gpu: nvgpu: support usermode submit buffers

Import userd and gpfifo buffers from userspace if provided via
NVGPU_IOCTL_CHANNEL_ALLOC_GPFIFO_EX. Also supply the work submit token
(i.e., the hw channel id) to userspace.

To keep the buffers alive, store their dmabuf and attachment/sgt handles
in nvgpu_channel_linux. Our nvgpu_mem doesn't provide such data for
buffers that are mainly in kernel use. The buffers are freed via a new
API in the os_channel interface.

Fix a bug in gk20a_channel_free_usermode_buffers: also unmap the
usermode gpfifo buffer.

Bug 200145225
Bug 200541476

Change-Id: I8416af7085c91b044ac8ccd9faa38e2a6d0c3946
Signed-off-by: Konsta Holtta <kholtta@nvidia.com>
Reviewed-on: https://git-master.nvidia.com/r/1795821
Signed-off-by: Debarshi Dutta <ddutta@nvidia.com>
(cherry picked from commit 99b1c6dcdf328efcfe47338ad1b71a114ab7f272
in dev-main)
Reviewed-on: https://git-master.nvidia.com/r/2170603
GVS: Gerrit_Virtual_Submit
Reviewed-by: Bibek Basu <bbasu@nvidia.com>
Reviewed-by: mobile promotions <svcmobile_promotions@nvidia.com>
Tested-by: mobile promotions <svcmobile_promotions@nvidia.com>

7 files changed, 205 insertions, 4 deletions

diff --git a/drivers/gpu/nvgpu/common/fifo/channel.c b/drivers/gpu/nvgpu/common/fifo/channel.c
index adb59ac4..6ebeb596 100644
--- a/drivers/gpu/nvgpu/common/fifo/channel.c
+++ b/drivers/gpu/nvgpu/common/fifo/channel.c
@@ -1228,7 +1228,9 @@ int nvgpu_channel_setup_bind(struct channel_gk20a *c,
         nvgpu_log_info(g, "channel %d : gpfifo_base 0x%016llx, size %d",
                 c->chid, gpfifo_gpu_va, c->gpfifo.entry_num);
 
-        g->ops.fifo.setup_userd(c);
+        if (!c->usermode_submit_enabled) {
+                g->ops.fifo.setup_userd(c);
+        }
 
         if (g->aggressive_sync_destroy_thresh == 0U) {
                 nvgpu_mutex_acquire(&c->sync_lock);
@@ -1326,7 +1328,10 @@ void gk20a_channel_free_usermode_buffers(struct channel_gk20a *c)
                 nvgpu_dma_free(c->g, &c->usermode_userd);
         }
         if (nvgpu_mem_is_valid(&c->usermode_gpfifo)) {
-                nvgpu_dma_free(c->g, &c->usermode_gpfifo);
+                nvgpu_dma_unmap_free(c->vm, &c->usermode_gpfifo);
+        }
+        if (c->g->os_channel.free_usermode_buffers != NULL) {
+                c->g->os_channel.free_usermode_buffers(c);
         }
 }
 
diff --git a/drivers/gpu/nvgpu/include/nvgpu/channel.h b/drivers/gpu/nvgpu/include/nvgpu/channel.h
index d7bf7816..7ca60fe3 100644
--- a/drivers/gpu/nvgpu/include/nvgpu/channel.h
+++ b/drivers/gpu/nvgpu/include/nvgpu/channel.h
@@ -95,7 +95,9 @@ struct nvgpu_setup_bind_args {
         u32 num_gpfifo_entries;
         u32 num_inflight_jobs;
         u32 userd_dmabuf_fd;
+        u64 userd_dmabuf_offset;
         u32 gpfifo_dmabuf_fd;
+        u64 gpfifo_dmabuf_offset;
         u32 work_submit_token;
         u32 flags;
 };
diff --git a/drivers/gpu/nvgpu/include/nvgpu/gk20a.h b/drivers/gpu/nvgpu/include/nvgpu/gk20a.h
index feae7d23..873fa826 100644
--- a/drivers/gpu/nvgpu/include/nvgpu/gk20a.h
+++ b/drivers/gpu/nvgpu/include/nvgpu/gk20a.h
@@ -1605,7 +1605,8 @@ struct gk20a {
                                 struct nvgpu_gpfifo_userdata userdata,
                                 u32 start, u32 length);
                 int (*alloc_usermode_buffers)(struct channel_gk20a *c,
-                        struct nvgpu_setup_bind_args *gpfifo_args);
+                        struct nvgpu_setup_bind_args *args);
+                void (*free_usermode_buffers)(struct channel_gk20a *c);
         } os_channel;
 
