Sat Sep 20 21:00 1997 Gerard Roudier (groudier@club-internet.fr) * revision 2.5c - Several PCI configuration registers fix-ups for powerpc. (Patch sent by Cort). Thu Aug 28 10:00 1997 Gerard Roudier (groudier@club-internet.fr) * revision 2.5b - Add 'ncr53c8xx' char pointer variable. This variable allows to pass a boot command to the driver when it is loaded as a module. Option separator is ' ' instead of ','. Example: insmod /ncr53c8xx.o ncr53c8xx='verb:2 sync:0 specf:n' - Always use 'driver_setup.settle_delay' for internal resets. 2 seconds hardcoded is sometimes too short. Suggested by Richard W. This delay may be shortenned in order to avoid spurious timeouts. - Fix release module stuff that failed for more than 1 controller. - For linux versions > 1.3.70, trust the 'dev_id' parameter passed to the interrupt handler (dev_id = struct ncb *). - Fix up in 'ncr_log_hard_error()' when the DSP points outside scripts. Suggested by Stefan Esser. Tue Aug 23 23:43 1997 Gerard Roudier (groudier@club-internet.fr) * revision 2.5a - Update Configure.help for inclusion in linux-2.1.51/2/3 - Use BASE_2 address from PCI config space instead of some IO register for getting the on-board SRAM bus address. - Remove error testing of pcibios_read/write functions. These functions are intended to be used for successfully detected PCI devices. Expecting error condition from them is nothing but paranoia. Thu Aug 21 23:00 1997 Gerard Roudier (groudier@club-internet.fr) * revision 2.5 - 53C860 chip support fix. - Move the 'host_status' to the last DWORD of the CCB header. This header is copied back by the script processor. This guarantees that the header is entirely copied back over the PCI when the CPU completes a CCB. - (re)read ISTAT prior to scanning CCBs for completion. This ensure that any posted buffer are flushed prior CCBs scan. - Support for BIG ENDIAN cpu. Added by Cort . Initial patch did'nt support disconnections and tagged commands. I've completed the patch and it seems that all is ok now. Only some powerpc under 2.1.X is supported for the moment. - Misc. trivial fixes and cleanups. Sat July 26 18:00 1997 Gerard Roudier (groudier@club-internet.fr) * revision 2.4 Several clean-ups: - Asynchronous pre-scaler calculation. Synchronous divisor calculation. - Use FE_ as feature identifier prefix instead of _F_. - Change 'ns_sync' identifier to "minsync". - Some others. Apply some SPI2-R12 recommendations. - Use Slow, Fast-10, Fast-20, Fast-40 SCSI instead of SCSI-2, FAST SCSI-2, ULTRA, ULTRA-2. - Reset the SCSI on bus mode change. Wed July 02 22:58 1997 Gerard Roudier (groudier@club-internet.fr) * revision 2.3c - Add define SCSI_NCR_PCI_FIX_UP_SUPPORT for conditionnal compilation of the corresponding pci fix-up code when a small driver is needed. - Use "ncr53c8xx" as driver name for both request_irq() and request_region(). Using different names confused 'lsdev'. (Suggestion sent by Henrik Storner). Wed June 24 22:08 1997 Gerard Roudier (groudier@club-internet.fr) * revision 2.3b - Print an error message on unexpected boot command line option. - Switch to asynchronous data transfer mode after SCSI wide negotiation. Wed June 14 22:00 1997 Gerard Roudier (groudier@club-internet.fr) * revision 2.3a - Add PCI LATENCY TIMER fixup code. Increase it if necessary according to burst size. Boot option bit : 'pcifix:4' - On phase mismatch, calculate residual data size for all OUTPUT phases. That's only required for interrupted DATA OUT phase, but this information is usefull for problem solving. - Add KERN_INFO to some messages printed to the log. (Patch sent by Wolfram Kleff). Tue June 02 22:30 1997 Gerard Roudier (groudier@club-internet.fr) * revision 2.3 - NvRAM support code slightly improved (I think): Use IO or MMIO according to driver setup for reading the NvRAM. Use structures for NvRAM data instead of raw data. - Prevent from queuing more than 1 command to the scsi SCRIPT with negotiation attached when tagged command queueing is enabled. - Fix-up for old 53C8XX chips that support PCI READ LINE but not CACHE LINE SIZE. If the cache line size is unknown, set burst to 8 dwords and disable READ LINE, otherwise set burst max to the cache line size value. Sat May 24 12:30 1997 Gerard Roudier (groudier@club-internet.fr) * revision 2.2c (for linux-2.1.40) - Remove reference to 'x86' symbol when MODULE is defined, since this symbol is not exported for module loading. The value of 'x86' is used for fixing up the PCI CACHE LINE SIZE configuration register. - Bytes/words read one bit at a time from the serial NVRAM were'nt initialized with zero. - Some comments added. Minor cosmetic changes. Mon May 19 20:30 1997 Gerard Roudier (groudier@club-internet.fr) * revision 2.2b - Patch for NVRAM support by Richard Waltham applied. The code detects Symbios NVRAM format and Tekram NVRAM format. This enhancement allows to get hosts and devices user set up from the NVRAM. - Use the NVRAM contents when present to initialize user definable target parameters. - Update the README file. Sun May 11 22:30 1997 Gerard Roudier (groudier@club-internet.fr) * revision 2.1b - Cosmetic changes. - Some heavy testings under pre-linux-2.1.37-6 Sun May 4 22:30 1997 Gerard Roudier (groudier@club-internet.fr) * revision 2.1a - PFEN wrongly used for PREFETCH feature bit testing. Changed to _F_PFEN. - 2 SCR_COPY that need NO FLUSH bit to be removed had been missed in tp->getscr[] script (loads SXFER and SCNTL3 on reselection). Sat May 3 22:30 1997 Gerard Roudier (groudier@club-internet.fr) * revision 2.1 - Use the NO FLUSH option for MOVE MEMORY (COPY) each time it is possible. More than 100 COPY with NO FLUSH and 6 with FLUSH for my configuration (max queued command / device = 8). This option bit is removed from the script instance for chips that donnot support