IRQs on an dual-core AMD processor

[rdos]

I still have problems with missed IRQs on my Compaq Pressario CQ57 portable. First, it was only the missing AHCI PCI IRQ, but now it is also the MSI-based RTL8168 network chip driver that has the same problem. In both cases, the drivers work when I combine them with timers that does exactly the same thing as the IRQs do. I also checked the trigger-mode for PCI IRQs, and it is already set to level-triggered in the APIC module.

Here is the IRQ / MSI handlers (they are both essentially identical, "nr" stands for IRQ-number):

Code: Select all

mem_irq&nr:
    push ds
    push es
    push fs
    pushad
;
    EnterInt                           ; this call would lock the scheduler on the core the IRQ is executing on
    push fs
    sti
;       
    mov ax,irq_sys_sel
    mov es,ax
    mov bx,OFFSET irq_arr + nr * SIZE irq_struc
    mov ax,word ptr es:[bx+4].user_handler
    or ax,ax
    jz mem_irq_default_error&nr
;
    mov ds,es:[bx].user_data
    xor eax,eax
    mov ax,cs
    push eax
    mov ax,OFFSET mem_irq_handle_done&nr
    push eax
    push es:[bx+4].user_handler
    push es:[bx].user_handler
    xor ax,ax
    mov es,ax
    retf32                                    ; call the device-driver's registered IRQ handler

mem_irq_default_error&nr:
    or es:bad_irqs, 1 SHL nr

mem_irq_handle_done&nr:
    cli
;    
    mov ax,apic_mem_sel
    mov ds,ax
    xor eax,eax
    mov ds:APIC_EOI,eax                  ; send EOI to APIC, memory-mapped mode
    pop fs
    LeaveInt                                    ; unlock the scheduler
;
    popad
    pop fs
    pop es
    pop ds
    iretd

Here is the RTL8168 interrupt handler (the AHCI handler is similar):

Code: Select all

NetInt  Proc far
niLoop:
    mov dx,ds:IoBase
    add dx,REG_ISR
    in ax,dx                                                  ; read out the ISR bits
    or ds:Isr,ax
    out dx,ax                                                ; clear ISR bits
    test ax,IR_ROK OR IR_RDU
    jz niNotRx
;
    mov bx,ds:Handle
    or bx,bx
    jz niNotRx
;
    NetReceived                                           ; wake up the receiver
    jmp niLoop

niNotRx:
    test ax,IR_TOK OR IR_TER
    jz niNotTx
;
    mov bx,ds:TxThread
    or bx,bx
    jz niNotTx
;
    Signal                                                         ; wake up any thread waiting for send completion
    jmp niLoop

niNotTx:
    retf32
NetInt  Endp

[gerryg400]

Do you miss every interrupt or just some ? Randomly or under some condition ?

[rdos]

Do you miss every interrupt or just some ? Randomly or under some condition ?

Very seldom, and randomly AFAIK right now. Maybe one in 1,000 or even less, but since one lost interrupt means a server thread is locked-up, it causes severe effects when there is no timer. I'm actually not sure if it is specific to the AMD machine either, as I have no similar test-scripts I can run on my other computers. It mostly happens when I install our terminal software from a FTP-server, or when I remote debug the application. At least I know it doesn't happen on a single-core machine running with an ordinary PIC, an IDE-disc and another RTL-chip (RTL8139).

[rdos]

I'm now 99% certain that the problem with the realtek 8168-chip is not present on my 4-core AMD Athlon. I've stressed it with the TCP/IP remote debugger for 30-minutes, constantly single-stepping code + running several stress-tests in the background, but it doesn't lock up on the ethernet. I have no timer on the driver, so the IRQ obviously works there.