nv-tegra-mirror/linux-nvgpu
1
0
Fork 0
mirror of git://nv-tegra.nvidia.com/linux-nvgpu.git synced 2025-12-22 17:36:20 +03:00
Code Issues Packages Projects Releases Wiki Activity

gpu: nvgpu: move CE app logic under CONFIG_NVGPU_DGPU

Browse Source 
CE app functionality from nvgpu is non-safe for igpu. CE engines init
/reset/cg related functionality is required in safety. Hence move the
CE app logic under CONFIG_NVGPU_DGPU flag and update the sources
accordingly.

JIRA NVGPU-3814

Change-Id: I37aa00b1184baccd5fe569ec315be60ac42dac9b
Signed-off-by: Sagar Kamble <skamble@nvidia.com>
Reviewed-on: https://git-master.nvidia.com/r/2168956
GVS: Gerrit_Virtual_Submit
Reviewed-by: Deepak Nibade <dnibade@nvidia.com>
Reviewed-by: mobile promotions <svcmobile_promotions@nvidia.com>
Tested-by: mobile promotions <svcmobile_promotions@nvidia.com>
This commit is contained in:
Sagar Kamble
2019-08-06 10:07:21 +05:30
committed by mobile promotions
parent 6e2e4d0658
commit 2f95efd8d1
12 changed files with 682 additions and 647 deletions
Download Patch File Download Diff File Expand all files Collapse all files

9
arch/nvgpu-common.yaml
View File

@@ -31,9 +31,16 @@ bios:
ce: ce:
safe: yes safe: yes
gpu: dgpu
owner: Thomas F owner: Thomas F
sources: [ common/ce/ce.c, sources: [ common/ce/ce.c,
include/nvgpu/ce.h ]
deps:
ce_app:
safe: yes
gpu: dgpu
owner: Thomas F
sources: [ common/ce/ce_app.c,
common/ce/ce_priv.h, common/ce/ce_priv.h,
include/nvgpu/ce.h ] include/nvgpu/ce.h ]
deps: deps:

2
drivers/gpu/nvgpu/Makefile
View File

@@ -40,7 +40,6 @@ ccflags-y += -DCONFIG_NVGPU_GRAPHICS
ccflags-y += -DCONFIG_NVGPU_CHANNEL_TSG_SCHEDULING ccflags-y += -DCONFIG_NVGPU_CHANNEL_TSG_SCHEDULING
ccflags-y += -DCONFIG_NVGPU_CHANNEL_TSG_CONTROL ccflags-y += -DCONFIG_NVGPU_CHANNEL_TSG_CONTROL
ccflags-y += -DCONFIG_NVGPU_POWER_PG ccflags-y += -DCONFIG_NVGPU_POWER_PG
ccflags-y += -DCONFIG_NVGPU_CE
ccflags-y += -DCONFIG_NVGPU_KERNEL_MODE_SUBMIT ccflags-y += -DCONFIG_NVGPU_KERNEL_MODE_SUBMIT
ccflags-y += -DCONFIG_NVGPU_COMPRESSION ccflags-y += -DCONFIG_NVGPU_COMPRESSION
ccflags-y += -DCONFIG_NVGPU_SIM ccflags-y += -DCONFIG_NVGPU_SIM
@@ -515,6 +514,7 @@ nvgpu-y += \
common/fence/fence.o \ common/fence/fence.o \
common/ecc.o \ common/ecc.o \
common/ce/ce.o \ common/ce/ce.o \
common/ce/ce_app.o \
common/debugger.o common/debugger.o
nvgpu-$(CONFIG_TEGRA_GR_VIRTUALIZATION) += \ nvgpu-$(CONFIG_TEGRA_GR_VIRTUALIZATION) += \

3
drivers/gpu/nvgpu/Makefile.shared.configs
View File

@@ -172,9 +172,6 @@ NVGPU_COMMON_CFLAGS += -DCONFIG_NVGPU_SIM
CONFIG_NVGPU_COMPRESSION := 1 CONFIG_NVGPU_COMPRESSION := 1
NVGPU_COMMON_CFLAGS += -DCONFIG_NVGPU_COMPRESSION NVGPU_COMMON_CFLAGS += -DCONFIG_NVGPU_COMPRESSION
CONFIG_NVGPU_CE := 1
NVGPU_COMMON_CFLAGS += -DCONFIG_NVGPU_CE
# Enable non FUSA HALs for normal build # Enable non FUSA HALs for normal build
CONFIG_NVGPU_HAL_NON_FUSA := 1 CONFIG_NVGPU_HAL_NON_FUSA := 1
NVGPU_COMMON_CFLAGS += -DCONFIG_NVGPU_HAL_NON_FUSA NVGPU_COMMON_CFLAGS += -DCONFIG_NVGPU_HAL_NON_FUSA

6
drivers/gpu/nvgpu/Makefile.sources
View File

@@ -140,6 +140,7 @@ srcs += common/utils/assert.c \
common/fifo/pbdma_status.c \ common/fifo/pbdma_status.c \
common/mc/mc.c \ common/mc/mc.c \
common/rc/rc.c \ common/rc/rc.c \
common/ce/ce.c \
hal/init/hal_gv11b.c \ hal/init/hal_gv11b.c \
hal/init/hal_gv11b_litter.c \ hal/init/hal_gv11b_litter.c \
hal/init/hal_init.c \ hal/init/hal_init.c \
@@ -365,10 +366,6 @@ srcs += hal/regops/regops_gm20b.c \
endif endif
endif endif
ifeq ($(CONFIG_NVGPU_CE),1)
srcs += common/ce/ce.c
endif
ifeq ($(CONFIG_NVGPU_KERNEL_MODE_SUBMIT),1) ifeq ($(CONFIG_NVGPU_KERNEL_MODE_SUBMIT),1)
srcs += common/fifo/submit.c srcs += common/fifo/submit.c
endif endif
@@ -581,6 +578,7 @@ srcs += common/sec2/sec2.c \
common/mm/allocators/page_allocator.c \ common/mm/allocators/page_allocator.c \
common/mm/vidmem.c \ common/mm/vidmem.c \
common/pramin.c \ common/pramin.c \
common/ce/ce_app.c \
hal/mm/mm_gv100.c \ hal/mm/mm_gv100.c \
hal/mm/mm_tu104.c \ hal/mm/mm_tu104.c \
hal/mc/mc_gv100.c \ hal/mc/mc_gv100.c \

