start mmap'ing vram, with hugetlb if possible

10 files changed, 152 insertions, 30 deletions

diff --git a/frontend/common/plat.h b/frontend/common/plat.h
index 0a9fc0b..416f8ac 100644
--- a/frontend/common/plat.h
+++ b/frontend/common/plat.h
@@ -45,6 +45,10 @@ int  plat_is_dir(const char *path);
 int  plat_wait_event(int *fds_hnds, int count, int timeout_ms);
 void plat_sleep_ms(int ms);
 
+void *plat_mmap(unsigned long addr, size_t size, int need_exec);
+void *plat_mremap(void *ptr, size_t oldsize, size_t newsize);
+void  plat_munmap(void *ptr, size_t size);
+
 /* timers, to be used for time diff and must refer to the same clock */
 unsigned int plat_get_ticks_ms(void);
 unsigned int plat_get_ticks_us(void);
diff --git a/frontend/linux/plat.c b/frontend/linux/plat.c
index b7152b5..044084e 100644
--- a/frontend/linux/plat.c
+++ b/frontend/linux/plat.c
@@ -17,9 +17,17 @@
 #include <time.h>
 #include <unistd.h>
 #include <sys/mman.h>
+#include <errno.h>
 
 #include "../common/plat.h"
 
+/* XXX: maybe unhardcode pagesize? */
+#define HUGETLB_PAGESIZE (2 * 1024 * 1024)
+#define HUGETLB_THRESHOLD (HUGETLB_PAGESIZE / 2)
+#ifndef MAP_HUGETLB
+#define MAP_HUGETLB 0x40000 /* arch specific */
+#endif
+
 
 int plat_is_dir(const char *path)
 {
@@ -126,16 +134,34 @@ int plat_wait_event(int *fds_hnds, int count, int timeout_ms)
 	return ret;
 }
 
-void *plat_mmap(unsigned long addr, size_t size)
+void *plat_mmap(unsigned long addr, size_t size, int need_exec)
 {
+	static int hugetlb_disabled;
+	int prot = PROT_READ | PROT_WRITE;
+	int flags = MAP_PRIVATE | MAP_ANONYMOUS;
 	void *req, *ret;
 
 	req = (void *)addr;
-	ret = mmap(req, size, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_ANONYMOUS, -1, 0);
+	if (need_exec)
+		prot |= PROT_EXEC;
+	if (size >= HUGETLB_THRESHOLD && !hugetlb_disabled)
+		flags |= MAP_HUGETLB;
+
+	ret = mmap(req, size, prot, flags, -1, 0);
+	if (ret == MAP_FAILED && (flags & MAP_HUGETLB)) {
+		fprintf(stderr,
+			"warning: failed to do hugetlb mmap (%p, %zu): %d\n",
+			req, size, errno);
+		hugetlb_disabled = 1;
+		flags &= ~MAP_HUGETLB;
+		ret = mmap(req, size, prot, flags, -1, 0);
+	}
 	if (ret == MAP_FAILED)
 		return NULL;
-	if (ret != req)
-		printf("warning: mmaped to %p, requested %p\n", ret, req);
+
+	if (req != NULL && ret != req)
+		fprintf(stderr,
+			"warning: mmaped to %p, requested %p\n", ret, req);
 
 	return ret;
 }
@@ -155,7 +181,18 @@ void *plat_mremap(void *ptr, size_t oldsize, size_t newsize)
 
 void plat_munmap(void *ptr, size_t size)
 {
-	munmap(ptr, size);
+	int ret;
+
+	ret = munmap(ptr, size);
+	if (ret != 0 && (size & (HUGETLB_PAGESIZE - 1))) {
+		// prehaps an autorounded hugetlb mapping?
+		size = (size + HUGETLB_PAGESIZE - 1) & ~(HUGETLB_PAGESIZE - 1);
+		ret = munmap(ptr, size);
+	}
+	if (ret != 0) {
+		fprintf(stderr,
+			"munmap(%p, %zu) failed: %d\n", ptr, size, errno);
+	}
 }
 
 /* lprintf */
diff --git a/frontend/plugin_lib.c b/frontend/plugin_lib.c
index 4dbb9a7..c2e2ab4 100644
--- a/frontend/plugin_lib.c
+++ b/frontend/plugin_lib.c
@@ -21,6 +21,7 @@
 #include "linux/fbdev.h"
 #include "common/fonts.h"
 #include "common/input.h"
+#include "common/plat.h"
 #include "menu.h"
 #include "main.h"
 #include "plat.h"
@@ -484,12 +485,24 @@ static void pl_get_layer_pos(int *x, int *y, int *w, int *h)
 	*h = g_layer_h;
 }
 
