Re: [RFC PATCH] usb: optimize acpi companion search for usb port devices
On 01.06.2017 02:01, Rafael J. Wysocki wrote: On 5/25/2017 5:24 PM, Mathias Nyman wrote: On 24.05.2017 17:44, Greg KH wrote: On Wed, May 24, 2017 at 04:11:12PM +0300, Mathias Nyman wrote: This optimization significantly reduces xhci driver load time. In ACPI tables the acpi companion port devices are children of the hub device. The port devices are identified by their port number returned by the ACPI _ADR method. _ADR 0 is reserved for the root hub device. The current implementation to find a acpi companion port device loops through all acpi port devices under that parent hub, calling their _ADR method each time a new port device is added. for a xHC controller with 25 ports under its roothub it will end up invoking ACPI bytecode 625 times before all ports are ready, making it really slow. The _ADR values are already read and cached earler. So instead of running the bytecode again we can check the cached _ADR value first, and then fall back to the old way. As one of the more significant changes, the xhci load time on Intel kabylake reduced by 70%, (28ms) from initcall xhci_pci_init+0x0/0x49 returned 0 after 39537 usecs to initcall xhci_pci_init+0x0/0x49 returned 0 after 11270 usecs Signed-off-by: Mathias Nyman--- drivers/usb/core/usb-acpi.c | 26 +++--- 1 file changed, 23 insertions(+), 3 deletions(-) Why is this RFC? What's wrong with it as-is? Last minute doubt, nothing should be wrong, but I started to wonder if there is any particular reason the ACPI part was done the way it was. Or if maybe other drivers could benefit from checking cached _ADR value first as well, and this whole thing should be a part of drivers/acpi/glue.c instead? That or we should just evaluate _ADR if present during the very initialization and store the value in a filed under struct acpi_device. Yes, that is pretty much done, acpi_init_device_object() will end up evaluating _ADR and store it in acpi_device.pnp.bus_address Just considering if there should be something like acpi_get_adr(acpi_device *adev) that would first check if adev->pnp.bus_address exists, if not, then it would evaluate _ADR. But really just a thought, I think I'll send this forward as is. Thanks Mathias -- To unsubscribe from this list: send the line "unsubscribe linux-usb" in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Re: [RFC PATCH] usb: optimize acpi companion search for usb port devices
On 5/25/2017 5:24 PM, Mathias Nyman wrote: On 24.05.2017 17:44, Greg KH wrote: On Wed, May 24, 2017 at 04:11:12PM +0300, Mathias Nyman wrote: This optimization significantly reduces xhci driver load time. In ACPI tables the acpi companion port devices are children of the hub device. The port devices are identified by their port number returned by the ACPI _ADR method. _ADR 0 is reserved for the root hub device. The current implementation to find a acpi companion port device loops through all acpi port devices under that parent hub, calling their _ADR method each time a new port device is added. for a xHC controller with 25 ports under its roothub it will end up invoking ACPI bytecode 625 times before all ports are ready, making it really slow. The _ADR values are already read and cached earler. So instead of running the bytecode again we can check the cached _ADR value first, and then fall back to the old way. As one of the more significant changes, the xhci load time on Intel kabylake reduced by 70%, (28ms) from initcall xhci_pci_init+0x0/0x49 returned 0 after 39537 usecs to initcall xhci_pci_init+0x0/0x49 returned 0 after 11270 usecs Signed-off-by: Mathias Nyman--- drivers/usb/core/usb-acpi.c | 26 +++--- 1 file changed, 23 insertions(+), 3 deletions(-) Why is this RFC? What's wrong with it as-is? Last minute doubt, nothing should be wrong, but I started to wonder if there is any particular reason the ACPI part was done the way it was. Or if maybe other drivers could benefit from checking cached _ADR value first as well, and this whole thing should be a part of drivers/acpi/glue.c instead? That or we should just evaluate _ADR if present during the very initialization and store the value in a filed under struct acpi_device. Thanks, Rafael -- To unsubscribe from this list: send the line "unsubscribe linux-usb" in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Re: [RFC PATCH] usb: optimize acpi companion search for usb port devices
