remove loop mapping. refactor

This commit is contained in:
Ole Algoritme
2024-03-24 13:18:35 +01:00
parent 5fceb3766d
commit 1ca8020db0
3 changed files with 108 additions and 83 deletions

View File

@@ -18,27 +18,21 @@ void register_signal_handlers(void)
int main(int argc, char **argv) int main(int argc, char **argv)
{ {
register_signal_handlers(); register_signal_handlers();
struct device devices[32];
int num_devs = gddr6_detect_compatible_gpus(devices, 32); gddr6_init();
int num_devs = gddr6_detect_compatible_gpus();
if (num_devs == 0) if (num_devs == 0)
{ {
printf("No compatible GPU found.\n"); printf("No compatible GPU found.\n");
return 1; return 1;
} else {
printf("Found %d compatible GPU(s).\n", num_devs);
} }
for (int i = 0; i < num_devs; i++) gddr6_memory_map();
{ gddr6_monitor_temperatures();
struct device *device = &devices[i];
printf("Device: %s %s (%s / 0x%04x) pci=%x:%x:%x\n", device->name, device->vram,
device->arch, device->dev_id, device->bus, device->dev, device->func);
}
gddr6_initialize();
gddr6_monitor_temperatures(devices, num_devs);
gddr6_cleanup(0);
return 0; return 0;
} }

View File

@@ -4,7 +4,8 @@
#include <stdint.h> #include <stdint.h>
struct device { struct device
{
uint32_t bar0; uint32_t bar0;
uint8_t bus, dev, func; uint8_t bus, dev, func;
uint32_t offset; uint32_t offset;
@@ -12,11 +13,21 @@ struct device {
const char *vram; const char *vram;
const char *arch; const char *arch;
const char *name; const char *name;
void *mapped_addr;
void *phys_addr;
uint32_t base_offset;
}; };
void gddr6_initialize(void); struct gddr6_ctx {
struct device *devices;
int num_devices;
int fd;
};
void gddr6_init(void);
void gddr6_memory_map(void);
void gddr6_cleanup(int signal); void gddr6_cleanup(int signal);
void gddr6_monitor_temperatures(const struct device *devices, int num_devices); void gddr6_monitor_temperatures(void);
int gddr6_detect_compatible_gpus(struct device *devices, int max_devices); int gddr6_detect_compatible_gpus(void);
#endif // GDDR6_H #endif // GDDR6_H

