3d_job bin, render;
+
+       /* Fence for when the scheduler considers the binner to be
+        * done, for render to depend on.
+        */
+       struct dma_fence *bin_done_fence;
+
+       struct kref refcount;
+
+       /* This is the array of BOs that were looked up at the start of exec. */
+       struct v3d_bo **bo;
+       u32 bo_count;
+
+       /* List of overflow BOs used in the job that need to be
+        * released once the job is complete.
+        */
+       struct list_head unref_list;
+
+       /* Submitted tile memory allocation start/size, tile state. */
+       u32 qma, qms, qts;
+};
+
+/**
+ * _wait_for - magic (register) wait macro
+ *
+ * Does the right thing for modeset paths when run under kdgb or similar atomic
+ * contexts. Note that it's important that we check the condition again after
+ * having timed out, since the timeout could be due to preemption or similar and
+ * we've never had a chance to check the condition before the timeout.
+ */
+#define wait_for(COND, MS) ({ \
+       unsigned long timeout__ = jiffies + msecs_to_jiffies(MS) + 1;   \
+       int ret__ = 0;                                                  \
+       while (!(COND)) {                                               \
+               if (time_after(jiffies, timeout__)) {                   \
+                       if (!(COND))                                    \
+                               ret__ = -ETIMEDOUT;                     \
+                       break;                                          \
+               }                                                       \
+               msleep(1);                                      \
+       }                                                               \
+       ret__;                                                          \
+})
+
+static inline unsigned long nsecs_to_jiffies_timeout(const u64 n)
+{
+       /* nsecs_to_jiffies64() does not guard against overflow */
+       if (NSEC_PER_SEC % HZ &&
+           div_u64(n, NSEC_PER_SEC) >= MAX_JIFFY_OFFSET / HZ)
+               return MAX_JIFFY_OFFSET;
+
+       return min_t(u64, MAX_JIFFY_OFFSET, nsecs_to_jiffies64(n) + 1);
+}
+
+/* v3d_bo.c */
+void v3d_free_object(struct drm_gem_object *gem_obj);
+struct v3d_bo *v3d_bo_create(struct drm_device *dev, struct drm_file *file_priv,
+                            size_t size);
+int v3d_create_bo_ioctl(struct drm_device *dev, void *data,
+                       struct drm_file *file_priv);
+int v3d_mmap_bo_ioctl(struct drm_device *dev, void *data,
+                     struct drm_file *file_priv);
+int v3d_get_bo_offset_ioctl(struct drm_device *dev, void *data,
+                           struct drm_file *file_priv);
+int v3d_gem_fault(struct vm_fault *vmf);
+int v3d_mmap(struct file *filp, struct vm_area_struct *vma);
+struct reservation_object *v3d_prime_res_obj(struct drm_gem_object *obj);
+int v3d_prime_mmap(struct drm_gem_object *obj, struct vm_area_struct *vma);
+struct sg_table *v3d_prime_get_sg_table(struct drm_gem_object *obj);
+struct drm_gem_object *v3d_prime_import_sg_table(struct drm_device *dev,
+                                                struct dma_buf_attachment *attach,
+                                                struct sg_table *sgt);
+
+/* v3d_debugfs.c */
+int v3d_debugfs_init(struct drm_minor *minor);
+
+/* v3d_fence.c */
+extern const struct dma_fence_ops v3d_fence_ops;
+struct dma_fence *v3d_fence_create(struct v3d_dev *v3d, enum v3d_queue queue);
+
+/* v3d_gem.c */
+int v3d_gem_init(struct drm_device *dev);
+void v3d_gem_destroy(struct drm_device *dev);
+int v3d_submit_cl_ioctl(struct drm_device *dev, void *data,
+                       struct drm_file *file_priv);
+int v3d_wait_bo_ioctl(struct drm_device *dev, void *data,
+                     struct drm_file *file_priv);
+void v3d_exec_put(struct v3d_exec_info *exec);
+void v3d_reset(struct v3d_dev *v3d);
+void v3d_invalidate_caches(struct v3d_dev *v3d);
+void v3d_flush_caches(struct v3d_dev *v3d);
+
+/* v3d_irq.c */
+void v3d_irq_init(struct v3d_dev *v3d);
+void v3d_irq_enable(struct v3d_dev *v3d);
+void v3d_irq_disable(struct v3d_dev *v3d);
+void v3d_irq_reset(struct v3d_dev *v3d);
+
+/* v3d_mmu.c */
+int v3d_mmu_get_offset(struct drm_file *file_priv, struct v3d_bo *bo,
+                      u32 *offset);
+int v3d_mmu_set_page_table(struct v3d_dev *v3d);
+void v3d_mmu_insert_ptes(struct v3d_bo *bo);
+void v3d_mmu_remove_ptes(struct v3d_bo *bo);
+
+/* v3d_sched.c */
+int v3d_sched_init(struct v3d_dev *v3d);
+void v3d_sched_fini(struct v3d_dev *v3d);
diff --git a/drivers/gpu/drm/v3d/v3d_fence.c b/drivers/gpu/drm/v3d/v3d_fence.c
new file mode 100644 (file)
index 0000000..087d49c
--- /dev/null
@@ -0,0 +1,58 @@
+// SPDX-License-Identifier: GPL-2.0+
+/* Copyright (C) 2017-2018 Broadcom */
+
+#include "v3d_drv.h"
+
+struct dma_fence *v3d_fence_create(struct v3d_dev *v3d, enum v3d_queue queue)
+{
+       struct v3d_fence *fence;
+
+       fence = kzalloc(sizeof(*fence), GFP_KERNEL);
+       if (!fence)
+               return ERR_PTR(-ENOMEM);
+
+       fence->dev = &v3d->drm;
+       fence->queue = queue;
+       fence->seqno = ++v3d->queue[queue].emit_seqno;
+       dma_fence_init(&fence->base, &v3d_fence_ops, &v3d->job_lock,
+                      v3d->queue[queue].fence_context, fence->seqno);
+
+       return &fence->base;
+}
+
+static const char *v3d_fence_get_driver_name(struct dma_fence *fence)
+{
+       return "v3d";
+}
+
+static const char *v3d_fence_get_timeline_name(struct dma_fence *fence)
+{
+       struct v3d_fence *f = to_v3d_fence(fence);
+
+       if (f->queue == V3D_BIN)
+               return "v3d-bin";
+       else
+               return "v3d-render";
+}
+
+static bool v3d_fence_enable_signaling(struct dma_fence *fence)
+{
+       return true;
+}
+
+static bool v3d_fence_signaled(struct dma_fence *fence)
+{
+       struct v3d_fence *f = to_v3d_fence(fence);
+       struct v3d_dev *v3d = to_v3d_dev(f->dev);
+
+       return v3d->queue[f->queue].finished_seqno >= f->seqno;
+}
+
+const struct dma_fence_ops v3d_fence_ops = {
+       .get_driver_name = v3d_fence_get_driver_name,
+       .get_timeline_name = v3d_fence_get_timeline_name,
+       .enable_signaling = v3d_fence_enable_signaling,
+       .signaled = v3d_fence_signaled,
+       .wait = dma_fence_default_wait,
+       .release = dma_fence_free,
+};
diff --git a/drivers/gpu/drm/v3d/v3d_gem.c b/drivers/gpu/drm/v3d/v3d_gem.c
new file mode 100644 (file)
index 0000000..b513f91
--- /dev/null
@@ -0,0 +1,668 @@
+// SPDX-License-Identifier: GPL-2.0+
+/* Copyright (C) 2014-2018 Broadcom */
+
+#include <drm/drmP.h>
+#include <drm/drm_syncobj.h>
+#include <linux/module.h>
+#include <linux/platform_device.h>
+#include <linux/pm_runtime.h>
+#include <linux/device.h>
+#include <linux/io.h>
+#include <linux/sched/signal.h>
+
+#include "uapi/drm/v3d_drm.h"
+#include "v3d_drv.h"
+#include "v3d_regs.h"
+#include "v3d_trace.h"
+
+static void
+v3d_init_core(struct v3d_dev *v3d, int core)
+{
+       /* Set OVRTMUOUT, which means that the texture sampler uniform
+        * configuration's tmu output type field is used, instead of
+        * using the hardware default behavior based on the texture
+        * type.  If you want the default behavior, you can still put
+        * "2" in the indirect texture state's output_type field.
+        */
+       V3D_CORE_WRITE(core, V3D_CTL_MISCCFG, V3D_MISCCFG_OVRTMUOUT);
+
+       /* Whenever we flush the L2T cache, we always want to flush
+        * the whole thing.
+        */
+       V3D_CORE_WRITE(core, V3D_CTL_L2TFLSTA, 0);
+       V3D_CORE_WRITE(core, V3D_CTL_L2TFLEND, ~0);
+}
+
+/* Sets invariant state for the HW. */
+static void
+v3d_init_hw_state(struct v3d_dev *v3d)
+{
+       v3d_init_core(v3d, 0);
+}
+
+static void
+v3d_idle_axi(struct v3d_dev *v3d, int core)
+{
+       V3D_CORE_WRITE(core, V3D_GMP_CFG, V3D_GMP_CFG_STOP_REQ);
+
+       if (wait_for((V3D_CORE_READ(core, V3D_GMP_STATUS) &
+                     (V3D_GMP_STATUS_RD_COUNT_MASK |
+                      V3D_GMP_STATUS_WR_COUNT_MASK |
+                      V3D_GMP_STATUS_CFG_BUSY)) == 0, 100)) {
+               DRM_ERROR("Failed to wait for safe GMP shutdown\n");
+       }
+}
+
+static void
+v3d_idle_gca(struct v3d_dev *v3d)
+{
+       if (v3d->ver >= 41)
+               return;
+
+       V3D_GCA_WRITE(V3D_GCA_SAFE_SHUTDOWN, V3D_GCA_SAFE_SHUTDOWN_EN);
+
+       if (wait_for((V3D_GCA_READ(V3D_GCA_SAFE_SHUTDOWN_ACK) &
+                     V3D_GCA_SAFE_SHUTDOWN_ACK_ACKED) ==
+                    V3D_GCA_SAFE_SHUTDOWN_ACK_ACKED, 100)) {
+               DRM_ERROR("Failed to wait for safe GCA shutdown\n");
+       }
+}
+
+static void
+v3d_reset_v3d(struct v3d_dev *v3d)
+{
+       int version = V3D_BRIDGE_READ(V3D_TOP_GR_BRIDGE_REVISION);
+
+       if (V3D_GET_FIELD(version, V3D_TOP_GR_BRIDGE_MAJOR) == 2) {
+               V3D_BRIDGE_WRITE(V3D_TOP_GR_BRIDGE_SW_INIT_0,
+                                V3D_TOP_GR_BRIDGE_SW_INIT_0_V3D_CLK_108_SW_INIT);
+               V3D_BRIDGE_WRITE(V3D_TOP_GR_BRIDGE_SW_INIT_0, 0);
+
+               /* GFXH-1383: The SW_INIT may cause a stray write to address 0
+                * of the unit, so reset it to its power-on value here.
