Re: [PATCH v2] drivers: spi: spi.c: Convert statistics to per-cpu u64_stats_t

[David Jander <david@xxxxxxxxxxx>]

On Thu, 9 Jun 2022 11:32:51 +0200
Marc Kleine-Budde <mkl@xxxxxxxxxxxxxx> wrote:

> On 24.05.2022 11:18:08, David Jander wrote:
> > This change gives a dramatic performance improvement in the hot path,
> > since many costly spin_lock_irqsave() calls can be avoided.
> >
> > On an i.MX8MM system with a MCP2518FD CAN controller connected via SPI,
> > the time the driver takes to handle interrupts, or in other words the time
> > the IRQ line of the CAN controller stays low is mainly dominated by the
> > time it takes to do 3 relatively short sync SPI transfers. The effect of
> > this patch is a reduction of this time from 136us down to only 98us.
> >
> > Suggested-by: Andrew Lunn <andrew@xxxxxxx>
> > Signed-off-by: David Jander <david@xxxxxxxxxxx>  
> 
> This patch (cherry picked on top of v5.19-rc1) explodes on 32-bit SMP
> ARMv7 (imx6q) with:
> 
> | [    0.397493] BUG: using smp_processor_id() in preemptible [00000000] code: swapper/0/1
> | [    0.397514] caller is debug_smp_processor_id+0x18/0x24
> | [    0.397544] CPU: 1 PID: 1 Comm: swapper/0 Not tainted 5.19.0-rc1-00001-g6ae0aec8a366 #181
> | [    0.397559] Hardware name: Freescale i.MX6 Quad/DualLite (Device Tree)
> | [    0.397566] Backtrace:
> | [    0.397576]  dump_backtrace from show_stack+0x20/0x24
> | [    0.397616]  r7:81024ffd r6:00000000 r5:81024ffd r4:60000013
> | [    0.397621]  show_stack from dump_stack_lvl+0x60/0x78
> | [    0.397644]  dump_stack_lvl from dump_stack+0x14/0x1c
> | [    0.397664]  r7:81024ffd r6:80f652de r5:80bec180 r4:819a2500
> | [    0.397669]  dump_stack from check_preemption_disabled+0xc8/0xf0
> | [    0.397690]  check_preemption_disabled from debug_smp_processor_id+0x18/0x24
> | [    0.397714]  r8:8119b7e0 r7:81205534 r6:819f5c00 r5:819f4c00 r4:c083d724
> | [    0.397719]  debug_smp_processor_id from __spi_sync+0x78/0x220
> | [    0.397746]  __spi_sync from spi_sync+0x34/0x4c
> | [    0.397772]  r10:bb7bf4e0 r9:c083d724 r8:00000007 r7:81a068c0 r6:822a83c0 r5:c083d724
> | [    0.397779]  r4:819f4c00
> | [    0.397784]  spi_sync from spi_mem_exec_op+0x338/0x370
> | [    0.397810]  r5:000000b4 r4:c083d910
> | [    0.397815]  spi_mem_exec_op from spi_nor_read_id+0x98/0xdc
> | [    0.397846]  r10:bb7bf4e0 r9:00000000 r8:00000000 r7:00000000 r6:00000000 r5:82358040
> | [    0.397852]  r4:819f7c40
> | [    0.397856]  spi_nor_read_id from spi_nor_detect+0x38/0x114
> | [    0.397878]  r7:82358040 r6:00000000 r5:819f7c40 r4:819f7c40
> | [    0.397883]  spi_nor_detect from spi_nor_scan+0x11c/0xbec
> | [    0.397910]  r10:bb7bf4e0 r9:00000000 r8:00000000 r7:c083da4c r6:00000000 r5:00010101
> | [    0.397916]  r4:819f7c40
> | [    0.397921]  spi_nor_scan from spi_nor_probe+0x10c/0x2d0
> | [    0.397946]  r10:bb7bf4e0 r9:bb7bf4d0 r8:00000000 r7:819f4c00 r6:00000000 r5:00000000
> | [    0.397952]  r4:819f7c40

*ouch*. Thanks for reporting. Fix coming up asap... apparently I forgot to
get_cpu/put_cpu around the per-cpu operations :-(

Best regards,

-- 
David Jander