Subversion Repositories Kolibri OS

; Implementation of periodic transaction scheduler for USB.

; Bandwidth dedicated to periodic transactions is limited, so

; different pipes should be scheduled as uniformly as possible.

; USB1 scheduler.

; Algorithm is simple:

; when adding a pipe, optimize the following quantity:

;  * for every millisecond, take all bandwidth scheduled to periodic transfers,

;  * calculate maximum over all milliseconds,

;  * select a variant which minimizes that maximum;

; when removing a pipe, do nothing (except for bookkeeping).

; sanity check: structures in UHCI and OHCI should be the same

if (sizeof.ohci_static_ep=sizeof.uhci_static_ep)&(ohci_static_ep.SoftwarePart=uhci_static_ep.SoftwarePart)&(ohci_static_ep.NextList=uhci_static_ep.NextList)

; Select a list for a new pipe.

; in: esi -> usb_controller, maxpacket, type, interval can be found in the stack

; in: ecx = 2 * maximal interval = total number of periodic lists + 1

; in: edx -> {u|o}hci_static_ep for the first list

; in: eax -> byte past {u|o}hci_static_ep for the last list in the first group

; out: edx -> usb_static_ep for the selected list or zero if failed

proc usb1_select_interrupt_list

; inherit some variables from usb_open_pipe

virtual at ebp-8

.bandwidth      dd      ?

.target         dd      ?

                dd      ?

                dd      ?

.config_pipe    dd      ?

.endpoint       dd      ?

.maxpacket      dd      ?

.type           dd      ?

.interval       dd      ?

end virtual

        push    ebx edi         ; save used registers to be stdcall

        push    eax             ; save eax for checks in step 3

; 1. Only intervals 2^k ms can be supported.

; The core specification says that the real interval should not be greater

; than the interval given by the endpoint descriptor, but can be less.

; Determine the actual interval as 2^k ms.

        mov     eax, ecx

; 1a. Set [.interval] to 1 if it was zero; leave it as is otherwise

        cmp     [.interval], 1

        adc     [.interval], 0

; 1b. Divide ecx by two while it is strictly greater than [.interval].

@@:

        shr     ecx, 1

        cmp     [.interval], ecx

        jb      @b

; ecx = the actual interval

;

; For example, let ecx = 8, eax = 64.

; The scheduler space is 32 milliseconds,

; we need to schedule something every 8 ms;

; there are 8 variants: schedule at times 0,8,16,24,

; schedule at times 1,9,17,25,..., schedule at times 7,15,23,31.

; Now concentrate: there are three nested loops,

; * the innermost loop calculates the total periodic bandwidth scheduled

;   in the given millisecond,

; * the intermediate loop calculates the maximum over all milliseconds

;   in the given variant, that is the quantity we're trying to minimize,

; * the outermost loop checks all variants.

; 2. Calculate offset between the first list and the first list for the

; selected interval, in bytes; save in the stack for step 4.

        sub     eax, ecx

        sub     eax, ecx

        imul    eax, sizeof.ohci_static_ep

        push    eax

        imul    ebx, ecx, sizeof.ohci_static_ep

; 3. Select the best variant.

; 3a. The outermost loop.

; Prepare for the loop: set the current optimal bandwidth to maximum

; possible value (so that any variant will pass the first comparison),

; calculate delta for the intermediate loop.

        or      [.bandwidth], -1

.varloop:

; 3b. The intermediate loop.

; Prepare for the loop: set the maximum to be calculated to zero,

; save counter of the outermost loop.

        xor     edi, edi

        push    edx

virtual at esp

.cur_variant    dd      ?       ; step 3b

.result_delta   dd      ?       ; step 2

.group1_limit   dd      ?       ; function prolog

end virtual

.calc_max_bandwidth:

; 3c. The innermost loop. Sum over all lists.

        xor     eax, eax

        push    edx

.calc_bandwidth:

        add     eax, [edx+ohci_static_ep.SoftwarePart+usb_static_ep.Bandwidth]

        mov     edx, [edx+ohci_static_ep.NextList]

        test    edx, edx

        jnz     .calc_bandwidth

        pop     edx

; 3d. The intermediate loop continued: update maximum.

        cmp     eax, edi

        jb      @f

        mov     edi, eax

@@:

; 3e. The intermediate loop continued: advance counter.

        add     edx, ebx

        cmp     edx, [.group1_limit]

        jb      .calc_max_bandwidth

; 3e. The intermediate loop done: restore counter of the outermost loop.

        pop     edx

; 3f. The outermost loop continued: if the current variant is

; better (maybe not strictly) then the previous optimum, update

; the optimal bandwidth and resulting list.

        cmp     edi, [.bandwidth]

        ja      @f

        mov     [.bandwidth], edi

        mov     [.target], edx

@@:

; 3g. The outermost loop continued: advance counter.

        add     edx, sizeof.ohci_static_ep

        dec     ecx

        jnz     .varloop

; 4. Get the pointer to the best list.

        pop     edx             ; restore value from step 2

        pop     eax             ; purge stack var from prolog

        add     edx, [.target]

; 5. Calculate bandwidth for the new pipe.

        mov     eax, [.maxpacket]       ; TODO: calculate real bandwidth

        and     eax, (1 shl 11) - 1

; 6. TODO: check that bandwidth for the new pipe plus old bandwidth

; still fits to maximum allowed by the core specification.

