On Thu, 15 Feb 2018 13:38:00 +0200, cantabile wrote: > On 15/02/18 02:45, Jakub Kicinski wrote: > > On Wed, 14 Feb 2018 13:34:38 +0200, cantabile wrote: > >> The firmware running on the device sometimes survives a reboot > >> (firmware_running returns 1). When this happens the driver never calls > >> request_firmware, which means the kernel's firmware handling code > >> doesn't know this firmware should be cached before hibernating. Upon > >> resuming from several hours of hibernation, the firmware is no longer > >> running on the device, so the driver calls request_firmware. Since the > >> firmware was never cached, it needs to be loaded from disk, and this is > >> when the system freezes, somewhere in the xfs driver. Fix this by always > >> requesting the firmware, whether it's already running on the device or not. > >> > >> Signed-off-by: John Smith <cantabile.d...@gmail.com> > > > > Thanks for tracking this down, but this seems like the wrong > > direction. > > > > What's your hard drive? Is it some complex configuration which > > prevents the block device from coming online after resume? > > > > If it's really because of some peculiarities of XFS the fix should > > go there, no driver will be able to load FW on resume... > > > > My hard drive is a SATA SSD, with a regular MS-DOS partition table, with > three primary partitions: one ext4 mounted at /boot, one xfs mounted at > /, one xfs mounted at /home. No encryption, no software RAID, no logical > volumes, etc. > > The kernel has a firmware caching mechanism to make sure that no driver > needs to load firmware from disk during the resume process. Because > that's unreliable. See this bit of the documentation: [1] > > This is how it works: when the driver's probe callback calls > request_firmware, that function adds the firmware file's name in a > devres in your 'struct device'. When you suspend the system, the kernel > calls fw_pm_notify [2], which goes through all the 'struct device's and > looks for firmware file names previously added by request_firmware, and > loads into memory all the firmware files it finds. When you resume the > system, all calls to request_firmware ought to find the firmware files > already loaded in memory. After resuming, the kernel calls fw_pm_notify > again to release the cached firmware files (with some delay). > > This caching mechanism currently doesn't always work for your driver > because your driver doesn't always call request_firmware from the probe > callback. Because the firmware running on the device sometimes survives > a reboot.
Thanks for the info. Would it be cleaner to EXPORT fw_add_devm_name() and just call that in case driver sees FW is already loaded? That should inform the fw subsystem that we want the image around in case of hibernation, but there is no need to load it immediately? > You can get some very useful messages if you compile firmware_class.c > with -DDEBUG [3]. This is how I figured out the problem. > > [1] > https://www.kernel.org/doc/html/v4.13/driver-api/firmware/request_firmware.html#considerations-for-suspend-and-resume > [2] > https://github.com/torvalds/linux/blob/master/drivers/base/firmware_class.c#L1804 > [3] https://www.kernel.org/doc/local/pr_debug.txt