+                */
+               V3D_WRITE(V3D_HUB_AXICFG, V3D_HUB_AXICFG_MAX_LEN_MASK);
+       } else {
+               WARN_ON_ONCE(V3D_GET_FIELD(version,
+                                          V3D_TOP_GR_BRIDGE_MAJOR) != 7);
+               V3D_BRIDGE_WRITE(V3D_TOP_GR_BRIDGE_SW_INIT_1,
+                                V3D_TOP_GR_BRIDGE_SW_INIT_1_V3D_CLK_108_SW_INIT);
+               V3D_BRIDGE_WRITE(V3D_TOP_GR_BRIDGE_SW_INIT_1, 0);
+       }
+
+       v3d_init_hw_state(v3d);
+}
+
+void
+v3d_reset(struct v3d_dev *v3d)
+{
+       struct drm_device *dev = &v3d->drm;
+
+       DRM_ERROR("Resetting GPU.\n");
+       trace_v3d_reset_begin(dev);
+
+       /* XXX: only needed for safe powerdown, not reset. */
+       if (false)
+               v3d_idle_axi(v3d, 0);
+
+       v3d_idle_gca(v3d);
+       v3d_reset_v3d(v3d);
+
+       v3d_mmu_set_page_table(v3d);
+       v3d_irq_reset(v3d);
+
+       trace_v3d_reset_end(dev);
+}
+
+static void
+v3d_flush_l3(struct v3d_dev *v3d)
+{
+       if (v3d->ver < 41) {
+               u32 gca_ctrl = V3D_GCA_READ(V3D_GCA_CACHE_CTRL);
+
+               V3D_GCA_WRITE(V3D_GCA_CACHE_CTRL,
+                             gca_ctrl | V3D_GCA_CACHE_CTRL_FLUSH);
+
+               if (v3d->ver < 33) {
+                       V3D_GCA_WRITE(V3D_GCA_CACHE_CTRL,
+                                     gca_ctrl & ~V3D_GCA_CACHE_CTRL_FLUSH);
+               }
+       }
+}
+
+/* Invalidates the (read-only) L2 cache. */
+static void
+v3d_invalidate_l2(struct v3d_dev *v3d, int core)
+{
+       V3D_CORE_WRITE(core, V3D_CTL_L2CACTL,
+                      V3D_L2CACTL_L2CCLR |
+                      V3D_L2CACTL_L2CENA);
+}
+
+static void
+v3d_invalidate_l1td(struct v3d_dev *v3d, int core)
+{
+       V3D_CORE_WRITE(core, V3D_CTL_L2TCACTL, V3D_L2TCACTL_TMUWCF);
+       if (wait_for(!(V3D_CORE_READ(core, V3D_CTL_L2TCACTL) &
+                      V3D_L2TCACTL_L2TFLS), 100)) {
+               DRM_ERROR("Timeout waiting for L1T write combiner flush\n");
+       }
+}
+
+/* Invalidates texture L2 cachelines */
+static void
+v3d_flush_l2t(struct v3d_dev *v3d, int core)
+{
+       v3d_invalidate_l1td(v3d, core);
+
+       V3D_CORE_WRITE(core, V3D_CTL_L2TCACTL,
+                      V3D_L2TCACTL_L2TFLS |
+                      V3D_SET_FIELD(V3D_L2TCACTL_FLM_FLUSH, V3D_L2TCACTL_FLM));
+       if (wait_for(!(V3D_CORE_READ(core, V3D_CTL_L2TCACTL) &
+                      V3D_L2TCACTL_L2TFLS), 100)) {
+               DRM_ERROR("Timeout waiting for L2T flush\n");
+       }
+}
+
+/* Invalidates the slice caches.  These are read-only caches. */
+static void
+v3d_invalidate_slices(struct v3d_dev *v3d, int core)
+{
+       V3D_CORE_WRITE(core, V3D_CTL_SLCACTL,
+                      V3D_SET_FIELD(0xf, V3D_SLCACTL_TVCCS) |
+                      V3D_SET_FIELD(0xf, V3D_SLCACTL_TDCCS) |
+                      V3D_SET_FIELD(0xf, V3D_SLCACTL_UCC) |
+                      V3D_SET_FIELD(0xf, V3D_SLCACTL_ICC));
+}
+
+/* Invalidates texture L2 cachelines */
+static void
+v3d_invalidate_l2t(struct v3d_dev *v3d, int core)
+{
+       V3D_CORE_WRITE(core,
+                      V3D_CTL_L2TCACTL,
+                      V3D_L2TCACTL_L2TFLS |
+                      V3D_SET_FIELD(V3D_L2TCACTL_FLM_CLEAR, V3D_L2TCACTL_FLM));
+       if (wait_for(!(V3D_CORE_READ(core, V3D_CTL_L2TCACTL) &
+                      V3D_L2TCACTL_L2TFLS), 100)) {
+               DRM_ERROR("Timeout waiting for L2T invalidate\n");
+       }
+}
+
+void
+v3d_invalidate_caches(struct v3d_dev *v3d)
+{
+       v3d_flush_l3(v3d);
+
+       v3d_invalidate_l2(v3d, 0);
+       v3d_invalidate_slices(v3d, 0);
+       v3d_flush_l2t(v3d, 0);
+}
+
+void
+v3d_flush_caches(struct v3d_dev *v3d)
+{
+       v3d_invalidate_l1td(v3d, 0);
+       v3d_invalidate_l2t(v3d, 0);
+}
+
+static void
+v3d_attach_object_fences(struct v3d_exec_info *exec)
+{
+       struct dma_fence *out_fence = &exec->render.base.s_fence->finished;
+       struct v3d_bo *bo;
+       int i;
+
+       for (i = 0; i < exec->bo_count; i++) {
+               bo = to_v3d_bo(&exec->bo[i]->base);
+
+               /* XXX: Use shared fences for read-only objects. */
+               reservation_object_add_excl_fence(bo->resv, out_fence);
+       }
+}
+
+static void
+v3d_unlock_bo_reservations(struct drm_device *dev,
+                          struct v3d_exec_info *exec,
+                          struct ww_acquire_ctx *acquire_ctx)
+{
+       int i;
+
+       for (i = 0; i < exec->bo_count; i++) {
+               struct v3d_bo *bo = to_v3d_bo(&exec->bo[i]->base);
+
+               ww_mutex_unlock(&bo->resv->lock);
+       }
+
+       ww_acquire_fini(acquire_ctx);
+}
+
+/* Takes the reservation lock on all the BOs being referenced, so that
+ * at queue submit time we can update the reservations.
+ *
+ * We don't lock the RCL the tile alloc/state BOs, or overflow memory
+ * (all of which are on exec->unref_list).  They're entirely private
+ * to v3d, so we don't attach dma-buf fences to them.
+ */
+static int
+v3d_lock_bo_reservations(struct drm_device *dev,
+                        struct v3d_exec_info *exec,
+                        struct ww_acquire_ctx *acquire_ctx)
+{
+       int contended_lock = -1;
+       int i, ret;
+       struct v3d_bo *bo;
+
+       ww_acquire_init(acquire_ctx, &reservation_ww_class);
+
+retry:
+       if (contended_lock != -1) {
+               bo = to_v3d_bo(&exec->bo[contended_lock]->base);
+               ret = ww_mutex_lock_slow_interruptible(&bo->resv->lock,
+                                                      acquire_ctx);
+               if (ret) {
+                       ww_acquire_done(acquire_ctx);
+                       return ret;
+               }
+       }
+
+       for (i = 0; i < exec->bo_count; i++) {
+               if (i == contended_lock)
+                       continue;
+
+               bo = to_v3d_bo(&exec->bo[i]->base);
+
+               ret = ww_mutex_lock_interruptible(&bo->resv->lock, acquire_ctx);
+               if (ret) {
+                       int j;
+
+                       for (j = 0; j < i; j++) {
+                               bo = to_v3d_bo(&exec->bo[j]->base);
+                               ww_mutex_unlock(&bo->resv->lock);
+                       }
+
+                       if (contended_lock != -1 && contended_lock >= i) {
+                               bo = to_v3d_bo(&exec->bo[contended_lock]->base);
+
+                               ww_mutex_unlock(&bo->resv->lock);
+                       }
+
+                       if (ret == -EDEADLK) {
+                               contended_lock = i;
+                               goto retry;
+                       }
+
+                       ww_acquire_done(acquire_ctx);
+                       return ret;
+               }
+       }
+
+       ww_acquire_done(acquire_ctx);
+
+       /* Reserve space for our shared (read-only) fence references,
+        * before we commit the CL to the hardware.
+        */
+       for (i = 0; i < exec->bo_count; i++) {
+               bo = to_v3d_bo(&exec->bo[i]->base);
+
+               ret = reservation_object_reserve_shared(bo->resv);
+               if (ret) {
+                       v3d_unlock_bo_reservations(dev, exec, acquire_ctx);
+                       return ret;
+               }
+       }
+
+       return 0;
+}
+
+/**
+ * v3d_cl_lookup_bos() - Sets up exec->bo[] with the GEM objects
+ * referenced by the job.
+ * @dev: DRM device
+ * @file_priv: DRM file for this fd
+ * @exec: V3D job being set up
+ *
+ * The command validator needs to reference BOs by their index within
+ * the submitted job's BO list.  This does the validation of the job's
+ * BO list and reference counting for the lifetime of the job.
+ *
+ * Note that this function doesn't need to unreference the BOs on
+ * failure, because that will happen at v3d_exec_cleanup() time.
+ */
+static int
+v3d_cl_lookup_bos(struct drm_device *dev,
+                 struct drm_file *file_priv,
+                 struct drm_v3d_submit_cl *args,
+                 struct v3d_exec_info *exec)
+{
+       u32 *handles;
+       int ret = 0;
+       int i;
+
+       exec->bo_count = args->bo_handle_count;
+
+       if (!exec->bo_count) {
+               /* See comment on bo_index for why we have to check
+                * this.
