7 files changed, 33 insertions, 21 deletions

diff --git a/drivers/gpu/nvgpu/gv11b/acr_gv11b.c b/drivers/gpu/nvgpu/gv11b/acr_gv11b.c
index 799b2db4..4fa3f324 100644
--- a/drivers/gpu/nvgpu/gv11b/acr_gv11b.c
+++ b/drivers/gpu/nvgpu/gv11b/acr_gv11b.c
@@ -27,9 +27,10 @@
 #include <nvgpu/nvgpu_common.h>
 #include <nvgpu/kmem.h>
 #include <nvgpu/nvgpu_mem.h>
-#include <nvgpu/acr/nvgpu_acr.h>
 #include <nvgpu/firmware.h>
 #include <nvgpu/mm.h>
+#include <nvgpu/enabled.h>
+#include <nvgpu/acr/nvgpu_acr.h>
 
 #include "gk20a/gk20a.h"
 #include "acr_gv11b.h"
@@ -220,7 +221,9 @@ static int bl_bootstrap(struct nvgpu_pmu *pmu,
                 pwr_pmu_new_instblk_ptr_f(
                 nvgpu_inst_block_addr(g, &mm->pmu.inst_block) >> 12) |
                 pwr_pmu_new_instblk_valid_f(1) |
-                pwr_pmu_new_instblk_target_sys_ncoh_f());
+                     (nvgpu_is_enabled(g, NVGPU_USE_COHERENT_SYSMEM) ?
+                      pwr_pmu_new_instblk_target_sys_coh_f() :
+                      pwr_pmu_new_instblk_target_sys_ncoh_f())) ;
 
         /*copy bootloader interface structure to dmem*/
         nvgpu_flcn_copy_to_dmem(pmu->flcn, 0, (u8 *)pbl_desc,
diff --git a/drivers/gpu/nvgpu/gv11b/css_gr_gv11b.c b/drivers/gpu/nvgpu/gv11b/css_gr_gv11b.c
index 617ea61d..86977bb3 100644
--- a/drivers/gpu/nvgpu/gv11b/css_gr_gv11b.c
+++ b/drivers/gpu/nvgpu/gv11b/css_gr_gv11b.c
@@ -31,14 +31,14 @@
 #include <nvgpu/dma.h>
 #include <nvgpu/mm.h>
 #include <nvgpu/sizes.h>
+#include <nvgpu/enabled.h>
+#include <nvgpu/log.h>
+#include <nvgpu/bug.h>
 
 #include "gk20a/gk20a.h"
 #include "gk20a/css_gr_gk20a.h"
 #include "css_gr_gv11b.h"
 
-#include <nvgpu/log.h>
-#include <nvgpu/bug.h>
-
 #include <nvgpu/hw/gv11b/hw_perf_gv11b.h>
 #include <nvgpu/hw/gv11b/hw_mc_gv11b.h>
 
@@ -144,6 +144,7 @@ int gv11b_css_hw_enable_snapshot(struct channel_gk20a *ch,
                         perf_pmasys_mem_block_valid_true_f() |
                         nvgpu_aperture_mask(g, &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()));
 
 
diff --git a/drivers/gpu/nvgpu/gv11b/dbg_gpu_gv11b.c b/drivers/gpu/nvgpu/gv11b/dbg_gpu_gv11b.c
index e5d88e8c..562476ca 100644
--- a/drivers/gpu/nvgpu/gv11b/dbg_gpu_gv11b.c
+++ b/drivers/gpu/nvgpu/gv11b/dbg_gpu_gv11b.c
@@ -59,11 +59,12 @@ int gv11b_perfbuf_enable_locked(struct gk20a *g, u64 offset, u32 size)
         inst_pa_page = nvgpu_inst_block_addr(g, &mm->perfbuf.inst_block) >> 12;
 
         gk20a_writel(g, perf_pmasys_mem_block_r(),
-                        perf_pmasys_mem_block_base_f(inst_pa_page) |
-                        perf_pmasys_mem_block_valid_true_f() |
-                        nvgpu_aperture_mask(g, &mm->perfbuf.inst_block,
-+                               perf_pmasys_mem_block_target_sys_ncoh_f(),
-+                               perf_pmasys_mem_block_target_lfb_f()));
+                     perf_pmasys_mem_block_base_f(inst_pa_page) |
+                     perf_pmasys_mem_block_valid_true_f() |
+                     nvgpu_aperture_mask(g, &mm->perfbuf.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()));
 
