Document: PCI IDE Controllers for OS/2 Maintainer: Patrick Duffy, duffy@theory.chem.ubc.ca Last Revision Date: August 20, 1995 Archived at: ftp.netcom.com, in directory /pub/ab/abe/ CompuServe (GO BENCHMARK) Web pages: http://warp.eecs.berkeley.edu/os2/workbench/work.htm This document may be distributed freely, provided it is distributed in its entirety and is unmodified. There seems to be a proliferation of PCI IDE hardware both built in to PCI motherboards or available as separate cards. Some of it is reported to work well, and some of it is best left unused. One in particular, the PC Tech RZ1000, has been the cause of much discussion on UseNet, as it is essentially broken when its prefetch buffers are enabled, and it was used in boards by Intel and sold to many OEMs. The full story is below. I've therefore created this PCI IDE list so that people will know if their particular PCI IDE implementation will work with OS/2. I've tried to keep this information as accurate as possible, but if you spot any flaws or omissions, please do not hesitate to let me know. Much of this was received from sources within IBM. You may see square brackets beside certain chipsets. If the word "confirmed" appears therein, this means that the bug detailed in the description to follow along with the subsequent fix have been confirmed to me by IBM. I've placed a '*' beside the controllers which seem to work well. The vendor's name and PCI ID appear in HEX/decimal in brackets beside the controller name. Useful numbers: --------------- DTC: (408) 942-4010 (BBS) (408) 942-4005 (Faxback) (408) 942-4081 (sales) (408) 942-4027 (FAX) Chipsets -------- 1) CMD 640 [confirmed] (CMD Technology Inc.: 1095/4245) Channels (two IDE drives may be attached to each IDE channel) cannot be operated concurrently because there is a single I/O queue for both channels. Simultaneous reads to both channels will cause data corruption. This is not mentioned in the chipset errata. Their stand-alone PCI board does not have a BIOS and appears to come-up disabled per the PCI definition. Its really difficult to tell whether a PCI-IDE controller is operating in 'legacy' mode since (at least) the CMD chipset does not update the PCI config space with the legacy base port addresses. Also convincing PCI to route IRQ 14/15 to a PCI Int is problematic unless the motherboard BIOS explicitly supports this function. The CMD chipset also has a documented restriction that it will not support DWORD config write cycles. Of course the Award BIOS on the (Vobis) motherboard on which testing was being performed did not issue byte/word config cycles explicitly and instead implemented all config read/write services as read/modify/write of DWORD data. The net result was that the system PCI BIOS could not be used to program the motherboard chipset! On top of all that, it appears that at least one revision of the CMD chipset (the 640B) seems to suffer from the same problem (or at least one with similar symptoms) as the second of the two PC Tech RZ1000 problems, discussed below. 2) The Intel Triton PCI chipset (Intel: 8086/32902) Any board which uses the Intel Triton PCI chipset will have, via the Triton chipset, support for EIDE built in to the board. The EIDE is busmastering, and apparently allows for throughputs of up to 22 MB/s. This particular chipset is reported to be problem-free so far. The file triton.exe, available on the major FTP sites, contains drivers which take advantage of the busmastering capabilities of the Triton chipset's built-in IDE controller. One person has reported problems with this driver and fixpack 9. 3) PC Tech RZ1000 (PC Technology Inc.: 1042/4162) There are two bugs in this chip. The first has been known for about a year, and has a fix in fixpack 5. Its story is: Produces accelerated IRQ. IBM1S506.ADD actually tolerates this. However, when the driver reads 1x7 during interrupt processing to clear the interrupt, the chipset puts the status in its read-ahead FIFO resulting in data corruption. Intel used this chip on their motherboards which they sold to Gateway, AT&T, IBM and others. This feature is not mentioned in its chipset errata. The second bug was discovered only recently by PowerQuest. It has been the source of much discussion on UseNet. Here's how the now-famous bug occurs: IDE is (in general) a polled-mode affair. In a typical IDE I/O operation, the CPU sits in a tight loop reading data from the IDE controller and writing it out to memory. When the controller is done sending the data, it sends an interrupt to the CPU to tell it that it's done. The interrupt handler then