linux-nvgpu/drivers/gpu/nvgpu/gm20b/ltc_gm20b.c

/*
 * GM20B L2
 *
 * Copyright (c) 2014-2017 NVIDIA CORPORATION.  All rights reserved.
 *
 * This program is free software; you can redistribute it and/or modify it
 * under the terms and conditions of the GNU General Public License,
 * version 2, as published by the Free Software Foundation.
 *
 * This program is distributed in the hope it will be useful, but WITHOUT
 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
 * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
 * more details.
 */

#include <trace/events/gk20a.h>

#include "gk20a/gk20a.h"

#include <nvgpu/timers.h>
#include <nvgpu/enabled.h>
#include <nvgpu/bug.h>

#include <nvgpu/hw/gm20b/hw_mc_gm20b.h>
#include <nvgpu/hw/gm20b/hw_ltc_gm20b.h>
#include <nvgpu/hw/gm20b/hw_top_gm20b.h>
#include <nvgpu/hw/gm20b/hw_pri_ringmaster_gm20b.h>

#include "gk20a/ltc_common.c"
#include "ltc_gm20b.h"

static int gm20b_ltc_init_comptags(struct gk20a *g, struct gr_gk20a *gr)
{
	/* max memory size (MB) to cover */
	u32 max_size = gr->max_comptag_mem;
	/* one tag line covers 128KB */
	u32 max_comptag_lines = max_size << 3;

	u32 hw_max_comptag_lines =
		ltc_ltcs_ltss_cbc_ctrl3_clear_upper_bound_init_v();

	u32 cbc_param =
		gk20a_readl(g, ltc_ltcs_ltss_cbc_param_r());
	u32 comptags_per_cacheline =
		ltc_ltcs_ltss_cbc_param_comptags_per_cache_line_v(cbc_param);
	u32 cacheline_size =
		512 << ltc_ltcs_ltss_cbc_param_cache_line_size_v(cbc_param);
	u32 slices_per_ltc =
		ltc_ltcs_ltss_cbc_param_slices_per_ltc_v(cbc_param);

	u32 compbit_backing_size;

	int err;

	gk20a_dbg_fn("");

	if (max_comptag_lines == 0)
		return 0;

	if (max_comptag_lines > hw_max_comptag_lines)
		max_comptag_lines = hw_max_comptag_lines;

	compbit_backing_size =
		DIV_ROUND_UP(max_comptag_lines, comptags_per_cacheline) *
		cacheline_size * slices_per_ltc * g->ltc_count;

	/* aligned to 2KB * ltc_count */
	compbit_backing_size +=
		g->ltc_count << ltc_ltcs_ltss_cbc_base_alignment_shift_v();

	/* must be a multiple of 64KB */
	compbit_backing_size = roundup(compbit_backing_size, 64*1024);

	max_comptag_lines =
		(compbit_backing_size * comptags_per_cacheline) /
		(cacheline_size * slices_per_ltc * g->ltc_count);

	if (max_comptag_lines > hw_max_comptag_lines)
		max_comptag_lines = hw_max_comptag_lines;

	gk20a_dbg_info("compbit backing store size : %d",
		compbit_backing_size);
	gk20a_dbg_info("max comptag lines : %d",
		max_comptag_lines);

	if (nvgpu_is_enabled(g, NVGPU_IS_FMODEL))
		err = gk20a_ltc_alloc_phys_cbc(g, compbit_backing_size);
	else
		err = gk20a_ltc_alloc_virt_cbc(g, compbit_backing_size);

	if (err)
		return err;

	err = gk20a_comptag_allocator_init(&gr->comp_tags, max_comptag_lines);
	if (err)
		return err;

	gr->comptags_per_cacheline = comptags_per_cacheline;
	gr->slices_per_ltc = slices_per_ltc;
	gr->cacheline_size = cacheline_size;

