Re: [PATCH v2] dec: tulip: de4x5: Replace mdelay with usleep_range in de4x5_hw_init
On 2018/4/12 10:21, arvindY wrote: On Thursday 12 April 2018 07:00 AM, Jia-Ju Bai wrote: On 2018/4/12 0:16, James Bottomley wrote: On Wed, 2018-04-11 at 23:39 +0800, Jia-Ju Bai wrote: de4x5_hw_init() is never called in atomic context. de4x5_hw_init() is only called by de4x5_pci_probe(), which is only set as ".probe" in struct pci_driver. Despite never getting called from atomic context, de4x5_hw_init() calls mdelay() to busily wait. This is not necessary and can be replaced with usleep_range() to avoid busy waiting. This is found by a static analysis tool named DCNS written by myself. And I also manually check it. Did you actually test this? The usual reason for wanting m/udelay is that the timing must be exact. The driver is filled with mdelay()s for this reason. The one you've picked on is in the init path so it won't affect the runtime in any way. I also don't think we have the hrtimer machinery for usleep_range() to work properly on parisc, so I don't think the replacement works. James Hello, James. Thanks for your reply :) I agree that usleep_range() here will not much affect the real execution of this driver. But I think usleep_range() can more opportunity for other threads to use the CPU core to schedule during waiting. That is why I detect mdelay() that can be replaced with msleep() or usleep_range(). James is right, You have added all usleep_range() during system boot-up time. During boot-up system will run as single threaded. Where this change will not make much sense. System first priority is match the exact timing on each and every boot-up. Hello, Arvind. Thanks for your reply :) I admit I am not familiar with this driver. I did not know this driver is only loaded during system boot-up time, I thought this driver can be loaded as a kernel module (like many drivers) after system booting. After knowing this, I admit my patch is not proper, sorry... Best wishes, Jia-Ju Bai
Re: [PATCH v2] dec: tulip: de4x5: Replace mdelay with usleep_range in de4x5_hw_init
On Thursday 12 April 2018 07:00 AM, Jia-Ju Bai wrote: On 2018/4/12 0:16, James Bottomley wrote: On Wed, 2018-04-11 at 23:39 +0800, Jia-Ju Bai wrote: de4x5_hw_init() is never called in atomic context. de4x5_hw_init() is only called by de4x5_pci_probe(), which is only set as ".probe" in struct pci_driver. Despite never getting called from atomic context, de4x5_hw_init() calls mdelay() to busily wait. This is not necessary and can be replaced with usleep_range() to avoid busy waiting. This is found by a static analysis tool named DCNS written by myself. And I also manually check it. Did you actually test this? The usual reason for wanting m/udelay is that the timing must be exact. The driver is filled with mdelay()s for this reason. The one you've picked on is in the init path so it won't affect the runtime in any way. I also don't think we have the hrtimer machinery for usleep_range() to work properly on parisc, so I don't think the replacement works. James Hello, James. Thanks for your reply :) I agree that usleep_range() here will not much affect the real execution of this driver. But I think usleep_range() can more opportunity for other threads to use the CPU core to schedule during waiting. That is why I detect mdelay() that can be replaced with msleep() or usleep_range(). James is right, You have added all usleep_range() during system boot-up time. During boot-up system will run as single threaded. Where this change will not make much sense. System first priority is match the exact timing on each and every boot-up. ~arvind Best wishes, Jia-Ju Bai
Re: [PATCH v2] dec: tulip: de4x5: Replace mdelay with usleep_range in de4x5_hw_init
On 2018/4/12 0:16, James Bottomley wrote: On Wed, 2018-04-11 at 23:39 +0800, Jia-Ju Bai wrote: de4x5_hw_init() is never called in atomic context. de4x5_hw_init() is only called by de4x5_pci_probe(), which is only set as ".probe" in struct pci_driver. Despite never getting called from atomic context, de4x5_hw_init() calls mdelay() to busily wait. This is not necessary and can be replaced with usleep_range() to avoid busy waiting. This is found by a static analysis tool named DCNS written by myself. And I also manually check it. Did you actually test this? The usual reason for wanting m/udelay is that the timing must be exact. The driver is filled with mdelay()s for this reason. The one you've picked on is in the init path so it won't affect the runtime in any way. I also don't think we have the hrtimer machinery for usleep_range() to work properly on parisc, so I don't think the replacement works. James Hello, James. Thanks for your reply :) I agree that usleep_range() here will not much affect the real execution of this driver. But I think usleep_range() can more opportunity for other threads to use the CPU core to schedule during waiting. That is why I detect mdelay() that can be replaced with msleep() or usleep_range(). Best wishes, Jia-Ju Bai
Re: [PATCH v2] dec: tulip: de4x5: Replace mdelay with usleep_range in de4x5_hw_init
On Wed, 2018-04-11 at 23:39 +0800, Jia-Ju Bai wrote: > de4x5_hw_init() is never called in atomic context. > > de4x5_hw_init() is only called by de4x5_pci_probe(), which is only > set as ".probe" in struct pci_driver. > > Despite never getting called from atomic context, de4x5_hw_init() > calls mdelay() to busily wait. This is not necessary and can be > replaced with usleep_range() to avoid busy waiting. > > This is found by a static analysis tool named DCNS written by myself. > And I also manually check it. Did you actually test this? The usual reason for wanting m/udelay is that the timing must be exact. The driver is filled with mdelay()s for this reason. The one you've picked on is in the init path so it won't affect the runtime in any way. I also don't think we have the hrtimer machinery for usleep_range() to work properly on parisc, so I don't think the replacement works. James
