Skip to contentSkip to conversation

Consider moving the init of pl330 to earlier?

DISCUSSION Ray Jui 2 participants 3 messages Oct 9, 2014 – Oct 16, 2014

lore
Share & tools
lore

Ray Jui

author
#0 lore raw

Hi,

How do you guys like the idea of moving the driver initialization of the pl330 DMA driver to earlier? For example, to arch_initcall_sync or subsys_initcall? Currently the pl330 driver is registered through module_amba_driver call, which translates to device_initcall in the end. This is a bit late considering in many systems, DMA controller is one of the core components that may have many slave devices depending on. Most slave drivers are typically done at device_initcall. When done at the same level, DMA may not be ready while the slave devices are being initialized. In addition, under the drivers/dma directory, we currently have various other DMA drivers also done at subsys_initcall.

One issue that I'm seeing with the Broadcom Cygnus SoC is that, when we try to use DMA with SPI (PL022), the request for DMA channel would fail because the spi-pl022 driver we use is initialized at subsys_initcall. At the time the PL330 driver is not yet initialized. The reason why spi-pl022 is done at subsys_initcall is that some regulators communicate through the SPI bus so people decided to move it to earlier (commit 25c8e03b by Linus Walleij), which seems to be a valid reason to me.

Thanks,

Ray

#1 lore raw
Quoted 14 lines · Ray Jui

Hi,

How do you guys like the idea of moving the driver initialization of the pl330 DMA driver to earlier? For example, to arch_initcall_sync or subsys_initcall? Currently the pl330 driver is registered through module_amba_driver call, which translates to device_initcall in the end. This is a bit late considering in many systems, DMA controller is one of the core components that may have many slave devices depending on. Most slave drivers are typically done at device_initcall. When done at the same level, DMA may not be ready while the slave devices are being initialized. In addition, under the drivers/dma directory, we currently have various other DMA drivers also done at subsys_initcall.

Quoted 8 lines

One issue that I'm seeing with the Broadcom Cygnus SoC is that, when we try to use DMA with SPI (PL022), the request for DMA channel would fail because the spi-pl022 driver we use is initialized at subsys_initcall. At the time the PL330 driver is not yet initialized. The reason why spi-pl022 is done at subsys_initcall is that some regulators communicate through the SPI bus so people decided to move it to earlier (commit 25c8e03b by Linus Walleij), which seems to be a valid reason to me.

As a subsytem we encourage folks to use early init calls and then have clients use late_init so that dependency is solved.

So feel free to send the patch

Signature · 1 line
~Vinod
author
#2 lore raw
Quoted 23 lines · Vinod Koul
Quoted 14 lines · Ray Jui

Hi,

How do you guys like the idea of moving the driver initialization of the pl330 DMA driver to earlier? For example, to arch_initcall_sync or subsys_initcall? Currently the pl330 driver is registered through module_amba_driver call, which translates to device_initcall in the end. This is a bit late considering in many systems, DMA controller is one of the core components that may have many slave devices depending on. Most slave drivers are typically done at device_initcall. When done at the same level, DMA may not be ready while the slave devices are being initialized. In addition, under the drivers/dma directory, we currently have various other DMA drivers also done at subsys_initcall.

Quoted 8 lines

One issue that I'm seeing with the Broadcom Cygnus SoC is that, when we try to use DMA with SPI (PL022), the request for DMA channel would fail because the spi-pl022 driver we use is initialized at subsys_initcall. At the time the PL330 driver is not yet initialized. The reason why spi-pl022 is done at subsys_initcall is that some regulators communicate through the SPI bus so people decided to move it to earlier (commit 25c8e03b by Linus Walleij), which seems to be a valid reason to me.

Quoted 5 lines

As a subsytem we encourage folks to use early init calls and then have clients use late_init so that dependency is solved.

So feel free to send the patch

Will do. Thanks!

Ray