[PATCH] siimage.c - turning DMA on because of 'md' kernel thread.

[PATCH] siimage.c - turning DMA on because of 'md' kernel thread.

[Kresimir Kukulj]

Hi

	I'm trying to use Sil3112a controller with two Seagate 120Gb SATA
disks for RAID-1 (mirror) with 'md - multiple devices' driver. I am aware
that I need to use some tricks to get it working, like this:

 # hdparm -d1 -X69 /dev/xxx
 # echo "max_kb_per_request:15" > /proc/ide/hd?/settings

First one prevents freezing the system when enabling DMA.
Second one stops IO errors.

As far as I can tell, it works ok, but with reduced bandwidth (because of
max_kb_per_request). When using 'md' driver for RAID-1 mirror (rootfs, swap)
with persistent superblocks I have problems. If server crashes (for some
reason), 'md' tries to resync mirrors automatically. That is done by a kernel
thread that is activated before init(8) is started. I have put 'hdparm' cludge
very early in the boot process, but that happens _after_ 'md' thread starts to
resync.  That means that disks are busy (in PIO mode), and when hdparm -d1 -X69
executes, system freezes [if there is no disk/little activity hdparm cludge
passes ok - for example, if RAID-1 is clean so there is no resync].

For past two days I tried to find a way to force ide driver to initialize disk
in dma mode automatically by the kernel itself although disks are set to PIO by
the bios for some reason. I devised a patch that does hdparm + echo cludge. I
have never seen linux kernel, so I am unsure if I did it correctly. Patch is
attached.  Please, if anyone uses it, test it to see if  it works ok for you! I
am not familiar with linux ide driver so I did this with a lot of printk's and
trial & error.

With this patch, disks are initialized by kernel to UDMA100, with
max_kb_per_request:15.  Now, RAID-1 is resynced at boot with DMA already
activated (so there is no need for hdparm) and it completed successfully.
At least it works for me. I need to do more disk stress tests to be sure it is
stable.

This is all done with vanilla 2.4.21.
2.4.22-pre4 didn't work at all. Copying a few MB of data freezes the kernel.

My motherboard is Asus P4G8X deluxe.
http://www.asus.com/products/mb/socket478/p4g8x-d/overview.htm

--- siimage.c.orig	Fri Jul 18 23:36:58 2003
+++ siimage.c	Fri Jul 18 23:36:59 2003
@@ -97,7 +97,7 @@
 
 	switch(hwif->pci_dev->device) {
 		case PCI_DEVICE_ID_SII_3112:
-			return 4;
+			return 3;
 		case PCI_DEVICE_ID_SII_680:
 			if ((scsc & 0x30) == 0x10)	/* 133 */
 				mode = 4;
@@ -297,7 +297,7 @@
 	struct hd_driveid *id	= drive->id;
 
 	if ((id->capability & 1) != 0 && drive->autodma) {
-		if (!(hwif->atapi_dma))
+		if ((hwif->atapi_dma))
 			goto fast_ata_pio;
 		/* Consult the list of known "bad" drives */
 		if (hwif->ide_dma_bad_drive(drive))
@@ -803,7 +803,7 @@
 	pci_read_config_dword(dev, PCI_CLASS_REVISION, &class_rev);
 	class_rev &= 0xff;
 
-	hwif->rqsize = 128;
+	hwif->rqsize = 15;
 	if ((dev->device == PCI_DEVICE_ID_SII_3112) && (!(class_rev)))
 		hwif->rqsize = 16;
 
-----------------------

I'm confused why is this check there:
 	if ((id->capability & 1) != 0 && drive->autodma) {
 		if (!(hwif->atapi_dma))
 			goto fast_ata_pio;

Is it intentional ?

-- 
Kresimir Kukulj                      madmax@iskon.hr
+--------------------------------------------------+
Old PC's never die. They just become Unix terminals.

Re: [PATCH] siimage.c - turning DMA on because of 'md' kernel thread.

[Alan Cox]

On Gwe, 2003-07-18 at 23:34, Kresimir Kukulj wrote:
> resync.  That means that disks are busy (in PIO mode), and when hdparm -d1 -X69
> executes, system freezes [if there is no disk/little activity hdparm cludge
> passes ok - for example, if RAID-1 is clean so there is no resync].

The newest driver should put them into DMA automatically and set the
limit. The other changes I'll check over but certainly look like they
may be needed.

Re: [PATCH] siimage.c - turning DMA on because of 'md' kernel thread.

