On Tue, Mar 09, 2021 at 02:07:36PM -0800, Linus Torvalds wrote: > On Tue, Mar 9, 2021 at 1:17 PM Rasmus Villemoes > <linux@xxxxxxxxxxxxxxxxxx> wrote: > > > > So add an initramfs_async= kernel parameter, allowing the main init > > process to proceed to handling device_initcall()s without waiting for > > populate_rootfs() to finish. > > I like this smaller second version of the patch, but am wondering why > we even need the parameter. > > It sounds mostly like a "maybe I didn't think of all cases" thing - > and one that will mean that this code will not see a lot of actual > test coverage.. > > And because of the lack of test coverage, I'd rather reverse the > meaning, and have the async case on by default (without even the > Kconfig option), and have the kernel command line purely as a "oops, > it's buggy, easy to ask people to test if this is what ails them". If we're going to set this as default it might be good to document on init.h that components that need content in initramfs need the wait call. > What *can* happen early boot outside of firmware loading and usermodehelpers? *In practice* the only thing I can think of at this time is races with other rootfs_initcall() calls, granted ordering among these is done at linker time, but I can't think of a issue with them: arch/x86/kernel/pci-dma.c:rootfs_initcall(pci_iommu_init); drivers/iommu/intel/irq_remapping.c:rootfs_initcall(ir_dev_scope_init); drivers/mfd/sta2x11-mfd.c:rootfs_initcall(sta2x11_mfd_init); drivers/thunderbolt/nhi.c:rootfs_initcall(nhi_init); But Cc'ing the maintainers of these components just in case. Luis
