On Sun, 2022-06-19 at 09:32 -0400, John David Anglin wrote: > On 2022-06-18 9:18 p.m., John David Anglin wrote: > > Shared anonymous mappings only work when the mappings are equivalen > > t or meet the > > requirements for multi reader, single writer. The problem is we do > > n't in general know what mappings are equivalent on PA8800/PA8900 p > > rocessors. > This change introduced the possibility that shared mappings might not > be equivalent on parisc: > https://marc.info/?l=git-commits-head&m=139776619924910&w=2 > > The shmat() man page describes SHMLBA as follows: I wouldn't read what the manpage says, I'd read what the code does. SHMLBA at page size allows a mapping to start anywhere as opposed to only on 4M boundaries. SHM_COLOUR at 4M should ensure that any additional mappings of the area are all aliased correctly. However, the theory of that seems to rest on the ability of GET_LAST_MMAP/SET_LAST_MMAP to stash the mmaped address, which appears to be broken. So I'd say rather that over flushing we should fix that, either by stashing properly or falling back to 4MB SHMLBA. James