prefetching. - Rewrite the ncr_exception() routine more simple (I think) and remove useless code. - Change the data_in and data_out script management. Use the bottom part of these scripts instead of the beginning. That avoids to zero the scatter/gather array when a command is queued (1k) and to deal with some weird IID on MOVE 0 bytes when a target wants to transfer more bytes than expected. - Misc. improvements in the init code. - Remove IOMAPPED/MMIO automatic switching option. Was useless and reported not reliable. - Fix a double read of DSTAT and remove DFE testing in the Phase mismatch service routine. - Etc... Fri Apr 26 20:00 1997 Gerard Roudier (groudier@club-internet.fr) * revision 2.0a - Add support if the Diamond FirePort 40 (SYM53C875J chip) Mon Apr 22 22:00 1997 Gerard Roudier (groudier@club-internet.fr) * revision 2.0 - incorporate __initdata and __initfunc directives in order to allow 'init' to free unused memory after driver initialisations. Patch sent by Roberto Fichera. - rewrite the init code of the driver. Now a feature descriptor is used for each real chip types. The code is a lot more clean, since the driver uses device and revision ids only in the detection procedure. - add 'pcifix' boot command line. This command allows to fix up PCI config space for new chips which support features based on the cache line size and 'write and invalidate'. - incorporate in the driver, the code used for error recovery testing. This code is normally not compiled; have to define SCSI_NCR_DEBUG_ERROR_RECOVERY in order to compile it. - take into account actual SCSI bus mode for 53C895 LVD/SE controller. In single ended mode only fast20 is supported. (Just to not be late since such controllers are not yet available) Sat Apr 20 21:00 1997 Gerard Roudier (groudier@club-internet.fr) * revision 1.18f - fix an old bug included in the initial port (version 0.0). The driver allocated 10 bytes of static data and uses 12 bytes. No danger, since data are generally aligned on 4 bytes boundary and so byte 10 and 11 are free (I hope ...) Wed Apr 16 12:00 1997 Gerard Roudier (groudier@club-internet.fr) * revision 1.18e - reset all when an unexpected data cycle is detected while disconnecting. - make changes to abort() ans reset() functions according to Leonard's documentation. - small fix in some message for hard errors. Sat Apr 5 13:00 1997 Gerard Roudier (groudier@club-internet.fr) * revision 1.18d - Probe NCR pci device ids in reverse order if asked by user from the boot command line. Suggested by Richard Waltham. - Make a separate function that prints out verbose information on severe error (assumed from hardware). - Add the transfer period factor and the max commands per lun value to the proc info data. If debug flags are set or verbosity is greater than 1, debug flags and verbosity are returned in proc info data. - Update the documentation. Thu Mar 20 23:00 1997 Gerard Roudier (groudier@club-internet.fr) * revision 1.18c - Add special features support for NCR53C885 and NCR53C896 chip. Quite obvious, but untested, and based on the fact that: The 885 supports same features as the 875. The 896 is a 64 bits PCI version of the 895. - Improve recovery from SCSI GROSS ERRORS. I can get such errors by making the driver negotiate offset 8 with a disk and setting the ncr chip to a lower offset value. I got bunches of errors that have been gracefully recovered by the driver. The driver now uses its timer handler in order to wait 2 sec. for devices to settle after SCSI reset and so does not uselessly freeze the system with interrupt masked for seconds. - Enable 'burst op code fetch' and 'read line' for 815 chips. - Use a 2 commands queue depth instead of 1 for devices that does not support tagged command queuing. - The ULTRA timing flag setting was based on the output resulting period factor of the ncr and not on the negotiated one. This flag setting was wrong only for 24 ns negotiated period factor. - Some other minor changes and cleanups. Thu Feb 27 23:00 1997 Gerard Roudier (groudier@club-internet.fr) * ncr53c8xx.c ncr53c8xx.h revision 1.18b - 'On paper' support of the NCR53C895 Ultra-2 chip. (Clock quadrupler + 7 clock divisors) - Load the main part of the script into the on-board RAM. - 810A rev. 0x11 PCI problem fixed. This chip is now supported with all PCI features enabled and 16 dwords burst transfers. - Align on 32 boundary some internal structures. That fixes the 810A problem and allows cache line bursting when moving the global header (64 bytes) from/to CCBs to/from NCB. - Synchronous parameters calculation rewritten. The driver now uses all available clock divisors and will be able to support clock frequencies that are not multiple of 40 Mhz if necessary. Sat Feb 8 22:00 1997 Gerard Roudier (groudier@club-internet.fr) * ncr53c8xx.c - revision 1.17a - IRQ mode set up from boot setup command. irqm:0 open drain (default) irqm:1 preserve initial setting (assumed from BIOS) irqm:2 totem pole - DIFF mode set up from boot setup command. Suggested by Richard Waltham. diff:0 never set up diff mode (default) diff:1 set up diff mode according to initial setting (BIOS?) diff:2 always set up diff mode diff:3 set up diff mode if GPIO3 is zero (SYMBIOS boards) - Change CONFIG option for LED support. CONFIG_SCSI_NCR53C8XX_SYMBIOS_COMPAT allows LED support and DIFF support for SYMBIOS boards and compatibles (clones?). - Set 16 DWORD bursts for 810A rev. >= 0x12 since my SC200 with such a chip have no problem with it (MB with Triton 2 HX). 