Potential Regression in 3.6-rc1 - Kirkwood SATA

Him

On Tue, Aug 14, 2012 at 07:59:00PM -0400, Josh Coombs wrote:

I finally got a chance to test 3.6-rc1 after having my GoFlex Net
support patch accepted for the 3.6 release train.  Included in 3.6-rc1
was an update for Kirkwoods switching SATA to DT which was not part of
my original testing.  It seems something with this change has
partially broken the GoFlex and a couple other Kirkwood based devices.

The key factor is the number of SATA ports defined in the dts:

		sata at 80000 {
			status = "okay";
			nr-ports = <2>;
		};

If set at the correct number for my device, 2, my GFN does not
complete kernel init, hanging here:

<SNIP>
[   15.287832] Dquot-cache hash table entries: 1024 (order 0, 4096 bytes)
[   15.296545] jffs2: version 2.2. (NAND) ?? 2001-2006 Red Hat, Inc.
[   15.303202] msgmni has been set to 240
[   15.308503] Block layer SCSI generic (bsg) driver version 0.4 loaded (major )
[   15.316021] io scheduler noop registered
[   15.320149] io scheduler deadline registered
[   15.324558] io scheduler cfq registered (default)
[   15.329462] mv_xor_shared mv_xor_shared.0: Marvell shared XOR driver
[   15.335962] mv_xor_shared mv_xor_shared.1: Marvell shared XOR driver
[   15.376751] mv_xor mv_xor.0: Marvell XOR: ( xor cpy )
[   15.416736] mv_xor mv_xor.1: Marvell XOR: ( xor fill cpy )
[   15.456735] mv_xor mv_xor.2: Marvell XOR: ( xor cpy )
[   15.496734] mv_xor mv_xor.3: Marvell XOR: ( xor fill cpy )
[   15.506309] Serial: 8250/16550 driver, 2 ports, IRQ sharing disabled
[   15.509111] serial8250.0: ttyS0 at MMIO 0xf1012000 (irq = 33) is a 16550A
[   15.509141] console [ttyS0] enabled, bootconsole disabled
[   15.518967] brd: module loaded
[   15.524991] loop: module loaded
[   15.528584] sata_mv sata_mv.0: cannot get optional clkdev
[   15.534180] sata_mv sata_mv.0: slots 32 ports 2

If you set nr-ports to 1 the unit boots cleanly, save for only
detecting one functional SATA port.  Another user has confirmed this
behavior on an Iomega IX2-200.

Try booting with "coherent_pool=1M" (or bigger) kernel parameter. I think
with the recent DMA mapping changes, the default 256 KB coherent pool may
be too small, mv_xor and sata_mv together needs more. (I'm not sure how
it actually worked before commit e9da6e9905e639b0f842a244bc770b48ad0523e9,
but it seems this is the cuase).

It should be noted that dma_pool_alloc() uses GFP_ATOMIC always, and if
drivers exhaust the coherent pool already during the boot that function
just keeps looping forever. Users only see boot hanging with no clue
what to do. I would say it's quite a poor error handling...

A.