; 7. Convert {o|u}hci_static_ep to usb_static_ep, update bandwidth and return.

        add     edx, ohci_static_ep.SoftwarePart

        add     [edx+usb_static_ep.Bandwidth], eax

        pop     edi ebx         ; restore used registers to be stdcall

        ret

endp

; sanity check, part 2

else

.err select_interrupt_list must be different for UHCI and OHCI

end if

; Pipe is removing, update the corresponding lists.

; We do not reorder anything, so just update book-keeping variable

; in the list header.

proc usb1_interrupt_list_unlink

virtual at esp

                dd      ?       ; return address

.maxpacket      dd      ?

.lowspeed       db      ?

.direction      db      ?

                rb      2

end virtual

; find list header

        mov     edx, ebx

@@:

        mov     edx, [edx+usb_pipe.NextVirt]

        cmp     [edx+usb_pipe.Controller], esi

        jnz     @b

; subtract pipe bandwidth

; TODO: calculate real bandwidth

        mov     eax, [.maxpacket]

        and     eax, (1 shl 11) - 1

        sub     [edx+usb_static_ep.Bandwidth], eax

        ret     8

endp

; USB2 scheduler.

; There are two parts: high-speed pipes and split-transaction pipes.

; Split-transaction scheduler is currently a stub.

; High-speed scheduler uses the same algorithm as USB1 scheduler:

; when adding a pipe, optimize the following quantity:

;  * for every microframe, take all bandwidth scheduled to periodic transfers,

;  * calculate maximum over all microframe,

;  * select a variant which minimizes that maximum;

; when removing a pipe, do nothing (except for bookkeeping).

; in: esi -> usb_controller

; out: edx -> usb_static_ep, eax = S-Mask

proc ehci_select_hs_interrupt_list

; inherit some variables from usb_open_pipe

virtual at ebp-12

.targetsmask    dd      ?

.bandwidth      dd      ?

.target         dd      ?

                dd      ?

                dd      ?

.config_pipe    dd      ?

.endpoint       dd      ?

.maxpacket      dd      ?

.type           dd      ?

.interval       dd      ?

end virtual

; prolog, initialize local vars

        or      [.bandwidth], -1

        or      [.target], -1

        or      [.targetsmask], -1

        push    ebx edi         ; save used registers to be stdcall

; 1. In EHCI, every list describes one millisecond = 8 microframes.

; Thus, there are two significantly different branches:

; for pipes with interval >= 8 microframes, advance to 2,

; for pipes which should be planned in every frame (one or more microframes),

; go to 9.

; Note: the actual interval for high-speed devices is 2^([.interval]-1),

; (the core specification forbids [.interval] == 0)

        mov     ecx, [.interval]

        dec     ecx

        cmp     ecx, 3

        jb      .every_frame

; 2. Determine the actual interval in milliseconds.

        sub     ecx, 3

        cmp     ecx, 5  ; maximum 32ms

        jbe     @f

        push    5

        pop     ecx

@@:

; There are four nested loops,

; * Loop #4 (the innermost one) calculates the total periodic bandwidth

;   scheduled in the given microframe of the given millisecond.

; * Loop #3 calculates the maximum over all milliseconds

;   in the given variant, that is the quantity we're trying to minimize.

; * Loops #1 and #2 check all variants;

;   loop #1 is responsible for the target millisecond,

;   loop #2 is responsible for the microframe within millisecond.

; 3. Prepare for loops.

; ebx = number of iterations of loop #1

; [esp] = delta of counter for loop #3, in bytes

; [esp+4] = delta between the first group and the target group, in bytes

        push    1

        pop     ebx

        push    sizeof.ehci_static_ep

        pop     edx

        shl     ebx, cl

        shl     edx, cl

        mov     eax, 64*sizeof.ehci_static_ep

        sub     eax, edx

        sub     eax, edx

        push    eax

        push    edx

; 4. Select the best variant.

; 4a. Loop #1: initialize counter = pointer to ehci_static_ep for

; the target millisecond in the first group.

        lea     edx, [esi+ehci_controller.IntEDs-sizeof.ehci_controller]

.varloop0:

; 4b. Loop #2: initialize counter = microframe within the target millisecond.

        xor     ecx, ecx

.varloop:

; 4c. Loop #3: save counter of loop #1,

; initialize counter with the value of loop #1 counter,

; initialize maximal bandwidth = zero.

        xor     edi, edi

        push    edx

virtual at esp

.saved_counter1         dd      ?       ; step 4c

.loop3_delta            dd      ?       ; step 3

.target_delta           dd      ?       ; step 3

end virtual

.calc_max_bandwidth:

; 4d. Loop #4: initialize counter with the value of loop #3 counter,

; initialize total bandwidth = zero.

        xor     eax, eax

        push    edx