	return 0;
}

int gm20b_ltc_cbc_ctrl(struct gk20a *g, enum gk20a_cbc_op op,
		       u32 min, u32 max)
{
	struct gr_gk20a *gr = &g->gr;
	struct nvgpu_timeout timeout;
	int err = 0;
	u32 ltc, slice, ctrl1, val, hw_op = 0;
	u32 slices_per_ltc = ltc_ltcs_ltss_cbc_param_slices_per_ltc_v(
				gk20a_readl(g, ltc_ltcs_ltss_cbc_param_r()));
	u32 ltc_stride = nvgpu_get_litter_value(g, GPU_LIT_LTC_STRIDE);
	u32 lts_stride = nvgpu_get_litter_value(g, GPU_LIT_LTS_STRIDE);

	gk20a_dbg_fn("");

	trace_gk20a_ltc_cbc_ctrl_start(g->name, op, min, max);

	if (gr->compbit_store.mem.size == 0)
		return 0;

	nvgpu_mutex_acquire(&g->mm.l2_op_lock);

	if (op == gk20a_cbc_op_clear) {
		gk20a_writel(g, ltc_ltcs_ltss_cbc_ctrl2_r(),
			ltc_ltcs_ltss_cbc_ctrl2_clear_lower_bound_f(min));
		gk20a_writel(g, ltc_ltcs_ltss_cbc_ctrl3_r(),
			ltc_ltcs_ltss_cbc_ctrl3_clear_upper_bound_f(max));
		hw_op = ltc_ltcs_ltss_cbc_ctrl1_clear_active_f();
	} else if (op == gk20a_cbc_op_clean) {
		hw_op = ltc_ltcs_ltss_cbc_ctrl1_clean_active_f();
	} else if (op == gk20a_cbc_op_invalidate) {
		hw_op = ltc_ltcs_ltss_cbc_ctrl1_invalidate_active_f();
	} else {
		BUG_ON(1);
	}
	gk20a_writel(g, ltc_ltcs_ltss_cbc_ctrl1_r(),
		     gk20a_readl(g, ltc_ltcs_ltss_cbc_ctrl1_r()) | hw_op);

	for (ltc = 0; ltc < g->ltc_count; ltc++) {
		for (slice = 0; slice < slices_per_ltc; slice++) {

			ctrl1 = ltc_ltc0_lts0_cbc_ctrl1_r() +
				ltc * ltc_stride + slice * lts_stride;

			nvgpu_timeout_init(g, &timeout, 2000,
					   NVGPU_TIMER_RETRY_TIMER);
			do {
				val = gk20a_readl(g, ctrl1);
				if (!(val & hw_op))
					break;
				nvgpu_udelay(5);
			} while (!nvgpu_timeout_expired(&timeout));

			if (nvgpu_timeout_peek_expired(&timeout)) {
				nvgpu_err(g, "comp tag clear timeout");
				err = -EBUSY;
				goto out;
			}
		}
	}
out:
	trace_gk20a_ltc_cbc_ctrl_done(g->name);
	nvgpu_mutex_release(&g->mm.l2_op_lock);
	return err;
}

void gm20b_ltc_init_fs_state(struct gk20a *g)
{
	u32 reg;

	gk20a_dbg_info("initialize gm20b l2");

	g->max_ltc_count = gk20a_readl(g, top_num_ltcs_r());
	g->ltc_count = gk20a_readl(g, pri_ringmaster_enum_ltc_r());
	gk20a_dbg_info("%d ltcs out of %d", g->ltc_count, g->max_ltc_count);

	gk20a_writel(g, ltc_ltcs_ltss_cbc_num_active_ltcs_r(),
	g->ltc_count);
	gk20a_writel(g, ltc_ltcs_misc_ltc_num_active_ltcs_r(),
	g->ltc_count);

	gk20a_writel(g, ltc_ltcs_ltss_dstg_cfg0_r(),
		     gk20a_readl(g, ltc_ltc0_lts0_dstg_cfg0_r()) |
		     ltc_ltcs_ltss_dstg_cfg0_vdc_4to2_disable_m());