810A rev. 0x11 are set to 8 DWORD bursts since they may give problems with PCI read multiple and Triton 2 HX. Thanks to Stefan for this information. Sat Jan 25 22:00 1997 Gerard Roudier (groudier@club-internet.fr) * ncr53c8xx.c - revision 1.17 - Controller LED support. Only works with LED pin wired to GPIO_FETCHN, so probably with all boards using SMDS BIOS. This option can be enabled only if CONFIG_EXPERIMENTAL is set. - Assume clock doubler for 875 chip when clock frequency measurement result is 40 MHz. May help when some old stuff as SDMS BIOS 3.0 or some old driver has broken the normal BIOS settings. - Add wide negotiation control from boot setup command. May be usefull with systems using a 875 based board connected to a wide device through a 50 pins to 68 pins converter. - Add a "boot fail safe option" to the boot setup command line. - Rewrite the "reset_command" routine. Low-level driver are responsible to keep the involved command alive. The new code seems to behave correctly. - Change some variables used by the script from u_long to u_int32. - Remove some useless code. Sun Jan 12 12:00 1997 Gerard Roudier (groudier@club-internet.fr) * ncr53c8xx.c - revision 1.16e - Add support of PCI burst length control from boot setup command. burst:0 disable burst burst:255 get burst from initial settings (BIOS settings?) burst:#x set burst transfers to 1<<#x - Only check xfer direction for common op-codes. For all device specific / vendor specific opcodes the driver now uses the xfer direction decided by the target. Sun Jan 05 12:00 1997 Gerard Roudier (groudier@club-internet.fr) * ncr53c8xx.c - revision 1.16d - The driver is now able to process scsi commands without knowledge of xfer data direction. Stefan agreed with this change for Linux. This change is not needed under FreeBSD since low-level drivers receive the expected data direction for each scsi request. - Save ctest5 features bits at start-up and restore them at module release step. Avoid side effects when a ncr driver which trusts bios settings is reloaded (could be the ncr53c8xx itself). Wed Jan 01 23:30 1997 Gerard Roudier (groudier@club-internet.fr) * ncr53c8xx.c - revision 1.16c - Bad decision about 20MHz for 13 ns period factor. Was wrong, so I restore the previous algorithm. - Burst length 128 not correctly set in dmode. Thu Dec 26 22:00 1996 Gerard Roudier (groudier@club-internet.fr) * ncr53c8xx.c ncr53c8xx.h README.ncr53c8xx - revision 1.16b - Remove useless code. - Try to improve error recovery in case of abort and reset. - Remove DEBUG_NEGO by default. - Add boot setup command support. Now, all experimental config options can be removed. - Update README file. Mon Dec 23 23:00 1996 Gerard Roudier (groudier@club-internet.fr) * ncr53c8xx.c ncr53c8xx.h - revision 1.16a New display for speed ##.# MB/s (From Stefan) - I add "WIDE" qualifier after ULTRA and FAST - I get "FAST WIDE SCSI-2 20 MB/s" with my Atlas. That's nice. Richard Waltham reports SYMBIOS set the 875 to 20 MB/s for 13 ns period factor. I decide to trust SYMBIOS. 20 MB/s output speed instead of 19.2 MB/s should not cause problem. The ncr is only able to use 16.67 MB/s when 20 MB/s is not possible. Fix from Markus Kossman: "Ultra SCSI enabled" wrongly printed when not enabled. Set DEBUG_NEGO by default in order to get reports about sync nego. Will remove it in the next patch. Thu Dec 19 21:00 1996 Gerard Roudier (groudier@club-internet.fr) * ncr53c8xx.c ncr53c8xx.h README.ncr53c8xx - revision 1.16 Incorporate new definitions in ncr53c8xx.h (From Stefan). Check changes against Stefan's current version of the driver. All seems ok. Sat Nov 30 21:00 1996 Gerard Roudier (groudier@club-internet.fr) * ncr53c8xx.c ncr53c8xx.h Make changes in order to support: - Clock doubler and so 80 Mhz scsi clock for 875 chips. - Sync transfers below 7.5 MB/sec. Use Clock/2 between 5 and 10 Mega-transfers/s and Clock/4 below 5. - Ultra SCSI data transfers. - Offset 16. Works with my configuration. However I cannot test Ultra transfers, since my disks are only fast scsi-2. Tue Nov 28 21:00 1996 Gerard Roudier (groudier@club-internet.fr) * ncr53c8xx.c I received yesterday my Promise SCSI Ultra board. NCR53C875 rev. 3 with clock doubler. Add the code to support some bus features, the large 536 dma fifo and burst 128. Works. Mon Nov 4 21:00 1996 Gerard Roudier (groudier@club-internet.fr) * ncr53c8xx.c ncr53c8xx.h - revision 1.14c Severall control command improvements: - Allow to specify "all" to commands that apply to #target. For example: "setsync all 255" sets asynchronous data transfers for all targets on a bus. - Allow to control disconnection privilege per device, as follow: "setflag #target no_sync" disables disconnection for #target. "setflag #target" with no flag specified reenables it. Obviously #target may be specified as "all" in order to control disconnection for all targets with a single control command. - README file updated and some hints about SCSI problems solving added. Sun Oct 27 22:00 1996 Gerard Roudier (groudier@club-internet.fr) * ncr53c8xx.c ncr53c8xx.h - revision 1.14b Add the following config parameters: - CONFIG_SCSI_NCR53C8XX_MAX_TAGS Max number of queued tagged commands. Allow from 2 to 12, default 4. - CONFIG_SCSI_NCR53C8XX_SYNC Synchronous transfers frequency in MHz. Allow from 5 to 10, default 5, 0 means asynchronous. (And so remove CONFIG_SCSI_NCR53C8XX_FORCE_ASYNCHRONOUS) Sun Oct 20 16:00 1996 Gerard Roudier (groudier@club-internet.fr) * ncr53c8xx.c ncr_scatter() rewritten. remove "ncr dead" detection. Sun Oct 13 19:00 1996 Gerard Roudier (groudier@club-internet.fr) * ncr53c8xx.c ncr53c8xx.h - revision 1.14a Enabling some special features makes problems with some hardware. So, disable them by default. Add SCSI_NCR_SPECIAL_FEATURES define to play with. Sun Oct 13 14:00 1996 Gerard Roudier (groudier@club-internet.fr) * ncr53c8xx.c ncr53c8xx.h Incorporate Stefan's patch for clock frequency detection. (Committed in FreeBSD/ncr.c rev. 1.81). The driver then does about the following: Assume 40 MHz clock for all ncr chips except: - NCR53C860 chips: Assume 80 Mhz