Include nvgpu headers

These are needed to build on NVIDIA's Jetson boards for the time
being. Only a couple structs are required, so it should be fairly
easy to remove this dependency at some point in the future.

1 files changed, 636 insertions, 0 deletions

diff --git a/include/gk20a/css_gr_gk20a.c b/include/gk20a/css_gr_gk20a.c
new file mode 100644
index 0000000..28a3d49
--- /dev/null
+++ b/include/gk20a/css_gr_gk20a.c
@@ -0,0 +1,636 @@
+/*
+ * GK20A Cycle stats snapshots support (subsystem for gr_gk20a).
+ *
+ * Copyright (c) 2015-2018, 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/bitops.h>
+#include <nvgpu/kmem.h>
+#include <nvgpu/lock.h>
+#include <nvgpu/dma.h>
+#include <nvgpu/mm.h>
+#include <nvgpu/sizes.h>
+#include <nvgpu/barrier.h>
+#include <nvgpu/log.h>
+#include <nvgpu/bug.h>
+#include <nvgpu/io.h>
+#include <nvgpu/utils.h>
+#include <nvgpu/channel.h>
+#include <nvgpu/unit.h>
+
+#include "gk20a.h"
+#include "css_gr_gk20a.h"
+
+#include <nvgpu/hw/gk20a/hw_perf_gk20a.h>
+
+/* check client for pointed perfmon ownership */
+#define CONTAINS_PERFMON(cl, pm)                                \
+                ((cl)->perfmon_start <= (pm) &&                 \
+                ((pm) - (cl)->perfmon_start) < (cl)->perfmon_count)
+
+/* address of fifo entry by offset */
+#define CSS_FIFO_ENTRY(fifo, offs)                              \
+        ((struct gk20a_cs_snapshot_fifo_entry *)(((char *)(fifo)) + (offs)))
+
+/* calculate area capacity in number of fifo entries */
+#define CSS_FIFO_ENTRY_CAPACITY(s)                              \
+        (((s) - sizeof(struct gk20a_cs_snapshot_fifo))          \
+                / sizeof(struct gk20a_cs_snapshot_fifo_entry))
+
+/* reserved to indicate failures with data */
+#define CSS_FIRST_PERFMON_ID    32
+/* should correlate with size of gk20a_cs_snapshot_fifo_entry::perfmon_id */
+#define CSS_MAX_PERFMON_IDS     256
+
+/* reports whether the hw queue overflowed */
+bool css_hw_get_overflow_status(struct gk20a *g)
+{
+        const u32 st = perf_pmasys_control_membuf_status_overflowed_f();
+        return st == (gk20a_readl(g, perf_pmasys_control_r()) & st);
+}
+
+/* returns how many pending snapshot entries are pending */
+u32 css_hw_get_pending_snapshots(struct gk20a *g)
+{
+        return gk20a_readl(g, perf_pmasys_mem_bytes_r()) /
+                        sizeof(struct gk20a_cs_snapshot_fifo_entry);
+}
+
+/* informs hw how many snapshots have been processed (frees up fifo space) */
+void css_hw_set_handled_snapshots(struct gk20a *g, u32 done)
+{
+        if (done > 0) {
+                gk20a_writel(g, perf_pmasys_mem_bump_r(),
+                     done * sizeof(struct gk20a_cs_snapshot_fifo_entry));
+        }
+}
+
+/* disable streaming to memory */
+static void css_hw_reset_streaming(struct gk20a *g)
+{
+        u32 engine_status;
+
+        /* reset the perfmon */
+        g->ops.mc.reset(g, g->ops.mc.reset_mask(g, NVGPU_UNIT_PERFMON));
+
+        /* RBUFEMPTY must be set -- otherwise we'll pick up */
+        /* snapshot that have been queued up from earlier   */
+        engine_status = gk20a_readl(g, perf_pmasys_enginestatus_r());
+        WARN_ON(0 == (engine_status
+                        & perf_pmasys_enginestatus_rbufempty_empty_f()));
+
+        /* turn off writes */
+        gk20a_writel(g, perf_pmasys_control_r(),
+                        perf_pmasys_control_membuf_clear_status_doit_f());
+
+        /* pointing all pending snapshots as handled */
+        css_hw_set_handled_snapshots(g, css_hw_get_pending_snapshots(g));
+}
+
+/*
+ * WARNING: all css_gr_XXX functions are local and expected to be called
+ * from locked context (protected by cs_lock)