	/* Disable LTC interrupts */
	reg = gk20a_readl(g, ltc_ltcs_ltss_intr_r());
	reg &= ~ltc_ltcs_ltss_intr_en_evicted_cb_m();
	reg &= ~ltc_ltcs_ltss_intr_en_illegal_compstat_access_m();
	reg &= ~ltc_ltcs_ltss_intr_en_illegal_compstat_m();
	gk20a_writel(g, ltc_ltcs_ltss_intr_r(), reg);
}

void gm20b_ltc_isr(struct gk20a *g)
{
	u32 mc_intr, ltc_intr;
	unsigned int ltc, slice;
	u32 ltc_stride = nvgpu_get_litter_value(g, GPU_LIT_LTC_STRIDE);
	u32 lts_stride = nvgpu_get_litter_value(g, GPU_LIT_LTS_STRIDE);

	mc_intr = gk20a_readl(g, mc_intr_ltc_r());
	nvgpu_err(g, "mc_ltc_intr: %08x", mc_intr);
	for (ltc = 0; ltc < g->ltc_count; ltc++) {
		if ((mc_intr & 1 << ltc) == 0)
			continue;
		for (slice = 0; slice < g->gr.slices_per_ltc; slice++) {
			ltc_intr = gk20a_readl(g, ltc_ltc0_lts0_intr_r() +
					   ltc_stride * ltc +
					   lts_stride * slice);
			nvgpu_err(g, "ltc%d, slice %d: %08x",
				  ltc, slice, ltc_intr);
			gk20a_writel(g, ltc_ltc0_lts0_intr_r() +
					   ltc_stride * ltc +
					   lts_stride * slice,
				     ltc_intr);
		}
	}
}

u32 gm20b_ltc_cbc_fix_config(struct gk20a *g, int base)
{
	u32 val = gk20a_readl(g, ltc_ltcs_ltss_cbc_num_active_ltcs_r());
	if (val == 2) {
		return base * 2;
	} else if (val != 1) {
		nvgpu_err(g, "Invalid number of active ltcs: %08x", val);
	}

	return base;
}

/*
 * Performs a full flush of the L2 cache.
 */
void gm20b_flush_ltc(struct gk20a *g)
{
	struct nvgpu_timeout timeout;
	unsigned int ltc;
	u32 ltc_stride = nvgpu_get_litter_value(g, GPU_LIT_LTC_STRIDE);

	/* Clean... */
	gk20a_writel(g, ltc_ltcs_ltss_tstg_cmgmt1_r(),
		ltc_ltcs_ltss_tstg_cmgmt1_clean_pending_f() |
		ltc_ltcs_ltss_tstg_cmgmt1_max_cycles_between_cleans_3_f() |
		ltc_ltcs_ltss_tstg_cmgmt1_clean_wait_for_fb_to_pull_true_f() |
		ltc_ltcs_ltss_tstg_cmgmt1_clean_evict_last_class_true_f() |
		ltc_ltcs_ltss_tstg_cmgmt1_clean_evict_normal_class_true_f() |
		ltc_ltcs_ltss_tstg_cmgmt1_clean_evict_first_class_true_f());

	/* Wait on each LTC individually. */
	for (ltc = 0; ltc < g->ltc_count; ltc++) {
		u32 op_pending;

		/*
		 * Use 5ms - this should be sufficient time to flush the cache.
		 * On tegra, rough EMC BW available for old tegra chips (newer
		 * chips are strictly faster) can be estimated as follows:
		 *
		 * Lowest reasonable EMC clock speed will be around 102MHz on
		 * t124 for display enabled boards and generally fixed to max
		 * for non-display boards (since they are generally plugged in).
		 *
		 * Thus, the available BW is 64b * 2 * 102MHz = 1.3GB/s. Of that
		 * BW the GPU will likely get about half (display and overhead/
		 * utilization inefficiency eating the rest) so 650MB/s at
		 * worst. Assuming at most 1MB of GPU L2 cache (less for most
		 * chips) worst case is we take 1MB/650MB/s = 1.5ms.
		 *
		 * So 5ms timeout here should be more than sufficient.
		 */
		nvgpu_timeout_init(g, &timeout, 5, NVGPU_TIMER_CPU_TIMER);

		do {
			int cmgmt1 = ltc_ltc0_ltss_tstg_cmgmt1_r() +
				     ltc * ltc_stride;
			op_pending = gk20a_readl(g, cmgmt1);
		} while ((op_pending &
			  ltc_ltc0_ltss_tstg_cmgmt1_clean_pending_f()) &&
			 !nvgpu_timeout_expired_msg(&timeout,
						    "L2 flush timeout!"));
	}