+                */
+               DRM_DEBUG("Rendering requires BOs\n");
+               return -EINVAL;
+       }
+
+       exec->bo = kvmalloc_array(exec->bo_count,
+                                 sizeof(struct drm_gem_cma_object *),
+                                 GFP_KERNEL | __GFP_ZERO);
+       if (!exec->bo) {
+               DRM_DEBUG("Failed to allocate validated BO pointers\n");
+               return -ENOMEM;
+       }
+
+       handles = kvmalloc_array(exec->bo_count, sizeof(u32), GFP_KERNEL);
+       if (!handles) {
+               ret = -ENOMEM;
+               DRM_DEBUG("Failed to allocate incoming GEM handles\n");
+               goto fail;
+       }
+
+       if (copy_from_user(handles,
+                          (void __user *)(uintptr_t)args->bo_handles,
+                          exec->bo_count * sizeof(u32))) {
+               ret = -EFAULT;
+               DRM_DEBUG("Failed to copy in GEM handles\n");
+               goto fail;
+       }
+
+       spin_lock(&file_priv->table_lock);
+       for (i = 0; i < exec->bo_count; i++) {
+               struct drm_gem_object *bo = idr_find(&file_priv->object_idr,
+                                                    handles[i]);
+               if (!bo) {
+                       DRM_DEBUG("Failed to look up GEM BO %d: %d\n",
+                                 i, handles[i]);
+                       ret = -ENOENT;
+                       spin_unlock(&file_priv->table_lock);
+                       goto fail;
+               }
+               drm_gem_object_get(bo);
+               exec->bo[i] = to_v3d_bo(bo);
+       }
+       spin_unlock(&file_priv->table_lock);
+
+fail:
+       kvfree(handles);
+       return ret;
+}
+
+static void
+v3d_exec_cleanup(struct kref *ref)
+{
+       struct v3d_exec_info *exec = container_of(ref, struct v3d_exec_info,
+                                                 refcount);
+       struct v3d_dev *v3d = exec->v3d;
+       unsigned int i;
+       struct v3d_bo *bo, *save;
+
+       dma_fence_put(exec->bin.in_fence);
+       dma_fence_put(exec->render.in_fence);
+
+       dma_fence_put(exec->bin.done_fence);
+       dma_fence_put(exec->render.done_fence);
+
+       dma_fence_put(exec->bin_done_fence);
+
+       for (i = 0; i < exec->bo_count; i++)
+               drm_gem_object_put_unlocked(&exec->bo[i]->base);
+       kvfree(exec->bo);
+
+       list_for_each_entry_safe(bo, save, &exec->unref_list, unref_head) {
+               drm_gem_object_put_unlocked(&bo->base);
+       }
+
+       pm_runtime_mark_last_busy(v3d->dev);
+       pm_runtime_put_autosuspend(v3d->dev);
+
+       kfree(exec);
+}
+
+void v3d_exec_put(struct v3d_exec_info *exec)
+{
+       kref_put(&exec->refcount, v3d_exec_cleanup);
+}
+
+int
+v3d_wait_bo_ioctl(struct drm_device *dev, void *data,
+                 struct drm_file *file_priv)
+{
+       int ret;
+       struct drm_v3d_wait_bo *args = data;
+       struct drm_gem_object *gem_obj;
+       struct v3d_bo *bo;
+       ktime_t start = ktime_get();
+       u64 delta_ns;
+       unsigned long timeout_jiffies =
+               nsecs_to_jiffies_timeout(args->timeout_ns);
+
+       if (args->pad != 0)
+               return -EINVAL;
+
+       gem_obj = drm_gem_object_lookup(file_priv, args->handle);
+       if (!gem_obj) {
+               DRM_DEBUG("Failed to look up GEM BO %d\n", args->handle);
+               return -EINVAL;
+       }
+       bo = to_v3d_bo(gem_obj);
+
+       ret = reservation_object_wait_timeout_rcu(bo->resv,
+                                                 true, true,
+                                                 timeout_jiffies);
+
+       if (ret == 0)
+               ret = -ETIME;
+       else if (ret > 0)
+               ret = 0;
+
+       /* Decrement the user's timeout, in case we got interrupted
+        * such that the ioctl will be restarted.
+        */
+       delta_ns = ktime_to_ns(ktime_sub(ktime_get(), start));
+       if (delta_ns < args->timeout_ns)
+               args->timeout_ns -= delta_ns;
+       else
+               args->timeout_ns = 0;
+
+       /* Asked to wait beyond the jiffie/scheduler precision? */
+       if (ret == -ETIME && args->timeout_ns)
+               ret = -EAGAIN;
+
+       drm_gem_object_put_unlocked(gem_obj);
+
+       return ret;
+}
+
+/**
+ * v3d_submit_cl_ioctl() - Submits a job (frame) to the V3D.
+ * @dev: DRM device
+ * @data: ioctl argument
+ * @file_priv: DRM file for this fd
+ *
+ * This is the main entrypoint for userspace to submit a 3D frame to
+ * the GPU.  Userspace provides the binner command list (if
+ * applicable), and the kernel sets up the render command list to draw
+ * to the framebuffer described in the ioctl, using the command lists
+ * that the 3D engine's binner will produce.
+ */
+int
+v3d_submit_cl_ioctl(struct drm_device *dev, void *data,
+                   struct drm_file *file_priv)
+{
+       struct v3d_dev *v3d = to_v3d_dev(dev);
+       struct v3d_file_priv *v3d_priv = file_priv->driver_priv;
+       struct drm_v3d_submit_cl *args = data;
+       struct v3d_exec_info *exec;
+       struct ww_acquire_ctx acquire_ctx;
+       struct drm_syncobj *sync_out;
+       int ret = 0;
+
+       if (args->pad != 0) {
+               DRM_INFO("pad must be zero: %d\n", args->pad);
+               return -EINVAL;
+       }
+
+       exec = kcalloc(1, sizeof(*exec), GFP_KERNEL);
+       if (!exec)
+               return -ENOMEM;
+
+       ret = pm_runtime_get_sync(v3d->dev);
+       if (ret < 0) {
+               kfree(exec);
+               return ret;
+       }
+
+       kref_init(&exec->refcount);
+
+       ret = drm_syncobj_find_fence(file_priv, args->in_sync_bcl,
+                                    &exec->bin.in_fence);
+       if (ret == -EINVAL)
+               goto fail;
+
+       ret = drm_syncobj_find_fence(file_priv, args->in_sync_rcl,
+                                    &exec->render.in_fence);
+       if (ret == -EINVAL)
+               goto fail;
+
+       exec->qma = args->qma;
+       exec->qms = args->qms;
+       exec->qts = args->qts;
+       exec->bin.exec = exec;
+       exec->bin.start = args->bcl_start;
+       exec->bin.end = args->bcl_end;
+       exec->render.exec = exec;
+       exec->render.start = args->rcl_start;
+       exec->render.end = args->rcl_end;
+       exec->v3d = v3d;
+       INIT_LIST_HEAD(&exec->unref_list);
+
+       ret = v3d_cl_lookup_bos(dev, file_priv, args, exec);
+       if (ret)
+               goto fail;
+
+       ret = v3d_lock_bo_reservations(dev, exec, &acquire_ctx);
+       if (ret)
+               goto fail;
+
+       if (exec->bin.start != exec->bin.end) {
+               ret = drm_sched_job_init(&exec->bin.base,
+                                        &v3d->queue[V3D_BIN].sched,
+                                        &v3d_priv->sched_entity[V3D_BIN],
+                                        v3d_priv);
+               if (ret)
+                       goto fail_unreserve;
+
+               exec->bin_done_fence =
+                       dma_fence_get(&exec->bin.base.s_fence->finished);
+
+               kref_get(&exec->refcount); /* put by scheduler job completion */
+               drm_sched_entity_push_job(&exec->bin.base,
+                                         &v3d_priv->sched_entity[V3D_BIN]);
+       }
+
+       ret = drm_sched_job_init(&exec->render.base,
+                                &v3d->queue[V3D_RENDER].sched,
+                                &v3d_priv->sched_entity[V3D_RENDER],
+                                v3d_priv);
+       if (ret)
+               goto fail_unreserve;
+
+       kref_get(&exec->refcount); /* put by scheduler job completion */
+       drm_sched_entity_push_job(&exec->render.base,
+                                 &v3d_priv->sched_entity[V3D_RENDER]);
+
+       v3d_attach_object_fences(exec);
+
+       v3d_unlock_bo_reservations(dev, exec, &acquire_ctx);
+
+       /* Update the return sync object for the */
+       sync_out = drm_syncobj_find(file_priv, args->out_sync);
+       if (sync_out) {
+               drm_syncobj_replace_fence(sync_out,
+                                         &exec->render.base.s_fence->finished);
+               drm_syncobj_put(sync_out);
+       }
+
+       v3d_exec_put(exec);
+
+       return 0;
+
+fail_unreserve:
+       v3d_unlock_bo_reservations(dev, exec, &acquire_ctx);
+fail:
+       v3d_exec_put(exec);
+
+       return ret;
+}
+
+int
+v3d_gem_init(struct drm_device *dev)
+{
+       struct v3d_dev *v3d = to_v3d_dev(dev);
+       u32 pt_size = 4096 * 1024;
+       int ret, i;
+
+       for (i = 0; i < V3D_MAX_QUEUES; i++)
+               v3d->queue[i].fence_context = dma_fence_context_alloc(1);
+
+       spin_lock_init(&v3d->mm_lock);
+       spin_lock_init(&v3d->job_lock);
+       mutex_init(&v3d->bo_lock);
+       mutex_init(&v3d->reset_lock);
+
+       /* Note: We don't allocate address 0.  Various bits of HW
+        * treat 0 as special, such as the occlusion query counters
+        * where 0 means "disabled".
+        */
+       drm_mm_init(&v3d->mm, 1, pt_size / sizeof(u32) - 1);
+
+       v3d->pt = dma_alloc_wc(v3d->dev, pt_size,
+                              &v3d->pt_paddr,
+                              GFP_KERNEL | __GFP_NOWARN | __GFP_ZERO);
+       if (!v3d->pt) {
+               drm_mm_takedown(&v3d->mm);
+               dev_err(v3d->dev,
+                       "Failed to allocate page tables. "
+                       "Please ensure you have CMA enabled.\n");
+               return -ENOMEM;
+       }
+
+       v3d_init_hw_state(v3d);
+       v3d_mmu_set_page_table(v3d);
+
+       ret = v3d_sched_init(v3d);
+       if (ret) {
+               drm_mm_takedown(&v3d->mm);
+               dma_free_coherent(v3d->dev, 4096 * 1024, (void *)v3d->pt,
+                                 v3d->pt_paddr);
+       }
+
+       return 0;
+}
+
+void
+v3d_gem_destroy(struct drm_device *dev)
+{
+       struct v3d_dev *v3d = to_v3d_dev(dev);
+       enum v3d_queue q;
+
+       v3d_sched_fini(v3d);
+
+       /* Waiting for exec to finish would need to be done before
+        * unregistering V3D.