+static void *pl_mmap(unsigned int size)
+{
+	return plat_mmap(0, size, 0);
+}
+
+static void pl_munmap(void *ptr, unsigned int size)
+{
+	plat_munmap(ptr, size);
+}
+
 struct rearmed_cbs pl_rearmed_cbs = {
 	pl_get_layer_pos,
 	pl_vout_open,
 	pl_vout_set_mode,
 	pl_vout_flip,
 	pl_vout_close,
+	pl_mmap,
+	pl_munmap,
 };
 
 /* watchdog */
diff --git a/frontend/plugin_lib.h b/frontend/plugin_lib.h
index 7687bf8..1701d06 100644
--- a/frontend/plugin_lib.h
+++ b/frontend/plugin_lib.h
@@ -44,6 +44,8 @@ struct rearmed_cbs {
 	void *(*pl_vout_set_mode)(int w, int h, int bpp);
 	void *(*pl_vout_flip)(void);
 	void  (*pl_vout_close)(void);
+	void *(*mmap)(unsigned int size);
+	void  (*munmap)(void *ptr, unsigned int size);
 	// these are only used by some frontends
 	void  (*pl_vout_raw_flip)(int x, int y);
 	void  (*pl_vout_set_raw_vram)(void *vram);
diff --git a/plugins/dfxvideo/gpulib_if.c b/plugins/dfxvideo/gpulib_if.c
index 6f425bd..50130f6 100644
--- a/plugins/dfxvideo/gpulib_if.c
+++ b/plugins/dfxvideo/gpulib_if.c
@@ -265,9 +265,9 @@ long           lLowerpart;
 
 /////////////////////////////////////////////////////////////////////////////
 
-int renderer_init(void)
+static void set_vram(void *vram)
 {
- psxVub=(void *)gpu.vram;
+ psxVub=vram;
 
  psxVsb=(signed char *)psxVub;                         // different ways of accessing PSX VRAM
  psxVsw=(signed short *)psxVub;
@@ -276,6 +276,11 @@ int renderer_init(void)
  psxVul=(uint32_t *)psxVub;
 
  psxVuw_eom=psxVuw+1024*512;                           // pre-calc of end of vram
+}
+
+int renderer_init(void)
+{
+ set_vram(gpu.vram);
 
  PSXDisplay.RGB24        = FALSE;                      // init some stuff
  PSXDisplay.Interlaced   = FALSE;
@@ -416,4 +421,5 @@ void renderer_set_config(const struct rearmed_cbs *cbs)
 {
  iUseDither = cbs->gpu_peops.iUseDither;
  dwActFixes = cbs->gpu_peops.dwActFixes;
+ set_vram(gpu.vram);
 }
diff --git a/plugins/gpu-gles/gpulib_if.c b/plugins/gpu-gles/gpulib_if.c
index 09dc200..d245c44 100644
--- a/plugins/gpu-gles/gpulib_if.c
+++ b/plugins/gpu-gles/gpulib_if.c
@@ -479,10 +479,15 @@ switch((gdata>>24)&0xff)
 
 static int is_opened;
 
-int renderer_init(void)
+static void set_vram(void *vram)
 {
- psxVub=(void *)gpu.vram;
+ psxVub=vram;
  psxVuw=(unsigned short *)psxVub;
+}
+
+int renderer_init(void)
+{
+ set_vram(gpu.vram);
 
  PSXDisplay.RGB24        = FALSE;                      // init some stuff
  PSXDisplay.Interlaced   = FALSE;
@@ -710,6 +715,8 @@ void renderer_set_config(const struct rearmed_cbs *cbs_)
  bUseFastMdec = cbs->gpu_peopsgl.bUseFastMdec;
  iTexGarbageCollection = cbs->gpu_peopsgl.iTexGarbageCollection;
  iVRamSize = cbs->gpu_peopsgl.iVRamSize;
+
+ set_vram(gpu.vram);
 }
 
 void SetAspectRatio(void)
diff --git a/plugins/gpu_neon/psx_gpu_if.c b/plugins/gpu_neon/psx_gpu_if.c
index ca76fe2..1b4dcc5 100644
--- a/plugins/gpu_neon/psx_gpu_if.c
+++ b/plugins/gpu_neon/psx_gpu_if.c
@@ -15,6 +15,7 @@ extern const unsigned char cmd_lengths[256];
 #define command_lengths cmd_lengths
 
 static unsigned int *ex_regs;
+static int initialized;
 