	/* And invalidate. */
	gk20a_writel(g, ltc_ltcs_ltss_tstg_cmgmt0_r(),
	     ltc_ltcs_ltss_tstg_cmgmt0_invalidate_pending_f() |
	     ltc_ltcs_ltss_tstg_cmgmt0_max_cycles_between_invalidates_3_f() |
	     ltc_ltcs_ltss_tstg_cmgmt0_invalidate_evict_last_class_true_f() |
	     ltc_ltcs_ltss_tstg_cmgmt0_invalidate_evict_normal_class_true_f() |
	     ltc_ltcs_ltss_tstg_cmgmt0_invalidate_evict_first_class_true_f());

	/* Wait on each LTC individually. */
	for (ltc = 0; ltc < g->ltc_count; ltc++) {
		u32 op_pending;

		/* Again, 5ms. */
		nvgpu_timeout_init(g, &timeout, 5, NVGPU_TIMER_CPU_TIMER);

		do {
			int cmgmt0 = ltc_ltc0_ltss_tstg_cmgmt0_r() +
				     ltc * ltc_stride;
			op_pending = gk20a_readl(g, cmgmt0);
		} while ((op_pending &
			  ltc_ltc0_ltss_tstg_cmgmt0_invalidate_pending_f()) &&
			 !nvgpu_timeout_expired_msg(&timeout,
						    "L2 flush timeout!"));
	}
}

static int gm20b_determine_L2_size_bytes(struct gk20a *g)
{
	u32 lts_per_ltc;
	u32 ways;
	u32 sets;
	u32 bytes_per_line;
	u32 active_ltcs;
	u32 cache_size;

	u32 tmp;
	u32 active_sets_value;

	tmp = gk20a_readl(g, ltc_ltc0_lts0_tstg_cfg1_r());
	ways = hweight32(ltc_ltc0_lts0_tstg_cfg1_active_ways_v(tmp));

	active_sets_value = ltc_ltc0_lts0_tstg_cfg1_active_sets_v(tmp);
	if (active_sets_value == ltc_ltc0_lts0_tstg_cfg1_active_sets_all_v()) {
		sets = 64;
	} else if (active_sets_value ==
		 ltc_ltc0_lts0_tstg_cfg1_active_sets_half_v()) {
		sets = 32;
	} else if (active_sets_value ==
		 ltc_ltc0_lts0_tstg_cfg1_active_sets_quarter_v()) {
		sets = 16;
	} else {
		nvgpu_err(g, "Unknown constant %u for active sets",
		       (unsigned)active_sets_value);
		sets = 0;
	}

	active_ltcs = g->gr.num_fbps;

	/* chip-specific values */
	lts_per_ltc = 2;
	bytes_per_line = 128;
	cache_size = active_ltcs * lts_per_ltc * ways * sets * bytes_per_line;

	return cache_size;
}

void gm20b_init_ltc(struct gpu_ops *gops)
{
	/* Gk20a reused ops. */
	gops->ltc.determine_L2_size_bytes = gm20b_determine_L2_size_bytes;
	gops->ltc.set_zbc_color_entry = gk20a_ltc_set_zbc_color_entry;
	gops->ltc.set_zbc_depth_entry = gk20a_ltc_set_zbc_depth_entry;
	gops->ltc.init_cbc = gk20a_ltc_init_cbc;

	/* GM20b specific ops. */
	gops->ltc.init_fs_state = gm20b_ltc_init_fs_state;
	gops->ltc.init_comptags = gm20b_ltc_init_comptags;
	gops->ltc.cbc_ctrl = gm20b_ltc_cbc_ctrl;
	gops->ltc.isr = gm20b_ltc_isr;
	gops->ltc.cbc_fix_config = gm20b_ltc_cbc_fix_config;
	gops->ltc.flush = gm20b_flush_ltc;
#ifdef CONFIG_DEBUG_FS
	gops->ltc.sync_debugfs = gk20a_ltc_sync_debugfs;
#endif
}