+        */
+       for (q = 0; q < V3D_MAX_QUEUES; q++) {
+               WARN_ON(v3d->queue[q].emit_seqno !=
+                       v3d->queue[q].finished_seqno);
+       }
+
+       drm_mm_takedown(&v3d->mm);
+
+       dma_free_coherent(v3d->dev, 4096 * 1024, (void *)v3d->pt, v3d->pt_paddr);
+}
diff --git a/drivers/gpu/drm/v3d/v3d_irq.c b/drivers/gpu/drm/v3d/v3d_irq.c
new file mode 100644 (file)
index 0000000..77e1fa0
--- /dev/null
@@ -0,0 +1,206 @@
+// SPDX-License-Identifier: GPL-2.0+
+/* Copyright (C) 2014-2018 Broadcom */
+
+/**
+ * DOC: Interrupt management for the V3D engine
+ *
+ * When we take a binning or rendering flush done interrupt, we need
+ * to signal the fence for that job so that the scheduler can queue up
+ * the next one and unblock any waiters.
+ *
+ * When we take the binner out of memory interrupt, we need to
+ * allocate some new memory and pass it to the binner so that the
+ * current job can make progress.
+ */
+
+#include "v3d_drv.h"
+#include "v3d_regs.h"
+
+#define V3D_CORE_IRQS ((u32)(V3D_INT_OUTOMEM | \
+                            V3D_INT_FLDONE |   \
+                            V3D_INT_FRDONE |   \
+                            V3D_INT_GMPV))
+
+#define V3D_HUB_IRQS ((u32)(V3D_HUB_INT_MMU_WRV |      \
+                           V3D_HUB_INT_MMU_PTI |       \
+                           V3D_HUB_INT_MMU_CAP))
+
+static void
+v3d_overflow_mem_work(struct work_struct *work)
+{
+       struct v3d_dev *v3d =
+               container_of(work, struct v3d_dev, overflow_mem_work);
+       struct drm_device *dev = &v3d->drm;
+       struct v3d_bo *bo = v3d_bo_create(dev, NULL /* XXX: GMP */, 256 * 1024);
+       unsigned long irqflags;
+
+       if (IS_ERR(bo)) {
+               DRM_ERROR("Couldn't allocate binner overflow mem\n");
+               return;
+       }
+
+       /* We lost a race, and our work task came in after the bin job
+        * completed and exited.  This can happen because the HW
+        * signals OOM before it's fully OOM, so the binner might just
+        * barely complete.
+        *
+        * If we lose the race and our work task comes in after a new
+        * bin job got scheduled, that's fine.  We'll just give them
+        * some binner pool anyway.
+        */
+       spin_lock_irqsave(&v3d->job_lock, irqflags);
+       if (!v3d->bin_job) {
+               spin_unlock_irqrestore(&v3d->job_lock, irqflags);
+               goto out;
+       }
+
+       drm_gem_object_get(&bo->base);
+       list_add_tail(&bo->unref_head, &v3d->bin_job->unref_list);
+       spin_unlock_irqrestore(&v3d->job_lock, irqflags);
+
+       V3D_CORE_WRITE(0, V3D_PTB_BPOA, bo->node.start << PAGE_SHIFT);
+       V3D_CORE_WRITE(0, V3D_PTB_BPOS, bo->base.size);
+
+out:
+       drm_gem_object_put_unlocked(&bo->base);
+}
+
+static irqreturn_t
+v3d_irq(int irq, void *arg)
+{
+       struct v3d_dev *v3d = arg;
+       u32 intsts;
+       irqreturn_t status = IRQ_NONE;
+
+       intsts = V3D_CORE_READ(0, V3D_CTL_INT_STS);
+
+       /* Acknowledge the interrupts we're handling here. */
+       V3D_CORE_WRITE(0, V3D_CTL_INT_CLR, intsts);
+
+       if (intsts & V3D_INT_OUTOMEM) {
+               /* Note that the OOM status is edge signaled, so the
+                * interrupt won't happen again until the we actually
+                * add more memory.
+                */
+               schedule_work(&v3d->overflow_mem_work);
+               status = IRQ_HANDLED;
+       }
+
+       if (intsts & V3D_INT_FLDONE) {
+               v3d->queue[V3D_BIN].finished_seqno++;
+               dma_fence_signal(v3d->bin_job->bin.done_fence);
+               status = IRQ_HANDLED;
+       }
+
+       if (intsts & V3D_INT_FRDONE) {
+               v3d->queue[V3D_RENDER].finished_seqno++;
+               dma_fence_signal(v3d->render_job->render.done_fence);
+
+               status = IRQ_HANDLED;
+       }
+
+       /* We shouldn't be triggering these if we have GMP in
+        * always-allowed mode.
+        */
+       if (intsts & V3D_INT_GMPV)
+               dev_err(v3d->dev, "GMP violation\n");
+
+       return status;
+}
+
+static irqreturn_t
+v3d_hub_irq(int irq, void *arg)
+{
+       struct v3d_dev *v3d = arg;
+       u32 intsts;
+       irqreturn_t status = IRQ_NONE;
+
+       intsts = V3D_READ(V3D_HUB_INT_STS);
+
+       /* Acknowledge the interrupts we're handling here. */
+       V3D_WRITE(V3D_HUB_INT_CLR, intsts);
+
+       if (intsts & (V3D_HUB_INT_MMU_WRV |
+                     V3D_HUB_INT_MMU_PTI |
+                     V3D_HUB_INT_MMU_CAP)) {
+               u32 axi_id = V3D_READ(V3D_MMU_VIO_ID);
+               u64 vio_addr = (u64)V3D_READ(V3D_MMU_VIO_ADDR) << 8;
+
+               dev_err(v3d->dev, "MMU error from client %d at 0x%08llx%s%s%s\n",
+                       axi_id, (long long)vio_addr,
+                       ((intsts & V3D_HUB_INT_MMU_WRV) ?
+                        ", write violation" : ""),
+                       ((intsts & V3D_HUB_INT_MMU_PTI) ?
+                        ", pte invalid" : ""),
+                       ((intsts & V3D_HUB_INT_MMU_CAP) ?
+                        ", cap exceeded" : ""));
+               status = IRQ_HANDLED;
+       }
+
+       return status;
+}
+
+void
+v3d_irq_init(struct v3d_dev *v3d)
+{
+       int ret, core;
+
+       INIT_WORK(&v3d->overflow_mem_work, v3d_overflow_mem_work);
+
+       /* Clear any pending interrupts someone might have left around
+        * for us.
+        */
+       for (core = 0; core < v3d->cores; core++)
+               V3D_CORE_WRITE(core, V3D_CTL_INT_CLR, V3D_CORE_IRQS);
+       V3D_WRITE(V3D_HUB_INT_CLR, V3D_HUB_IRQS);
+
+       ret = devm_request_irq(v3d->dev, platform_get_irq(v3d->pdev, 0),
+                              v3d_hub_irq, IRQF_SHARED,
+                              "v3d_hub", v3d);
+       ret = devm_request_irq(v3d->dev, platform_get_irq(v3d->pdev, 1),
+                              v3d_irq, IRQF_SHARED,
+                              "v3d_core0", v3d);
+       if (ret)
+               dev_err(v3d->dev, "IRQ setup failed: %d\n", ret);
+
+       v3d_irq_enable(v3d);
+}
+
+void
+v3d_irq_enable(struct v3d_dev *v3d)
+{
+       int core;
+
+       /* Enable our set of interrupts, masking out any others. */
+       for (core = 0; core < v3d->cores; core++) {
+               V3D_CORE_WRITE(core, V3D_CTL_INT_MSK_SET, ~V3D_CORE_IRQS);
+               V3D_CORE_WRITE(core, V3D_CTL_INT_MSK_CLR, V3D_CORE_IRQS);
+       }
+
+       V3D_WRITE(V3D_HUB_INT_MSK_SET, ~V3D_HUB_IRQS);
+       V3D_WRITE(V3D_HUB_INT_MSK_CLR, V3D_HUB_IRQS);
+}
+
+void
+v3d_irq_disable(struct v3d_dev *v3d)
+{
+       int core;
+
+       /* Disable all interrupts. */
+       for (core = 0; core < v3d->cores; core++)
+               V3D_CORE_WRITE(core, V3D_CTL_INT_MSK_SET, ~0);
+       V3D_WRITE(V3D_HUB_INT_MSK_SET, ~0);
+
+       /* Clear any pending interrupts we might have left. */
+       for (core = 0; core < v3d->cores; core++)
+               V3D_CORE_WRITE(core, V3D_CTL_INT_CLR, V3D_CORE_IRQS);
+       V3D_WRITE(V3D_HUB_INT_CLR, V3D_HUB_IRQS);
+
+       cancel_work_sync(&v3d->overflow_mem_work);
+}
+
+/** Reinitializes interrupt registers when a GPU reset is performed. */
+void v3d_irq_reset(struct v3d_dev *v3d)
+{
+       v3d_irq_enable(v3d);
+}
diff --git a/drivers/gpu/drm/v3d/v3d_mmu.c b/drivers/gpu/drm/v3d/v3d_mmu.c
new file mode 100644 (file)
index 0000000..b00f97c
--- /dev/null
@@ -0,0 +1,122 @@
+// SPDX-License-Identifier: GPL-2.0+
+/* Copyright (C) 2017-2018 Broadcom */
+
+/**
+ * DOC: Broadcom V3D MMU
+ *
+ * The V3D 3.x hardware (compared to VC4) now includes an MMU.  It has
+ * a single level of page tables for the V3D's 4GB address space to
+ * map to AXI bus addresses, thus it could need up to 4MB of
+ * physically contiguous memory to store the PTEs.
+ *
+ * Because the 4MB of contiguous memory for page tables is precious,
+ * and switching between them is expensive, we load all BOs into the
+ * same 4GB address space.
+ *
+ * To protect clients from each other, we should use the GMP to
+ * quickly mask out (at 128kb granularity) what pages are available to
+ * each client.  This is not yet implemented.
