Re: 2.6.9-rc1-mm3

[Peter Osterlund]

Andrew Morton <akpm@xxxxxxxx> writes:

> Peter Osterlund <petero2@xxxxxxxxx> wrote:
> >
> > One problem that does remain though, is that when dumping huge amounts
> >  of data to a CD or DVD disc (so that you get memory pressure), the
> >  effective writing speed of other block devices (like IDE hard disks)
> >  is reduced to the same speed as the packet device.
> > 
> >  I have posted a patch that fixes this problem by limiting the amount
> >  of writeback data in the packet driver, but unfortunately it makes the
> >  effective writing speed of the packet device suffer a lot. The proper
> >  fix is probably to improve the filesystem and/or VM code to start I/O
> >  operations in sequential order a lot more often than it currently
> >  does.
> 
> If you decrease /proc/sys/vm/dirty_ratio and dirty_background_ratio to much
> smaller levels, does that fix things up? 

No, it seems to slow the packet device down without improving the hard
disk speed.

I wrote a patch that logs what I/O operations are received by the
packet driver. For my test case, which does

        time cp -av /testdata /udf ; time sync

where /testdata contains 120MB of data in 29 files, the I/O pattern
looks like this:

    10:06:03   421 - 1905
    10:06:03   1906 - 2482
    10:06:03   4255 - 6031
    10:06:03   6032 - 6320
    10:06:03   8514 - 10277
    10:06:03   10278 - 10566
    10:06:03   13773 - 15821
    10:06:03   16520 - 18568
    10:06:03   19021 - 21068
    10:06:03   21069 - 23117
    10:06:03   23208 - 24869
    10:06:03   24870 - 25286
    10:06:03   27120 - 28714
    10:06:03   28715 - 29195
    10:06:03   30307 - 32355
    10:06:03   32732 - 34704
    10:06:03   34705 - 34801
    10:06:03   36905 - 38953
    10:06:03   39390 - 41283
    10:06:03   41284 - 41444
    10:06:03   43663 - 45008
    10:06:03   45009 - 45713
    10:06:03   46719 - 48678
    10:06:03   48679 - 48775
    10:06:03   51004 - 53052
    10:06:03   53497 - 55545
    10:06:03   2482 - 4254
    10:06:03   55581 - 55869
    10:06:03   6320 - 8368
    10:06:03   57527 - 59529
    10:06:03   10566 - 10630
    10:06:03   59530 - 61578
    10:06:03   15821 - 16466
    10:06:03   16474 - 16519
    10:06:03   18568 - 19020
    10:06:03   23117 - 23207
    10:06:03   25286 - 26118
    10:06:03   29195 - 30306
    10:06:03   32355 - 32731
    10:06:03   34801 - 35377
    10:06:03   38953 - 39389
    10:06:03   41444 - 43076
    10:06:03   45713 - 46718
    10:06:03   48775 - 49831
    10:06:03   53052 - 53496
    10:06:03   55545 - 55579
    10:06:03   55869 - 57469
    10:06:03   8368 - 8513
    10:06:03   10630 - 12550
    10:06:03   61578 - 62086
    10:06:03   26118 - 27119
    10:06:03   35377 - 35921
    10:06:03   43076 - 43662
    10:06:04   49831 - 51003
    10:06:04   57469 - 57526
    10:06:04   12550 - 12806
    10:06:04   35921 - 36904
    10:06:04   12806 - 13772
    10:06:04   273 - 278
    10:06:04   417 - 418
    10:06:04   419 - 421
    10:06:04   1905 - 1906
    10:06:04   4254 - 4255
    10:06:04   6031 - 6032
    10:06:04   8513 - 8514
    10:06:04   10277 - 10278
    10:06:04   13772 - 13773
    10:06:04   16519 - 16520
    10:06:04   19020 - 19021
    10:06:04   21068 - 21069
    10:06:04   23207 - 23208
    10:06:04   24869 - 24870
    10:06:04   27119 - 27120
    10:06:04   28714 - 28715
    10:06:04   30306 - 30307
    10:06:04   32731 - 32732
    10:06:04   34704 - 34705
    10:06:04   36904 - 36905
    10:06:04   39389 - 39390
    10:06:04   41283 - 41284
    10:06:04   43662 - 43663
    10:06:04   45008 - 45009
    10:06:04   46718 - 46719
    10:06:04   48678 - 48679
    10:06:04   51003 - 51004
    10:06:04   53496 - 53497
    10:06:04   55579 - 55581
    10:06:04   57526 - 57527
    10:06:31   59529 - 59530
    10:06:31   273 - 274

See attachment for a graph of this data. The X axis corresponds to the
I/O submission order and the Y axis corresponds to the sector numbers.

Attachment: iopattern.png
Description: IO_pattern.png

diff -puN drivers/block/pktcdvd.c~packet-io-log drivers/block/pktcdvd.c
--- linux/drivers/block/pktcdvd.c~packet-io-log	2004-09-05 10:35:24.697900200 +0200
+++ linux-petero/drivers/block/pktcdvd.c	2004-09-05 10:35:24.701899592 +0200
@@ -896,6 +896,27 @@ static inline void pkt_set_state(struct 
 	pkt->state = state;
 }
 
+static void pkt_io_log(struct bio *bio)
+{
+	static sector_t start, stop;
+	typedef unsigned long long ull;
+
+	if (bio) {
+		if (stop == bio->bi_sector) {
+			stop = stop + bio_sectors(bio);
+		} else {
+			if (start != stop)
+				printk("pkt: %lld - %lld\n", (ull)(start >> 2), (ull)(stop >> 2));
+			start = bio->bi_sector;
+			stop = start + bio_sectors(bio);
+		}
+	} else {
+		if (start != stop)
+			printk("pkt: %lld - %lld\n", (ull)(start >> 2), (ull)(stop >> 2));
+		start = stop = 0;
+	}
+}
+
 /*
  * Scan the work queue to see if we can start a new packet.
  * returns non-zero if any work was done.
@@ -949,6 +970,8 @@ try_next_bio:
 	spin_unlock(&pd->lock);
 	if (!bio) {
 		VPRINTK("handle_queue: no bio\n");
+		if (list_empty(&pd->cdrw.pkt_active_list))
+			pkt_io_log(NULL);
 		return 0;
 	}
 
@@ -2197,6 +2220,8 @@ static int pkt_make_request(request_queu
 		}
 	}
 
+	pkt_io_log(bio);
+
 	/*
 	 * If we find a matching packet in state WAITING or READ_WAIT, we can
 	 * just append this bio to that packet.
_

-- 
Peter Osterlund - petero2@xxxxxxxxx
http://w1.894.telia.com/~u89404340