604
drivers/gpu/nvgpu/common/ce/ce.c
View File

@@ -23,404 +23,17 @@
#include <nvgpu/types.h> #include <nvgpu/types.h>
#include <nvgpu/gk20a.h> #include <nvgpu/gk20a.h>
#include <nvgpu/engines.h> #include <nvgpu/engines.h>
#include <nvgpu/os_sched.h>
#include <nvgpu/channel.h>
#include <nvgpu/dma.h>
#include <nvgpu/utils.h>
#include <nvgpu/fence.h>
#include <nvgpu/ce.h> #include <nvgpu/ce.h>
#include <nvgpu/power_features/cg.h> #include <nvgpu/power_features/cg.h>
#include "common/ce/ce_priv.h"
static inline u32 nvgpu_ce_get_valid_launch_flags(struct gk20a *g,
u32 launch_flags)
{
#ifdef CONFIG_NVGPU_DGPU
/*
* there is no local memory available,
* don't allow local memory related CE flags
*/
if (g->mm.vidmem.size == 0ULL) {
launch_flags &= ~(NVGPU_CE_SRC_LOCATION_LOCAL_FB |
NVGPU_CE_DST_LOCATION_LOCAL_FB);
}
#endif
return launch_flags;
}
int nvgpu_ce_execute_ops(struct gk20a *g,
u32 ce_ctx_id,
u64 src_buf,
u64 dst_buf,
u64 size,
unsigned int payload,
u32 launch_flags,
u32 request_operation,
u32 submit_flags,
struct nvgpu_fence_type **fence_out)
{
int ret = -EPERM;
struct nvgpu_ce_app *ce_app = g->ce_app;
struct nvgpu_ce_gpu_ctx *ce_ctx, *ce_ctx_save;
bool found = false;
u32 *cmd_buf_cpu_va;
u64 cmd_buf_gpu_va = 0UL;
u32 method_size;
u32 cmd_buf_read_offset;
u32 dma_copy_class;
struct nvgpu_gpfifo_entry gpfifo;
struct nvgpu_channel_fence fence = {0U, 0U};
struct nvgpu_fence_type *ce_cmd_buf_fence_out = NULL;
if (!ce_app->initialised || ce_app->app_state != NVGPU_CE_ACTIVE) {
goto end;
}
nvgpu_mutex_acquire(&ce_app->app_mutex);
nvgpu_list_for_each_entry_safe(ce_ctx, ce_ctx_save,
&ce_app->allocated_contexts, nvgpu_ce_gpu_ctx, list) {
if (ce_ctx->ctx_id == ce_ctx_id) {
found = true;
break;
}
}
nvgpu_mutex_release(&ce_app->app_mutex);
if (!found) {
ret = -EINVAL;
goto end;
}
if (ce_ctx->gpu_ctx_state != NVGPU_CE_GPU_CTX_ALLOCATED) {
ret = -ENODEV;
goto end;
}
nvgpu_mutex_acquire(&ce_ctx->gpu_ctx_mutex);
ce_ctx->cmd_buf_read_queue_offset %= NVGPU_CE_MAX_INFLIGHT_JOBS;
cmd_buf_read_offset = (ce_ctx->cmd_buf_read_queue_offset *
(NVGPU_CE_MAX_COMMAND_BUFF_BYTES_PER_KICKOFF /
U32(sizeof(u32))));
cmd_buf_cpu_va = (u32 *)ce_ctx->cmd_buf_mem.cpu_va;
if (ce_ctx->postfences[ce_ctx->cmd_buf_read_queue_offset] != NULL) {
struct nvgpu_fence_type **prev_post_fence =
&ce_ctx->postfences[ce_ctx->cmd_buf_read_queue_offset];
ret = nvgpu_fence_wait(g, *prev_post_fence,
nvgpu_get_poll_timeout(g));
nvgpu_fence_put(*prev_post_fence);
*prev_post_fence = NULL;
if (ret != 0) {
goto noop;
}
}
cmd_buf_gpu_va = (ce_ctx->cmd_buf_mem.gpu_va +
(u64)(cmd_buf_read_offset * sizeof(u32)));
dma_copy_class = g->ops.get_litter_value(g, GPU_LIT_DMA_COPY_CLASS);
method_size = nvgpu_ce_prepare_submit(src_buf,
dst_buf,
size,
&cmd_buf_cpu_va[cmd_buf_read_offset],
NVGPU_CE_MAX_COMMAND_BUFF_BYTES_PER_KICKOFF,
payload,
nvgpu_ce_get_valid_launch_flags(g, launch_flags),
request_operation,
dma_copy_class);
if (method_size != 0U) {
/* store the element into gpfifo */
g->ops.pbdma.format_gpfifo_entry(g, &gpfifo,
cmd_buf_gpu_va, method_size);
/*
* take always the postfence as it is needed for protecting the
* ce context
*/
submit_flags |= NVGPU_SUBMIT_FLAGS_FENCE_GET;
nvgpu_smp_wmb();
ret = nvgpu_submit_channel_gpfifo_kernel(ce_ctx->ch, &gpfifo,
1, submit_flags, &fence, &ce_cmd_buf_fence_out);
if (ret == 0) {
ce_ctx->postfences[ce_ctx->cmd_buf_read_queue_offset] =
ce_cmd_buf_fence_out;
if (fence_out != NULL) {
nvgpu_fence_get(ce_cmd_buf_fence_out);
*fence_out = ce_cmd_buf_fence_out;
}
/* Next available command buffer queue Index */
++ce_ctx->cmd_buf_read_queue_offset;
}
} else {
ret = -ENOMEM;
}
noop:
nvgpu_mutex_release(&ce_ctx->gpu_ctx_mutex);
end:
return ret;
}
/* static CE app api */
static void nvgpu_ce_put_fences(struct nvgpu_ce_gpu_ctx *ce_ctx)
{
u32 i;
for (i = 0U; i < NVGPU_CE_MAX_INFLIGHT_JOBS; i++) {
struct nvgpu_fence_type **fence = &ce_ctx->postfences[i];
if (*fence != NULL) {
nvgpu_fence_put(*fence);
}
*fence = NULL;
}
}
/* caller must hold ce_app->app_mutex */
static void nvgpu_ce_delete_gpu_context_locked(struct nvgpu_ce_gpu_ctx *ce_ctx)
{
struct nvgpu_list_node *list = &ce_ctx->list;
ce_ctx->gpu_ctx_state = NVGPU_CE_GPU_CTX_DELETED;
ce_ctx->tsg->abortable = true;
nvgpu_mutex_acquire(&ce_ctx->gpu_ctx_mutex);
if (nvgpu_mem_is_valid(&ce_ctx->cmd_buf_mem)) {
nvgpu_ce_put_fences(ce_ctx);
nvgpu_dma_unmap_free(ce_ctx->vm, &ce_ctx->cmd_buf_mem);
}
/*
* free the channel
* nvgpu_channel_close() will also unbind the channel from TSG
*/
nvgpu_channel_close(ce_ctx->ch);
nvgpu_ref_put(&ce_ctx->tsg->refcount, nvgpu_tsg_release);
/* housekeeping on app */
if ((list->prev != NULL) && (list->next != NULL)) {
nvgpu_list_del(list);
}
nvgpu_mutex_release(&ce_ctx->gpu_ctx_mutex);
nvgpu_mutex_destroy(&ce_ctx->gpu_ctx_mutex);
nvgpu_kfree(ce_ctx->g, ce_ctx);
}
static inline unsigned int nvgpu_ce_get_method_size(u32 request_operation,
u64 size)
{
/* failure size */
unsigned int methodsize = UINT_MAX;
unsigned int iterations = 0;
u32 shift;
u64 chunk = size;
u32 height, width;
while (chunk != 0ULL) {
iterations++;
shift = (MAX_CE_ALIGN(chunk) != 0ULL) ?