+ */
+
+#include "v3d_drv.h"
+#include "v3d_regs.h"
+
+#define V3D_MMU_PAGE_SHIFT 12
+
+/* Note: All PTEs for the 1MB superpage must be filled with the
+ * superpage bit set.
+ */
+#define V3D_PTE_SUPERPAGE BIT(31)
+#define V3D_PTE_WRITEABLE BIT(29)
+#define V3D_PTE_VALID BIT(28)
+
+static int v3d_mmu_flush_all(struct v3d_dev *v3d)
+{
+       int ret;
+
+       /* Make sure that another flush isn't already running when we
+        * start this one.
+        */
+       ret = wait_for(!(V3D_READ(V3D_MMU_CTL) &
+                        V3D_MMU_CTL_TLB_CLEARING), 100);
+       if (ret)
+               dev_err(v3d->dev, "TLB clear wait idle pre-wait failed\n");
+
+       V3D_WRITE(V3D_MMU_CTL, V3D_READ(V3D_MMU_CTL) |
+                 V3D_MMU_CTL_TLB_CLEAR);
+
+       V3D_WRITE(V3D_MMUC_CONTROL,
+                 V3D_MMUC_CONTROL_FLUSH |
+                 V3D_MMUC_CONTROL_ENABLE);
+
+       ret = wait_for(!(V3D_READ(V3D_MMU_CTL) &
+                        V3D_MMU_CTL_TLB_CLEARING), 100);
+       if (ret) {
+               dev_err(v3d->dev, "TLB clear wait idle failed\n");
+               return ret;
+       }
+
+       ret = wait_for(!(V3D_READ(V3D_MMUC_CONTROL) &
+                        V3D_MMUC_CONTROL_FLUSHING), 100);
+       if (ret)
+               dev_err(v3d->dev, "MMUC flush wait idle failed\n");
+
+       return ret;
+}
+
+int v3d_mmu_set_page_table(struct v3d_dev *v3d)
+{
+       V3D_WRITE(V3D_MMU_PT_PA_BASE, v3d->pt_paddr >> V3D_MMU_PAGE_SHIFT);
+       V3D_WRITE(V3D_MMU_CTL,
+                 V3D_MMU_CTL_ENABLE |
+                 V3D_MMU_CTL_PT_INVALID |
+                 V3D_MMU_CTL_PT_INVALID_ABORT |
+                 V3D_MMU_CTL_WRITE_VIOLATION_ABORT |
+                 V3D_MMU_CTL_CAP_EXCEEDED_ABORT);
+       V3D_WRITE(V3D_MMU_ILLEGAL_ADDR,
+                 (v3d->mmu_scratch_paddr >> V3D_MMU_PAGE_SHIFT) |
+                 V3D_MMU_ILLEGAL_ADDR_ENABLE);
+       V3D_WRITE(V3D_MMUC_CONTROL, V3D_MMUC_CONTROL_ENABLE);
+
+       return v3d_mmu_flush_all(v3d);
+}
+
+void v3d_mmu_insert_ptes(struct v3d_bo *bo)
+{
+       struct v3d_dev *v3d = to_v3d_dev(bo->base.dev);
+       u32 page = bo->node.start;
+       u32 page_prot = V3D_PTE_WRITEABLE | V3D_PTE_VALID;
+       unsigned int count;
+       struct scatterlist *sgl;
+
+       for_each_sg(bo->sgt->sgl, sgl, bo->sgt->nents, count) {
+               u32 page_address = sg_dma_address(sgl) >> V3D_MMU_PAGE_SHIFT;
+               u32 pte = page_prot | page_address;
+               u32 i;
+
+               BUG_ON(page_address + (sg_dma_len(sgl) >> V3D_MMU_PAGE_SHIFT) >=
+                      BIT(24));
+
+               for (i = 0; i < sg_dma_len(sgl) >> V3D_MMU_PAGE_SHIFT; i++)
+                       v3d->pt[page++] = pte + i;
+       }
+
+       WARN_ON_ONCE(page - bo->node.start !=
+                    bo->base.size >> V3D_MMU_PAGE_SHIFT);
+
+       if (v3d_mmu_flush_all(v3d))
+               dev_err(v3d->dev, "MMU flush timeout\n");
+}
+
+void v3d_mmu_remove_ptes(struct v3d_bo *bo)
+{
+       struct v3d_dev *v3d = to_v3d_dev(bo->base.dev);
+       u32 npages = bo->base.size >> V3D_MMU_PAGE_SHIFT;
+       u32 page;
+
+       for (page = bo->node.start; page < bo->node.start + npages; page++)
+               v3d->pt[page] = 0;
+
+       if (v3d_mmu_flush_all(v3d))
+               dev_err(v3d->dev, "MMU flush timeout\n");
+}
diff --git a/drivers/gpu/drm/v3d/v3d_regs.h b/drivers/gpu/drm/v3d/v3d_regs.h
new file mode 100644 (file)
index 0000000..fc13282
--- /dev/null
@@ -0,0 +1,295 @@
+// SPDX-License-Identifier: GPL-2.0+
+/* Copyright (C) 2017-2018 Broadcom */
+
+#ifndef V3D_REGS_H
+#define V3D_REGS_H
+
+#include <linux/bitops.h>
+
+#define V3D_MASK(high, low) ((u32)GENMASK(high, low))
+/* Using the GNU statement expression extension */
+#define V3D_SET_FIELD(value, field)                                    \
+       ({                                                              \
+               u32 fieldval = (value) << field##_SHIFT;                \
+               WARN_ON((fieldval & ~field##_MASK) != 0);               \
+               fieldval & field##_MASK;                                \
+        })
+
+#define V3D_GET_FIELD(word, field) (((word) & field##_MASK) >>         \
+                                   field##_SHIFT)
+
+/* Hub registers for shared hardware between V3D cores. */
+
+#define V3D_HUB_AXICFG                                 0x00000
+# define V3D_HUB_AXICFG_MAX_LEN_MASK                   V3D_MASK(3, 0)
+# define V3D_HUB_AXICFG_MAX_LEN_SHIFT                  0
+#define V3D_HUB_UIFCFG                                 0x00004
+#define V3D_HUB_IDENT0                                 0x00008
+
+#define V3D_HUB_IDENT1                                 0x0000c
+# define V3D_HUB_IDENT1_WITH_MSO                       BIT(19)
+# define V3D_HUB_IDENT1_WITH_TSY                       BIT(18)
+# define V3D_HUB_IDENT1_WITH_TFU                       BIT(17)
+# define V3D_HUB_IDENT1_WITH_L3C                       BIT(16)
+# define V3D_HUB_IDENT1_NHOSTS_MASK                    V3D_MASK(15, 12)
+# define V3D_HUB_IDENT1_NHOSTS_SHIFT                   12
+# define V3D_HUB_IDENT1_NCORES_MASK                    V3D_MASK(11, 8)
+# define V3D_HUB_IDENT1_NCORES_SHIFT                   8
+# define V3D_HUB_IDENT1_REV_MASK                       V3D_MASK(7, 4)
+# define V3D_HUB_IDENT1_REV_SHIFT                      4
+# define V3D_HUB_IDENT1_TVER_MASK                      V3D_MASK(3, 0)
+# define V3D_HUB_IDENT1_TVER_SHIFT                     0
+
+#define V3D_HUB_IDENT2                                 0x00010
+# define V3D_HUB_IDENT2_WITH_MMU                       BIT(8)
+# define V3D_HUB_IDENT2_L3C_NKB_MASK                   V3D_MASK(7, 0)
+# define V3D_HUB_IDENT2_L3C_NKB_SHIFT                  0
+
+#define V3D_HUB_IDENT3                                 0x00014
+# define V3D_HUB_IDENT3_IPREV_MASK                     V3D_MASK(15, 8)
+# define V3D_HUB_IDENT3_IPREV_SHIFT                    8
+# define V3D_HUB_IDENT3_IPIDX_MASK                     V3D_MASK(7, 0)
+# define V3D_HUB_IDENT3_IPIDX_SHIFT                    0
+
+#define V3D_HUB_INT_STS                                0x00050
+#define V3D_HUB_INT_SET                                0x00054
+#define V3D_HUB_INT_CLR                                0x00058
+#define V3D_HUB_INT_MSK_STS                            0x0005c
+#define V3D_HUB_INT_MSK_SET                            0x00060
+#define V3D_HUB_INT_MSK_CLR                            0x00064
+# define V3D_HUB_INT_MMU_WRV                           BIT(5)
+# define V3D_HUB_INT_MMU_PTI                           BIT(4)
+# define V3D_HUB_INT_MMU_CAP                           BIT(3)
+# define V3D_HUB_INT_MSO                               BIT(2)
+# define V3D_HUB_INT_TFUC                              BIT(1)
+# define V3D_HUB_INT_TFUF                              BIT(0)
+
+#define V3D_GCA_CACHE_CTRL                             0x0000c
+# define V3D_GCA_CACHE_CTRL_FLUSH                      BIT(0)
+
+#define V3D_GCA_SAFE_SHUTDOWN                          0x000b0
+# define V3D_GCA_SAFE_SHUTDOWN_EN                      BIT(0)
+
+#define V3D_GCA_SAFE_SHUTDOWN_ACK                      0x000b4
+# define V3D_GCA_SAFE_SHUTDOWN_ACK_ACKED               3
+
+# define V3D_TOP_GR_BRIDGE_REVISION                    0x00000
+# define V3D_TOP_GR_BRIDGE_MAJOR_MASK                  V3D_MASK(15, 8)
+# define V3D_TOP_GR_BRIDGE_MAJOR_SHIFT                 8
+# define V3D_TOP_GR_BRIDGE_MINOR_MASK                  V3D_MASK(7, 0)
+# define V3D_TOP_GR_BRIDGE_MINOR_SHIFT                 0
+
+/* 7268 reset reg */
+# define V3D_TOP_GR_BRIDGE_SW_INIT_0                   0x00008
+# define V3D_TOP_GR_BRIDGE_SW_INIT_0_V3D_CLK_108_SW_INIT BIT(0)
+/* 7278 reset reg */
+# define V3D_TOP_GR_BRIDGE_SW_INIT_1                   0x0000c
+# define V3D_TOP_GR_BRIDGE_SW_INIT_1_V3D_CLK_108_SW_INIT BIT(0)
+
+/* Per-MMU registers. */
+
+#define V3D_MMUC_CONTROL                               0x01000
+# define V3D_MMUC_CONTROL_CLEAR                        BIT(3)
+# define V3D_MMUC_CONTROL_FLUSHING                     BIT(2)
+# define V3D_MMUC_CONTROL_FLUSH                        BIT(1)
+# define V3D_MMUC_CONTROL_ENABLE                       BIT(0)
+
+#define V3D_MMU_CTL                                    0x01200
+# define V3D_MMU_CTL_CAP_EXCEEDED                      BIT(27)
+# define V3D_MMU_CTL_CAP_EXCEEDED_ABORT                BIT(26)
+# define V3D_MMU_CTL_CAP_EXCEEDED_INT                  BIT(25)
+# define V3D_MMU_CTL_CAP_EXCEEDED_EXCEPTION            BIT(24)