On 24.05.2017 17:44, Greg KH wrote: On Wed, May 24, 2017 at 04:11:12PM +0300, Mathias Nyman wrote: This optimization significantly reduces xhci driver load time. In ACPI tables the acpi companion port devices are children of the hub device. The port devices are identified by their port number returned by the ACPI _ADR method. _ADR 0 is reserved for the root hub device. The current implementation to find a acpi companion port device loops through all acpi port devices under that parent hub, calling their _ADR method each time a new port device is added. for a xHC controller with 25 ports under its roothub it will end up invoking ACPI bytecode 625 times before all ports are ready, making it really slow. The _ADR values are already read and cached earler. So instead of running the bytecode again we can check the cached _ADR value first, and then fall back to the old way. As one of the more significant changes, the xhci load time on Intel kabylake reduced by 70%, (28ms) from initcall xhci_pci_init+0x0/0x49 returned 0 after 39537 usecs to initcall xhci_pci_init+0x0/0x49 returned 0 after 11270 usecs Signed-off-by: Mathias Nyman--- drivers/usb/core/usb-acpi.c | 26 +++--- 1 file changed, 23 insertions(+), 3 deletions(-) Why is this RFC? What's wrong with it as-is? Last minute doubt, nothing should be wrong, but I started to wonder if there is any particular reason the ACPI part was done the way it was. Or if maybe other drivers could benefit from checking cached _ADR value first as well, and this whole thing should be a part of drivers/acpi/glue.c instead? (adding acpi mailing list, not just Rafael) Thanks -Mathias -- To unsubscribe from this list: send the line "unsubscribe linux-usb" in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Re: [RFC PATCH] usb: optimize acpi companion search for usb port devices
On Wed, May 24, 2017 at 04:11:12PM +0300, Mathias Nyman wrote: > This optimization significantly reduces xhci driver load time. > > In ACPI tables the acpi companion port devices are children of > the hub device. The port devices are identified by their port number > returned by the ACPI _ADR method. > _ADR 0 is reserved for the root hub device. > > The current implementation to find a acpi companion port device > loops through all acpi port devices under that parent hub, calling > their _ADR method each time a new port device is added. > > for a xHC controller with 25 ports under its roothub it > will end up invoking ACPI bytecode 625 times before all ports > are ready, making it really slow. > > The _ADR values are already read and cached earler. So instead of > running the bytecode again we can check the cached _ADR value first, > and then fall back to the old way. > > As one of the more significant changes, the xhci load time on > Intel kabylake reduced by 70%, (28ms) from > initcall xhci_pci_init+0x0/0x49 returned 0 after 39537 usecs > to > initcall xhci_pci_init+0x0/0x49 returned 0 after 11270 usecs > > Signed-off-by: Mathias Nyman> --- > drivers/usb/core/usb-acpi.c | 26 +++--- > 1 file changed, 23 insertions(+), 3 deletions(-) Why is this RFC? What's wrong with it as-is? thanks, greg k-h -- To unsubscribe from this list: send the line "unsubscribe linux-usb" in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
[RFC PATCH] usb: optimize acpi companion search for usb port devices
This optimization significantly reduces xhci driver load time. In ACPI tables the acpi companion port devices are children of the hub device. The port devices are identified by their port number returned by the ACPI _ADR method. _ADR 0 is reserved for the root hub device. The current implementation to find a acpi companion port device loops through all acpi port devices under that parent hub, calling their _ADR method each time a new port device is added. for a xHC controller with 25 ports under its roothub it will end up invoking ACPI bytecode 625 times before all ports are ready, making it really slow. The _ADR values are already read and cached earler. So instead of running the bytecode again we can check the cached _ADR value first, and then fall back to the old way. As one of the more significant changes, the xhci load time on Intel kabylake reduced by 70%, (28ms) from initcall xhci_pci_init+0x0/0x49 returned 0 after 39537 usecs to initcall xhci_pci_init+0x0/0x49 returned 0 after 11270 usecs Signed-off-by: Mathias Nyman--- drivers/usb/core/usb-acpi.c | 26 +++--- 1 file changed, 23 insertions(+), 3 deletions(-) diff --git a/drivers/usb/core/usb-acpi.c b/drivers/usb/core/usb-acpi.c index 2776cfe..ef9cf4a 100644 --- a/drivers/usb/core/usb-acpi.c +++ b/drivers/usb/core/usb-acpi.c @@ -127,6 +127,22 @@ static enum usb_port_connect_type usb_acpi_get_connect_type(acpi_handle handle, */ #define USB_ACPI_LOCATION_VALID (1 << 31) +static struct acpi_device *usb_acpi_find_port(struct acpi_device *parent, + int raw) +{ + struct acpi_device *adev; + + if (!parent) + return NULL; + + list_for_each_entry(adev, >children, node) { + if (acpi_device_adr(adev) == raw) + return adev; + } + + return acpi_find_child_device(parent, raw, false); +} + static struct acpi_device *usb_acpi_find_companion(struct device *dev) { struct usb_device *udev; @@ -174,8 +190,10 @@ static struct acpi_device *usb_acpi_find_companion(struct device *dev) int raw; raw = usb_hcd_find_raw_port_number(hcd, port1); - adev = acpi_find_child_device(ACPI_COMPANION(>dev), - raw, false); + + adev = usb_acpi_find_port(ACPI_COMPANION(>dev), + raw); + if (!adev) return NULL; } else { @@ -186,7 +204,9 @@ static struct acpi_device *usb_acpi_find_companion(struct device *dev) return NULL; acpi_bus_get_device(parent_handle, ); - adev = acpi_find_child_device(adev, port1, false); + + adev = usb_acpi_find_port(adev, port1); + if (!adev) return NULL; } -- 1.9.1 -- To unsubscribe from this list: send the line "unsubscribe linux-usb" in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html