02:28 <marex> arnd: the PCI patch looks awful, but oh well ... it does its job

09:13 <arnd> marex: can't you just use a .fixup section hack in the inline asm that sets the return value to all ones?

16:07 <marex> arnd: a .fixup section hack ? What's that ?

16:14 <arnd> marex: see get_user()

16:15 <j`ey> and arch/arm64/include/asm/asm-extable.h

16:15 <arnd> you have a small bit of fixup code that loads the return code when an exception happens in the .text.fixup section, and a pair of pointers in the __ex_table section that points to the (potentially) faulting instruction and the fixup

16:16 <arnd> the generic exception handler checks if an exception came from an address listed in the __ex_table section, and continues in the corresponding .text.fixup if it did, instead of trying to handle the exception as normal

16:22 <marex> arnd: the fixup code I have also brings the PCIe link back up, so it isn't small

16:23 <marex> arnd: is that also available on arm32 , because that's where I have this problem

16:26 <arnd> The fixup mechanism works on all architectures, but of course each one needs customer asm instructions

16:27 <arnd> marex: you should be able to hook up the link retraining by using an asm with two outputs, then check for an error that gets flagged in the second one

16:40 <marex> arnd: oh, all right, I'll have to revisit that and grep around when I have more time