From: Peter Zijlstra > Sent: 12 January 2022 15:34 > > On Wed, Jan 12, 2022 at 03:15:27PM +0000, David Laight wrote: > > From: Mathieu Desnoyers > > > Sent: 12 January 2022 15:06 > > > > > > ----- On Jan 12, 2022, at 9:58 AM, David Laight David.Laight@xxxxxxxxxx wrote: > > > > > > >> * [*] The openrisc, powerpc64 and x86-64 architectures define a "redzone" as a > > > >> * stack area beyond the stack pointer which can be used by the compiler > > > >> * to store local variables in leaf functions. > > > > > > > > I wonder if that is really worth the trouble it causes! > > > > By the time a function is spilling values to stack the cost > > > > of a %sp update is almost certainly noise. > > > > > > > > Someone clearly thought it was a 'good idea (tm)'. > > > > > > I must admit that I've been surprised to learn about these redzones. Thanks for > > > pointing them out to me, it was clearly a blind spot. I suspect it would be useful > > > to introduce per-architecture KERNEL_REDZONE, USER_REDZONE and COMPAT_USER_REDZONE > > > with a asm-generic version defining them to 0, with proper documentation. It would > > > make it clearer to kernel developers working on stuff similar to signal handler > > > delivery that they need to consider these carefully. > > > > They can never be used in kernel - any ISR would overwrite them. > > That depends on how the architecture does exceptions; True, many newer ones don't actually write anything to the stack. Makes the cpu simpler. > also consider: > > https://www.intel.com/content/www/us/en/develop/download/flexible-return-and-event-delivery- > specification.html That contains the snippet: The SWAPGS instruction supports efficient updates of the GS base address. Which is just so horribly not true... Even FRED is always doing a GS swap - so you can easily lose the kernel GS value. I remember fixing all the 'in kernel' faults in the netbsd x86-64 return to user path. Entirely horrid... David - Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK Registration No: 1397386 (Wales)