clears the interrupt and the whole transaction comes to an end. However, sometimes another I/O device can interrupt the CPU to do DMA (a floppy drive, for example) just as the IDE I/O operation is coming to an end. This is still no problem, until the device which interrupted the I/O operation reads its own status from an address different than the IDE I/O status register. What should happen then is that the hardware should provide two unique data items: one for the device just queried and one for the IDE drive. This apparently does not happen when the PC Tech RZ1000 is in use, and as a result data corruption occurs. (This is somewhat vague, I know, but it's the best I can make of what I've read. If someone would care to clarify -- say Sam Detweiler -- I'd be happy to fix it up.) 4) Promise (Promise Technology: 105A/4186) [confirmed/software fix implemented in Warp fixpack 5] Promise uses a proprietary implementation to allow 4 drives per channel. In normal compatibility mode IBM1S506.ADD can only see the 1st two drives per channel. In addition their current controller products do not operate properly when Set-Multiple support is enabled. It is therefore necessary to detect Promise controllers (which Promise is trying to figure out how to do) and, upon detection of a promise controller, disable multiple mode support in IBM1S506.ADD. (Promise has their own drivers for their cards.) Controllers ----------- BusLogic (104B/4171) BT-910 BusLogic makes what is apparently a good, fast, and reliable controller which has driver support built into OS/2 (via IBM1S506.ADD), or via their own driver. The card apparently has an on-board 80286 processor which supports up to a 16 MB cache (RAM must be purchased separately), and support for disk mirroring (for drives on separate cables) and disk linking (for disks on the same cable). DTC 2130S This is a single-port controller which is reported to work well with Warp when using the latest drivers (available from the DTC BBS). It comes with its own drivers for DOS, Windows, and OS/2 (among others), but will work with IBM1S506.ADD or WDCIDEOS.ADD (the driver from Western Digital). It can, apparently, be operated with a 16-bit paddle card (which will use the ISA bus to grab a second IRQ for the second IDE channel) for compatibility. It supports mode 3 operation, multiple sector transfers, LBA addressing, and ATA rev. 4.0 devices. (Data Technology Corporation: 107F/4223) Setup tips: After you've installed the driver, modify its line in config.sys to read: BASEDEV=DTCIDE.ADD /V /D0:P3 /D0:M16 /D0:L1 to enable PIO mode 3, multiple (16) sector transfers, and LRA addressing. /D0:M32 may yield better performance. GigaByte GA-0108 This card not only has the IDE controller (CMD 640), but also has support for serial, parallel, and game ports (2/1/1) on board. It is reported to work well with Warp (the drivers were supplied with the card). (CMD Technology Inc.: 1095/4245) Promise 5030* This card works with or without a paddleboard, and is reported to work well with Warp. Apparently it has options which will allow it to work with buggy or older PCI hardware (like register config disable, etc.), which make it a good choice for OS/2. I don't know if it busmasters or what modes of PIO it supports. (Promise Technology: 105A/4186) Tekram DC290N* I don't have many specifics on this card, but it is apparently reported to work well under Warp with the supplied drivers. The one report of success I've had with this board also indicated that use of the supplied paddle board may not be necessary. (Tekram Technology Co., Ltd.: 10E1/4321) Tekram DC290S This is the busmastering version of the DC290N. I have been told that the card works well, but that the necessary driver to get CD-ROM support in OS/2 does not come with the card. (The same setup gave CD-ROM support when hooked up to the on-baord EIDE controller on the motherboard.) (Tekram Technology Co., Ltd.: 10E1/4321) Tyan S1336 This card, like the Gigabyte above, is based on the CMD IDE chipset and has support (via the SMC37C665) for serial/parallel/game ports. I would not recommend this card, since apparently the documentation contains errors, and the card is difficult to set up. (CMD Technology Inc.: 1095/4245) Vision Technologies QDI6500 This card uses NS16550s for its serial port and comes with drivers for DOS, Windows, OS/2 and NT. It is reported to work well with OS/2. (IDE controller mfr. unknown.) That's what I know. E-Mail corrections/suggestions and I'll post again. -- Patrick Duffy, duffy@theory.chem.ubc.ca -- I am Homer of Borg. Resistance is fu-- mmmmm... donuts...