(nvgpu_ffs(MAX_CE_ALIGN(chunk)) - 1UL) :
MAX_CE_SHIFT;
width = chunk >> shift;
height = BIT32(shift);
width = MAX_CE_ALIGN(width);
chunk -= (u64) height * width;
}
if ((request_operation & NVGPU_CE_PHYS_MODE_TRANSFER) != 0U) {
methodsize = (2U + (16U * iterations)) *
(unsigned int)sizeof(u32);
} else if ((request_operation & NVGPU_CE_MEMSET) != 0U) {
methodsize = (2U + (15U * iterations)) *
(unsigned int)sizeof(u32);
}
return methodsize;
}
u32 nvgpu_ce_prepare_submit(u64 src_buf,
u64 dst_buf,
u64 size,
u32 *cmd_buf_cpu_va,
u32 max_cmd_buf_size,
unsigned int payload,
u32 launch_flags,
u32 request_operation,
u32 dma_copy_class)
{
u32 launch = 0;
u32 methodSize = 0;
u64 offset = 0;
u64 chunk_size = 0;
u64 chunk = size;
/* failure case handling */
if ((nvgpu_ce_get_method_size(request_operation, size) >
max_cmd_buf_size) || (size == 0ULL) ||
(request_operation > NVGPU_CE_MEMSET)) {
return 0;
}
/* set the channel object */
cmd_buf_cpu_va[methodSize++] = 0x20018000;
cmd_buf_cpu_va[methodSize++] = dma_copy_class;
/*
* The purpose clear the memory in 2D rectangles. We get the ffs to
* determine the number of lines to copy. The only constraint is that
* maximum number of pixels per line is 4Gpix - 1, which is awkward for
* calculation, so we settle to 2Gpix per line to make calculatione
* more agreable
*/
/* The copy engine in 2D mode can have (2^32 - 1) x (2^32 - 1) pixels in
* a single submit, we are going to try to clear a range of up to 2Gpix
* multiple lines. Because we want to copy byte aligned we will be
* setting 1 byte pixels */
/*
* per iteration
* <------------------------- 40 bits ------------------------------>
* 1 <------ ffs ------->
* <-----------up to 30 bits----------->
*/
while (chunk != 0ULL) {
u32 width, height, shift;
/*
* We will be aligning to bytes, making the maximum number of
* pix per line 2Gb
*/
shift = (MAX_CE_ALIGN(chunk) != 0ULL) ?
(nvgpu_ffs(MAX_CE_ALIGN(chunk)) - 1UL) :
MAX_CE_SHIFT;
height = chunk >> shift;
width = BIT32(shift);
height = MAX_CE_ALIGN(height);
chunk_size = (u64) height * width;
/* reset launch flag */
launch = 0;
if ((request_operation & NVGPU_CE_PHYS_MODE_TRANSFER) != 0U) {
/* setup the source */
cmd_buf_cpu_va[methodSize++] = 0x20028100;
cmd_buf_cpu_va[methodSize++] = (u64_hi32(src_buf +
offset) & NVGPU_CE_UPPER_ADDRESS_OFFSET_MASK);
cmd_buf_cpu_va[methodSize++] = (u64_lo32(src_buf +
offset) & NVGPU_CE_LOWER_ADDRESS_OFFSET_MASK);
cmd_buf_cpu_va[methodSize++] = 0x20018098;
if ((launch_flags &
NVGPU_CE_SRC_LOCATION_LOCAL_FB) != 0U) {
cmd_buf_cpu_va[methodSize++] = 0x00000000;
} else if ((launch_flags &
NVGPU_CE_SRC_LOCATION_NONCOHERENT_SYSMEM) != 0U) {
cmd_buf_cpu_va[methodSize++] = 0x00000002;
} else {
cmd_buf_cpu_va[methodSize++] = 0x00000001;
}
launch |= 0x00001000U;
} else if ((request_operation & NVGPU_CE_MEMSET) != 0U) {
/* Remap from component A on 1 byte wide pixels */
cmd_buf_cpu_va[methodSize++] = 0x200181c2;
cmd_buf_cpu_va[methodSize++] = 0x00000004;
cmd_buf_cpu_va[methodSize++] = 0x200181c0;
cmd_buf_cpu_va[methodSize++] = payload;
launch |= 0x00000400U;
} else {
/* Illegal size */
return 0;
}
/* setup the destination/output */
cmd_buf_cpu_va[methodSize++] = 0x20068102;
cmd_buf_cpu_va[methodSize++] = (u64_hi32(dst_buf +
offset) & NVGPU_CE_UPPER_ADDRESS_OFFSET_MASK);
cmd_buf_cpu_va[methodSize++] = (u64_lo32(dst_buf +
offset) & NVGPU_CE_LOWER_ADDRESS_OFFSET_MASK);
/* Pitch in/out */
cmd_buf_cpu_va[methodSize++] = width;
cmd_buf_cpu_va[methodSize++] = width;
/* width and line count */
cmd_buf_cpu_va[methodSize++] = width;
cmd_buf_cpu_va[methodSize++] = height;
cmd_buf_cpu_va[methodSize++] = 0x20018099;
if ((launch_flags & NVGPU_CE_DST_LOCATION_LOCAL_FB) != 0U) {
cmd_buf_cpu_va[methodSize++] = 0x00000000;
} else if ((launch_flags &
NVGPU_CE_DST_LOCATION_NONCOHERENT_SYSMEM) != 0U) {
cmd_buf_cpu_va[methodSize++] = 0x00000002;
} else {
cmd_buf_cpu_va[methodSize++] = 0x00000001;
}
launch |= 0x00002005U;
if ((launch_flags &
NVGPU_CE_SRC_MEMORY_LAYOUT_BLOCKLINEAR) != 0U) {
launch |= 0x00000000U;
} else {
launch |= 0x00000080U;
}
if ((launch_flags &
NVGPU_CE_DST_MEMORY_LAYOUT_BLOCKLINEAR) != 0U) {
launch |= 0x00000000U;
} else {
launch |= 0x00000100U;
}
cmd_buf_cpu_va[methodSize++] = 0x200180c0;
cmd_buf_cpu_va[methodSize++] = launch;
offset += chunk_size;
chunk -= chunk_size;
}
return methodSize;
}
/* global CE app related apis */
int nvgpu_ce_init_support(struct gk20a *g) int nvgpu_ce_init_support(struct gk20a *g)
{ {
struct nvgpu_ce_app *ce_app = g->ce_app;
u32 ce_reset_mask; u32 ce_reset_mask;
if (g->ops.ce.set_pce2lce_mapping != NULL) { if (g->ops.ce.set_pce2lce_mapping != NULL) {
g->ops.ce.set_pce2lce_mapping(g); g->ops.ce.set_pce2lce_mapping(g);
} }
if (unlikely(ce_app == NULL)) {
ce_app = nvgpu_kzalloc(g, sizeof(*ce_app));
if (ce_app == NULL) {
return -ENOMEM;
}
g->ce_app = ce_app;
}
ce_reset_mask = nvgpu_engine_get_all_ce_reset_mask(g); ce_reset_mask = nvgpu_engine_get_all_ce_reset_mask(g);
g->ops.mc.reset(g, ce_reset_mask); g->ops.mc.reset(g, ce_reset_mask);
@@ -429,226 +42,9 @@ int nvgpu_ce_init_support(struct gk20a *g)
nvgpu_cg_blcg_ce_load_enable(g); nvgpu_cg_blcg_ce_load_enable(g);