         struct gk20a_scale_profile *scale_profile;
diff --git a/drivers/gpu/nvgpu/os/linux/channel.h b/drivers/gpu/nvgpu/os/linux/channel.h
index 2210678d..e6326fad 100644
--- a/drivers/gpu/nvgpu/os/linux/channel.h
+++ b/drivers/gpu/nvgpu/os/linux/channel.h
@@ -63,6 +63,19 @@ struct nvgpu_os_fence_framework {
         struct sync_timeline *timeline;
 };
 
+struct nvgpu_usermode_bufs_linux {
+        /*
+         * Common low level info of these is stored in nvgpu_mems in
+         * channel_gk20a; these hold lifetimes for the actual dmabuf and its
+         * dma mapping.
+         */
+        struct nvgpu_usermode_buf_linux {
+                struct dma_buf *dmabuf;
+                struct dma_buf_attachment *attachment;
+                struct sg_table *sgt;
+        } gpfifo, userd;
+};
+
 struct nvgpu_channel_linux {
         struct channel_gk20a *ch;
 
@@ -72,6 +85,8 @@ struct nvgpu_channel_linux {
         struct nvgpu_error_notifier error_notifier;
 
         struct dma_buf *cyclestate_buffer_handler;
+
+        struct nvgpu_usermode_bufs_linux usermode;
 };
 
 u32 nvgpu_submit_gpfifo_user_flags_to_common_flags(u32 user_flags);
diff --git a/drivers/gpu/nvgpu/os/linux/ioctl_channel.c b/drivers/gpu/nvgpu/os/linux/ioctl_channel.c
index d243c425..22177171 100644
--- a/drivers/gpu/nvgpu/os/linux/ioctl_channel.c
+++ b/drivers/gpu/nvgpu/os/linux/ioctl_channel.c
@@ -590,6 +590,9 @@ static u32 nvgpu_setup_bind_user_flags_to_common_flags(u32 user_flags)
         if (user_flags & NVGPU_CHANNEL_SETUP_BIND_FLAGS_REPLAYABLE_FAULTS_ENABLE)
                 flags |= NVGPU_SETUP_BIND_FLAGS_REPLAYABLE_FAULTS_ENABLE;
 
+        if (user_flags & NVGPU_CHANNEL_SETUP_BIND_FLAGS_USERMODE_SUPPORT)
+                flags |= NVGPU_SETUP_BIND_FLAGS_USERMODE_SUPPORT;
+
         return flags;
 }
 
@@ -601,6 +604,14 @@ static void nvgpu_get_setup_bind_args(
                 channel_setup_bind_args->num_gpfifo_entries;
         setup_bind_args->num_inflight_jobs =
                 channel_setup_bind_args->num_inflight_jobs;
+        setup_bind_args->userd_dmabuf_fd =
+                channel_setup_bind_args->userd_dmabuf_fd;
+        setup_bind_args->userd_dmabuf_offset =
+                channel_setup_bind_args->userd_dmabuf_offset;
+        setup_bind_args->gpfifo_dmabuf_fd =
+                channel_setup_bind_args->gpfifo_dmabuf_fd;
+        setup_bind_args->gpfifo_dmabuf_offset =
+                channel_setup_bind_args->gpfifo_dmabuf_offset;
         setup_bind_args->flags = nvgpu_setup_bind_user_flags_to_common_flags(
                         channel_setup_bind_args->flags);
 }
@@ -1156,6 +1167,8 @@ long gk20a_channel_ioctl(struct file *filp,
                         break;
                 }
                 err = nvgpu_channel_setup_bind(ch, &setup_bind_args);
+                channel_setup_bind_args->work_submit_token =
+                        setup_bind_args.work_submit_token;
                 gk20a_idle(ch->g);
                 break;
         }
diff --git a/drivers/gpu/nvgpu/os/linux/linux-channel.c b/drivers/gpu/nvgpu/os/linux/linux-channel.c
index 1fec3604..d035baf7 100644
--- a/drivers/gpu/nvgpu/os/linux/linux-channel.c
+++ b/drivers/gpu/nvgpu/os/linux/linux-channel.c
@@ -20,6 +20,7 @@
 #include <nvgpu/os_sched.h>
 #include <nvgpu/gk20a.h>
 #include <nvgpu/channel.h>
+#include <nvgpu/dma.h>
 