clock. - NCR53C875 chips: If clock doubler enabled, disable it and assume 40 Mhz clock. Else if (scntl3&7)=0 measure scsi clock frequency. Else trust bios setting of scntl3&7 (3=40 Mhz, 5=80Mhz). Wed Oct 9 22:00 1996 Gerard Roudier (groudier@club-internet.fr) * ncr53c8xx.c - release 1.14 For now, just change the clock detection as follow: - If clock doubler selected by BIOS, assume 40 MHz clock since clock doubler will be disabled by chip reset. - Else if NCR53C860 assume 80 MHz clock. - Else trust BIOS setting if (scntl3&7 >= 3) - Else assume 40 MHz clock. Sat Oct 05 17:00 1996 Gerard Roudier (groudier@club-internet.fr) * ncr53c8xx.c Stefan sent me a patch that improves the clock frequency detection of the driver. Stefan uses the general timer register stime1 in order to measure as accurately as possible the scsi clock. Works ok with my 825, but needs still testing. So will be released later. Sun Sep 29 17:00 1996 Gerard Roudier (groudier@club-internet.fr) * ncr53c8xx.c Preserve dcntl/dmode/ctest3/ctest4 features bits at start-up. Add the define option SCSI_NCR_TRUST_BIOS_SETTING. - If this option is defined, the driver will preserve the corresponding bits of io registers. - Else, the driver will set features bits according to chip and revision ids. Sun Sep 22 17:00 1996 Gerard Roudier (groudier@club-internet.fr) * ncr53c8xx.c Remove useless fields and code and so spare cpu: - profile data are accumulated in jiffies ticks and converted to milli-seconds when read through proc fs. - when IOMAPPED is not defined, try only MMIO. (avoid testing a value in order to choose between IO and MMIO) Sun Sep 01 20:00 1996 Gerard Roudier (groudier@club-internet.fr) * ncr53c8xx.h, ncr53c8xx.c - Version 1.13 Adaptation of the tagged command queuing depth control of the FreeBSD driver to Linux. Now, tagged command queueing can be disabled at run time by a "settags N 0" control command. Add the following heuristic in order to manage intelligently (perhaps) QUEUE_FULL status: - Each time a QUEUE FULL status is returned by a device, disable tagged command queuing for that device. - Every 100 successfully complete commands, increment the maximum queuable commands (up to the allowed limit). Fri Aug 30 10:00 1996 Gerard Roudier (groudier@club-internet.fr) * ncr53c8xx.c - Version 1.12c Incorporate the changes of FreeBSD/ncr.c revision 1.76. The changes add support for the 53c860 and 53c875, but without taking advantage of the new features. Those chips are used exactly as the old 53c810. Sun Jul 21 00:00 1996 Gerard Roudier (groudier@club-internet.fr) * ncr53c8xx.c, README.ncr53c8xx Add the ncr53c8xx_select_queue_depths() function. Set queue_depth to SCSI_NCR_MAX_TAGS (4 by default) for devices that support tagged command queueing. For other devices, set queue_depth to 1. No need to queue a command to the driver if this command cannot be sent to the device. Each time the driver hide io requests from the kernel and/or from the driver, it may break a little (or a lot) optimization algorithms that try to increase throughput by reordering io requests. It is better to enable the disk write caching to reduce latencies for write operations, and to trust asynchronous read ahead from the device and from the kernel that can reduce latencies for read operations, even when tagged command queuing is not supported or enabled. Sat Jul 20 20:00 1996 Gerard Roudier (groudier@club-internet.fr) * ncr53c8xx.c Minor changes: - Problem of "CCB address mismatch" that happens with the 3 versions of the driver. The CCB is correct and Stefan Esser suggests a little patch that seems to be a bypass. Stefan says he will change that in a future version of the BSD driver. - Set burst transfers to 8 for 815 chips. Sun Jul 14 15:00 1996 Gerard Roudier (groudier@club-internet.fr) * ncr53c8xx.c, Configure.help Memory mapped io donnot work under linux/Alpha for the driver. For the moment it is better to not support this feature for this architecture. Tue Jul 09 20:00 1996 Gerard Roudier (groudier@club-internet.fr) * ncr53c8xx.c Garbage printed out with the following command (fixed): - cat /proc/scsi/ncr53c8xx/0 Sun Jul 07 20:00 1996 Gerard Roudier (groudier@club-internet.fr) * ncr53c8xx.c Tagged command queueing cannot be disabled at run time. I probably never try that because I felt the risk. Shortest patch sent to Linus. I have to plan something better. Wed Jul 03 23:00 1996 Gerard Roudier (groudier@club-internet.fr) Release 1.12a Tested linux releases: 1.2.13, 2.0.0, 2.0.1 Mon Jul 01 21:00 1996 Gerard Roudier (groudier@club-internet.fr) ncr53c8xx.h, ncr53c8xx.c Add "clearprof" user command that clear the profile counters. Automatically clear profile counters when num_kbytes=1000000000 in order to avoid ugly overflow. Donnot compile user command code and profile data with 1.2.13. Wed Jun 29 20:38 1996 Gerard Roudier (groudier@club-internet.fr) Matthew Geier reported to me a weird problem of unexpected disconnection while asynchronous negotiation. The message sent by the driver is 1-3-1-ff-00. I sent a patch to Matthew that change the message to 1-3-1-00-00. The sync msgout was correct however some devices might be to happy with ff. Wed Jun 26 22:57 1996 Gerard Roudier (groudier@club-internet.fr) Patch no 4 sent to Harald. The drived used "wtime" for timeouts adn time measurements. I change for jiffies. Work with my P133. Mon Jun 24 23:05 1996 Gerard Roudier (groudier@club-internet.fr) Patch no 3 sent to Harald. Sun Jun 23 22:29 1996 Gerard Roudier (groudier@club-internet.fr) Patch no 2 sent to Harald. I think that the driver have some chance to work. Sun Jun 23 15:00 1996 Gerard Roudier (groudier@club-internet.fr) Harald Koenig is interested in the adaptation of the driver to Linux/Alpha. I have prepared a patch and sent it to Harald. Sun Jun 16 19:00 1996 