AARCH64 port using CMA, based on juj#261

Author: hakmo

.gitignore

@@ -6,3 +6,4 @@
 *.S
 *.symvers
 *.order
+build/*

CMakeLists.txt

@@ -24,6 +24,7 @@ set(DEFAULT_TO_SINGLE_CORE_BOARD OFF)
 set(DEFAULT_TO_ARMV6Z OFF)
 set(DEFAULT_TO_ARMV7A OFF)
 set(DEFAULT_TO_ARMV8A OFF)
+set(DEFAULT_USE_VCSM_CMA OFF)
 
 # http://ozzmaker.com/check-raspberry-software-hardware-version-command-line/
 if (BOARD_REVISION MATCHES "(0002)|(0003)|(0004)|(0005)|(0006)|(0007)|(0008)|(0009)" OR BOARD_REVISION MATCHES "(000d)|(000e)|(000f)|(0010)|(0011)|(0012)" OR BOARD_REVISION MATCHES "(900092)|(900093)|(9000c1)")
@@ -46,7 +47,20 @@ if (SINGLE_CORE_BOARD)
   set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -DSINGLE_CORE_BOARD=1")
 endif()
 
-set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -marm -mabi=aapcs-linux -mhard-float -mfloat-abi=hard -mlittle-endian -mtls-dialect=gnu2 -funsafe-math-optimizations")
+option(AARCH64 "Target a Raspberry Pi with aarch64 architecture" NO)
+if (AARCH64)
+	message(STATUS "Enable AARCH64 build")
+	set(DEFAULT_USE_VCSM_CMA ON)
+	set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -mlittle-endian -funsafe-math-optimizations -DTIMER_32BIT")
+else()
+	set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -marm -mabi=aapcs-linux -mhard-float -mfloat-abi=hard -mlittle-endian -mtls-dialect=gnu2 -funsafe-math-optimizations")
+endif()
+
+option(USE_VCSM_CMA "Map Memory from CPU instead of GPU" ${DEFAULT_USE_VCSM_CMA})
+if (USE_VCSM_CMA)
+	message(STATUS "Enabling Map Memory from CPU instead of GPU")
+	set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -DUSE_VCSM_CMA=1")
+endif()
 
 option(ARMV6Z "Target a Raspberry Pi with ARMv6Z instruction set (Pi 1A, 1A+, 1B, 1B+, Zero, Zero W)" ${DEFAULT_TO_ARMV6Z})
 if (ARMV6Z)

cma.cpp

@@ -0,0 +1,63 @@
+#ifdef USE_VCSM_CMA
+
+#include "config.h"
+#include "cma.h"
+#include "util.h"
+#include <sys/ioctl.h>
+#include <fcntl.h>
+#include <syslog.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <unistd.h>
+
+static int cma_fd = -1;
+#define PAGE_SIZE 4096
+
+void OpenVCSM(void) {
+  cma_fd = open("/dev/vcsm-cma", O_RDWR|O_SYNC);
+  if (cma_fd < 0) FATAL_ERROR("can't open /dev/vcsm-cma");
+}
+
+void CloseVCSM(void) {
+  if (cma_fd >= 0) {
+    close(cma_fd);
+  }
+}
+
+const int NAME_LENGTH = 32;
+
+struct Allocate {
+  /* user -> kernel */
+  uint32_t size;
+  uint32_t num;
+  uint32_t flags;
+  uint32_t pad;
+  char name[NAME_LENGTH];
+
+  /* kernel -> user */
+  int32_t fd;
+  uint32_t vcHandle;
+  uint64_t dmaAddr;
+};
+
+int AllocateCMA(const char* reason, size_t req, CMAInfo* res) {
+  if (res == NULL) {
+    return -1;
+  }
+  Allocate ctx;
+  memset(&ctx, 0, sizeof(ctx));
+  ctx.size = ALIGN_UP(req, PAGE_SIZE);
+  ctx.flags = 0; // NO cache
+  strncpy((char*)ctx.name, reason, NAME_LENGTH -1);
+  ctx.num = 1;
+  if (ioctl(cma_fd, _IOR('J', 0x5A, struct Allocate), &ctx) < 0 || ctx.fd < 0) { // allocate cmd
+    return -1;
+  }
+  res->size = ctx.size;
+  res->vcHandle = ctx.vcHandle;
+  res->dmaAddr = ctx.dmaAddr;
+  res->fd = ctx.fd;
+  return 0;
+}
+
+#endif