 /*
  * This is required for nvgpu_vm_find_buf() which is used in the tracing
@@ -31,6 +32,7 @@
 #include "channel.h"
 #include "ioctl_channel.h"
 #include "os_linux.h"
+#include "dmabuf.h"
 
 #include <nvgpu/hw/gk20a/hw_pbdma_gk20a.h>
 
@@ -383,6 +385,147 @@ static int nvgpu_channel_copy_user_gpfifo(struct nvgpu_gpfifo_entry *dest,
         return n == 0 ? 0 : -EFAULT;
 }
 
+int nvgpu_usermode_buf_from_dmabuf(struct gk20a *g, int dmabuf_fd,
+                struct nvgpu_mem *mem, struct nvgpu_usermode_buf_linux *buf)
+{
+        struct device *dev = dev_from_gk20a(g);
+        struct dma_buf *dmabuf;
+        struct sg_table *sgt;
+        struct dma_buf_attachment *attachment;
+        int err;
+
+        dmabuf = dma_buf_get(dmabuf_fd);
+        if (IS_ERR(dmabuf)) {
+                return PTR_ERR(dmabuf);
+        }
+
+        if (gk20a_dmabuf_aperture(g, dmabuf) == APERTURE_INVALID) {
+                err = -EINVAL;
+                goto put_dmabuf;
+        }
+
+        err = gk20a_dmabuf_alloc_drvdata(dmabuf, dev);
+        if (err != 0) {
+                goto put_dmabuf;
+        }
+
+        sgt = gk20a_mm_pin(dev, dmabuf, &attachment);
+        if (IS_ERR(sgt)) {
+                nvgpu_warn(g, "Failed to pin dma_buf!");
+                err = PTR_ERR(sgt);
+                goto put_dmabuf;
+        }
+
+        buf->dmabuf = dmabuf;
+        buf->attachment = attachment;
+        buf->sgt = sgt;
+
+        /*
+         * This mem is unmapped and freed in a common path; for Linux, we'll
+         * also need to unref the dmabuf stuff (above) but the sgt here is only
+         * borrowed, so it cannot be freed by nvgpu_mem_*.
+         */
+        mem->mem_flags  = NVGPU_MEM_FLAG_FOREIGN_SGT;
+        mem->aperture   = APERTURE_SYSMEM;
+        mem->skip_wmb   = 0;
+        mem->size       = dmabuf->size;
+
+        mem->priv.flags = 0;
+        mem->priv.pages = NULL;
+        mem->priv.sgt   = sgt;
+
+        return 0;
+put_dmabuf:
+        dma_buf_put(dmabuf);
+        return err;
+}
+
+void nvgpu_channel_free_usermode_buffers(struct channel_gk20a *c)
+{
+        struct nvgpu_channel_linux *priv = c->os_priv;
+        struct gk20a *g = c->g;
+        struct device *dev = dev_from_gk20a(g);
+
+        if (priv->usermode.gpfifo.dmabuf != NULL) {
+                gk20a_mm_unpin(dev, priv->usermode.gpfifo.dmabuf,
+                               priv->usermode.gpfifo.attachment,
+                               priv->usermode.gpfifo.sgt);
+                dma_buf_put(priv->usermode.gpfifo.dmabuf);
+                priv->usermode.gpfifo.dmabuf = NULL;
+        }
+
+        if (priv->usermode.userd.dmabuf != NULL) {
+                gk20a_mm_unpin(dev, priv->usermode.userd.dmabuf,
+                       priv->usermode.userd.attachment,
+                       priv->usermode.userd.sgt);
+                dma_buf_put(priv->usermode.userd.dmabuf);
+                priv->usermode.userd.dmabuf = NULL;
+        }
+}
+
+static int nvgpu_channel_alloc_usermode_buffers(struct channel_gk20a *c,
+                struct nvgpu_setup_bind_args *args)
+{
+        struct nvgpu_channel_linux *priv = c->os_priv;
+        struct gk20a *g = c->g;
+        struct device *dev = dev_from_gk20a(g);
+        size_t gpfifo_size;
+        int err;
+
+        if (args->gpfifo_dmabuf_fd == 0 || args->userd_dmabuf_fd == 0) {
+                return -EINVAL;
+        }
+
+        if (args->gpfifo_dmabuf_offset != 0 ||
+                        args->userd_dmabuf_offset != 0) {
+                /* TODO - not yet supported */
+                return -EINVAL;
+        }
+
+        err = nvgpu_usermode_buf_from_dmabuf(g, args->gpfifo_dmabuf_fd,
+                        &c->usermode_gpfifo, &priv->usermode.gpfifo);
+        if (err < 0) {
+                return err;
+        }
+
+        gpfifo_size = max_t(u32, SZ_4K,
+                        args->num_gpfifo_entries *
+                        nvgpu_get_gpfifo_entry_size());
+
+        if (c->usermode_gpfifo.size < gpfifo_size) {
+                err = -EINVAL;
+                goto free_gpfifo;
+        }
+
+        c->usermode_gpfifo.gpu_va = nvgpu_gmmu_map(c->vm, &c->usermode_gpfifo,
+                        c->usermode_gpfifo.size, 0, gk20a_mem_flag_none,
+                        false, c->usermode_gpfifo.aperture);
+
+        if (c->usermode_gpfifo.gpu_va == 0) {
+                err = -ENOMEM;
+                goto unmap_free_gpfifo;
+        }
+
+        err = nvgpu_usermode_buf_from_dmabuf(g, args->userd_dmabuf_fd,
+                        &c->usermode_userd, &priv->usermode.userd);
+        if (err < 0) {
+                goto unmap_free_gpfifo;
+        }
+
+        args->work_submit_token = g->fifo.channel_base + c->chid;
+
+        return 0;
+unmap_free_gpfifo:
+        nvgpu_dma_unmap_free(c->vm, &c->usermode_gpfifo);
+free_gpfifo:
+        gk20a_mm_unpin(dev, priv->usermode.gpfifo.dmabuf,
+                       priv->usermode.gpfifo.attachment,
+                       priv->usermode.gpfifo.sgt);
+        dma_buf_put(priv->usermode.gpfifo.dmabuf);
+        priv->usermode.gpfifo.dmabuf = NULL;
+        return err;
+}
+
 int nvgpu_init_channel_support_linux(struct nvgpu_os_linux *l)
 {
         struct gk20a *g = &l->g;
@@ -417,6 +560,12 @@ int nvgpu_init_channel_support_linux(struct nvgpu_os_linux *l)
         g->os_channel.copy_user_gpfifo =
                 nvgpu_channel_copy_user_gpfifo;
 