if (ce_app->initialised) {
/* assume this happen during poweron/poweroff GPU sequence */
ce_app->app_state = NVGPU_CE_ACTIVE;
return 0;
}
nvgpu_log(g, gpu_dbg_fn, "ce: init");
nvgpu_mutex_init(&ce_app->app_mutex);
nvgpu_mutex_acquire(&ce_app->app_mutex);
nvgpu_init_list_node(&ce_app->allocated_contexts);
ce_app->ctx_count = 0;
ce_app->next_ctx_id = 0;
ce_app->initialised = true;
ce_app->app_state = NVGPU_CE_ACTIVE;
nvgpu_mutex_release(&ce_app->app_mutex);
if (g->ops.ce.init_prod_values != NULL) { if (g->ops.ce.init_prod_values != NULL) {
g->ops.ce.init_prod_values(g); g->ops.ce.init_prod_values(g);
} }
nvgpu_log(g, gpu_dbg_cde_ctx, "ce: init finished");
return 0; return 0;
} }
void nvgpu_ce_destroy(struct gk20a *g)
{
struct nvgpu_ce_app *ce_app = g->ce_app;
struct nvgpu_ce_gpu_ctx *ce_ctx, *ce_ctx_save;
if (ce_app == NULL) {
return;
}
if (ce_app->initialised == false) {
goto free;
}
ce_app->app_state = NVGPU_CE_SUSPEND;
ce_app->initialised = false;
nvgpu_mutex_acquire(&ce_app->app_mutex);
nvgpu_list_for_each_entry_safe(ce_ctx, ce_ctx_save,
&ce_app->allocated_contexts, nvgpu_ce_gpu_ctx, list) {
nvgpu_ce_delete_gpu_context_locked(ce_ctx);
}
nvgpu_init_list_node(&ce_app->allocated_contexts);
ce_app->ctx_count = 0;
ce_app->next_ctx_id = 0;
nvgpu_mutex_release(&ce_app->app_mutex);
nvgpu_mutex_destroy(&ce_app->app_mutex);
free:
nvgpu_kfree(g, ce_app);
g->ce_app = NULL;
}
void nvgpu_ce_suspend(struct gk20a *g)
{
struct nvgpu_ce_app *ce_app = g->ce_app;
if (ce_app == NULL || !ce_app->initialised) {
return;
}
ce_app->app_state = NVGPU_CE_SUSPEND;
}
/* CE app utility functions */
u32 nvgpu_ce_create_context(struct gk20a *g,
u32 runlist_id,
int timeslice,
int runlist_level)
{
struct nvgpu_ce_gpu_ctx *ce_ctx;
struct nvgpu_ce_app *ce_app = g->ce_app;
struct nvgpu_setup_bind_args setup_bind_args;
u32 ctx_id = NVGPU_CE_INVAL_CTX_ID;
int err = 0;
if (!ce_app->initialised || ce_app->app_state != NVGPU_CE_ACTIVE) {
return ctx_id;
}
ce_ctx = nvgpu_kzalloc(g, sizeof(*ce_ctx));
if (ce_ctx == NULL) {
return ctx_id;
}
nvgpu_mutex_init(&ce_ctx->gpu_ctx_mutex);
ce_ctx->g = g;
ce_ctx->cmd_buf_read_queue_offset = 0;
ce_ctx->vm = g->mm.ce.vm;
/* allocate a tsg if needed */
ce_ctx->tsg = nvgpu_tsg_open(g, nvgpu_current_pid(g));
if (ce_ctx->tsg == NULL) {
nvgpu_err(g, "ce: gk20a tsg not available");
goto end;
}
/* this TSG should never be aborted */
ce_ctx->tsg->abortable = false;
/* always kernel client needs privileged channel */
ce_ctx->ch = nvgpu_channel_open_new(g, runlist_id, true,
nvgpu_current_pid(g), nvgpu_current_tid(g));
if (ce_ctx->ch == NULL) {
nvgpu_err(g, "ce: gk20a channel not available");
goto end;
}
#ifdef CONFIG_NVGPU_CHANNEL_WDT
ce_ctx->ch->wdt.enabled = false;
#endif
/* bind the channel to the vm */
err = g->ops.mm.vm_bind_channel(g->mm.ce.vm, ce_ctx->ch);
if (err != 0) {
nvgpu_err(g, "ce: could not bind vm");
goto end;
}
err = nvgpu_tsg_bind_channel(ce_ctx->tsg, ce_ctx->ch);
if (err != 0) {
nvgpu_err(g, "ce: unable to bind to tsg");
goto end;
}
setup_bind_args.num_gpfifo_entries = 1024;
setup_bind_args.num_inflight_jobs = 0;
setup_bind_args.flags = 0;
err = nvgpu_channel_setup_bind(ce_ctx->ch, &setup_bind_args);
if (err != 0) {
nvgpu_err(g, "ce: unable to setup and bind channel");
goto end;
}
/* allocate command buffer from sysmem */
err = nvgpu_dma_alloc_map_sys(ce_ctx->vm,
NVGPU_CE_MAX_INFLIGHT_JOBS *
NVGPU_CE_MAX_COMMAND_BUFF_BYTES_PER_KICKOFF,
&ce_ctx->cmd_buf_mem);
if (err != 0) {
nvgpu_err(g,
"ce: alloc command buffer failed");
goto end;
}
(void) memset(ce_ctx->cmd_buf_mem.cpu_va, 0x00,
ce_ctx->cmd_buf_mem.size);
#ifdef CONFIG_NVGPU_CHANNEL_TSG_SCHEDULING
/* -1 means default channel timeslice value */
if (timeslice != -1) {
err = g->ops.tsg.set_timeslice(ce_ctx->tsg, timeslice);
if (err != 0) {
nvgpu_err(g, "ce: set timesliced failed for CE context");
goto end;
}
}
/* -1 means default channel runlist level */
if (runlist_level != -1) {
err = nvgpu_tsg_set_interleave(ce_ctx->tsg, runlist_level);
if (err != 0) {
nvgpu_err(g, "ce: set runlist interleave failed");
goto end;
}
}
#endif
nvgpu_mutex_acquire(&ce_app->app_mutex);
ctx_id = ce_ctx->ctx_id = ce_app->next_ctx_id;
nvgpu_list_add(&ce_ctx->list, &ce_app->allocated_contexts);
++ce_app->next_ctx_id;
++ce_app->ctx_count;
nvgpu_mutex_release(&ce_app->app_mutex);
ce_ctx->gpu_ctx_state = NVGPU_CE_GPU_CTX_ALLOCATED;
end:
if (ctx_id == NVGPU_CE_INVAL_CTX_ID) {
nvgpu_mutex_acquire(&ce_app->app_mutex);
nvgpu_ce_delete_gpu_context_locked(ce_ctx);
nvgpu_mutex_release(&ce_app->app_mutex);
}
return ctx_id;
}
void nvgpu_ce_delete_context(struct gk20a *g,
u32 ce_ctx_id)
{
struct nvgpu_ce_app *ce_app = g->ce_app;
struct nvgpu_ce_gpu_ctx *ce_ctx, *ce_ctx_save;
if (ce_app == NULL || !ce_app->initialised ||
ce_app->app_state != NVGPU_CE_ACTIVE) {
return;
}
nvgpu_mutex_acquire(&ce_app->app_mutex);
nvgpu_list_for_each_entry_safe(ce_ctx, ce_ctx_save,
&ce_app->allocated_contexts, nvgpu_ce_gpu_ctx, list) {
if (ce_ctx->ctx_id == ce_ctx_id) {
nvgpu_ce_delete_gpu_context_locked(ce_ctx);
--ce_app->ctx_count;
break;
}
}
nvgpu_mutex_release(&ce_app->app_mutex);
}