Gerard Roudier (groudier@club-internet.fr) Release 1.11 Tested linux releases: 1.2.13, 2.0.0 Sat Jun 15 23:00 1996 Gerard Roudier (groudier@club-internet.fr) * ncr53c8xx.h, Configure.help, scsi/Config.in Add CONFIG_SCSI_NCR53C8XX_IOMAPPED config option. Prepare the 2.0.0 with the new version of the driver. Wed Jun 12 23:00 1996 Gerard Roudier (groudier@club-internet.fr) Rewrite the README file. Add some documentations of the proc file system support. Sun Jun 9 18:00 1996 Gerard Roudier (groudier@club-internet.fr) * ncr53c8xx.c Add proc filesystem support of the driver. Read operations returns profile information. Write operations send control commands to the host adapter driver. Wed Jun 5 22:00 1996 Gerard Roudier (groudier@club-internet.fr) Change xfer direction for SCAN command to write. Was bogus. Tue May 30 18:00 1996 Gerard Roudier (groudier@club-internet.fr) * ncr53c8xx.c Set the DMA FIFO to 88 for 825A and 875 boards. The previous value of 536 is bogus since the script only read 7 bits for the fifo size (thanks to Stefan). Mon May 27 18:00 1996 Gerard Roudier (groudier@club-internet.fr) * ncr53c8xx.c Correct the xfer direction guessing for scanner SCAN command (write). Mon May 27 18:00 1996 Gerard Roudier (groudier@club-internet.fr) * ncr53c8xx.c Add the following config options: SCSI_NCR_DISABLE_MPARITY_CHECK : disable master parity checking. SCSI_NCR_DISABLE_PARITY_CHECK : disable scsi parity checking. SCSI_NCR_FORCE_SYNC_NEGO : force sync nego for all scsi 2 devices. Sat May 25 22:00 1996 Gerard Roudier (groudier@club-internet.fr) Release 1.10 Tested linux releases: 1.2.13, 1.3.45, 1.3.71, 1.3.90, 1.3.100 1.99.6, 1.99.7 Switch between Drew's driver and Bsd driver tested for 1.99.7. Both driver was made as modules. Sat May 25 16:00 1996 Gerard Roudier (groudier@club-internet.fr) * ncr53c8xx.c Some weird problems happen with multi-lun configurations and HDs. SDTR seems to be sent with TEST UNIT READY of lun 1. Tagged Queue cannot be enabled. It seems that inqdata are filled with garbage probably due to some INQUIRY command to lun 1. I have fixed the problem as follow: - negotiation are initiated only with a command to lun 0. - inquiry data are store only for lun 0. Wed May 22 22:00 1996 Gerard Roudier (groudier@club-internet.fr) * ncr53c8xx.c, ncr53c8xx.h Have prepared the patch that allow to install the driver in the kernel tree, without moving Drew's one. Seems to work. However, I have to check that nothing has been broken for 1.2.13 and 1.3.45 to 1.3.100. Sun May 4 22:00 1996 Gerard Roudier (groudier@club-internet.fr) * ncr53c8xx.h Adapt the source to some modifications of the linux tree of 1.3.98. (include/linux/scsicam.h moved to include/scsi/scsicam.h) Thu Apr 25 21:00 1996 Gerard Roudier (groudier@club-internet.fr) * ncr53c8xx.h, ncr53c8xx.c Release 1.9 Prepare conditionnal compilations for the future Linux version(s). Assume that these versions will be nicely compatible with current one. 1.3.255, 1.4.0 or 2.0.0 ? I suggest 3.0.0 for some obvious reason. Wed Apr 24 23:15 1996 Gerard Roudier (groudier@club-internet.fr) * Install.ncr53c8xx Add Patch-Current.ncr53c8xx to the distribution. This patch is applied to the scsi Makefile at installation time for Linux release V.P.S (V*1000000000+P*100000000+S > 1300000094). Each time it'll be necessary I will send the patch corresponding to the current Linux release to the linux-scsi@vger.rutgers.edu. Sun Apr 21 19:00 1996 Gerard Roudier (groudier@club-internet.fr) * README.ncr53c8xx Update Bonnie results of linux-1.3.92 + prepatch read-ahead 1.3.93. Results are so good that I must remove FreeBSD-2.0.5 results from the README file, otherwise I should cause trouble for myself. Sun Apr 07 21:00 1996 Gerard Roudier (groudier@club-internet.fr) * ncr53c8xx.h Define SCSI_NCR_MAX_LUN (8) inconditionaly. Previous releases did not work for multi-lun devices. This definition was wrongly conditionned: (SCSI_CONFIG_MULTI_LUN instead of CONFIG_SCSI_MULTI_LUN). No luck, since I donnot have multi-lun devices and could'nt test it. Some tests under linux-1.3.84 with an experimental patch that try to do asynchronous read-ahead. Wed Apr 03 23:15 1996 Gerard Roudier (groudier@club-internet.fr) * ncr53c8xx.h, ncr53c8xx.c Change some wrong "assert (target == cmd->target & 7)" to "assert (target == (cmd->target & 0xf)". Remove NCR_TIMEOUT_ALERT from ncr53c8xx.c Add SCSI_NCR_TIMEOUT_ALERT to ncr53c8xx.h Sun Mar 24 21:15 1996 Gerard Roudier (groudier@club-internet.fr) * ncr53c8xx.h During "make dep" of linux-1.2.13, LINUX_VERSION_CODE is undefined. Have to assume 1.2.13 in such situation. Release 1.8 Sun Mar 24 21:00 1996 Gerard Roudier (groudier@club-internet.fr) * README.ncr53c8xx Make changes according to the new uninstallation procedure. Sun Mar 24 20:00 1996 Gerard Roudier (groudier@club-internet.fr) * Uninstall.ncr53c8xx Add an uninstallation script to the distribution. This shell script restore the standard driver. Very usefull for people who prefers to use a driver that does not support: - Master parity check - Tagged command queuing - Fast Wide Scsi-2 features (up to 20 MB/sec) and that would be capable to reject a Wide Negotiation that it had previously initiated. Sat Mar 23 22:00 1996 Gerard Roudier (groudier@club-internet.fr) * README.ncr53c8xx Make changes according to the new installation procedure. Fri Mar 22 23:00 1996 Gerard Roudier (groudier@club-internet.fr) * Install.ncr53c8xx Add an installation script to the distribution. 