View File

@@ -20,10 +20,9 @@
__LINE__, __FILE__, errno, strerror(errno)); exit(1); \ __LINE__, __FILE__, errno, strerror(errno)); exit(1); \
} while(0) } while(0)
static int fd = -1; #define MAX_DEVICES 32
static void *map_base = MAP_FAILED;
static struct device devices[32];
struct gddr6_ctx ctx = {0};
struct device dev_table[] = struct device dev_table[] =
{ {
{ .offset = 0x0000E2A8, .dev_id = 0x2684, .vram = "GDDR6X", .arch = "AD102", .name = "RTX 4090" }, { .offset = 0x0000E2A8, .dev_id = 0x2684, .vram = "GDDR6X", .arch = "AD102", .name = "RTX 4090" },
@@ -46,94 +45,92 @@ struct device dev_table[] =
{ .offset = 0x0000E2A8, .dev_id = 0x2236, .vram = "GDDR6", .arch = "GA102", .name = "A10" }, { .offset = 0x0000E2A8, .dev_id = 0x2236, .vram = "GDDR6", .arch = "GA102", .name = "A10" },
}; };
void gddr6_init(void)
void gddr6_initialize(void)
{ {
const char *MEM = "/dev/mem"; ctx.fd = open("/dev/mem", O_RDONLY);
if ((fd = open(MEM, O_RDONLY)) == -1) { if (ctx.fd == -1) {
PRINT_ERROR(); PRINT_ERROR();
} }
} }
int gddr6_detect_compatible_gpus(struct device *devices, int max_devices) int gddr6_detect_compatible_gpus(void)
{ {
ctx.devices = NULL;
ctx.num_devices = 0;
struct pci_access *pacc = NULL; struct pci_access *pacc = NULL;
struct pci_dev *pci_dev = NULL; struct pci_dev *pci_dev = NULL;
int num_devs = 0;
ssize_t dev_table_size = (sizeof(dev_table)/sizeof(struct device)); ssize_t dev_table_size = (sizeof(dev_table)/sizeof(struct device));
pacc = pci_alloc(); pacc = pci_alloc();
pci_init(pacc); pci_init(pacc);
pci_scan_bus(pacc); pci_scan_bus(pacc);
for (pci_dev = pacc->devices; pci_dev; pci_dev = pci_dev->next) for (pci_dev = pacc->devices; pci_dev != NULL; pci_dev = pci_dev->next)
{ {
pci_fill_info(pci_dev, PCI_FILL_IDENT | PCI_FILL_BASES | PCI_FILL_CLASS); pci_fill_info(pci_dev, PCI_FILL_IDENT | PCI_FILL_BASES | PCI_FILL_CLASS);
for (uint32_t i = 0; i < dev_table_size; ++i)
for (uint32_t i = 0; i < dev_table_size; i++)
{ {
if (pci_dev->device_id == dev_table[i].dev_id) if (pci_dev->device_id == dev_table[i].dev_id)
{ {
devices[num_devs] = dev_table[i]; struct device *new_devices = realloc(ctx.devices, (ctx.num_devices + 1) * sizeof(struct device));
devices[num_devs].bar0 = (pci_dev->base_addr[0] & 0xFFFFFFFF); if (new_devices == NULL)
devices[num_devs].bus = pci_dev->bus; {
devices[num_devs].dev = pci_dev->dev; fprintf(stderr, "Memory allocation failed\n");
devices[num_devs].func = pci_dev->func; pci_cleanup(pacc);
num_devs++; free(ctx.devices);
ctx.devices = NULL;
return 0;
}
ctx.devices = new_devices;
ctx.devices[i] = dev_table[i];
ctx.devices[i].bar0 = (pci_dev->base_addr[0] & 0xffffffff);
ctx.devices[i].bus = pci_dev->bus;
ctx.devices[i].dev = pci_dev->dev;
ctx.devices[i].func = pci_dev->func;
ctx.num_devices++;
} }
} }
} }
pci_cleanup(pacc); pci_cleanup(pacc);
return num_devs; return ctx.num_devices;
} }
void gddr6_cleanup(int signal) void gddr6_memory_map(void)
{ {
if (map_base != MAP_FAILED) for (uint32_t i = 0; i < ctx.num_devices; i++)
{ {
munmap(map_base, PG_SZ); ctx.devices[i].phys_addr = (void *) (ctx.devices[i].bar0 + ctx.devices[i].offset);
map_base = MAP_FAILED; ctx.devices[i].base_offset = ctx.devices[i].phys_addr & ~(PG_SZ - 1);
}
if (fd != -1) ctx.devices[i].mapped_addr = mmap(0, PG_SZ, PROT_READ, MAP_SHARED, ctx.fd, ctx.devices[i].base_offset);
if (ctx.devices[i].mapped_addr == MAP_FAILED)
{ {
close(fd); ctx.devices[i].mapped_addr = NULL;
fd = -1; fprintf(stderr, "Memory mapping failed for pci=%x:%x:%x\n", ctx.devices[i].bus, ctx.devices[i].dev, ctx.devices[i].func);
} else {
printf("Device: %s %s (%s / 0x%04x) pci=%x:%x:%x\n", ctx.devices[i].name, ctx.devices[i].vram,
ctx.devices[i].arch, ctx.devices[i].dev_id, ctx.devices[i].bus, ctx.devices[i].dev, ctx.devices[i].func);
}
} }
exit(signal);
} }
void gddr6_monitor_temperatures(const struct device *devices, int num_devices) void gddr6_monitor_temperatures(void)
{ {
void *virt_addr; while (1) {
uint32_t temp;
uint32_t phys_addr;
uint32_t read_result;
uint32_t base_offset;
while (1)
{
printf("\rVRAM Temps: |"); printf("\rVRAM Temps: |");
for (int i = 0; i < num_devices; i++) for (uint32_t i = 0; i < ctx.num_devices; i++)
{ {
const struct device *device = &devices[i]; if (ctx.devices[i].mapped_addr == NULL || ctx.devices[i].mapped_addr == MAP_FAILED)
phys_addr = (device->bar0 + device->offset);
base_offset = phys_addr & ~(PG_SZ-1);
map_base = mmap(0, PG_SZ, PROT_READ, MAP_SHARED, fd, base_offset);
if(map_base == (void *) MAP_FAILED)
{ {
if (fd != -1) continue;
{
close(fd);
} }
printf("Can't read memory. If you are root, enable kernel parameter iomem=relaxed\n");
PRINT_ERROR(); void *virt_addr = (uint8_t *) ctx.devices[i].mapped_addr + (ctx.devices[i].phys_addr - ctx.devices[i].base_offset);
} uint32_t read_result = *((uint32_t *)virt_addr);
virt_addr = (uint8_t *) map_base + (phys_addr - base_offset); uint32_t temp = ((read_result & 0x00000fff) / 0x20);
read_result = *((uint32_t *) virt_addr);
temp = ((read_result & 0x00000fff) / 0x20);
printf(" %3u°C |", temp); printf(" %3u°C |", temp);
} }
@@ -141,3 +138,26 @@ void gddr6_monitor_temperatures(const struct device *devices, int num_devices)
sleep(1); sleep(1);
} }
} }
void gddr6_cleanup(int signal)
{
for (uint32_t i = 0; i < ctx.num_devices; i++)
{
if (ctx.devices[i].mapped_addr != NULL && ctx.devices[i].mapped_addr != MAP_FAILED)
{
munmap(ctx.devices[i].mapped_addr, PG_SZ);
ctx.devices[i].mapped_addr = NULL;
}
}
if (ctx.fd != -1)
{
close(ctx.fd);
ctx.fd = -1;
}
if (ctx.devices)
{
free(ctx.devices);
ctx.devices = NULL;
}
exit(signal);
}