+# define V3D_MMU_CTL_PT_INVALID                        BIT(20)
+# define V3D_MMU_CTL_PT_INVALID_ABORT                  BIT(19)
+# define V3D_MMU_CTL_PT_INVALID_INT                    BIT(18)
+# define V3D_MMU_CTL_PT_INVALID_EXCEPTION              BIT(17)
+# define V3D_MMU_CTL_WRITE_VIOLATION                   BIT(16)
+# define V3D_MMU_CTL_WRITE_VIOLATION_ABORT             BIT(11)
+# define V3D_MMU_CTL_WRITE_VIOLATION_INT               BIT(10)
+# define V3D_MMU_CTL_WRITE_VIOLATION_EXCEPTION         BIT(9)
+# define V3D_MMU_CTL_TLB_CLEARING                      BIT(7)
+# define V3D_MMU_CTL_TLB_STATS_CLEAR                   BIT(3)
+# define V3D_MMU_CTL_TLB_CLEAR                         BIT(2)
+# define V3D_MMU_CTL_TLB_STATS_ENABLE                  BIT(1)
+# define V3D_MMU_CTL_ENABLE                            BIT(0)
+
+#define V3D_MMU_PT_PA_BASE                             0x01204
+#define V3D_MMU_HIT                                    0x01208
+#define V3D_MMU_MISSES                                 0x0120c
+#define V3D_MMU_STALLS                                 0x01210
+
+#define V3D_MMU_ADDR_CAP                               0x01214
+# define V3D_MMU_ADDR_CAP_ENABLE                       BIT(31)
+# define V3D_MMU_ADDR_CAP_MPAGE_MASK                   V3D_MASK(11, 0)
+# define V3D_MMU_ADDR_CAP_MPAGE_SHIFT                  0
+
+#define V3D_MMU_SHOOT_DOWN                             0x01218
+# define V3D_MMU_SHOOT_DOWN_SHOOTING                   BIT(29)
+# define V3D_MMU_SHOOT_DOWN_SHOOT                      BIT(28)
+# define V3D_MMU_SHOOT_DOWN_PAGE_MASK                  V3D_MASK(27, 0)
+# define V3D_MMU_SHOOT_DOWN_PAGE_SHIFT                 0
+
+#define V3D_MMU_BYPASS_START                           0x0121c
+#define V3D_MMU_BYPASS_END                             0x01220
+
+/* AXI ID of the access that faulted */
+#define V3D_MMU_VIO_ID                                 0x0122c
+
+/* Address for illegal PTEs to return */
+#define V3D_MMU_ILLEGAL_ADDR                           0x01230
+# define V3D_MMU_ILLEGAL_ADDR_ENABLE                   BIT(31)
+
+/* Address that faulted */
+#define V3D_MMU_VIO_ADDR                               0x01234
+
+/* Per-V3D-core registers */
+
+#define V3D_CTL_IDENT0                                 0x00000
+# define V3D_IDENT0_VER_MASK                           V3D_MASK(31, 24)
+# define V3D_IDENT0_VER_SHIFT                          24
+
+#define V3D_CTL_IDENT1                                 0x00004
+/* Multiples of 1kb */
+# define V3D_IDENT1_VPM_SIZE_MASK                      V3D_MASK(31, 28)
+# define V3D_IDENT1_VPM_SIZE_SHIFT                     28
+# define V3D_IDENT1_NSEM_MASK                          V3D_MASK(23, 16)
+# define V3D_IDENT1_NSEM_SHIFT                         16
+# define V3D_IDENT1_NTMU_MASK                          V3D_MASK(15, 12)
+# define V3D_IDENT1_NTMU_SHIFT                         12
+# define V3D_IDENT1_QUPS_MASK                          V3D_MASK(11, 8)
+# define V3D_IDENT1_QUPS_SHIFT                         8
+# define V3D_IDENT1_NSLC_MASK                          V3D_MASK(7, 4)
+# define V3D_IDENT1_NSLC_SHIFT                         4
+# define V3D_IDENT1_REV_MASK                           V3D_MASK(3, 0)
+# define V3D_IDENT1_REV_SHIFT                          0
+
+#define V3D_CTL_IDENT2                                 0x00008
+# define V3D_IDENT2_BCG_INT                            BIT(28)
+
+#define V3D_CTL_MISCCFG                                0x00018
+# define V3D_MISCCFG_OVRTMUOUT                         BIT(0)
+
+#define V3D_CTL_L2CACTL                                0x00020
+# define V3D_L2CACTL_L2CCLR                            BIT(2)
+# define V3D_L2CACTL_L2CDIS                            BIT(1)
+# define V3D_L2CACTL_L2CENA                            BIT(0)
+
+#define V3D_CTL_SLCACTL                                0x00024
+# define V3D_SLCACTL_TVCCS_MASK                        V3D_MASK(27, 24)
+# define V3D_SLCACTL_TVCCS_SHIFT                       24
+# define V3D_SLCACTL_TDCCS_MASK                        V3D_MASK(19, 16)
+# define V3D_SLCACTL_TDCCS_SHIFT                       16
+# define V3D_SLCACTL_UCC_MASK                          V3D_MASK(11, 8)
+# define V3D_SLCACTL_UCC_SHIFT                         8
+# define V3D_SLCACTL_ICC_MASK                          V3D_MASK(3, 0)
+# define V3D_SLCACTL_ICC_SHIFT                         0
+
+#define V3D_CTL_L2TCACTL                               0x00030
+# define V3D_L2TCACTL_TMUWCF                           BIT(8)
+# define V3D_L2TCACTL_L2T_NO_WM                        BIT(4)
+# define V3D_L2TCACTL_FLM_FLUSH                        0
+# define V3D_L2TCACTL_FLM_CLEAR                        1
+# define V3D_L2TCACTL_FLM_CLEAN                        2
+# define V3D_L2TCACTL_FLM_MASK                         V3D_MASK(2, 1)
+# define V3D_L2TCACTL_FLM_SHIFT                        1
+# define V3D_L2TCACTL_L2TFLS                           BIT(0)
+#define V3D_CTL_L2TFLSTA                               0x00034
+#define V3D_CTL_L2TFLEND                               0x00038
+
+#define V3D_CTL_INT_STS                                0x00050
+#define V3D_CTL_INT_SET                                0x00054
+#define V3D_CTL_INT_CLR                                0x00058
+#define V3D_CTL_INT_MSK_STS                            0x0005c
+#define V3D_CTL_INT_MSK_SET                            0x00060
+#define V3D_CTL_INT_MSK_CLR                            0x00064
+# define V3D_INT_QPU_MASK                              V3D_MASK(27, 16)
+# define V3D_INT_QPU_SHIFT                             16
+# define V3D_INT_GMPV                                  BIT(5)
+# define V3D_INT_TRFB                                  BIT(4)
+# define V3D_INT_SPILLUSE                              BIT(3)
+# define V3D_INT_OUTOMEM                               BIT(2)
+# define V3D_INT_FLDONE                                BIT(1)
+# define V3D_INT_FRDONE                                BIT(0)
+
+#define V3D_CLE_CT0CS                                  0x00100
+#define V3D_CLE_CT1CS                                  0x00104
+#define V3D_CLE_CTNCS(n) (V3D_CLE_CT0CS + 4 * n)
+#define V3D_CLE_CT0EA                                  0x00108
+#define V3D_CLE_CT1EA                                  0x0010c
+#define V3D_CLE_CTNEA(n) (V3D_CLE_CT0EA + 4 * n)
+#define V3D_CLE_CT0CA                                  0x00110
+#define V3D_CLE_CT1CA                                  0x00114
+#define V3D_CLE_CTNCA(n) (V3D_CLE_CT0CA + 4 * n)
+#define V3D_CLE_CT0RA                                  0x00118
+#define V3D_CLE_CT1RA                                  0x0011c
+#define V3D_CLE_CT0LC                                  0x00120
+#define V3D_CLE_CT1LC                                  0x00124
+#define V3D_CLE_CT0PC                                  0x00128
+#define V3D_CLE_CT1PC                                  0x0012c
+#define V3D_CLE_PCS                                    0x00130
+#define V3D_CLE_BFC                                    0x00134
+#define V3D_CLE_RFC                                    0x00138
+#define V3D_CLE_TFBC                                   0x0013c
+#define V3D_CLE_TFIT                                   0x00140
+#define V3D_CLE_CT1CFG                                 0x00144
+#define V3D_CLE_CT1TILECT                              0x00148
+#define V3D_CLE_CT1TSKIP                               0x0014c
+#define V3D_CLE_CT1PTCT                                0x00150
+#define V3D_CLE_CT0SYNC                                0x00154
+#define V3D_CLE_CT1SYNC                                0x00158
+#define V3D_CLE_CT0QTS                                 0x0015c
+# define V3D_CLE_CT0QTS_ENABLE                         BIT(1)
+#define V3D_CLE_CT0QBA                                 0x00160
+#define V3D_CLE_CT1QBA                                 0x00164
+#define V3D_CLE_CTNQBA(n) (V3D_CLE_CT0QBA + 4 * n)
+#define V3D_CLE_CT0QEA                                 0x00168
+#define V3D_CLE_CT1QEA                                 0x0016c
+#define V3D_CLE_CTNQEA(n) (V3D_CLE_CT0QEA + 4 * n)
+#define V3D_CLE_CT0QMA                                 0x00170
+#define V3D_CLE_CT0QMS                                 0x00174
+#define V3D_CLE_CT1QCFG                                0x00178
+/* If set without ETPROC, entirely skip tiles with no primitives. */
+# define V3D_CLE_QCFG_ETFILT                           BIT(7)
+/* If set with ETFILT, just write the clear color to tiles with no
+ * primitives.