cma.h

@@ -0,0 +1,16 @@
+#pragma once
+#ifdef USE_VCSM_CMA
+
+#include <memory.h>
+#include <inttypes.h>
+struct CMAInfo {
+  size_t size;
+  uintptr_t dmaAddr;
+  uint32_t fd;
+  uint32_t vcHandle;
+};
+
+void OpenVCSM(void);
+void CloseVCSM(void);
+int AllocateCMA(const char* reason, size_t req, CMAInfo* res);
+#endif

dma.cpp

@@ -14,6 +14,10 @@
 #include "util.h"
 #include "mailbox.h"
 
+#ifdef USE_VCSM_CMA
+#include "cma.h"
+#endif
+
 #ifdef USE_DMA_TRANSFERS
 
 #define BCM2835_PERI_BASE               0x3F000000
@@ -36,6 +40,9 @@ struct GpuMemory
   void *virtualAddr;
   uintptr_t busAddress;
   uint32_t sizeBytes;
+#ifdef USE_VCSM_CMA
+  uint32_t vcHandle;
+#endif
 };
 
 #define NUM_DMA_CBS 1024
@@ -127,7 +134,33 @@ void FreeDMAChannel(int channel)
 #define VIRT_TO_BUS(block, x) ((uintptr_t)(x) - (uintptr_t)((block).virtualAddr) + (block).busAddress)
 
 uint64_t totalGpuMemoryUsed = 0;
+#ifdef USE_VCSM_CMA
 
+void FreeUncachedGpuMemory(GpuMemory mem) {
+    munmap(mem.virtualAddr, mem.sizeBytes);
+    close(mem.allocationHandle);
+}
+
+GpuMemory AllocateUncachedGpuMemory(uint32_t numBytes, const char *reason) {
+    GpuMemory mem;
+    CMAInfo ctx;
+    if (AllocateCMA(reason, numBytes, &ctx) != 0) {
+      FATAL_ERROR("alloc cma failed");
+    }
+    mem.sizeBytes = ctx.size;
+    mem.busAddress = ctx.dmaAddr;
+    mem.allocationHandle = ctx.fd;
+    mem.vcHandle = ctx.vcHandle;
+    mem.virtualAddr = mmap(0, mem.sizeBytes, PROT_READ | PROT_WRITE, MAP_SHARED, ctx.fd, 0);
+    totalGpuMemoryUsed += mem.sizeBytes;
+    if (mem.virtualAddr == MAP_FAILED) {
+        FreeUncachedGpuMemory(mem);
+        FATAL_ERROR("Failed to mmap CMA memory!");
+    }
+    printf("Allocated %u bytes of GPU memory for %s (bus address=%p). Total GPU memory used: %llu bytes\n", mem.sizeBytes, reason, (void*)mem.busAddress, totalGpuMemoryUsed);
+    return mem;
+}
+#else
 // Allocates the given number of bytes in GPU side memory, and returns the virtual address and physical bus address of the allocated memory block.
 // The virtual address holds an uncached view to the allocated memory, so writes and reads to that memory address bypass the L1 and L2 caches. Use
 // this kind of memory to pass data blocks over to the DMA controller to process.
@@ -154,7 +187,7 @@ void FreeUncachedGpuMemory(GpuMemory mem)
   Mailbox(MEM_UNLOCK_MESSAGE, mem.allocationHandle);
   Mailbox(MEM_FREE_MESSAGE, mem.allocationHandle);
 }
-
+#endif
 volatile DMAChannelRegisterFile *GetDMAChannel(int channelNumber)
 {
   if (channelNumber < 0 || channelNumber >= BCM2835_NUM_DMA_CHANNELS)
@@ -720,8 +753,8 @@ void SPIDMATransfer(SPITask *task)
   while((dmaTx->cs & BCM2835_DMA_CS_ACTIVE))
   {
     CheckSPIDMAChannelsNotStolen();
-    if (tick() - dmaTaskStart > 5000000)
-      FATAL_ERROR("DMA TX channel has stalled!");
+    if (tick() - dmaTaskStart > 5000000) 
+		FATAL_ERROR("DMA TX channel has stalled!");
   }
   while((dmaRx->cs & BCM2835_DMA_CS_ACTIVE))
   {

spi.cpp

@@ -33,6 +33,10 @@ void ChipSelectHigh();
 #define TOGGLE_CHIP_SELECT_LINE() ((void)0)
 #endif
 
