Linux with mmview extensions

The goal of the mmview extensions is to provide means to create additional address spaces, called views in the userspace. The views are selectable on the thread level and are generally kept in sync with the original address space. Users of the extension can designate specific regions as copy-on-write, allowing them to diverge between views or threads.

Usage

• long mmview(MMVIEW_CREATE)

  Creates an additional address space view.

  Returns id on success. On failure, -1 is returned and errno is set appropriately:

  ENOMEM

  Out of memory.

• long mmview(MMVIEW_DELETE, long id)

  Marks the address space view referenced by id for deletion. Its resources will be freed after the last thread stops using it.

  Returns 0 on success. On failure, -1 is returned and errno is set appropriately:

  EPERM	Address space view referenced by id is the base view.
  EINVAL	Address space view referenced by id does not exist.

• long mmview(MMVIEW_CURRENT)

  Returns id of the current address space. No errors.

• long mmview(MMVIEW_MIGRATE, long id)

  Migrates current thread to the address space view referenced by id.

  Returns id of the previous address space view on success. On failure, -1 is returned and errno is set appropriately:

  EINVAL

  Address space view referenced by id does not exist.

• long mmview(MMVIEW_UNSHARE, void *addr, size_t len)

  long mmview(MMVIEW_SHARE, void *addr, size_t len)

  Designates memory pages containing any part of the address range in the interval [addr, addr+len-1] as shared or unshared. Must only be called with single view present. addr must be aligned to a page boundary.

  Returns 0 on success. On failure, -1 is returned and errno is set appropriately:

  EPERM	Calling process has additional views.
  EINVAL	An invalid address range has been specified.
  EACCES	Memory contained in the interval cannot be unshared.

• long mmview(MMVIEW_SWITCH_BASE)

  Appoints the current view to be the new base view of the process. This can be used to optimize memory accesses, or to free the memory belonging to the previous base view by allowing its deletion.

  Returns id of the current view on success. On failure, -1 is returned and errno is set appropriately:

  ENOMEM

  Out of memory.

System call wrappers along with the test suite can be found in the mmview-tests repository.

News

BREAKING: Redesigned the system call interface

Due to the system call interface getting out of control after introduction of new features, the mmview functionality is now re-exported under a single system call with the first parameter specifying the requested procedure.

Added possibility to designate areas as shared

After fork() or deletion of other mmviews, it might be desired to redesignate some areas as shared again. This is now possible.

Enabled migration of the base mmview

The base mmview is the initial address space, used as the master copy for shared areas in other views. Changing the base view is transparent to the application, but enables some techniques, which were not possible before (e.g., deletion of the original address space). Another side effect is the reduced overhead of page faults in the base mmview.

Updated kernel to 5.15

Updating the kernel tree revealed a few bugs, some of which were previously non-reproducible:

• Page fault handler returning with VM_FAULT_RETRY drops the mmap_lock. In case the fault is initiated in base mm the lock has to be retaken.
• Streamlined creation of new mms, ensuring that mms without common are not exposed to the outside world.
• Fixed broken anonymous VM_SHARED mappings, where a separate file was created for each view in mmap_region.
• Zapping instead of refaulting after swap-in of shared mmview pages achieved a modest speed improvement during swapping.

Revamped zapping of shared mmview pages on copy

• Fixed race condition where a concurrent fault could restore the old page after zapping.
• No more redundant zapping such as when reusing a page or managing software dirty/young bits.
• A new reference counter in struct page tracks the amount of page table entries referencing the page from a non-base view. This allows reliably detecting cases where a shared mmview page can be reused on write fault.

Added general multiarch support (ARM)

• Converted types to allow clean compilation on 32-bit architectures.
• Fixed invalid usage of high memory mappings in the pagefault handler.
• Modified pagefault handler to account for architectures with software emulated dirty/young bits.

Revisited copy_page_range

copy_page_range called during fork or mmview_create system calls now uses pagewalk API to write-protect the views in the COW-case. This change has allowed proper usage of the secondary mmu notifier, which is required when modifying protection bits of the PTEs.

Added hugepages support

Coredumping support

Structures facilitating coredumping need to be shared among threads which wasn’t the case if threads used different views.

Swapping support for anonymous mappings

Reverse mapping structures (anon_vma) were missing for non-base views, which prevented swap-out of pages referenced by these views.