         gk20a_idle(g);
         return 0;
diff --git a/drivers/gpu/nvgpu/gv11b/fifo_gv11b.c b/drivers/gpu/nvgpu/gv11b/fifo_gv11b.c
index feed2002..9e60d9f7 100644
--- a/drivers/gpu/nvgpu/gv11b/fifo_gv11b.c
+++ b/drivers/gpu/nvgpu/gv11b/fifo_gv11b.c
@@ -101,12 +101,14 @@ void gv11b_get_ch_runlist_entry(struct channel_gk20a *c, u32 *runlist)
                                                 c->runqueue_sel) |
                         ram_rl_entry_chan_userd_target_f(
                                 nvgpu_aperture_mask(g, &g->fifo.userd,
-                                ram_rl_entry_chan_userd_target_sys_mem_ncoh_v(),
-                                ram_rl_entry_chan_userd_target_vid_mem_v())) |
+                                        ram_rl_entry_chan_userd_target_sys_mem_ncoh_v(),
+                                        ram_rl_entry_chan_userd_target_sys_mem_coh_v(),
+                                        ram_rl_entry_chan_userd_target_vid_mem_v())) |
                         ram_rl_entry_chan_inst_target_f(
                                 nvgpu_aperture_mask(g, &c->inst_block,
-                                ram_rl_entry_chan_inst_target_sys_mem_ncoh_v(),
-                                ram_rl_entry_chan_inst_target_vid_mem_v()));
+                                        ram_rl_entry_chan_inst_target_sys_mem_ncoh_v(),
+                                        ram_rl_entry_chan_inst_target_sys_mem_coh_v(),
+                                        ram_rl_entry_chan_inst_target_vid_mem_v()));
 
         addr_lo = u64_lo32(c->userd_iova) >>
                         ram_rl_entry_chan_userd_ptr_align_shift_v();
diff --git a/drivers/gpu/nvgpu/gv11b/mm_gv11b.c b/drivers/gpu/nvgpu/gv11b/mm_gv11b.c
index ade1d9fe..b46ecb0a 100644
--- a/drivers/gpu/nvgpu/gv11b/mm_gv11b.c
+++ b/drivers/gpu/nvgpu/gv11b/mm_gv11b.c
@@ -26,6 +26,7 @@
 #include <nvgpu/dma.h>
 #include <nvgpu/log.h>
 #include <nvgpu/mm.h>
+#include <nvgpu/enabled.h>
 
 #include "gk20a/gk20a.h"
 #include "gk20a/mm_gk20a.h"
@@ -292,8 +293,9 @@ int gv11b_init_bar2_mm_hw_setup(struct gk20a *g)
 
         gk20a_writel(g, bus_bar2_block_r(),
                      nvgpu_aperture_mask(g, inst_block,
-                                bus_bar2_block_target_sys_mem_ncoh_f(),
-                                bus_bar2_block_target_vid_mem_f()) |
+                                         bus_bar2_block_target_sys_mem_ncoh_f(),
+                                         bus_bar2_block_target_sys_mem_coh_f(),
+                                         bus_bar2_block_target_vid_mem_f()) |
                      bus_bar2_block_mode_virtual_f() |
                      bus_bar2_block_ptr_f(inst_pa));
 
diff --git a/drivers/gpu/nvgpu/gv11b/pmu_gv11b.c b/drivers/gpu/nvgpu/gv11b/pmu_gv11b.c
index 32e751d9..c1b519d0 100644
--- a/drivers/gpu/nvgpu/gv11b/pmu_gv11b.c
+++ b/drivers/gpu/nvgpu/gv11b/pmu_gv11b.c
@@ -195,9 +195,11 @@ int gv11b_pmu_bootstrap(struct nvgpu_pmu *pmu)
 
         gk20a_writel(g, pwr_pmu_new_instblk_r(),
                 pwr_pmu_new_instblk_ptr_f(
-                nvgpu_inst_block_addr(g, &mm->pmu.inst_block) >> ALIGN_4KB)
-                | pwr_pmu_new_instblk_valid_f(1)
-                | pwr_pmu_new_instblk_target_sys_ncoh_f());
+                nvgpu_inst_block_addr(g, &mm->pmu.inst_block) >> ALIGN_4KB) |
+                     pwr_pmu_new_instblk_valid_f(1) |
+                     (nvgpu_is_enabled(g, NVGPU_USE_COHERENT_SYSMEM) ?
+                      pwr_pmu_new_instblk_target_sys_coh_f() :
+                      pwr_pmu_new_instblk_target_sys_ncoh_f()));
 
         /* TBD: load all other surfaces */
         g->ops.pmu_ver.set_pmu_cmdline_args_trace_size(
diff --git a/drivers/gpu/nvgpu/gv11b/subctx_gv11b.c b/drivers/gpu/nvgpu/gv11b/subctx_gv11b.c
index 12a606bf..61fb8954 100644
--- a/drivers/gpu/nvgpu/gv11b/subctx_gv11b.c
+++ b/drivers/gpu/nvgpu/gv11b/subctx_gv11b.c
@@ -178,8 +178,9 @@ void gv11b_subctx_commit_pdb(struct vm_gk20a *vm,
         u64 pdb_addr;
         u32 max_subctx_count = gr_pri_fe_chip_def_info_max_veid_count_init_v();
         u32 aperture = nvgpu_aperture_mask(g, vm->pdb.mem,
-                        ram_in_sc_page_dir_base_target_sys_mem_ncoh_v(),
-                        ram_in_sc_page_dir_base_target_vid_mem_v());
+                                ram_in_sc_page_dir_base_target_sys_mem_ncoh_v(),
+                                ram_in_sc_page_dir_base_target_sys_mem_coh_v(),
+                                ram_in_sc_page_dir_base_target_vid_mem_v());
 
         pdb_addr = nvgpu_mem_get_addr(g, vm->pdb.mem);
         pdb_addr_lo = u64_lo32(pdb_addr >> ram_in_base_shift_v());