3 differents patches are necessary: - linux-1.2.13 - linux-1.3.45 to linux-1.3.49 - linux-1.3.50 to linux-1.3.77 Wed Mar 13 22:00 1996 Gerard Roudier (groudier@club-internet.fr) * ncr53c8xx.c, ncr53c8xx.h Add share irq support. This facility appears with linux-1.3.70. It seems that the corresponding code of the kernel was questionnable before 1.3.72. I decide to support this options from linux-1.3.72 and above. (Add define option SCSI_NCR_SHARE_IRQ) Release 1.7 Tue Mar 12 22:00 1996 Gerard Roudier (groudier@club-internet.fr) * ncr53c8xx.c Download BSD ncr.c 1.67 and apply the correction of negotiations order. Now the Wide and Sync negotiation are done in the proper order. (Problem reported by Johannes Plass). Round up correctly the announced speed in MB/sec. Tue Mar 05 22:00 1996 Gerard Roudier (groudier@club-internet.fr) * README.ncr53c8xx Have to upload release 1.6 for users of linux-1.3.70-71 Mon Mar 04 16:00 1996 Gerard Roudier (groudier@club-internet.fr) * README.ncr53c8xx Add some Bonnie results to the README file. Sun Mar 03 20:00 1996 Gerard Roudier (groudier@club-internet.fr) * ncr53c8xx.c Make changes for linux-1.3.70 according to the new specification of irq services interface (request_irq(), free_irq()). With 26 letters, "_", and 10 digits we can build about: 53x62**9 C names of 10 characters. Why to use SAME function names with DIFFERENT parameters ? Sat Mar 02 22:30 1996 Gerard Roudier (groudier@club-internet.fr) * ncr53c8xx.c Using SIMPLE QUEUE TAG for all operations is good for performances, but may be bad for assumed timeout values. Under heavy disk load (Bonnie), the drive may start IO process of a command, then disconnect, then execute lots of other commands before completing the interrupted command. The standard value of SD_TIMEOUT (6 sec or 7 sec) seems (is) too short. I fix the problem by forcing an ORDERED QUEUE TAG for the next command when I found some "old" pending command. "Old" means that they might be timeouted in a few seconds. Add NCR_TIMEOUT_ALERT and set it to 3 seconds. Fri Mar 01 22:30 1996 Gerard Roudier (groudier@club-internet.fr) * ncr53c8xx.h, ncr53c8xx.c Add define SCSI_NCR_SEGMENT_SIZE and set it by default to 512. If undefined, the driver use the scatter list given by the upper scsi driver, else it tries to split segments into shorter one according to the value of SCSI_NCR_SEGMENT_SIZE. Tue Feb 27 21:30 1996 Gerard Roudier (groudier@club-internet.fr) * ncr53c8xx.h Set sg_tablesize to 127 = SCSI_NCR_MAX_SCATTER-1. (was 64 = SCSI_NCR_MAX_SCATTER/2). May increase the speed (about 20%) for linear read/write operations. Bonnie results may be better, but I prefered the previous value. Tue Feb 27 21:00 1996 Gerard Roudier (groudier@club-internet.fr) * ncr54c8xx.c, ncr53c8xx.h Tagged command queueing seems to affect linux-1.3.XY kernels. I decide to disable tagged queue by default and to provide a command tool to enable it per device after boot-up. Add scsitag.c to the distribution. Usage: scsitag device Examples: scsitag /dev/sda scsitag /dev/sdb Sun Feb 25 14:00 1996 Gerard Roudier (groudier@club-internet.fr) * ncr53c8xx.c, ncr53c8xx.h Add INQ7_Default definition and set the target capabilities to this value by default. Add some code to reject a synchronous negotiation request from a target that is defined as not capable of Sync in the table of capabilities. Sat Feb 24 23:00 1996 Gerard Roudier (groudier@club-internet.fr) * ncr53c8xx.c Add some code to fill some write-only host instance fields: - base - io_port - n_io_port - dma_channel This fields may be used with some "standard" drivers that do not process correctly the release function. Fri Feb 23 23:00 1996 Gerard Roudier (groudier@club-internet.fr) I receive a mail from Jason Duerstock. A new bug in ncrBsd2Linux which is probably outside the driver code and inside Linux kernel code. The kernel memory is quite corrupted and we cannot get any information by looking into the messages about the crash. However Linus see a bug in ncrBsd2Linux because it is a "non standard" driver. I think too that ncrBsd2Linux is not a standard driver of Linux kernel, because it has no bugs enough to become standard for Linux. Drew's driver is quite standard. Wed Feb 21 23:00 1996 Gerard Roudier (groudier@club-internet.fr) * ncr53c8xx.c, ncr53c8xx.h I incorporate the differences between FreeBSD ncr.c revision 1.62 and revision 1.64. - Some cosmetic changes. - Use M_SIMPLE_TAG by default (even for write operations). I seems to me that SIMPLE TAG is safe only if the flag "queue algorithm modifier" is set to zero. I will ask some questions to Stefan Esser about this. Add option SCSI_NCR_ALWAYS_SIMPLE_TAG in ncr53c8xx.h. Fri Feb 16 23:00 1996 Gerard Roudier (groudier@club-internet.fr) I have found the bug. It is a recursion in __get_free_pages(). I will send a mail to Linus about this. Sat Feb 10 20:00 1996 Gerard Roudier (groudier@club-internet.fr) I am sure that the kernel stack overflow is due to a severe bug in the Linux kernel. I decide to try to find the bug by myself. Fri Feb 09 20:00 1996 Gerard Roudier (groudier@club-internet.fr) * ncr53c8xx.c, ncr53c8xx.h Limit commands per lun to 2 for linux-1.3.XY. The patch-1.3.60 does not correct the kernel stack overflow problem. I decide to make some tests with Drew's driver and Bsd2Linux with same working conditions (3 commands per lun, FAST SCSI, no command queueing). I get the stack overflow problem with the 2 drivers at the same frequency. With only 2 commands per lun, I dont have the problem with any driver. It seems that the madness of recursion and the recent introduction of the silly generic read function have broken performance and reliability of scsi drivers. Thu Feb 08 23:00 1996 Gerard Roudier (groudier@club-internet.fr) * ncr53c8xx.c, ncr_attach() Release memory mapped region and io port if initialisation does not succeed. Thu Feb 08 22:00 1996 Gerard Roudier (groudier@club-internet.fr) * ncr53c8xx.c, ncr53c8xx.c Try to spare some CPU time in queue_command() function and interrupt handler: Add SCSI_NCR_PARANOIA define option in ncr53c8xx.h. Add SCSI_NCR_PROFILE define option in ncr53c8xx.h. Avoid