639
drivers/gpu/nvgpu/common/ce/ce_app.c Normal file
View File

@@ -0,0 +1,639 @@
/*
* Copyright (c) 2018-2019, NVIDIA CORPORATION. All rights reserved.
*
* 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 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.
*/
#include <nvgpu/types.h>
#include <nvgpu/gk20a.h>
#include <nvgpu/engines.h>
#include <nvgpu/os_sched.h>
#include <nvgpu/channel.h>
#include <nvgpu/dma.h>
#include <nvgpu/utils.h>
#include <nvgpu/fence.h>
#include <nvgpu/ce.h>
#include <nvgpu/power_features/cg.h>
#include "common/ce/ce_priv.h"
static inline u32 nvgpu_ce_get_valid_launch_flags(struct gk20a *g,
u32 launch_flags)
{
#ifdef CONFIG_NVGPU_DGPU
/*
* there is no local memory available,
* don't allow local memory related CE flags
*/
if (g->mm.vidmem.size == 0ULL) {
launch_flags &= ~(NVGPU_CE_SRC_LOCATION_LOCAL_FB |
NVGPU_CE_DST_LOCATION_LOCAL_FB);
}
#endif
return launch_flags;
}
int nvgpu_ce_execute_ops(struct gk20a *g,
u32 ce_ctx_id,
u64 src_buf,
u64 dst_buf,
u64 size,
unsigned int payload,
u32 launch_flags,
u32 request_operation,
u32 submit_flags,
struct nvgpu_fence_type **fence_out)
{
int ret = -EPERM;
struct nvgpu_ce_app *ce_app = g->ce_app;
struct nvgpu_ce_gpu_ctx *ce_ctx, *ce_ctx_save;
bool found = false;
u32 *cmd_buf_cpu_va;
u64 cmd_buf_gpu_va = 0UL;
u32 method_size;
u32 cmd_buf_read_offset;
u32 dma_copy_class;
struct nvgpu_gpfifo_entry gpfifo;
struct nvgpu_channel_fence fence = {0U, 0U};
struct nvgpu_fence_type *ce_cmd_buf_fence_out = NULL;
if (!ce_app->initialised || ce_app->app_state != NVGPU_CE_ACTIVE) {
goto end;
}
nvgpu_mutex_acquire(&ce_app->app_mutex);
nvgpu_list_for_each_entry_safe(ce_ctx, ce_ctx_save,
&ce_app->allocated_contexts, nvgpu_ce_gpu_ctx, list) {
if (ce_ctx->ctx_id == ce_ctx_id) {
found = true;
break;
}
}
nvgpu_mutex_release(&ce_app->app_mutex);
if (!found) {
ret = -EINVAL;
goto end;
}
if (ce_ctx->gpu_ctx_state != NVGPU_CE_GPU_CTX_ALLOCATED) {
ret = -ENODEV;
goto end;
}
nvgpu_mutex_acquire(&ce_ctx->gpu_ctx_mutex);
ce_ctx->cmd_buf_read_queue_offset %= NVGPU_CE_MAX_INFLIGHT_JOBS;
cmd_buf_read_offset = (ce_ctx->cmd_buf_read_queue_offset *
(NVGPU_CE_MAX_COMMAND_BUFF_BYTES_PER_KICKOFF /
U32(sizeof(u32))));
cmd_buf_cpu_va = (u32 *)ce_ctx->cmd_buf_mem.cpu_va;
if (ce_ctx->postfences[ce_ctx->cmd_buf_read_queue_offset] != NULL) {
struct nvgpu_fence_type **prev_post_fence =
&ce_ctx->postfences[ce_ctx->cmd_buf_read_queue_offset];
ret = nvgpu_fence_wait(g, *prev_post_fence,
nvgpu_get_poll_timeout(g));
nvgpu_fence_put(*prev_post_fence);
*prev_post_fence = NULL;
if (ret != 0) {
goto noop;
}
}
cmd_buf_gpu_va = (ce_ctx->cmd_buf_mem.gpu_va +
(u64)(cmd_buf_read_offset * sizeof(u32)));
dma_copy_class = g->ops.get_litter_value(g, GPU_LIT_DMA_COPY_CLASS);
method_size = nvgpu_ce_prepare_submit(src_buf,
dst_buf,
size,
&cmd_buf_cpu_va[cmd_buf_read_offset],
NVGPU_CE_MAX_COMMAND_BUFF_BYTES_PER_KICKOFF,
payload,
nvgpu_ce_get_valid_launch_flags(g, launch_flags),
request_operation,
dma_copy_class);
if (method_size != 0U) {
/* store the element into gpfifo */
g->ops.pbdma.format_gpfifo_entry(g, &gpfifo,
cmd_buf_gpu_va, method_size);
/*
* take always the postfence as it is needed for protecting the
* ce context
*/
submit_flags |= NVGPU_SUBMIT_FLAGS_FENCE_GET;
nvgpu_smp_wmb();
ret = nvgpu_submit_channel_gpfifo_kernel(ce_ctx->ch, &gpfifo,
1, submit_flags, &fence, &ce_cmd_buf_fence_out);
if (ret == 0) {
ce_ctx->postfences[ce_ctx->cmd_buf_read_queue_offset] =
ce_cmd_buf_fence_out;
if (fence_out != NULL) {
nvgpu_fence_get(ce_cmd_buf_fence_out);
*fence_out = ce_cmd_buf_fence_out;
}
/* Next available command buffer queue Index */
++ce_ctx->cmd_buf_read_queue_offset;
}
} else {
ret = -ENOMEM;
}
noop:
nvgpu_mutex_release(&ce_ctx->gpu_ctx_mutex);
end:
return ret;
}
/* static CE app api */
static void nvgpu_ce_put_fences(struct nvgpu_ce_gpu_ctx *ce_ctx)
{
u32 i;
for (i = 0U; i < NVGPU_CE_MAX_INFLIGHT_JOBS; i++) {
struct nvgpu_fence_type **fence = &ce_ctx->postfences[i];
if (*fence != NULL) {
nvgpu_fence_put(*fence);
}
*fence = NULL;
}
}
/* caller must hold ce_app->app_mutex */
static void nvgpu_ce_delete_gpu_context_locked(struct nvgpu_ce_gpu_ctx *ce_ctx)
{
struct nvgpu_list_node *list = &ce_ctx->list;
ce_ctx->gpu_ctx_state = NVGPU_CE_GPU_CTX_DELETED;
ce_ctx->tsg->abortable = true;
nvgpu_mutex_acquire(&ce_ctx->gpu_ctx_mutex);
if (nvgpu_mem_is_valid(&ce_ctx->cmd_buf_mem)) {
nvgpu_ce_put_fences(ce_ctx);
nvgpu_dma_unmap_free(ce_ctx->vm, &ce_ctx->cmd_buf_mem);
}
/*
* free the channel
* nvgpu_channel_close() will also unbind the channel from TSG
*/
nvgpu_channel_close(ce_ctx->ch);
nvgpu_ref_put(&ce_ctx->tsg->refcount, nvgpu_tsg_release);
/* housekeeping on app */
if ((list->prev != NULL) && (list->next != NULL)) {
nvgpu_list_del(list);
}
nvgpu_mutex_release(&ce_ctx->gpu_ctx_mutex);
nvgpu_mutex_destroy(&ce_ctx->gpu_ctx_mutex);
nvgpu_kfree(ce_ctx->g, ce_ctx);
}
static inline unsigned int nvgpu_ce_get_method_size(u32 request_operation,
u64 size)
{
/* failure size */
unsigned int methodsize = UINT_MAX;
unsigned int iterations = 0;
u32 shift;
u64 chunk = size;
u32 height, width;
while (chunk != 0ULL) {
iterations++;
shift = (MAX_CE_ALIGN(chunk) != 0ULL) ?
(nvgpu_ffs(MAX_CE_ALIGN(chunk)) - 1UL) :
MAX_CE_SHIFT;
width = chunk >> shift;
height = BIT32(shift);
width = MAX_CE_ALIGN(width);
chunk -= (u64) height * width;
}
if ((request_operation & NVGPU_CE_PHYS_MODE_TRANSFER) != 0U) {
methodsize = (2U + (16U * iterations)) *
(unsigned int)sizeof(u32);
} else if ((request_operation & NVGPU_CE_MEMSET) != 0U) {
methodsize = (2U + (15U * iterations)) *
(unsigned int)sizeof(u32);
}
return methodsize;
}
u32 nvgpu_ce_prepare_submit(u64 src_buf,
u64 dst_buf,
u64 size,
u32 *cmd_buf_cpu_va,
u32 max_cmd_buf_size,
unsigned int payload,
u32 launch_flags,
u32 request_operation,
u32 dma_copy_class)
{
u32 launch = 0;
u32 methodSize = 0;
u64 offset = 0;
u64 chunk_size = 0;
u64 chunk = size;
/* failure case handling */
if ((nvgpu_ce_get_method_size(request_operation, size) >
max_cmd_buf_size) || (size == 0ULL) ||
(request_operation > NVGPU_CE_MEMSET)) {
return 0;
}
/* set the channel object */
cmd_buf_cpu_va[methodSize++] = 0x20018000;
cmd_buf_cpu_va[methodSize++] = dma_copy_class;
/*
* The purpose clear the memory in 2D rectangles. We get the ffs to
* determine the number of lines to copy. The only constraint is that
* maximum number of pixels per line is 4Gpix - 1, which is awkward for
* calculation, so we settle to 2Gpix per line to make calculatione
* more agreable
*/
/*
* The copy engine in 2D mode can have (2^32 - 1) x (2^32 - 1) pixels in
* a single submit, we are going to try to clear a range of up to 2Gpix
* multiple lines. Because we want to copy byte aligned we will be
* setting 1 byte pixels
*/
/*
* per iteration
* <------------------------- 40 bits ------------------------------>
* 1 <------ ffs ------->
* <-----------up to 30 bits----------->
*/
while (chunk != 0ULL) {
u32 width, height, shift;
/*
* We will be aligning to bytes, making the maximum number of
* pix per line 2Gb
*/
shift = (MAX_CE_ALIGN(chunk) != 0ULL) ?
(nvgpu_ffs(MAX_CE_ALIGN(chunk)) - 1UL) :
MAX_CE_SHIFT;
height = chunk >> shift;
width = BIT32(shift);
height = MAX_CE_ALIGN(height);
chunk_size = (u64) height * width;
/* reset launch flag */
launch = 0;
if ((request_operation & NVGPU_CE_PHYS_MODE_TRANSFER) != 0U) {
/* setup the source */