+ */
+# define V3D_CLE_QCFG_ETPROC                           BIT(6)
+# define V3D_CLE_QCFG_ETSFLUSH                         BIT(1)
+# define V3D_CLE_QCFG_MCDIS                            BIT(0)
+
+#define V3D_PTB_BPCA                                   0x00300
+#define V3D_PTB_BPCS                                   0x00304
+#define V3D_PTB_BPOA                                   0x00308
+#define V3D_PTB_BPOS                                   0x0030c
+
+#define V3D_PTB_BXCF                                   0x00310
+# define V3D_PTB_BXCF_RWORDERDISA                      BIT(1)
+# define V3D_PTB_BXCF_CLIPDISA                         BIT(0)
+
+#define V3D_GMP_STATUS                                 0x00800
+# define V3D_GMP_STATUS_GMPRST                         BIT(31)
+# define V3D_GMP_STATUS_WR_COUNT_MASK                  V3D_MASK(30, 24)
+# define V3D_GMP_STATUS_WR_COUNT_SHIFT                 24
+# define V3D_GMP_STATUS_RD_COUNT_MASK                  V3D_MASK(22, 16)
+# define V3D_GMP_STATUS_RD_COUNT_SHIFT                 16
+# define V3D_GMP_STATUS_WR_ACTIVE                      BIT(5)
+# define V3D_GMP_STATUS_RD_ACTIVE                      BIT(4)
+# define V3D_GMP_STATUS_CFG_BUSY                       BIT(3)
+# define V3D_GMP_STATUS_CNTOVF                         BIT(2)
+# define V3D_GMP_STATUS_INVPROT                        BIT(1)
+# define V3D_GMP_STATUS_VIO                            BIT(0)
+
+#define V3D_GMP_CFG                                    0x00804
+# define V3D_GMP_CFG_LBURSTEN                          BIT(3)
+# define V3D_GMP_CFG_PGCRSEN                           BIT()
+# define V3D_GMP_CFG_STOP_REQ                          BIT(1)
+# define V3D_GMP_CFG_PROT_ENABLE                       BIT(0)
+
+#define V3D_GMP_VIO_ADDR                               0x00808
+#define V3D_GMP_VIO_TYPE                               0x0080c
+#define V3D_GMP_TABLE_ADDR                             0x00810
+#define V3D_GMP_CLEAR_LOAD                             0x00814
+#define V3D_GMP_PRESERVE_LOAD                          0x00818
+#define V3D_GMP_VALID_LINES                            0x00820
+
+#endif /* V3D_REGS_H */
diff --git a/drivers/gpu/drm/v3d/v3d_sched.c b/drivers/gpu/drm/v3d/v3d_sched.c
new file mode 100644 (file)
index 0000000..b07bece
--- /dev/null
@@ -0,0 +1,228 @@
+// SPDX-License-Identifier: GPL-2.0+
+/* Copyright (C) 2018 Broadcom */
+
+/**
+ * DOC: Broadcom V3D scheduling
+ *
+ * The shared DRM GPU scheduler is used to coordinate submitting jobs
+ * to the hardware.  Each DRM fd (roughly a client process) gets its
+ * own scheduler entity, which will process jobs in order.  The GPU
+ * scheduler will round-robin between clients to submit the next job.
+ *
+ * For simplicity, and in order to keep latency low for interactive
+ * jobs when bulk background jobs are queued up, we submit a new job
+ * to the HW only when it has completed the last one, instead of
+ * filling up the CT[01]Q FIFOs with jobs.  Similarly, we use
+ * v3d_job_dependency() to manage the dependency between bin and
+ * render, instead of having the clients submit jobs with using the
+ * HW's semaphores to interlock between them.
+ */
+
+#include <linux/kthread.h>
+
+#include "v3d_drv.h"
+#include "v3d_regs.h"
+#include "v3d_trace.h"
+
+static struct v3d_job *
+to_v3d_job(struct drm_sched_job *sched_job)
+{
+       return container_of(sched_job, struct v3d_job, base);
+}
+
+static void
+v3d_job_free(struct drm_sched_job *sched_job)
+{
+       struct v3d_job *job = to_v3d_job(sched_job);
+
+       v3d_exec_put(job->exec);
+}
+
+/**
+ * Returns the fences that the bin job depends on, one by one.
+ * v3d_job_run() won't be called until all of them have been signaled.
+ */
+static struct dma_fence *
+v3d_job_dependency(struct drm_sched_job *sched_job,
+                  struct drm_sched_entity *s_entity)
+{
+       struct v3d_job *job = to_v3d_job(sched_job);
+       struct v3d_exec_info *exec = job->exec;
+       enum v3d_queue q = job == &exec->bin ? V3D_BIN : V3D_RENDER;
+       struct dma_fence *fence;
+
+       fence = job->in_fence;
+       if (fence) {
+               job->in_fence = NULL;
+               return fence;
+       }
+
+       if (q == V3D_RENDER) {
+               /* If we had a bin job, the render job definitely depends on
+                * it. We first have to wait for bin to be scheduled, so that
+                * its done_fence is created.
+                */
+               fence = exec->bin_done_fence;
+               if (fence) {
+                       exec->bin_done_fence = NULL;
+                       return fence;
+               }
+       }
+
+       /* XXX: Wait on a fence for switching the GMP if necessary,
+        * and then do so.
+        */
+
+       return fence;
+}
+
+static struct dma_fence *v3d_job_run(struct drm_sched_job *sched_job)
+{
+       struct v3d_job *job = to_v3d_job(sched_job);
+       struct v3d_exec_info *exec = job->exec;
+       enum v3d_queue q = job == &exec->bin ? V3D_BIN : V3D_RENDER;
+       struct v3d_dev *v3d = exec->v3d;
+       struct drm_device *dev = &v3d->drm;
+       struct dma_fence *fence;
+       unsigned long irqflags;
+
+       if (unlikely(job->base.s_fence->finished.error))
+               return NULL;
+
+       /* Lock required around bin_job update vs
+        * v3d_overflow_mem_work().
+        */
+       spin_lock_irqsave(&v3d->job_lock, irqflags);
+       if (q == V3D_BIN) {
+               v3d->bin_job = job->exec;
+
+               /* Clear out the overflow allocation, so we don't
+                * reuse the overflow attached to a previous job.
+                */
+               V3D_CORE_WRITE(0, V3D_PTB_BPOS, 0);
+       } else {
+               v3d->render_job = job->exec;
+       }
+       spin_unlock_irqrestore(&v3d->job_lock, irqflags);
+
+       /* Can we avoid this flush when q==RENDER?  We need to be
+        * careful of scheduling, though -- imagine job0 rendering to
+        * texture and job1 reading, and them being executed as bin0,
+        * bin1, render0, render1, so that render1's flush at bin time
+        * wasn't enough.
+        */
+       v3d_invalidate_caches(v3d);
+
+       fence = v3d_fence_create(v3d, q);
+       if (!fence)
+               return fence;
+
+       if (job->done_fence)
+               dma_fence_put(job->done_fence);
+       job->done_fence = dma_fence_get(fence);
+
+       trace_v3d_submit_cl(dev, q == V3D_RENDER, to_v3d_fence(fence)->seqno,
+                           job->start, job->end);
+
+       if (q == V3D_BIN) {
+               if (exec->qma) {
+                       V3D_CORE_WRITE(0, V3D_CLE_CT0QMA, exec->qma);
+                       V3D_CORE_WRITE(0, V3D_CLE_CT0QMS, exec->qms);
+               }
+               if (exec->qts) {
+                       V3D_CORE_WRITE(0, V3D_CLE_CT0QTS,
+                                      V3D_CLE_CT0QTS_ENABLE |
+                                      exec->qts);
+               }
+       } else {
+               /* XXX: Set the QCFG */
+       }
+
+       /* Set the current and end address of the control list.
+        * Writing the end register is what starts the job.
+        */
+       V3D_CORE_WRITE(0, V3D_CLE_CTNQBA(q), job->start);
+       V3D_CORE_WRITE(0, V3D_CLE_CTNQEA(q), job->end);
+
+       return fence;
+}
+
+static void
+v3d_job_timedout(struct drm_sched_job *sched_job)
+{
+       struct v3d_job *job = to_v3d_job(sched_job);
+       struct v3d_exec_info *exec = job->exec;
+       struct v3d_dev *v3d = exec->v3d;
+       enum v3d_queue q;
+
+       mutex_lock(&v3d->reset_lock);
+
+       /* block scheduler */
+       for (q = 0; q < V3D_MAX_QUEUES; q++) {
+               struct drm_gpu_scheduler *sched = &v3d->queue[q].sched;
+
+               kthread_park(sched->thread);
+               drm_sched_hw_job_reset(sched, (sched_job->sched == sched ?