useless code and function calls. Tue Feb 06 21:00 1996 Gerard Roudier (groudier@club-internet.fr) * ncr53c8xx.c, ncr53c8xx.h, ncr_timeout() Add SCSI_NCR_BROKEN_INTR define option in ncr53c8xx.h. If this option is set, the timeout handler polls the interrupt status register every tick (10 ms). So, boards with broken interrupt can work. Mon Feb 05 21:30 1996 Gerard Roudier (groudier@club-internet.fr) * ncr53c8xx.c Print the correct speed, for devices with successfull wide negotiation. For same period, such devices are two times faster then narrow ones. Mon Feb 05 21:00 1996 Gerard Roudier (groudier@club-internet.fr) * ncr53c8xx.h, ncr53c8xx.c, ncr_attach() Add define SCSI_NCR_SETTLE_TIME in header file and set it by default to 2 seconds. Sat Jan 27 14:00 1996 Gerard Roudier (groudier@club-internet.fr) Upload release 1.3 Wed Jan 24 24:00 1996 Gerard Roudier (groudier@club-internet.fr) * ncr53c8xx.c Update from ncr Bsd 1.60 (Stefan Esser): The handshake timeout is disabled, since a few devices will delay ACK for more than a second: Scanner, CDROM writer and a few old hard disk drives. Wed Jan 24 22:30 1996 Gerard Roudier (groudier@club-internet.fr) * ncr53c8xx.c Set the selection timeout to 0.4 sec, since 0.25 sec is recommended for scsi-1 devices. Note that the Bsd ncr driver sets this timeout to 0.1 sec and the linux standard ncr driver sets it to 0.8 sec. Wed Jan 24 22:00 1996 Gerard Roudier (groudier@club-internet.fr) * ncr53c8xx.c Add a 5 seconds delay after chip initialization, waiting for scsi devices to settle their stomach, as FreeBSD generic scsi driver does. Tue Jan 23 23:00 1996 Gerard Roudier (groudier@club-internet.fr) * ncr53c8xx.c Set burst length value according to chip type. The original Bsd ncr driver sets burst length to 16 for all chip types. Tue Jan 16 23:00 1996 Gerard Roudier (groudier@club-internet.fr) * ncr53c8xx.c, ncr53c8xx.h Add comments to linux specific glue code. Mon Jan 15 22:00 1996 Gerard Roudier (groudier@club-internet.fr) io mapped versus memory mapped. * ncr53c8xx.c Add some code to dynamicaly switch to io mapped if memory mapped does not work. No more need to reconfigure, compile and link the kernel if memory mapped is not possible. Sun Jan 14 18:00 1996 Gerard Roudier (groudier@club-internet.fr) Patch sent to ncr mailing list by Jason Duerstock I have omitted to provide the proc_dir_entry to the middle scsi driver. * ncr53c8xxx.c Add the declaration of the ncr53c8xx proc_dir_entry and return the pointer to middle scsi driver. Sat Jan 13 01:00 1996 Gerard Roudier (groudier@club-internet.fr) ncrBsd2Linux 1.1 is ready. Upload to sunsite immediatly. Fri Jan 12 23:45 1996 Gerard Roudier (groudier@club-internet.fr) It seems that scsi-2 devices too may cause problems because they have flawes in the firmware. * ncr53c8xx.h I add a table of capabilities per target. This table contains one byte per target. The value of this byte is anded with byte 7 of INQUIRY data. Edit ncr53c8xx.h and read the corresponding comments for more details. Wed Jan 10 22:35 1996 Gerard Roudier (groudier@club-internet.fr) I have some time to read the scsi-1 specifications. Some very old scsi devices may cause problems with the ncr Bsd driver for the following raisons: The Synchronous Negotiation protocol was optional. The standardized INQUIRY data does not include the byte 7 of the scsi-2 specifications which contains the capabilities of the device. I think that negotiation with such devices are very questionnable. * ncr53c8xx.c ncrBsd2Linux 1.1 does not negotiate with scsi-1 devices. Sat Jan 06 21:00 1996 Gerard Roudier (groudier@club-internet.fr) * ncr53c8xx.c function ncr_attach() Disable chip interrupt before soft reset in attach procedure. When loadlin is used to boot the system, the state of the NCR chip is unpredicable. This modification avoid unexpected interrupts. Thu Jan 04 23:45 1996 Gerard Roudier (groudier@club-internet.fr) ncrBsd2Linux 1.0 is ready. Upload to sunsite immediatly. Tue Jan 02 23:00 1996 Gerard Roudier (groudier@club-internet.fr) * ncr53c8xx.c Add a waiting list for Scsi Commands than can be inserted into the start queue immediatly. When a command complete, the waiting commands are requeued by calling the queuecommand() function. Sun Dec 31 23:59 1995 Gerard Roudier (groudier@club-internet.fr) * ncr53c8xx.c Use kmalloc()/kfree() for internal data structures allocation to avoid scsi memory pool shortage. Sat Dec 30 23:00 1995 Gerard Roudier (groudier@club-internet.fr) ncrBsd2Linux can now use memory mapped IO. Works fine. * ncr53c8xx.c Call vremap() from ncr_attach() to map the physical page which contains the memory IO window. Call vfree() from ncr_detach() (release module). Fri Dec 29 23:45 1995 Gerard Roudier (groudier@club-internet.fr) ncrBsd2Linux can now be configured as a module. Works fine. * ncr53c8xx.c: add new functions ncr53c8xx_release() and ncr_detach() Add the code to (per host): Stop the timer. Stop the chip. Free allocated memory. Fri Dec 29 23:00 1995 Gerard Roudier (groudier@club-internet.fr) Problem: detection routine returns 0 and can detect only one host. * ncr53c8xx.c function ncr_attach() ncr_attach() now returns 0 on success and -1 on error. ncr53c8xx_detect() returns the number of detected hosts. Thu Dec 28 22:00 1995 Gerard Roudier (groudier@club-internet.fr) I must upload the new version which corrects the severe problem with WRITE_10 command. Release 0.5 Known or probable problems with this ncr driver release: -------------------------------------------------------- Same as the previous release. Wed Dec 27 23:00 1995 Gerard Roudier (groudier@club-internet.fr) Problem: System CRASH or scsi ERROR "extra data disgarded" on WRITE(10) command. * ncr53c8xx.c function guess_xfer_direction() I add the WRITE(10) (0x2A) into the list of Data Out scsi