cmd_buf_cpu_va[methodSize++] = 0x20028100;
cmd_buf_cpu_va[methodSize++] = (u64_hi32(src_buf +
offset) & NVGPU_CE_UPPER_ADDRESS_OFFSET_MASK);
cmd_buf_cpu_va[methodSize++] = (u64_lo32(src_buf +
offset) & NVGPU_CE_LOWER_ADDRESS_OFFSET_MASK);
cmd_buf_cpu_va[methodSize++] = 0x20018098;
if ((launch_flags &
NVGPU_CE_SRC_LOCATION_LOCAL_FB) != 0U) {
cmd_buf_cpu_va[methodSize++] = 0x00000000;
} else if ((launch_flags &
NVGPU_CE_SRC_LOCATION_NONCOHERENT_SYSMEM) != 0U) {
cmd_buf_cpu_va[methodSize++] = 0x00000002;
} else {
cmd_buf_cpu_va[methodSize++] = 0x00000001;
}
launch |= 0x00001000U;
} else if ((request_operation & NVGPU_CE_MEMSET) != 0U) {
/* Remap from component A on 1 byte wide pixels */
cmd_buf_cpu_va[methodSize++] = 0x200181c2;
cmd_buf_cpu_va[methodSize++] = 0x00000004;
cmd_buf_cpu_va[methodSize++] = 0x200181c0;
cmd_buf_cpu_va[methodSize++] = payload;
launch |= 0x00000400U;
} else {
/* Illegal size */
return 0;
}
/* setup the destination/output */
cmd_buf_cpu_va[methodSize++] = 0x20068102;
cmd_buf_cpu_va[methodSize++] = (u64_hi32(dst_buf +
offset) & NVGPU_CE_UPPER_ADDRESS_OFFSET_MASK);
cmd_buf_cpu_va[methodSize++] = (u64_lo32(dst_buf +
offset) & NVGPU_CE_LOWER_ADDRESS_OFFSET_MASK);
/* Pitch in/out */
cmd_buf_cpu_va[methodSize++] = width;
cmd_buf_cpu_va[methodSize++] = width;
/* width and line count */
cmd_buf_cpu_va[methodSize++] = width;
cmd_buf_cpu_va[methodSize++] = height;
cmd_buf_cpu_va[methodSize++] = 0x20018099;
if ((launch_flags & NVGPU_CE_DST_LOCATION_LOCAL_FB) != 0U) {
cmd_buf_cpu_va[methodSize++] = 0x00000000;
} else if ((launch_flags &
NVGPU_CE_DST_LOCATION_NONCOHERENT_SYSMEM) != 0U) {
cmd_buf_cpu_va[methodSize++] = 0x00000002;
} else {
cmd_buf_cpu_va[methodSize++] = 0x00000001;
}
launch |= 0x00002005U;
if ((launch_flags &
NVGPU_CE_SRC_MEMORY_LAYOUT_BLOCKLINEAR) != 0U) {
launch |= 0x00000000U;
} else {
launch |= 0x00000080U;
}
if ((launch_flags &
NVGPU_CE_DST_MEMORY_LAYOUT_BLOCKLINEAR) != 0U) {
launch |= 0x00000000U;
} else {
launch |= 0x00000100U;
}
cmd_buf_cpu_va[methodSize++] = 0x200180c0;
cmd_buf_cpu_va[methodSize++] = launch;
offset += chunk_size;
chunk -= chunk_size;
}
return methodSize;
}
/* global CE app related apis */
int nvgpu_ce_app_init_support(struct gk20a *g)
{
struct nvgpu_ce_app *ce_app = g->ce_app;
if (unlikely(ce_app == NULL)) {
ce_app = nvgpu_kzalloc(g, sizeof(*ce_app));
if (ce_app == NULL) {
return -ENOMEM;
}
g->ce_app = ce_app;
}
if (ce_app->initialised) {
/* assume this happen during poweron/poweroff GPU sequence */
ce_app->app_state = NVGPU_CE_ACTIVE;
return 0;
}
nvgpu_log(g, gpu_dbg_fn, "ce: init");
nvgpu_mutex_init(&ce_app->app_mutex);
nvgpu_mutex_acquire(&ce_app->app_mutex);
nvgpu_init_list_node(&ce_app->allocated_contexts);
ce_app->ctx_count = 0;
ce_app->next_ctx_id = 0;
ce_app->initialised = true;
ce_app->app_state = NVGPU_CE_ACTIVE;
nvgpu_mutex_release(&ce_app->app_mutex);
nvgpu_log(g, gpu_dbg_cde_ctx, "ce: init finished");
return 0;
}
void nvgpu_ce_app_destroy(struct gk20a *g)
{
struct nvgpu_ce_app *ce_app = g->ce_app;
struct nvgpu_ce_gpu_ctx *ce_ctx, *ce_ctx_save;
if (ce_app == NULL) {
return;
}
if (ce_app->initialised == false) {
goto free;
}
ce_app->app_state = NVGPU_CE_SUSPEND;
ce_app->initialised = false;
nvgpu_mutex_acquire(&ce_app->app_mutex);
nvgpu_list_for_each_entry_safe(ce_ctx, ce_ctx_save,
&ce_app->allocated_contexts, nvgpu_ce_gpu_ctx, list) {
nvgpu_ce_delete_gpu_context_locked(ce_ctx);
}
nvgpu_init_list_node(&ce_app->allocated_contexts);
ce_app->ctx_count = 0;
ce_app->next_ctx_id = 0;
nvgpu_mutex_release(&ce_app->app_mutex);
nvgpu_mutex_destroy(&ce_app->app_mutex);
free:
nvgpu_kfree(g, ce_app);
g->ce_app = NULL;
}
void nvgpu_ce_app_suspend(struct gk20a *g)
{
struct nvgpu_ce_app *ce_app = g->ce_app;
if (ce_app == NULL || !ce_app->initialised) {
return;
}
ce_app->app_state = NVGPU_CE_SUSPEND;
}
/* CE app utility functions */
u32 nvgpu_ce_app_create_context(struct gk20a *g,
u32 runlist_id,
int timeslice,
int runlist_level)
{
struct nvgpu_ce_gpu_ctx *ce_ctx;
struct nvgpu_ce_app *ce_app = g->ce_app;
struct nvgpu_setup_bind_args setup_bind_args;
u32 ctx_id = NVGPU_CE_INVAL_CTX_ID;
int err = 0;
if (!ce_app->initialised || ce_app->app_state != NVGPU_CE_ACTIVE) {
return ctx_id;
}
ce_ctx = nvgpu_kzalloc(g, sizeof(*ce_ctx));
if (ce_ctx == NULL) {
return ctx_id;
}
nvgpu_mutex_init(&ce_ctx->gpu_ctx_mutex);
ce_ctx->g = g;
ce_ctx->cmd_buf_read_queue_offset = 0;
ce_ctx->vm = g->mm.ce.vm;
/* allocate a tsg if needed */
ce_ctx->tsg = nvgpu_tsg_open(g, nvgpu_current_pid(g));
if (ce_ctx->tsg == NULL) {
nvgpu_err(g, "ce: gk20a tsg not available");
goto end;
}
/* this TSG should never be aborted */
ce_ctx->tsg->abortable = false;
/* always kernel client needs privileged channel */
ce_ctx->ch = nvgpu_channel_open_new(g, runlist_id, true,
nvgpu_current_pid(g), nvgpu_current_tid(g));
if (ce_ctx->ch == NULL) {
nvgpu_err(g, "ce: gk20a channel not available");
goto end;
}
#ifdef CONFIG_NVGPU_CHANNEL_WDT
ce_ctx->ch->wdt.enabled = false;
#endif
/* bind the channel to the vm */
err = g->ops.mm.vm_bind_channel(g->mm.ce.vm, ce_ctx->ch);
if (err != 0) {
nvgpu_err(g, "ce: could not bind vm");
goto end;
}
err = nvgpu_tsg_bind_channel(ce_ctx->tsg, ce_ctx->ch);
if (err != 0) {
nvgpu_err(g, "ce: unable to bind to tsg");
goto end;
}
setup_bind_args.num_gpfifo_entries = 1024;
setup_bind_args.num_inflight_jobs = 0;
setup_bind_args.flags = 0;
err = nvgpu_channel_setup_bind(ce_ctx->ch, &setup_bind_args);
if (err != 0) {
nvgpu_err(g, "ce: unable to setup and bind channel");
goto end;
}
/* allocate command buffer from sysmem */
err = nvgpu_dma_alloc_map_sys(ce_ctx->vm,
NVGPU_CE_MAX_INFLIGHT_JOBS *
NVGPU_CE_MAX_COMMAND_BUFF_BYTES_PER_KICKOFF,
&ce_ctx->cmd_buf_mem);
if (err != 0) {
nvgpu_err(g,
"ce: alloc command buffer failed");
goto end;
}
(void) memset(ce_ctx->cmd_buf_mem.cpu_va, 0x00,
ce_ctx->cmd_buf_mem.size);
#ifdef CONFIG_NVGPU_CHANNEL_TSG_SCHEDULING
/* -1 means default channel timeslice value */
if (timeslice != -1) {
err = g->ops.tsg.set_timeslice(ce_ctx->tsg, timeslice);
if (err != 0) {
nvgpu_err(g, "ce: set timesliced failed for CE context");
goto end;
}
}
/* -1 means default channel runlist level */
if (runlist_level != -1) {
err = nvgpu_tsg_set_interleave(ce_ctx->tsg, runlist_level);
if (err != 0) {
nvgpu_err(g, "ce: set runlist interleave failed");
goto end;
}
}
#endif
nvgpu_mutex_acquire(&ce_app->app_mutex);
ctx_id = ce_ctx->ctx_id = ce_app->next_ctx_id;
nvgpu_list_add(&ce_ctx->list, &ce_app->allocated_contexts);
++ce_app->next_ctx_id;
++ce_app->ctx_count;
nvgpu_mutex_release(&ce_app->app_mutex);
ce_ctx->gpu_ctx_state = NVGPU_CE_GPU_CTX_ALLOCATED;
end:
if (ctx_id == NVGPU_CE_INVAL_CTX_ID) {
nvgpu_mutex_acquire(&ce_app->app_mutex);
nvgpu_ce_delete_gpu_context_locked(ce_ctx);
nvgpu_mutex_release(&ce_app->app_mutex);
}
return ctx_id;
}
void nvgpu_ce_app_delete_context(struct gk20a *g,
u32 ce_ctx_id)
{
struct nvgpu_ce_app *ce_app = g->ce_app;
struct nvgpu_ce_gpu_ctx *ce_ctx, *ce_ctx_save;
if (ce_app == NULL || !ce_app->initialised ||
ce_app->app_state != NVGPU_CE_ACTIVE) {
return;
}
nvgpu_mutex_acquire(&ce_app->app_mutex);
nvgpu_list_for_each_entry_safe(ce_ctx, ce_ctx_save,
&ce_app->allocated_contexts, nvgpu_ce_gpu_ctx, list) {
if (ce_ctx->ctx_id == ce_ctx_id) {
nvgpu_ce_delete_gpu_context_locked(ce_ctx);
--ce_app->ctx_count;
break;
}
}
nvgpu_mutex_release(&ce_app->app_mutex);
}

