October 26, 2005
X server 的 low-level 觀點
"low-level" 這個詞很妙，通常程度比較粗淺的開發者 (Low-level Programmer) 會從高階應用 (High-level application) 切入，而經驗豐富後，就稱為資深開發者 (High-level Programmer)，卻往往可能是擔任 Low-level Application 的主力。在 Linux Kernel 中，driver 如果要存取 I/O memory，如 X server 為了效能考量，通常會直接存取 Graphics Devices，那就須在有 CPU 能處理的 virtual address，而這也涉及 virtual memory 中 mapping I/O memory 的議題。Kernel API 中，有個 ioremap 函式：
#include <asm/io.h>
void *ioremap(unsigned long phys_addr,
unsigned long size);
void iounmap(void *address);
而在 user-space 的應用程式如果需要對 physical address 作存取，可以透過 /dev/mem 這個特別的 device：$ ls -l /dev/mem crw-r----- 1 root kmem 1, 1 2005-10-26 17:42 /dev/mem不過 /dev/mem 事實上只能針對 non-RAM (I/O memory) address 作處理，透過 Kernel 設定的特殊 flags 來標示可存取的範圍，簡單來說，不會動到 physical RAM address，但這個設計對於 X server 是相當重要的。在 TinyX / FeeDesktop.org Xserver 中，[hw/kdrive/vesa/vm86.c] 處理了這部份的細節，以下是 X server 配置 video memory I/O 的 Vm86Setup 函式實作：
Vm86InfoPtr
Vm86Setup(int mapHoles)
{
int devmem = -1, devzero = -1;
void *magicMem, *loMem, *hiMem;
void *hole1, *hole2;
U32 stack_base, ret_code;
Vm86InfoPtr vi = NULL;
devmem = open("/dev/mem", O_RDWR);
if(devmem < 0) {
perror("open /dev/mem");
goto fail;
}
devzero = open("/dev/zero", O_RDWR);
if(devzero < 0) {
perror("open /dev/zero");
goto fail;
}
magicMem = MAP_FAILED;
loMem = MAP_FAILED;
hiMem = MAP_FAILED;
hole1 = MAP_FAILED;
hole2 = MAP_FAILED;
magicMem = mmap((void*)MAGICMEM_BASE, MAGICMEM_SIZE,
PROT_READ | PROT_WRITE | PROT_EXEC,
MAP_PRIVATE | MAP_FIXED, devmem, MAGICMEM_BASE);
if(magicMem == MAP_FAILED) {
ErrorF("Couldn't map magic memory\n");
goto unmapfail;
}
if(mapHoles) {
hole1 = mmap((void*)HOLE1_BASE, HOLE1_SIZE,
PROT_READ | PROT_WRITE | PROT_EXEC,
MAP_PRIVATE | MAP_FIXED, devzero, HOLE1_BASE);
if(hole1 == MAP_FAILED) {
ErrorF("Couldn't map first hole\n");
goto unmapfail;
}
}
loMem = mmap((void*)LOMEM_BASE, LOMEM_SIZE,
PROT_READ | PROT_WRITE | PROT_EXEC,
MAP_PRIVATE | MAP_FIXED, devzero, LOMEM_BASE);
if(loMem == MAP_FAILED) {
ErrorF("Couldn't map low memory\n");
munmap(magicMem, MAGICMEM_SIZE);
goto unmapfail;
}
if(mapHoles) {
hole2 = mmap((void*)HOLE2_BASE, HOLE2_SIZE,
PROT_READ | PROT_WRITE | PROT_EXEC,
MAP_PRIVATE | MAP_FIXED, devzero, HOLE2_BASE);
if(hole2 == MAP_FAILED) {
ErrorF("Couldn't map first hole\n");
goto unmapfail;
}
}
hiMem = mmap((void*)HIMEM_BASE, HIMEM_SIZE,
PROT_READ | PROT_WRITE | PROT_EXEC,
MAP_SHARED | MAP_FIXED,
devmem, HIMEM_BASE);
if(hiMem == MAP_FAILED) {
ErrorF("Couldn't map high memory\n");
goto unmapfail;
}
vi = xalloc(sizeof(Vm86InfoRec));
if (!vi)
goto unmapfail;
vi->magicMem = magicMem;
vi->hole1 = hole1;
vi->loMem = loMem;
vi->hole2 = hole2;
vi->hiMem = hiMem;
vi->brk = LOMEM_BASE;
stack_base = Vm86AllocateMemory(vi, STACK_SIZE);
if(stack_base == ALLOC_FAIL)
goto unmapfail;
ret_code = Vm86AllocateMemory(vi, sizeof(retcode_data));
if(ret_code == ALLOC_FAIL)
goto unmapfail;
vi->stack_base = stack_base;
vi->ret_code = ret_code;
memset(&vi->vms, 0, sizeof(struct vm86_struct));
vi->vms.flags = 0;
vi->vms.screen_bitmap = 0;
vi->vms.cpu_type = CPU_586;
memcpy(&vi->vms.int_revectored, rev_ints, sizeof(rev_ints));