commands. It was a big mistake. This BUG was introduced in the release 0.3 and is obvious present in the release 0.4. Wed Dec 27 22:00 1995 Gerard Roudier (groudier@club-internet.fr) Problem: When I was testing tagged command queueing and disconnections with one hard disk at a time (IBM S12), the script process hung every 5 minutes with a non empty stall queue. * ncr53c8xx.c function ncr_exception() I replace "OUTB (nc_istat, INTF)" by "OUTB (nc_istat, (istat & SIGP) | INTF)". This statement cleared the INTF condition, but cleared the SIGP flag too. (This bug is in the original FreeBSD ncr driver). Mon Dec 25 22:00 1995 Gerard Roudier (groudier@club-internet.fr) Release 0.4 Known or probable problems with this ncr driver release: -------------------------------------------------------- Hardware (or software) conflicts with some ethernet cards. See release 0.2 above. Crash with Intel saturn chipset with write-back cache enabled. The SCSI SCRIPT access the internal registers of the NCR chip by memory addressing. Ensure that the memory area of the NCR chip is not cacheable. Use scanpci to get the base memory address of the ncr chip. The 128 bytes following this address must not be cached. Sat Dec 23 22:00 1995 Gerard Roudier (groudier@club-internet.fr) Problem: FreeBSD driver important comments --------------------------------- We try to reduce the number of interrupts caused by unexpected phase changes due to disconnects. A typical harddisk may disconnect before ANY block. If we wanted to avoid unexpected phase changes at all we had to use a break point every 512 bytes. Of course the number of scatter/gather blocks is limited. * ncr53c8xx.c function ncr_scatter() This function has been rewritten according to the above comments. The Linux scatter list is scanned, and blocks are broken as possible into 512 bytes chunks. Wed Dec 22 22:00 1995 Gerard Roudier (groudier@club-internet.fr) Problem: ensure that there are enough allocated Command Control Blocks for each unit to enqueue commands according to cmd_per_lun. * ncr53c8xx.c function ncr_allocate_ccb() Allocate all required ccb(s) instead of one at a time. Mon Dec 18 23:00 1995 Gerard Roudier (groudier@club-internet.fr) Problem: A copy of the system time is used to compute timeouts. When the system time is changed , we can get spurious timeouts if scsi commands are pending. * ncr53c8xx.c function ncr_timeout() In FreeBSD the kernel time (volatile struct timeval time) is not affected by settimeofday() or other change time functions. For Linux, I replace "time" by "xtime". But "xtime" is the system time, and is affected by change time functions. If we detect a change <=-1s or >1s we assume system time has been changed. For all active ccb(s), we recompute tlimit. We set heartbeat to thistime to prevent spurious chip reset. Sun Dec 17 23:00 1995 Gerard Roudier (groudier@club-internet.fr) Release 0.3. Sun Dec 17 11:00 1995 Gerard Roudier (groudier@club-internet.fr) Problem: Linux middle-level scsi driver does not provide the direction of transfert in scsi command parameters. FreeBSD ncr driver need this information to patch the SCSI script for SAVE DATA POINTER and to check actual data transfer direction. * ncr53c8xx.c I add the function guess_xfer_direction(int opcode) which try to guess the transfer direction. Unfortunately my documentation about SCSI-II standard is very old. It does not contain PHOTO-CD command specifications. I assume input transfer direction, for unknown command. (not perfect) Wed Dec 15 23:00 1995 Gerard Roudier (groudier@club-internet.fr) It's time to schedule the release 0.2 Known or probable problems with this ncr driver release: -------------------------------------------------------- Scsi tapes do not work. scsi-config-1.5 does not work too. Hardware (or software) conflicts with some ethernet cards. The linux native ncr53c810 scsi driver does'not use memory mapped IO at all. The BSD driver can use memory mapped IO. Under Linux, i seems to be difficult (or impossible) to map a PCI memory area. So I decide to use normal IO in the code of the ported driver. However, the SCSI SCRIPT access the internal registers of the NCR chip by memory addressing. EThernet cards use a memory area to communicate with the system. I think that this memory area conflicts with the memory area used by the NCR chip. The configuration diskette of your ethernet card can help you to fix the problem (try sofware configurations). Wed Dec 15 22:00 1995 Gerard Roudier (groudier@club-internet.fr) Problem: detection of CD change did not work. * ncr53c8xx.c Signal a CHECK_CONDITION (S_CHECK_COND) to the middle-level scsi drivers when the scsi completion status = S_SENSE|S_GOOD. Problem: System hang with status <4/82> after mounting the root partition. System disk is a QUANTUM EMPIRE 1080 S. Submitted by rwilhelm@Physik.TU-Muenchen.DE (Robert Wilhelm) Thu, 14 Dec 1995 10:18:43 +0100 (MET) * ncr53c8xx.c Signal a CHECK_CONDITION (S_CHECK_COND) to the middle-level scsi drivers when the scsi completion status = S_SENSE|S_CHECK_COND. (HS_COMPLETE=4, S_SENSE|S_CHECK_COND=x82). I can't test this fix. I think that it should work. Thu Dec 14 22:00 1995 Gerard Roudier (groudier@club-internet.fr) Problem submitted by Bill Dyess Tue 12 Dec 1995 Only one lun detected on a Pioneer DRM-602X 6-disk CD-ROM changer. * ncr53c8xx.h, ncr53c8xx.c Define NCR_SCSI_MAX_LUN to 8 if the Linux configuration flag SCSI_CONFIG_MULTI_LUN is set. My CD device has only one lun. I can't test multi-lun device, but I think that it should work. * ncr53c8xx.c Write the ncr_abort_command() and ncr_reset_command() functions. I intend to test these functions next week. Sun Dec 10 22:00 1995 Gerard Roudier (groudier@club-internet.fr) * Release 0.1 Works fine with hard disks under Linux 1.2.13 and Linux 1.3.45.