21
drivers/gpu/nvgpu/common/init/nvgpu_init.c
View File

@@ -47,9 +47,6 @@
#include <nvgpu/channel_sync.h> #include <nvgpu/channel_sync.h>
#include <nvgpu/gr/gr.h> #include <nvgpu/gr/gr.h>
#include <nvgpu/nvgpu_init.h> #include <nvgpu/nvgpu_init.h>
#ifndef CONFIG_NVGPU_CE
#include <nvgpu/engines.h>
#endif
#ifdef CONFIG_NVGPU_TRACE #ifdef CONFIG_NVGPU_TRACE
#include <trace/events/gk20a.h> #include <trace/events/gk20a.h>
@@ -136,10 +133,8 @@ int nvgpu_prepare_poweroff(struct gk20a *g)
nvgpu_falcon_sw_free(g, FALCON_ID_GSPLITE); nvgpu_falcon_sw_free(g, FALCON_ID_GSPLITE);
nvgpu_falcon_sw_free(g, FALCON_ID_NVDEC); nvgpu_falcon_sw_free(g, FALCON_ID_NVDEC);
nvgpu_falcon_sw_free(g, FALCON_ID_SEC2); nvgpu_falcon_sw_free(g, FALCON_ID_SEC2);
#endif
#ifdef CONFIG_NVGPU_CE nvgpu_ce_app_suspend(g);
nvgpu_ce_suspend(g);
#endif #endif
#ifdef CONFIG_NVGPU_DGPU #ifdef CONFIG_NVGPU_DGPU
@@ -494,17 +489,19 @@ int nvgpu_finalize_poweron(struct gk20a *g)
g->ops.fb.set_debug_mode(g, g->mmu_debug_ctrl); g->ops.fb.set_debug_mode(g, g->mmu_debug_ctrl);
#endif #endif
#ifdef CONFIG_NVGPU_CE
err = nvgpu_ce_init_support(g); err = nvgpu_ce_init_support(g);
if (err != 0) { if (err != 0) {
nvgpu_err(g, "failed to init ce"); nvgpu_err(g, "failed to init ce");
goto done; goto done;
} }
#else
g->ops.mc.reset(g, nvgpu_engine_get_all_ce_reset_mask(g));
#endif
#ifdef CONFIG_NVGPU_DGPU #ifdef CONFIG_NVGPU_DGPU
err = nvgpu_ce_app_init_support(g);
if (err != 0) {
nvgpu_err(g, "failed to init ce app");
goto done;
}
if (g->ops.xve.available_speeds != NULL) { if (g->ops.xve.available_speeds != NULL) {
u32 speed; u32 speed;
@@ -655,8 +652,8 @@ static void gk20a_free_cb(struct nvgpu_ref *refcount)
nvgpu_log(g, gpu_dbg_shutdown, "Freeing GK20A struct!"); nvgpu_log(g, gpu_dbg_shutdown, "Freeing GK20A struct!");
#ifdef CONFIG_NVGPU_CE #ifdef CONFIG_NVGPU_DGPU
nvgpu_ce_destroy(g); nvgpu_ce_app_destroy(g);
#endif #endif
#ifdef CONFIG_NVGPU_COMPRESSION #ifdef CONFIG_NVGPU_COMPRESSION