+                                              sched_job : NULL));
+       }
+
+       /* get the GPU back into the init state */
+       v3d_reset(v3d);
+
+       /* Unblock schedulers and restart their jobs. */
+       for (q = 0; q < V3D_MAX_QUEUES; q++) {
+               drm_sched_job_recovery(&v3d->queue[q].sched);
+               kthread_unpark(v3d->queue[q].sched.thread);
+       }
+
+       mutex_unlock(&v3d->reset_lock);
+}
+
+static const struct drm_sched_backend_ops v3d_sched_ops = {
+       .dependency = v3d_job_dependency,
+       .run_job = v3d_job_run,
+       .timedout_job = v3d_job_timedout,
+       .free_job = v3d_job_free
+};
+
+int
+v3d_sched_init(struct v3d_dev *v3d)
+{
+       int hw_jobs_limit = 1;
+       int job_hang_limit = 0;
+       int hang_limit_ms = 500;
+       int ret;
+
+       ret = drm_sched_init(&v3d->queue[V3D_BIN].sched,
+                            &v3d_sched_ops,
+                            hw_jobs_limit, job_hang_limit,
+                            msecs_to_jiffies(hang_limit_ms),
+                            "v3d_bin");
+       if (ret) {
+               dev_err(v3d->dev, "Failed to create bin scheduler: %d.", ret);
+               return ret;
+       }
+
+       ret = drm_sched_init(&v3d->queue[V3D_RENDER].sched,
+                            &v3d_sched_ops,
+                            hw_jobs_limit, job_hang_limit,
+                            msecs_to_jiffies(hang_limit_ms),
+                            "v3d_render");
+       if (ret) {
+               dev_err(v3d->dev, "Failed to create render scheduler: %d.",
+                       ret);
+               drm_sched_fini(&v3d->queue[V3D_BIN].sched);
+               return ret;
+       }
+
+       return 0;
+}
+
+void
+v3d_sched_fini(struct v3d_dev *v3d)
+{
+       enum v3d_queue q;
+
+       for (q = 0; q < V3D_MAX_QUEUES; q++)
+               drm_sched_fini(&v3d->queue[q].sched);
+}
diff --git a/drivers/gpu/drm/v3d/v3d_trace.h b/drivers/gpu/drm/v3d/v3d_trace.h
new file mode 100644 (file)
index 0000000..85dd351
--- /dev/null
@@ -0,0 +1,82 @@
+// SPDX-License-Identifier: GPL-2.0+
+/* Copyright (C) 2015-2018 Broadcom */
+
+#if !defined(_V3D_TRACE_H_) || defined(TRACE_HEADER_MULTI_READ)
+#define _V3D_TRACE_H_
+
+#include <linux/stringify.h>
+#include <linux/types.h>
+#include <linux/tracepoint.h>
+
+#undef TRACE_SYSTEM
+#define TRACE_SYSTEM v3d
+#define TRACE_INCLUDE_FILE v3d_trace
+
+TRACE_EVENT(v3d_submit_cl,
+           TP_PROTO(struct drm_device *dev, bool is_render,
+                    uint64_t seqno,
+                    u32 ctnqba, u32 ctnqea),
+           TP_ARGS(dev, is_render, seqno, ctnqba, ctnqea),
+
+           TP_STRUCT__entry(
+                            __field(u32, dev)
+                            __field(bool, is_render)
+                            __field(u64, seqno)
+                            __field(u32, ctnqba)
+                            __field(u32, ctnqea)
+                            ),
+
+           TP_fast_assign(
+                          __entry->dev = dev->primary->index;
+                          __entry->is_render = is_render;
+                          __entry->seqno = seqno;
+                          __entry->ctnqba = ctnqba;
+                          __entry->ctnqea = ctnqea;
+                          ),
+
+           TP_printk("dev=%u, %s, seqno=%llu, 0x%08x..0x%08x",
+                     __entry->dev,
+                     __entry->is_render ? "RCL" : "BCL",
+                     __entry->seqno,
+                     __entry->ctnqba,
+                     __entry->ctnqea)
+);
+
+TRACE_EVENT(v3d_reset_begin,
+           TP_PROTO(struct drm_device *dev),
+           TP_ARGS(dev),
+
+           TP_STRUCT__entry(
+                            __field(u32, dev)
+                            ),
+
+           TP_fast_assign(
+                          __entry->dev = dev->primary->index;
+                          ),
+
+           TP_printk("dev=%u",
+                     __entry->dev)
+);
+
+TRACE_EVENT(v3d_reset_end,
+           TP_PROTO(struct drm_device *dev),
+           TP_ARGS(dev),
+
+           TP_STRUCT__entry(
+                            __field(u32, dev)
+                            ),
+
+           TP_fast_assign(
+                          __entry->dev = dev->primary->index;
+                          ),
+
+           TP_printk("dev=%u",
+                     __entry->dev)
+);
+
+#endif /* _V3D_TRACE_H_ */
+
+/* This part must be outside protection */
+#undef TRACE_INCLUDE_PATH
+#define TRACE_INCLUDE_PATH .
+#include <trace/define_trace.h>
diff --git a/drivers/gpu/drm/v3d/v3d_trace_points.c b/drivers/gpu/drm/v3d/v3d_trace_points.c
new file mode 100644 (file)
index 0000000..482922d
--- /dev/null
@@ -0,0 +1,9 @@
+// SPDX-License-Identifier: GPL-2.0+
+/* Copyright (C) 2015 Broadcom */
+
+#include "v3d_drv.h"
+
+#ifndef __CHECKER__
+#define CREATE_TRACE_POINTS
+#include "v3d_trace.h"
+#endif
diff --git a/include/uapi/drm/v3d_drm.h b/include/uapi/drm/v3d_drm.h
new file mode 100644 (file)
index 0000000..7b66277
--- /dev/null
@@ -0,0 +1,194 @@
+/*
+ * Copyright © 2014-2018 Broadcom
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ * IN THE SOFTWARE.
+ */
+
+#ifndef _V3D_DRM_H_
+#define _V3D_DRM_H_
+
+#include "drm.h"
+
+#if defined(__cplusplus)
+extern "C" {
+#endif
+
+#define DRM_V3D_SUBMIT_CL                         0x00
+#define DRM_V3D_WAIT_BO                           0x01
+#define DRM_V3D_CREATE_BO                         0x02
+#define DRM_V3D_MMAP_BO                           0x03
+#define DRM_V3D_GET_PARAM                         0x04
+#define DRM_V3D_GET_BO_OFFSET                     0x05
+
+#define DRM_IOCTL_V3D_SUBMIT_CL           DRM_IOWR(DRM_COMMAND_BASE + DRM_V3D_SUBMIT_CL, struct drm_v3d_submit_cl)
+#define DRM_IOCTL_V3D_WAIT_BO             DRM_IOWR(DRM_COMMAND_BASE + DRM_V3D_WAIT_BO, struct drm_v3d_wait_bo)
+#define DRM_IOCTL_V3D_CREATE_BO           DRM_IOWR(DRM_COMMAND_BASE + DRM_V3D_CREATE_BO, struct drm_v3d_create_bo)
+#define DRM_IOCTL_V3D_MMAP_BO             DRM_IOWR(DRM_COMMAND_BASE + DRM_V3D_MMAP_BO, struct drm_v3d_mmap_bo)
+#define DRM_IOCTL_V3D_GET_PARAM           DRM_IOWR(DRM_COMMAND_BASE + DRM_V3D_GET_PARAM, struct drm_v3d_get_param)
+#define DRM_IOCTL_V3D_GET_BO_OFFSET       DRM_IOWR(DRM_COMMAND_BASE + DRM_V3D_GET_BO_OFFSET, struct drm_v3d_get_bo_offset)
+
+/**
+ * struct drm_v3d_submit_cl - ioctl argument for submitting commands to the 3D
+ * engine.
+ *
+ * This asks the kernel to have the GPU execute an optional binner
+ * command list, and a render command list.
+ */
+struct drm_v3d_submit_cl {
+       /* Pointer to the binner command list.
+        *
+        * This is the first set of commands executed, which runs the
+        * coordinate shader to determine where primitives land on the screen,
+        * then writes out the state updates and draw calls necessary per tile
+        * to the tile allocation BO.
+        */
+       __u32 bcl_start;
+
+        /** End address of the BCL (first byte after the BCL) */
+       __u32 bcl_end;
+
+       /* Offset of the render command list.
+        *
+        * This is the second set of commands executed, which will either
+        * execute the tiles that have been set up by the BCL, or a fixed set
+        * of tiles (in the case of RCL-only blits).
+        */
+       __u32 rcl_start;
+
+        /** End address of the RCL (first byte after the RCL) */
+       __u32 rcl_end;
+
+       /** An optional sync object to wait on before starting the BCL. */
+       __u32 in_sync_bcl;
+       /** An optional sync object to wait on before starting the RCL. */
+       __u32 in_sync_rcl;
+       /** An optional sync object to place the completion fence in. */
+       __u32 out_sync;
+
+       /* Offset of the tile alloc memory
+        *
+        * This is optional on V3D 3.3 (where the CL can set the value) but
+        * required on V3D 4.1.
+        */
+       __u32 qma;
+
+       /** Size of the tile alloc memory. */
+       __u32 qms;
+
+       /** Offset of the tile state data array. */
+       __u32 qts;
+
+       /* Pointer to a u32 array of the BOs that are referenced by the job.
+        */
+       __u64 bo_handles;
+
+       /* Number of BO handles passed in (size is that times 4). */
+       __u32 bo_handle_count;
+
+       /* Pad, must be zero-filled. */
+       __u32 pad;
+};
+
+/**
+ * struct drm_v3d_wait_bo - ioctl argument for waiting for
+ * completion of the last DRM_V3D_SUBMIT_CL on a BO.
+ *
+ * This is useful for cases where multiple processes might be
+ * rendering to a BO and you want to wait for all rendering to be
+ * completed.
+ */
+struct drm_v3d_wait_bo {
+       __u32 handle;
+       __u32 pad;
+       __u64 timeout_ns;
+};
+
+/**
+ * struct drm_v3d_create_bo - ioctl argument for creating V3D BOs.
+ *
+ * There are currently no values for the flags argument, but it may be
+ * used in a future extension.
+ */
+struct drm_v3d_create_bo {
+       __u32 size;
+       __u32 flags;
+       /** Returned GEM handle for the BO. */
+       __u32 handle;
+       /**
+        * Returned offset for the BO in the V3D address space.  This offset
+        * is private to the DRM fd and is valid for the lifetime of the GEM
+        * handle.
+        *
+        * This offset value will always be nonzero, since various HW
+        * units treat 0 specially.
+        */
+       __u32 offset;
+};
+
+/**
+ * struct drm_v3d_mmap_bo - ioctl argument for mapping V3D BOs.
+ *
+ * This doesn't actually perform an mmap.  Instead, it returns the
+ * offset you need to use in an mmap on the DRM device node.  This
+ * means that tools like valgrind end up knowing about the mapped
+ * memory.
+ *
+ * There are currently no values for the flags argument, but it may be
+ * used in a future extension.
+ */
+struct drm_v3d_mmap_bo {
+       /** Handle for the object being mapped. */
+       __u32 handle;
+       __u32 flags;
+       /** offset into the drm node to use for subsequent mmap call. */
+       __u64 offset;
+};
+
+enum drm_v3d_param {
+       DRM_V3D_PARAM_V3D_UIFCFG,
+       DRM_V3D_PARAM_V3D_HUB_IDENT1,
+       DRM_V3D_PARAM_V3D_HUB_IDENT2,
+       DRM_V3D_PARAM_V3D_HUB_IDENT3,
+       DRM_V3D_PARAM_V3D_CORE0_IDENT0,
+       DRM_V3D_PARAM_V3D_CORE0_IDENT1,
+       DRM_V3D_PARAM_V3D_CORE0_IDENT2,
+};
+
+struct drm_v3d_get_param {
+       __u32 param;
+       __u32 pad;
+       __u64 value;
+};
+
+/**
+ * Returns the offset for the BO in the V3D address space for this DRM fd.
+ * This is the same value returned by drm_v3d_create_bo, if that was called
+ * from this DRM fd.
+ */
+struct drm_v3d_get_bo_offset {
+       __u32 handle;
+       __u32 offset;
+};
+
+#if defined(__cplusplus)
+}
+#endif
+
+#endif /* _V3D_DRM_H_ */