On 2022-06-18 9:18 p.m., John David Anglin wrote:
Shared anonymous mappings only work when the mappings are equivalent or meet the requirements for multi reader, single writer. The problem is we don't in general know what mappings are equivalent on PA8800/PA8900 processors.
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:
SHMLBA Segment low boundary address multiple. When explicitly specify- ing an attach address in a call to shmat(), the caller should ensure that the address is a multiple of this value. This is necessary on some architectures, in order either to ensure good CPU cache performance or to ensure that different attaches of the same segment have consistent views within the CPU cache. SHMLBA is normally some multiple of the system page size. (On many Linux architectures, SHMLBA is the same as the system page size.)
Further, the manpage states:
o If shmaddr is NULL, the system chooses a suitable (unused) page- aligned address to attach the segment. o If shmaddr isn't NULL and SHM_RND is specified in shmflg, the attach occurs at the address equal to shmaddr rounded down to the nearest multiple of SHMLBA. o Otherwise, shmaddr must be a page-aligned address at which the attach occurs.
Thus, shared mappings may not be equivalent and we need to flush all aliases. Dave -- John David Anglin dave.anglin@xxxxxxxx
