Change Details

Both arm64 and riscv have the same problem that we addressed in r348246 on amd64 and i386. However, they do not define a PG_PROMOTED bit in their level 2 PTEs. PG_PROMOTED makes no sense on arm64, where we are required to "make before break". And, on riscv, there are no spare bits in the PTE. (Only two bits are available to the OS, and we have used them both.) Instead, I propose to use the page table page's valid field. And, in fact, I propose to switch amd64 (and i386) to this approach. Aside from my belief that having a single, portable approach across the different architectures is a virtue, when the erratum 383 work-around is enabled on AMD processors, we are currently performing unnecessary page table page initializations on demotion because we don't set PG_PROMOTED in the PDE when the 383 work-around is enabled. In addition, remove a stale comment.

Both arm64 and riscv have the same problem that we addressed in r348246 on amd64 and i386. However, they do not define a PG_PROMOTED bit in their level 2 PTEs. PG_PROMOTED makes no sense on arm64, where we are required to "break before make". And, on riscv, there are no spare bits in the PTE. (Only two bits are available to the OS, and we have used them both.) Instead, I propose to use the page table page's valid field. And, in fact, I propose to switch amd64 (and i386) to this approach. Aside from my belief that having a single, portable approach across the different architectures is a virtue, when the erratum 383 work-around is enabled on AMD processors, we are currently performing unnecessary page table page initializations on demotion because we don't set PG_PROMOTED in the PDE when the 383 work-around is enabled. In addition, remove a stale comment.

Both arm64 and riscv have the same problem that we addressed in r348246 on amd64 and i386. However, they do not define a PG_PROMOTED bit in their level 2 PTEs. PG_PROMOTED makes no sense on arm64, where we are required to "make"break before break"make". And, on riscv, there are no spare bits in the PTE. (Only two bits are available to the OS, and we have used them both.) Instead, I propose to use the page table page's valid field. And, in fact, I propose to switch amd64 (and i386) to this approach. Aside from my belief that having a single, portable approach across the different architectures is a virtue, when the erratum 383 work-around is enabled on AMD processors, we are currently performing unnecessary page table page initializations on demotion because we don't set PG_PROMOTED in the PDE when the 383 work-around is enabled. In addition, remove a stale comment.