 #define PCSX
 #define SET_Ex(r, v) \
@@ -42,33 +43,38 @@ int do_cmd_list(uint32_t *list, int count, int *last_cmd)
 
 #define ENHANCEMENT_BUF_SIZE (1024 * 1024 * 2 * 4 + 4096)
 
-int renderer_init(void)
+static void map_enhancement_buffer(void)
 {
-  initialize_psx_gpu(&egpu, gpu.vram);
-  ex_regs = gpu.ex_regs;
+  // currently we use 4x 1024*1024 buffers instead of single 2048*1024
+  // to be able to reuse 1024-width code better (triangle setup,
+  // dithering phase, lines).
+  gpu.enhancement_bufer = gpu.mmap(ENHANCEMENT_BUF_SIZE);
+  if (gpu.enhancement_bufer == NULL)
+    fprintf(stderr, "failed to map enhancement buffer\n");
+  egpu.enhancement_buf_ptr = gpu.enhancement_bufer;
+}
 
-  if (gpu.enhancement_bufer == NULL) {
-    // currently we use 4x 1024*1024 buffers instead of single 2048*1024
-    // to be able to reuse 1024-width code better (triangle setup,
-    // dithering phase, lines).
-    gpu.enhancement_bufer = mmap(NULL, ENHANCEMENT_BUF_SIZE,
-      PROT_READ | PROT_WRITE, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
-    if (gpu.enhancement_bufer == MAP_FAILED) {
-      printf("OOM for enhancement buffer\n");
-      gpu.enhancement_bufer = NULL;
-    }
+int renderer_init(void)
+{
+  if (gpu.vram != NULL) {
+    initialize_psx_gpu(&egpu, gpu.vram);
+    initialized = 1;
   }
-  egpu.enhancement_buf_ptr = gpu.enhancement_bufer;
 
+  if (gpu.mmap != NULL && gpu.enhancement_bufer == NULL)
+    map_enhancement_buffer();
+
+  ex_regs = gpu.ex_regs;
   return 0;
 }
 
 void renderer_finish(void)
 {
   if (gpu.enhancement_bufer != NULL)
-    munmap(gpu.enhancement_bufer, ENHANCEMENT_BUF_SIZE);
+    gpu.munmap(gpu.enhancement_bufer, ENHANCEMENT_BUF_SIZE);
   gpu.enhancement_bufer = NULL;
   egpu.enhancement_buf_ptr = NULL;
+  initialized = 0;
 }
 
 static __attribute__((noinline)) void
@@ -148,4 +154,12 @@ void renderer_set_config(const struct rearmed_cbs *cbs)
     sync_enhancement_buffers(0, 0, 1024, 512);
   }
   enhancement_was_on = cbs->gpu_neon.enhancement_enable;
+
+  if (!initialized) {
+    initialize_psx_gpu(&egpu, gpu.vram);
+    initialized = 1;
+  }
+
+  if (gpu.enhancement_bufer == NULL)
+    map_enhancement_buffer();
 }
diff --git a/plugins/gpu_unai/gpulib_if.cpp b/plugins/gpu_unai/gpulib_if.cpp
index 2f741ad..de16721 100644
--- a/plugins/gpu_unai/gpulib_if.cpp
+++ b/plugins/gpu_unai/gpulib_if.cpp
@@ -533,6 +533,8 @@ void renderer_set_config(const struct rearmed_cbs *cbs)
   enableAbbeyHack = cbs->gpu_unai.abe_hack;
   light = !cbs->gpu_unai.no_light;
   blend = !cbs->gpu_unai.no_blend;
+
+  GPU_FrameBuffer = (u16 *)gpu.vram;
 }
 
 #endif
diff --git a/plugins/gpulib/gpu.c b/plugins/gpulib/gpu.c
index e133f07..b61bff6 100644
--- a/plugins/gpulib/gpu.c
+++ b/plugins/gpulib/gpu.c
@@ -24,7 +24,7 @@
 //#define log_anomaly gpu_log
 #define log_anomaly(...)
 
