On Fri, Dec 17, 2021 at 12:36:43PM -0800, Linus Torvalds wrote:
- Take a R/O pin (RDMA, VFIO, ...)
-> refcount > 1
- memset(mem, 0xff, pagesize);
-> Write fault -> COW
I do not believe this is actually a bug.
You asked for a R/O pin, and you got one.
Then somebody else modified that page, and you got exactly what you asked for - a COW event. The original R/O pin has the original page that it asked for, and can read it just fine.
To remind all, the GUP users, like RDMA, VFIO use FOLL_FORCE|FOLL_WRITE to get a 'r/o pin' specifically because of the COW breaking the coherence. In these case 'r/o pin' does not mean "snapshot the data", but its only a promise not to write to the pages and still desires coherence with the memory map.
Eg in RDMA we know of apps asking for a R/O pin of something in .bss then filling that something with data finally doing the actual DMA. Breaking COW after pin breaks those apps.
The above #5 can occur for O_DIRECT read and in that case the 'snapshot the data' is perfectly fine as racing the COW with the O_DIRECT read just resolves the race toward the read() direction.
IIRC there is some other scenario that motivated this patch?
Jason