From appli.appli.se!appli.se!owner-amiga-mach@appli.se Thu Oct 22 14:43:01 1992 Received: by bruce (5.57/1.34) id AA13355; Thu, 22 Oct 92 14:42:22 +1000 Received: by mail.swip.net (5.65c8/1.2) id AA23435; Thu, 22 Oct 1992 05:41:42 +0100 Received: by appli.appli.se (/\=-/\ Smail3.1.17.5 #17.26) id ; Thu, 22 Oct 92 06:54 MDT Received: by appli.appli.se (/\=-/\ Smail3.1.17.5 #17.26) id ; Thu, 22 Oct 92 06:54 MDT Return-Path: Received: by appli.appli.se (/\=-/\ Smail3.1.17.5 #17.26) id ; Thu, 22 Oct 92 06:54 MDT Received: from amiga.physik.unizh.ch by mail.swip.net (5.65c8/1.2) id AA21503; Thu, 22 Oct 1992 02:48:27 +0100 Received: by amiga.physik.unizh.ch id AA22991 (5.65c8/IDA-1.4.4 for amiga-mach@appli.se); Thu, 22 Oct 1992 01:45:22 GMT Received: by godzilla.uucp id AA00671 (5.65c8/IDA-1.4.4 for amiga-mach@appli.se); Thu, 22 Oct 1992 02:40:36 +0100 Message-Id: <199210220140.AA00671@godzilla.uucp> Subject: Re: SCSI IO success (was: My MAXTOR SCSI disk needs special treatment, To: amiga-mach@appli.se (The AmigaMach mailing list) Date: Thu, 22 Oct 92 2:40:46 MET From: Markus Wild Reply-To: X-Mailer: ELM [version 2.3 PL11 913261353] Content-Type: text Content-Length: 4508 Sender: owner-amiga-mach@appli.se Status: OR > Too bad, it must be something else. BUT... it seems I don't need this > MODE SELECT command! I just commented it out and... Well, blush, I found the bug(s)... a3000_dma.c had several small but important bugs, the one that caused the above problem is easy.. The function that returns the number of bytes to transfer returned TRUE instead of the count.. Apply this diff to a3000_dma.c, and things should look much better... : *** a3000_dma.c Thu Oct 22 00:42:28 1992 --- /home/mw/a3000_dma.c Thu Oct 22 00:40:32 1992 *************** *** 199,204 **** --- 199,207 ---- io_req_t ior = tgt->ior; register int len = ior->io_count; + printf ("SCSI ABOUT TO WRITE TO DISK!!\n"); + gimmeabreak(); + tgt->transient_state.out_count = len; /* avoid leaks ?? */ *************** *** 209,215 **** tgt->transient_state.out_count = len; } ! tgt->transient_state.dma_offset = 0; } else { --- 212,219 ---- tgt->transient_state.out_count = len; } ! /* see start_cmd & start_dataout */ ! tgt->transient_state.dma_offset = 1; } else { *************** *** 354,360 **** } ! boolean_t a3000_dma_start_dataout (dma, tgt, regp, cmd) opaque_t dma; register target_info_t *tgt; --- 358,365 ---- } ! /* boolean_t */ ! int a3000_dma_start_dataout (dma, tgt, regp, cmd) opaque_t dma; register target_info_t *tgt; *************** *** 362,373 **** { register char *dma_ptr; ! /* since we don't transmit the CDB explicitly, make sure no code depends ! on such dma stuff */ if (tgt->transient_state.dma_offset) ! panic ("a3000_dma_start_dataout"); - dma_ptr = tgt->ior->io_data; if (! dma_ptr) panic ("a3000_dma_start_dataout no data_ptr!"); --- 367,379 ---- { register char *dma_ptr; ! /* see dma_map, we have set dma_offset to 1 if transferring from ior, else ! we're transferring from the command buffer */ if (tgt->transient_state.dma_offset) ! dma_ptr = tgt->ior->io_data; ! else ! dma_ptr = tgt->cmd_ptr + tgt->transient_state.cmd_count; if (! dma_ptr) panic ("a3000_dma_start_dataout no data_ptr!"); *************** *** 376,382 **** tgt->transient_state.dma_offset = 0; tgt->dma_ptr = dma_ptr; ! return TRUE; } int --- 382,391 ---- tgt->transient_state.dma_offset = 0; tgt->dma_ptr = dma_ptr; ! /* return TRUE;*/ ! /* since we're using start_dataout as part of start_cmd, return the ! transfer length the SBIC should be loaded with */ ! return tgt->transient_state.out_count; } int *************** *** 477,483 **** opaque_t dma; target_info_t *tgt; { ! /* Nothing needed */ return FALSE; } --- 486,495 ---- opaque_t dma; target_info_t *tgt; { ! /* we have to stop DMA here, as the A3000-DMAC doesn't have a notion ! of a transfer length */ ! a3000_stop_dma (); ! return FALSE; } > ******************************************************************* > *** *** > *** SCSI Read works !!!!! *** > *** *** > ******************************************************************* Hmm... it doesn't on my system ;-( Granted, I have disconnect enabled (;-)), but the problem lies deeper... I did apply your changes to pcb.c, but I think something in the scheduler is still broken. When I disable the calling of iowait() in scsi_go_and_wait() and just have the driver busy loop until the command is finished, I get rather far (I could test some of the disconnect/reconnect protocol), however, if I allow Mach to enter iowait, the system freezes. I think this must have something to do with task-scheduling, as iowait sooner or later calls thread_block() which forces a context switch. So, since things work for you, what did you do what you didn't tell us;-) ? As a sidenote to the above diff, I think I know now why SIM crashed.. one of the bugs in the DMA driver resulted in a panic, this panic must just have happend while MACH was in spl7, so SIM was called with all interrupts disabled, and I could imagine it didn't like this situation... So.. anxious to learn what magic tweaks the kernel needs.. ;-) -Markus -- Markus M. Wild - wild@nessie.cs.id.ethz.ch | wild@amiga.physik.unizh.ch Vital papers will demonstrate their vitality by spontaneously moving from where you left them to where you can't find them.