-struct psx_gpu gpu __attribute__((aligned(2048)));
+struct psx_gpu gpu;
 
 static noinline int do_cmd_buffer(uint32_t *data, int count);
 static void finish_vram_transfer(int is_read);
@@ -133,6 +133,22 @@ static noinline void get_gpu_info(uint32_t data)
   }
 }
 
+// double, for overdraw guard
+#define VRAM_SIZE (1024 * 512 * 2 * 2)
+
+static int map_vram(void)
+{
+  gpu.vram = gpu.mmap(VRAM_SIZE);
+  if (gpu.vram != NULL) {
+    gpu.vram += 4096 / 2;
+    return 0;
+  }
+  else {
+    fprintf(stderr, "could not map vram, expect crashes\n");
+    return -1;
+  }
+}
+
 long GPUinit(void)
 {
   int ret;
@@ -145,13 +161,26 @@ long GPUinit(void)
   gpu.cmd_len = 0;
   do_reset();
 
+  if (gpu.mmap != NULL) {
+    if (map_vram() != 0)
+      ret = -1;
+  }
   return ret;
 }
 
 long GPUshutdown(void)
 {
+  long ret;
+
   renderer_finish();
-  return vout_finish();
+  ret = vout_finish();
+  if (gpu.vram != NULL) {
+    gpu.vram -= 4096 / 2;
+    gpu.munmap(gpu.vram, VRAM_SIZE);
+  }
+  gpu.vram = NULL;
+
+  return ret;
 }
 
 void GPUwriteStatus(uint32_t data)
@@ -584,13 +613,13 @@ long GPUfreeze(uint32_t type, struct GPUFreeze *freeze)
     case 1: // save
       if (gpu.cmd_len > 0)
         flush_cmd_buffer();
-      memcpy(freeze->psxVRam, gpu.vram, sizeof(gpu.vram));
+      memcpy(freeze->psxVRam, gpu.vram, 1024 * 512 * 2);
       memcpy(freeze->ulControl, gpu.regs, sizeof(gpu.regs));
       memcpy(freeze->ulControl + 0xe0, gpu.ex_regs, sizeof(gpu.ex_regs));
       freeze->ulStatus = gpu.status.reg;
       break;
     case 0: // load
-      memcpy(gpu.vram, freeze->psxVRam, sizeof(gpu.vram));
+      memcpy(gpu.vram, freeze->psxVRam, 1024 * 512 * 2);
       memcpy(gpu.regs, freeze->ulControl, sizeof(gpu.regs));
       memcpy(gpu.ex_regs, freeze->ulControl + 0xe0, sizeof(gpu.ex_regs));
       gpu.status.reg = freeze->ulStatus;
@@ -673,6 +702,13 @@ void GPUrearmedCallbacks(const struct rearmed_cbs *cbs)
   gpu.state.allow_interlace = cbs->gpu_neon.allow_interlace;
   gpu.state.enhancement_enable = cbs->gpu_neon.enhancement_enable;
 
+  gpu.mmap = cbs->mmap;
+  gpu.munmap = cbs->munmap;
+
+  // delayed vram mmap
+  if (gpu.vram == NULL)
+    map_vram();
+
   if (cbs->pl_vout_set_raw_vram)
     cbs->pl_vout_set_raw_vram(gpu.vram);
   renderer_set_config(cbs);
diff --git a/plugins/gpulib/gpu.h b/plugins/gpulib/gpu.h
index 5ad2a46..78a8990 100644
--- a/plugins/gpulib/gpu.h
+++ b/plugins/gpulib/gpu.h
@@ -17,10 +17,9 @@ extern "C" {
 #define CMD_BUFFER_LEN          1024
 
 struct psx_gpu {
-  uint16_t vram[1024 * 512];
-  uint16_t guard[1024 * 512]; // overdraw guard
   uint32_t cmd_buffer[CMD_BUFFER_LEN];
   uint32_t regs[16];
+  uint16_t *vram;
   union {
     uint32_t reg;
     struct {
@@ -90,6 +89,8 @@ struct psx_gpu {
     uint32_t pending_fill[3];
   } frameskip;
   uint16_t *enhancement_bufer;
+  void *(*mmap)(unsigned int size);
+  void  (*munmap)(void *ptr, unsigned int size);
 };
 
 extern struct psx_gpu gpu;