Hugetlbfs file mappings are handled differently than regular files:
- pager_req_create will tell us the file is in a hugetlbfs
- allocate memory upfront, we need to fail if not enough memory
- the memory needs to be given again if another process maps the same
file
This implementation still has some hacks, in particular, the memory
needs to be freed when all mappings are done and the file has been
deleted/closed by all processes.
We cannot know when the file is closed/unlinked easily, so clean up
memory when all processes have exited.
To test, install libhugetlbfs and link a program with the additional
LDFLAGS += -B /usr/share/libhugetlbfs -Wl,--hugetlbfs-align
Then run with HUGETLB_ELFMAP=RW set, you can check this works with
HUGETLB_DEBUG=1 HUGETLB_VERBOSE=2
Change-Id: I327920ff06efd82e91b319b27319f41912169af1
- remove unused MF_END (that only makes sense for enums without holes,
this one is a set of bits masks)
- remove useless goto in pager_req_create()
- init maxprot to 0 from the start, it's not used in the error cases
(except for debug print)
Change-Id: Ic56c0754824b99f8a7e45fa8e99b8fe3e7c7e592
We had a deadlock between:
- free_process_memory_range (take lock) -> ihk_mc_pt_free_range ->
... -> remote_flush_tlb_array_cpumask -> "/* Wait for all cores */"
and
- obj_list_lookup() under fileobj_list_lock that disabled irqs
and thus never ack'd the remote flush
The rework is quite big but removes the need for the big lock,
although devobj and shmobj needed a new smaller lock to be
introduced - the new locks are used much more locally and
should not cause problems.
On the bright side, refcounting being moved to memobj level means
we could remove refcounting implemented separately in all object
types and simplifies code a bit.
Change-Id: I6bc8438a98b1d8edddc91c4ac33c11b88e097ebb
Introduction of VR_WRITE_COMBINED, PTATTR_WRITE_COMBINED and modification
to the memobj's get_page() interface so that Linux communicates back mapping
flags (such as write-combined).
implemented:
- Pages can be shared between maps.
- A change made to a map is written to the file, at munmap().
not yet implemented:
- VM operation during page IO execution.
Because page IO is executed with VM's lock.
- Page IO, which does not change a file size with any case.
When munmap() races with truncate(), the file size may be changed
illegally.