[Kresimir Kukulj]

Quoting Alan Cox (alan@lxorguk.ukuu.org.uk):
> On Gwe, 2003-07-18 at 23:34, Kresimir Kukulj wrote:
> > resync.  That means that disks are busy (in PIO mode), and when hdparm -d1 -X69
> > executes, system freezes [if there is no disk/little activity hdparm cludge
> > passes ok - for example, if RAID-1 is clean so there is no resync].
> 
> The newest driver should put them into DMA automatically and set the
> limit. The other changes I'll check over but certainly look like they
> may be needed.

Hm, I tried 2.4.22-pre7 and it did not set disks to DMA mode, but it did set
the limit (max_kb_per_request). With the following modification, kernel
(2.4.22-pre7) initializes both disk to UDMA133, and as far as I can see, it
works ok.

In siimage.c there is a check for Maxtor disks - they are/should be set to
UDMA100.  Is this also necessary for Seagate SATA disks ?

--------------
--- siimage.c.orig      Mon Jul 21 15:13:16 2003
+++ siimage.c   Mon Jul 21 16:56:28 2003
@@ -488,7 +488,7 @@
        struct hd_driveid *id   = drive->id;
 
        if ((id->capability & 1) != 0 && drive->autodma) {
-               if (!(hwif->atapi_dma))
+               if ((hwif->atapi_dma))
                        goto fast_ata_pio;
                /* Consult the list of known "bad" drives */
                if (hwif->ide_dma_bad_drive(drive))
---------------


And it looks like:

SiI3112 Serial ATA: IDE controller at PCI slot 02:04.0
PCI: Found IRQ 5 for device 02:04.0
PCI: Sharing IRQ 5 with 02:00.0
SiI3112 Serial ATA: chipset revision 2
SiI3112 Serial ATA: not 100% native mode: will probe irqs later
    ide2: MMIO-DMA , BIOS settings: hde:pio, hdf:pio
    ide3: MMIO-DMA , BIOS settings: hdg:pio, hdh:pio
hda: SAMSUNG CD-ROM SC-152A, ATAPI CD/DVD-ROM drive
hde: ST3120023AS, ATA DISK drive
blk: queue c0326ef8, I/O limit 4095Mb (mask 0xffffffff)
hdg: ST3120023AS, ATA DISK drive
blk: queue c0327364, I/O limit 4095Mb (mask 0xffffffff)
ide0 at 0x1f0-0x1f7,0x3f6 on irq 14
ide2 at 0xf8800080-0xf8800087,0xf880008a on irq 5
ide3 at 0xf88000c0-0xf88000c7,0xf88000ca on irq 5
hde: attached ide-disk driver.
hde: host protected area => 1
hde: 234441648 sectors (120034 MB) w/8192KiB Cache, CHS=232581/16/63, UDMA(133)
hdg: attached ide-disk driver.
hdg: host protected area => 1
hdg: 234441648 sectors (120034 MB) w/8192KiB Cache, CHS=232581/16/63, UDMA(133)


/dev/hdg:
 multcount    = 16 (on)
 I/O support  =  0 (default 16-bit)
 unmaskirq    =  0 (off)
 using_dma    =  1 (on)
 keepsettings =  0 (off)
 nowerr       =  0 (off)
 readonly     =  0 (off)
 readahead    =  8 (on)
 geometry     = 14593/255/63, sectors = 234441648, start = 0
 busstate     =  1 (on)


Capabilities:
        LBA, IORDY(can be disabled)
        Buffer size: 8192.0kB   ECC bytes: 4    Queue depth: 1
        Standby timer values: spec'd by standard
        r/w multiple sector transfer: Max = 16  Current = 16
        Advanced power management level: 65278
        DMA: mdma0 mdma1 mdma2 udma0 udma1 udma2 udma3 udma4 udma5 *udma6 
             Cycle time: min=120ns recommended=120ns
        PIO: pio0 pio1 pio2 pio3 pio4 
             Cycle time: no flow control=240ns  IORDY flow control=120ns


[Mirror]
# hdparm -tT /dev/md3

/dev/md3:
 Timing buffer-cache reads:   128 MB in  0.24 seconds =533.33 MB/sec
 Timing buffered disk reads:  64 MB in  2.42 seconds = 26.45 MB/sec

[One plex]
# hdparm -tT /dev/hde3

/dev/hde3:
 Timing buffer-cache reads:   128 MB in  0.24 seconds =533.33 MB/sec
 Timing buffered disk reads:  64 MB in  1.82 seconds = 35.16 MB/sec

Poor results, but at least it works.

-- 
Kresimir Kukulj                      madmax@iskon.hr
+--------------------------------------------------+
Old PC's never die. They just become Unix terminals.