+        g->os_channel.alloc_usermode_buffers =
+                nvgpu_channel_alloc_usermode_buffers;
+
+        g->os_channel.free_usermode_buffers =
+                nvgpu_channel_free_usermode_buffers;
+
         return 0;
 
 err_clean:
diff --git a/include/uapi/linux/nvgpu.h b/include/uapi/linux/nvgpu.h
index 0fa04f6b..38139524 100644
--- a/include/uapi/linux/nvgpu.h
+++ b/include/uapi/linux/nvgpu.h
@@ -1514,8 +1514,24 @@ struct nvgpu_channel_setup_bind_args {
 #define NVGPU_CHANNEL_SETUP_BIND_FLAGS_DETERMINISTIC    (1 << 1)
 /* enable replayable gmmu faults for this channel */
 #define NVGPU_CHANNEL_SETUP_BIND_FLAGS_REPLAYABLE_FAULTS_ENABLE   (1 << 2)
+/*
+ * Enable usermode submits on this channel.
+ *
+ * Submits in usermode are supported in some environments. If supported and
+ * this flag is set + USERD and GPFIFO buffers are provided here, a submit
+ * token is passed back to be written in the doorbell register in the usermode
+ * region to notify the GPU for new work on this channel. Usermode and
+ * kernelmode submit modes are mutually exclusive; by passing this flag, the
+ * SUBMIT_GPFIFO IOCTL cannot be used.
+ */
+#define NVGPU_CHANNEL_SETUP_BIND_FLAGS_USERMODE_SUPPORT (1 << 3)
         __u32 flags;
-        __u32 reserved[16];
+        __s32 userd_dmabuf_fd;  /* in */
+        __s32 gpfifo_dmabuf_fd; /* in */
+        __u32 work_submit_token; /* out */
+        __u64 userd_dmabuf_offset; /* in */
+        __u64 gpfifo_dmabuf_offset; /* in */
+        __u32 reserved[9];
 };
 struct nvgpu_fence {
         __u32 id;        /* syncpoint id or sync fence fd */