spi/spi-pl022: Query on working of DMA mode

Hi Linus,

I am preparing few patches for spi-pl022 for "Allocating DMA channels as and when required". And i found something strange.

When we have many transfers in a single message, following routines are called in specified order:

tasklet_schedule(pump_transfers) pump_transfers(unsigned long data) configure_dma(pl022) sg_alloc_table(&pl022->sgt_rx, pages, GFP_KERNEL);

Here, allocation is requested with GFP_KERNEL flag, from a tasklet. Which gives following crash logs:

BUG: sleeping function called from invalid context at /data/csd_sw/spear/drives_os/vireshk/spear/kernel/linux-2.6/mm/slub.c:7
93
in_atomic(): 1, irqs_disabled(): 0, pid: 11, name: kworker/u:1
Backtrace: 
[<8003dabc>] (dump_backtrace+0x0/0x10c) from [<803aa4ec>] (dump_stack+0x18/0x1c)
 r6:00000001 r5:000000d0 r4:bf402200 r3:60000113
[<803aa4d4>] (dump_stack+0x0/0x1c) from [<8004f330>] (__might_sleep+0xec/0x10c)
[<8004f244>] (__might_sleep+0x0/0x10c) from [<800bbd2c>] (__kmalloc+0x74/0x110)
[<800bbcb8>] (__kmalloc+0x0/0x110) from [<8019f69c>] (sg_kmalloc+0x2c/0x30)
 r8:80541b40 r7:00000001 r6:00000001 r5:00000100 r4:bf47f1e8

r3:00000000
[<8019f670>] (sg_kmalloc+0x0/0x30) from [<8019f710>] (__sg_alloc_table+0x70/0x118) [<8019f6a0>] (__sg_alloc_table+0x0/0x118) from [<8019f7e0>] (sg_alloc_table+0x28/0x54) [<8019f7b8>] (sg_alloc_table+0x0/0x54) from [<802649a0>] (configure_dma+0x230/0x3fc)
r5:bf47def8 r4:bf47f150
[<80264770>] (configure_dma+0x0/0x3fc) from [<8026509c>] (pump_transfers+0xcc/0x100) [<80264fd0>] (pump_transfers+0x0/0x100) from [<8005ecb0>] (tasklet_action+0xb0/0x160)
r7:805203c4 r6:bf488000 r5:bf47f1a0 r4:bf47f19c
[<8005ec00>] (tasklet_action+0x0/0x160) from [<8005f104>] (__do_softirq+0xa4/0x13c) [<8005f060>] (__do_softirq+0x0/0x13c) from [<8005f590>] (irq_exit+0x4c/0x54) [<8005f544>] (irq_exit+0x0/0x54) from [<80034094>] (asm_do_IRQ+0x94/0xd0) [<80034000>] (asm_do_IRQ+0x0/0xd0) from [<800396f4>] (__irq_svc+0x34/0xc0)

Exception stack(0xbf489e90 to 0xbf489ed8)

Now, i didn't get them earlier due to a mistake of mine. I have updated pl022_ssp_controller.enable_dma = 1,
but didn't do pl022_config_chip.com_mode = DMA_TRANSFER

So, finally DMA channels are allocated but never used, as both above must have been set.
Today only I came to know of this mistake, and found it is actually not working, with above crash occurring.

Is DMA mode working in your case?

viresh

What happens if you replace it with GFP_NOWAIT?

That should work for everybody.

Or you could remove the tasklet allover and replace it with a workqueue.

Yes, magically :-/

Yours,
Linus Walleij

Ok. I will fix this and include in my patchset.

But how? Don't you have any message with more than one transfer in it?

viresh

No maybe not, my test cases are for single messages, and on production systems we mainly use IRQ-driven traffic. I need to have these tests added I believe :-/

If you fix this up it will be much appreciated!

Thanks,
Linus Walleij

Ya I am fixing this up. I have resolved this yesterday only, with GFP_NOWAIT. But it looks there are few more issues. Let me fix everything first and then will send patchset with updates.

viresh