iopl(3);
if(devmem >= 0)
close(devmem);
if(devzero >= 0)
close(devzero);
return vi;
unmapfail:
if(magicMem != MAP_FAILED) munmap(magicMem, MAGICMEM_SIZE);
if(hole1 != MAP_FAILED) munmap(hole1, HOLE1_SIZE);
if(loMem != MAP_FAILED) munmap(loMem, LOMEM_SIZE);
if(hole2 != MAP_FAILED) munmap(hole2, HOLE2_SIZE);
if(hiMem != MAP_FAILED) munmap(hiMem, HIMEM_SIZE);
fail:
if(devmem >= 0)
close(devmem);
if(devzero >= 0)
close(devzero);
if(vi)
xfree(vi);
return NULL;
}
在要到 /dev/mem 的 Mapping 後，我們來看 GLIBC 提供的 Memory I/O Mapping 的 API：
void * mmap(void *start, size_t length, int prot,
int flags, int fd, off_t offset);
int munmap(void *start, size_t length);
這兩個函式允許 user-space 的應用程式，也就是探討的 X server 直接存取 device memory，這對於需要高頻寬 I/O 處理的應用程式來說相當重要，因為效能會比直接讀寫 /dev devices 來得快許多。所謂的 VMA (Virtual Memory Area) 就是來描述在同樣的 permission flags 情況下的 process virtual memory 的連續區域，以系統第一個 process，也就是 /sbin/init，來說，我們可以看看其 VMA 配置：
$ cat /proc/1/maps 08048000-08050000 r-xp 00000000 03:01 592323 /sbin/init 08050000-08051000 rw-p 00007000 03:01 592323 /sbin/init 08051000-08072000 rw-p 08051000 00:00 0 [heap] b7dcb000-b7dec000 rw-p b7dcb000 00:00 0 b7dec000-b7dee000 r-xp 00000000 03:01 1533 /lib/tls/i686/cmov/libdl-2.3.5.so b7dee000-b7df0000 rw-p 00001000 03:01 1533 /lib/tls/i686/cmov/libdl-2.3.5.so b7df0000-b7f21000 r-xp 00000000 03:01 1531 /lib/tls/i686/cmov/libc-2.3.5.so b7f21000-b7f22000 r--p 00131000 03:01 1531 /lib/tls/i686/cmov/libc-2.3.5.so b7f22000-b7f25000 rw-p 00132000 03:01 1531 /lib/tls/i686/cmov/libc-2.3.5.so b7f25000-b7f27000 rw-p b7f25000 00:00 0 b7f27000-b7f37000 r-xp 00000000 03:01 337350 /lib/libselinux.so.1 b7f37000-b7f38000 rw-p 00010000 03:01 337350 /lib/libselinux.so.1 b7f38000-b7f5d000 r-xp 00000000 03:01 337351 /lib/libsepol.so.1 b7f5d000-b7f5e000 rw-p 00025000 03:01 337351 /lib/libsepol.so.1 b7f5e000-b7f68000 rw-p b7f5e000 00:00 0 b7f75000-b7f76000 rw-p b7f75000 00:00 0 b7f76000-b7f8b000 r-xp 00000000 03:01 30973 /lib/ld-2.3.5.so b7f8b000-b7f8d000 rw-p 00014000 03:01 30973 /lib/ld-2.3.5.so bfb75000-bfb8b000 rw-p bfb75000 00:00 0 [stack] ffffe000-fffff000 ---p 00000000 00:00 0 [vdso]那麼，執行中的 X server 其 VMA 又是如何呢？我們來看看：
$ cat /proc/`pgrep X`/maps 000a0000-000c0000 rwxs 000a0000 00:0b 2432 /dev/mem 000f0000-00100000 r-xs 000f0000 00:0b 2432 /dev/mem 08048000-081d6000 r-xp 00000000 03:01 27050 /usr/X11R6/bin/Xorg 081d6000-08209000 rwxp 0018d000 03:01 27050 /usr/X11R6/bin/Xorg 08209000-09be8000 rwxp 08209000 00:00 0 [heap] ... b3310000-b3510000 rwxs a4102000 00:0b 15299 /dev/dri/card0 b3510000-b39f0000 rwxs a4302000 00:0b 15299 /dev/dri/card0 b39f0000-b3bf0000 rwxs a4102000 00:0b 15299 /dev/dri/card0 b3bf0000-b3bf1000 rwxs a4101000 00:0b 15299 /dev/dri/card0 b3bf1000-b3cf2000 rwxs a4000000 00:0b 15299 /dev/dri/card0 b3cf2000-b3cf4000 rwxs f8c35000 00:0b 15299 /dev/dri/card0 b3cf4000-b7cf4000 rwxs 98000000 00:0b 2432 /dev/mem b7cf4000-b7d74000 rwxs 90000000 00:0b 2432 /dev/mem bfd7f000-bfd95000 rwxp bfd7f000 00:00 0 [stack] ffffe000-fffff000 ---p 00000000 00:00 0 [vdso]由 jserv 發表於 October 26, 2005 04:30 PM
迴響
發表迴響