format binary
use16

org 0x7c00

DMA_BUF = 0x7e00

cld

xor     ax, ax
mov     ds, ax
mov     es, ax

; clear DMA buffer
mov     di, DMA_BUF
xor     ax, ax
mov     cx, 256
repnz stosw

; two TRIM requests (both are the same: one sector, starting at sector index 1)
mov     di, DMA_BUF
mov     dword [di+ 0], 0x00000001
mov     dword [di+ 4], 0x00010000
mov     dword [di+ 8], 0x00000001
mov     dword [di+12], 0x00010000

; find IDE PCI device
mov     ax, 0xb102
mov     dx, 0x8086
mov     cx, 0x7010
xor     si, si
int     0x1a

; bx has PCI address
push    bx

; enable BM+MEM+IO

mov     di, 0x04 ; command/status
mov     ax, 0xb10a ; read config dword
int     0x1a

pop     bx
push    bx

or      cl, 0x7 ; BM+MEM+IO
mov     di, 0x04
mov     ax, 0xb10d ; write config dword
int     0x1a

pop     bx
push    bx

; read BAR4 (DMA I/O space)

mov     di, 0x20 ; bar4
mov     ax, 0xb10a
int     0x1a

and     cx, 0xfffc ; DMA I/O base
push    cx

mov     dx, cx

; set up DMA

xor     al, al ; status: 0
out     dx, al

mov     al, 0x04 ; clear pending interrupts
add     dx, 2
out     dx, al

mov     eax, prdt
add     dx, 2
out     dx, eax

; send TRIM command

mov     dx, 0x1f0
mov     si, dsm_trim_cmd
mov     cx, 8
out_loop:
lodsb
out     dx, al
inc     dx
loop    out_loop

; start DMA transfer

pop     dx
mov     al, 0x01
out     dx, al

; immediately reset device, cancelling ongoing TRIM

mov     dx, 0x3f6
mov     al, 0x04
out     dx, al

cli
hlt


dsm_trim_cmd:
db 0x00, 0x01, 0x01, 0x00, 0x00, 0x00, 0xc0, 0x06

pciaddr:
dw ?

align(8)
prdt:
dd DMA_BUF
dd 512 or 0x80000000

times 510-($-$$) db 0
dw 0xaa55

times 512 db 0