+ */
+
+static int css_gr_create_shared_data(struct gr_gk20a *gr)
+{
+        struct gk20a_cs_snapshot *data;
+
+        if (gr->cs_data)
+                return 0;
+
+        data = nvgpu_kzalloc(gr->g, sizeof(*data));
+        if (!data)
+                return -ENOMEM;
+
+        nvgpu_init_list_node(&data->clients);
+        gr->cs_data = data;
+
+        return 0;
+}
+
+int css_hw_enable_snapshot(struct channel_gk20a *ch,
+                                struct gk20a_cs_snapshot_client *cs_client)
+{
+        struct gk20a *g = ch->g;
+        struct mm_gk20a *mm = &g->mm;
+        struct gr_gk20a *gr = &g->gr;
+        struct gk20a_cs_snapshot *data = gr->cs_data;
+        u32 snapshot_size = cs_client->snapshot_size;
+        int ret;
+
+        u32 virt_addr_lo;
+        u32 virt_addr_hi;
+        u32 inst_pa_page;
+
+        if (data->hw_snapshot)
+                return 0;
+
+        if (snapshot_size < CSS_MIN_HW_SNAPSHOT_SIZE)
+                snapshot_size = CSS_MIN_HW_SNAPSHOT_SIZE;
+
+        ret = nvgpu_dma_alloc_map_sys(g->mm.pmu.vm, snapshot_size,
+                                                        &data->hw_memdesc);
+        if (ret)
+                return ret;
+
+        /* perf output buffer may not cross a 4GB boundary - with a separate */
+        /* va smaller than that, it won't but check anyway */
+        if (!data->hw_memdesc.cpu_va ||
+                data->hw_memdesc.size < snapshot_size ||
+                data->hw_memdesc.gpu_va + u64_lo32(snapshot_size) > SZ_4G) {
+                ret = -EFAULT;
+                goto failed_allocation;
+        }
+
+        data->hw_snapshot =
+                (struct gk20a_cs_snapshot_fifo_entry *)data->hw_memdesc.cpu_va;
+        data->hw_end = data->hw_snapshot +
+                snapshot_size / sizeof(struct gk20a_cs_snapshot_fifo_entry);
+        data->hw_get = data->hw_snapshot;
+        memset(data->hw_snapshot, 0xff, snapshot_size);
+
+        /* address and size are aligned to 32 bytes, the lowest bits read back
+         * as zeros */
+        virt_addr_lo = u64_lo32(data->hw_memdesc.gpu_va);
+        virt_addr_hi = u64_hi32(data->hw_memdesc.gpu_va);
+
+        css_hw_reset_streaming(g);
+
+        gk20a_writel(g, perf_pmasys_outbase_r(), virt_addr_lo);
+        gk20a_writel(g, perf_pmasys_outbaseupper_r(),
+                        perf_pmasys_outbaseupper_ptr_f(virt_addr_hi));
+        gk20a_writel(g, perf_pmasys_outsize_r(), snapshot_size);
+
+        /* this field is aligned to 4K */
+        inst_pa_page = nvgpu_inst_block_addr(g, &g->mm.hwpm.inst_block) >> 12;
+
+        /* A write to MEM_BLOCK triggers the block bind operation. MEM_BLOCK
+         * should be written last */
+        gk20a_writel(g, perf_pmasys_mem_block_r(),
+                        perf_pmasys_mem_block_base_f(inst_pa_page) |
+                        nvgpu_aperture_mask(g, &mm->hwpm.inst_block,
+                                perf_pmasys_mem_block_target_sys_ncoh_f(),
+                                perf_pmasys_mem_block_target_sys_coh_f(),
+                                perf_pmasys_mem_block_target_lfb_f()) |
+                        perf_pmasys_mem_block_valid_true_f());
+
+        nvgpu_log_info(g, "cyclestats: buffer for hardware snapshots enabled\n");
+
+        return 0;
+
+failed_allocation:
+        if (data->hw_memdesc.size) {
+                nvgpu_dma_unmap_free(g->mm.pmu.vm, &data->hw_memdesc);
+                memset(&data->hw_memdesc, 0, sizeof(data->hw_memdesc));
+        }
+        data->hw_snapshot = NULL;
+
+        return ret;
+}
+
+void css_hw_disable_snapshot(struct gr_gk20a *gr)
+{
+        struct gk20a *g = gr->g;