16
drivers/gpu/nvgpu/common/mm/mm.c
View File

@@ -114,19 +114,17 @@ static int nvgpu_alloc_sysmem_flush(struct gk20a *g)
return nvgpu_dma_alloc_sys(g, SZ_4K, &g->mm.sysmem_flush); return nvgpu_dma_alloc_sys(g, SZ_4K, &g->mm.sysmem_flush);
} }
#ifdef CONFIG_NVGPU_CE #ifdef CONFIG_NVGPU_DGPU
static void nvgpu_remove_mm_ce_support(struct mm_gk20a *mm) static void nvgpu_remove_mm_ce_support(struct mm_gk20a *mm)
{ {
#ifdef CONFIG_NVGPU_DGPU
struct gk20a *g = gk20a_from_mm(mm); struct gk20a *g = gk20a_from_mm(mm);
if (mm->vidmem.ce_ctx_id != NVGPU_CE_INVAL_CTX_ID) { if (mm->vidmem.ce_ctx_id != NVGPU_CE_INVAL_CTX_ID) {
nvgpu_ce_delete_context(g, mm->vidmem.ce_ctx_id); nvgpu_ce_app_delete_context(g, mm->vidmem.ce_ctx_id);
} }
mm->vidmem.ce_ctx_id = NVGPU_CE_INVAL_CTX_ID; mm->vidmem.ce_ctx_id = NVGPU_CE_INVAL_CTX_ID;
nvgpu_vm_put(mm->ce.vm); nvgpu_vm_put(mm->ce.vm);
#endif
} }
#endif #endif
@@ -303,14 +301,13 @@ static int nvgpu_init_mmu_debug(struct mm_gk20a *mm)
return -ENOMEM; return -ENOMEM;
} }
#ifdef CONFIG_NVGPU_CE #if defined(CONFIG_NVGPU_DGPU)
void nvgpu_init_mm_ce_context(struct gk20a *g) void nvgpu_init_mm_ce_context(struct gk20a *g)
{ {
#if defined(CONFIG_NVGPU_DGPU)
if (g->mm.vidmem.size > 0U && if (g->mm.vidmem.size > 0U &&
(g->mm.vidmem.ce_ctx_id == NVGPU_CE_INVAL_CTX_ID)) { (g->mm.vidmem.ce_ctx_id == NVGPU_CE_INVAL_CTX_ID)) {
g->mm.vidmem.ce_ctx_id = g->mm.vidmem.ce_ctx_id =
nvgpu_ce_create_context(g, nvgpu_ce_app_create_context(g,
nvgpu_engine_get_fast_ce_runlist_id(g), nvgpu_engine_get_fast_ce_runlist_id(g),
-1, -1,
-1); -1);
@@ -320,9 +317,8 @@ void nvgpu_init_mm_ce_context(struct gk20a *g)
"Failed to allocate CE context for vidmem page clearing support"); "Failed to allocate CE context for vidmem page clearing support");
} }
} }
#endif
} }
#endif /* NVGPU_FENCE_CE */ #endif
static int nvgpu_init_mm_reset_enable_hw(struct gk20a *g) static int nvgpu_init_mm_reset_enable_hw(struct gk20a *g)
{ {
@@ -523,7 +519,7 @@ static int nvgpu_init_mm_setup_sw(struct gk20a *g)
} }
mm->remove_support = nvgpu_remove_mm_support; mm->remove_support = nvgpu_remove_mm_support;
#ifdef CONFIG_NVGPU_CE #ifdef CONFIG_NVGPU_DGPU
mm->remove_ce_support = nvgpu_remove_mm_ce_support; mm->remove_ce_support = nvgpu_remove_mm_ce_support;
#endif #endif

11
drivers/gpu/nvgpu/common/mm/vidmem.c
View File

@@ -103,7 +103,7 @@ static int nvgpu_vidmem_do_clear_all(struct gk20a *g)
vidmem_dbg(g, "Clearing all VIDMEM:"); vidmem_dbg(g, "Clearing all VIDMEM:");
#ifdef CONFIG_NVGPU_CE #ifdef CONFIG_NVGPU_DGPU
err = nvgpu_ce_execute_ops(g, err = nvgpu_ce_execute_ops(g,
mm->vidmem.ce_ctx_id, mm->vidmem.ce_ctx_id,
0, 0,
@@ -119,6 +119,9 @@ static int nvgpu_vidmem_do_clear_all(struct gk20a *g)
"Failed to clear vidmem : %d", err); "Failed to clear vidmem : %d", err);
return err; return err;
} }
#else
/* fail due to lack of ce app support */
return -ENOSYS;
#endif #endif
if (fence_out != NULL) { if (fence_out != NULL) {
@@ -464,7 +467,7 @@ int nvgpu_vidmem_clear(struct gk20a *g, struct nvgpu_mem *mem)
nvgpu_fence_put(last_fence); nvgpu_fence_put(last_fence);
} }
#ifdef CONFIG_NVGPU_CE #ifdef CONFIG_NVGPU_DGPU
err = nvgpu_ce_execute_ops(g, err = nvgpu_ce_execute_ops(g,
g->mm.vidmem.ce_ctx_id, g->mm.vidmem.ce_ctx_id,
0, 0,
@@ -476,12 +479,12 @@ int nvgpu_vidmem_clear(struct gk20a *g, struct nvgpu_mem *mem)
0, 0,
&fence_out); &fence_out);
#else #else
/* fail due to lack of ce support */ /* fail due to lack of ce app support */
err = -ENOSYS; err = -ENOSYS;
#endif #endif
if (err != 0) { if (err != 0) {
#ifdef CONFIG_NVGPU_CE #ifdef CONFIG_NVGPU_DGPU
nvgpu_err(g, nvgpu_err(g,
"Failed nvgpu_ce_execute_ops[%d]", err); "Failed nvgpu_ce_execute_ops[%d]", err);
#endif #endif

12
drivers/gpu/nvgpu/include/nvgpu/ce.h
View File

@@ -69,17 +69,19 @@ enum {
NVGPU_CE_GPU_CTX_DELETED = (1 << 1), NVGPU_CE_GPU_CTX_DELETED = (1 << 1),
}; };
/* global CE app related apis */
int nvgpu_ce_init_support(struct gk20a *g); int nvgpu_ce_init_support(struct gk20a *g);
void nvgpu_ce_suspend(struct gk20a *g);
void nvgpu_ce_destroy(struct gk20a *g); /* global CE app related apis */
int nvgpu_ce_app_init_support(struct gk20a *g);
void nvgpu_ce_app_suspend(struct gk20a *g);
void nvgpu_ce_app_destroy(struct gk20a *g);
/* CE app utility functions */ /* CE app utility functions */
u32 nvgpu_ce_create_context(struct gk20a *g, u32 nvgpu_ce_app_create_context(struct gk20a *g,
u32 runlist_id, u32 runlist_id,
int timeslice, int timeslice,
int runlist_level); int runlist_level);
void nvgpu_ce_delete_context(struct gk20a *g, void nvgpu_ce_app_delete_context(struct gk20a *g,
u32 ce_ctx_id); u32 ce_ctx_id);
int nvgpu_ce_execute_ops(struct gk20a *g, int nvgpu_ce_execute_ops(struct gk20a *g,
u32 ce_ctx_id, u32 ce_ctx_id,

4
drivers/gpu/nvgpu/include/nvgpu/mm.h
View File

@@ -242,7 +242,7 @@ struct mm_gk20a {
struct nvgpu_mem hw_fault_buf[NVGPU_MMU_FAULT_TYPE_NUM]; struct nvgpu_mem hw_fault_buf[NVGPU_MMU_FAULT_TYPE_NUM];
struct mmu_fault_info fault_info[NVGPU_MMU_FAULT_TYPE_NUM]; struct mmu_fault_info fault_info[NVGPU_MMU_FAULT_TYPE_NUM];
struct nvgpu_mutex hub_isr_mutex; struct nvgpu_mutex hub_isr_mutex;
#ifdef CONFIG_NVGPU_CE #ifdef CONFIG_NVGPU_DGPU
/* /*
* Separate function to cleanup the CE since it requires a channel to * Separate function to cleanup the CE since it requires a channel to
* be closed which must happen before fifo cleanup. * be closed which must happen before fifo cleanup.
@@ -317,7 +317,7 @@ static inline u64 nvgpu_gmmu_va_small_page_limit(void)
return ((u64)SZ_1G * 56U); return ((u64)SZ_1G * 56U);
} }
#ifdef CONFIG_NVGPU_CE #ifdef CONFIG_NVGPU_DGPU
void nvgpu_init_mm_ce_context(struct gk20a *g); void nvgpu_init_mm_ce_context(struct gk20a *g);
#endif #endif
int nvgpu_init_mm_support(struct gk20a *g); int nvgpu_init_mm_support(struct gk20a *g);

2
drivers/gpu/nvgpu/os/linux/module.c
View File

@@ -469,7 +469,7 @@ int gk20a_pm_finalize_poweron(struct device *dev)
if (err) if (err)
goto done; goto done;
#ifdef CONFIG_NVGPU_CE #ifdef CONFIG_NVGPU_DGPU
nvgpu_init_mm_ce_context(g); nvgpu_init_mm_ce_context(g);
#endif #endif