+#ifdef USE_VCSM_CMA
+#include "cma.h"
+#endif
+
 static uint32_t writeCounter = 0;
 
 #define WRITE_FIFO(word) do { \
@@ -49,7 +53,11 @@ volatile SPIRegisterFile *spi = 0;
 
 // Points to the system timer register. N.B. spec sheet says this is two low and high parts, in an 32-bit aligned (but not 64-bit aligned) address. Profiling shows
 // that Pi 3 Model B does allow reading this as a u64 load, and even when unaligned, it is around 30% faster to do so compared to loading in parts "lo | (hi << 32)".
+#ifdef TIMER_32BIT
+volatile systemTimer *systemTimerRegister = 0;
+#else
 volatile uint64_t *systemTimerRegister = 0;
+#endif
 
 void DumpSPICS(uint32_t reg)
 {
@@ -510,13 +518,20 @@ int InitSPI()
   // Memory map GPIO and SPI peripherals for direct access
   mem_fd = open("/dev/mem", O_RDWR|O_SYNC);
   if (mem_fd < 0) FATAL_ERROR("can't open /dev/mem (run as sudo)");
+#ifdef USE_VCSM_CMA
+  OpenVCSM();
+#endif
   printf("bcm_host_get_peripheral_address: %p, bcm_host_get_peripheral_size: %u, bcm_host_get_sdram_address: %p\n", bcm_host_get_peripheral_address(), bcm_host_get_peripheral_size(), bcm_host_get_sdram_address());
   bcm2835 = mmap(NULL, bcm_host_get_peripheral_size(), (PROT_READ | PROT_WRITE), MAP_SHARED, mem_fd, bcm_host_get_peripheral_address());
   if (bcm2835 == MAP_FAILED) FATAL_ERROR("mapping /dev/mem failed");
   spi = (volatile SPIRegisterFile*)((uintptr_t)bcm2835 + BCM2835_SPI0_BASE);
   gpio = (volatile GPIORegisterFile*)((uintptr_t)bcm2835 + BCM2835_GPIO_BASE);
-  systemTimerRegister = (volatile uint64_t*)((uintptr_t)bcm2835 + BCM2835_TIMER_BASE + 0x04); // Generates an unaligned 64-bit pointer, but seems to be fine.
-  // TODO: On graceful shutdown, (ctrl-c signal?) close(mem_fd)
+#ifdef TIMER_32BIT
+  systemTimerRegister = (volatile TIMER_TYPE*)((uintptr_t)bcm2835 + BCM2835_TIMER_BASE);
+#else
+  systemTimerRegister = (volatile TIMER_TYPE*)((uintptr_t)bcm2835 + BCM2835_TIMER_BASE + 0x04); // Generates an unaligned 64-bit pointer, but seems to be fine.
+#endif
+// TODO: On graceful shutdown, (ctrl-c signal?) close(mem_fd)
 #endif
 
   uint32_t currentBcmCoreSpeed = MailboxRet2(0x00030002/*Get Clock Rate*/, 0x4/*CORE*/);
@@ -658,6 +673,9 @@ void DeinitSPI()
     close(mem_fd);
     mem_fd = -1;
   }
+#ifdef USE_VCSM_CMA
+  CloseVCSM();
+#endif
 
 #ifndef KERNEL_MODULE_CLIENT
 

tick.h

@@ -5,8 +5,21 @@
 #include <unistd.h>
 
 // Initialized in spi.cpp along with the rest of the BCM2835 peripheral:
+#ifdef TIMER_32BIT
+struct __attribute__((packed, aligned(4))) systemTimer {
+    volatile uint32_t cs;
+    volatile uint32_t clo;
+    volatile uint32_t chi;
+    volatile uint32_t c[4];
+};
+#define TIMER_TYPE systemTimer
+extern volatile systemTimer* systemTimerRegister;
+#define tick() (((uint64_t)systemTimerRegister->clo) | ((uint64_t)(systemTimerRegister->chi) << 32))
+#else
+#define TIMER_TYPE uint64_t
 extern volatile uint64_t *systemTimerRegister;
 #define tick() (*systemTimerRegister)
+#endif
 
 #endif