qsnet-rhel4-2.6.patch

lustre 1.6.5 source code

+environment where paging occurs and do not require memory to be locked
+down. The advantage of this is that the user process can expose large
+portions of it's address space without having to worry about physical
+memory constraints.
+
+However should the operating system decide to swap a page to disk,
+then the NIC must be made aware that it should no longer read/write
+from this memory, but should generate a translation fault instead.
+
+The ioproc patch has been developed to provide a mechanism whereby the
+device driver for a NIC can be aware of when a user process's address
+translations change, either by paging or by explicitly mapping or
+unmapping memory.
+
+The patch involves inserting callbacks where translations are being
+invalidated to notify the NIC that the memory behind those
+translations is no longer visible to the application (and so should
+not be visible to the NIC). This callback is then responsible for
+ensuring that the NIC will not access the physical memory that was
+being mapped.
+
+An ioproc invalidate callback in the kswapd code could be utilised to
+prevent memory from being paged out if the NIC is unable to support
+network page faulting.
+
+For NICs which support network page faulting, there is no requirement
+for a user level pin down cache, since they are able to page-in their
+translations on the first communication using a buffer. However this
+is likely to be inefficient, resulting in slow first use of the
+buffer. If the communication buffers were continually allocated and
+freed using mmap based malloc() calls then this would lead to all
+communications being slower than desirable.
+
+To optimise these warm-up cases the ioproc patch adds calls to
+ioproc_update wherever the kernel is creating translations for a user
+process. These then allows the device driver to preload translations
+so that they are already present for the first network communication
+from a buffer.
+
+Linux 2.6 IOPROC implementation details
+=======================================
+
+The Linux IOPROC patch adds hooks to the Linux VM code whenever page
+table entries are being created and/or invalidated. IOPROC device
+drivers can register their interest in being informed of such changes
+by registering an ioproc_ops structure which is defined as follows;
+
+extern int ioproc_register_ops(struct mm_struct *mm, struct ioproc_ops *ip);
+extern int ioproc_unregister_ops(struct mm_struct *mm, struct ioproc_ops *ip);
+
+typedef struct ioproc_ops {
+	struct ioproc_ops *next;
+	void *arg;
+
+	void (*release)(void *arg, struct mm_struct *mm);
+	void (*sync_range)(void *arg, struct vm_area_struct *vma, unsigned long start, unsigned long end);
+	void (*invalidate_range)(void *arg, struct vm_area_struct *vma, unsigned long start, unsigned long end);
+	void (*update_range)(void *arg, struct vm_area_struct *vma, unsigned long start, unsigned long end);
+
+	void (*change_protection)(void *arg, struct vm_area_struct *vma, unsigned long start, unsigned long end, pgprot_t newprot);
+
+	void (*sync_page)(void *arg, struct vm_area_struct *vma, unsigned long address);
+	void (*invalidate_page)(void *arg, struct vm_area_struct *vma, unsigned long address);
+	void (*update_page)(void *arg, struct vm_area_struct *vma, unsigned long address);
+
+} ioproc_ops_t;
+
+ioproc_register_ops
+===================
+This function should be called by the IOPROC device driver to register
+it's interest in PTE changes for the process associated with the passed
+in mm_struct.
+
+The ioproc registration is not inherited across fork() and should be
+called once for each process that IOPROC is interested in.
+
+This function must be called whilst holding the mm->page_table_lock.
+
+ioproc_unregister_ops
+=====================
+This function should be called by the IOPROC device driver when it no
+longer requires informing of PTE changes in the process associated
+with the supplied mm_struct.
+
+This function is not normally needed to be called as the ioproc_ops
+struct is unlinked from the associated mm_struct during the
+ioproc_release() call.
+
+This function must be called whilst holding the mm->page_table_lock.
+
+ioproc_ops struct
+=================
+A linked list ioproc_ops structures is hung off the user process
+mm_struct (linux/sched.h). At each hook point in the patched kernel,
+the ioproc patch will call the associated ioproc_ops callback function
+pointer in turn for each registered structure.
+
+The intention of the callbacks is to allow the IOPROC device driver to
+inspect the new or modified PTE entry via the Linux kernel
+(e.g. find_pte_map()). These callbacks should not modify the Linux
+kernel VM state or PTE entries.
+
+The ioproc_ops callback function pointers are:
+
+ioproc_release
+==============
+The release hook is called when a program exits and all it's vma areas
+are torn down and unmapped, i.e. during exit_mmap(). Before each
+release hook is called the ioproc_ops structure is unlinked from the
+mm_struct.
+
+No locks are required as the process has the only reference to the mm
+at this point.
+
+ioproc_sync_[range|page]
+========================
+The sync hooks are called when a memory map is synchronised with its
+disk image i.e. when the msync() syscall is invoked. Any future read
+or write by the IOPROC device to the associated pages should cause the
+page to be marked as referenced or modified.
+
+Called holding the mm->page_table_lock.
+
+ioproc_invalidate_[range|page]
+==============================
+The invalidate hooks are called whenever a valid PTE is unloaded
+e.g. when a page is unmapped by the user or paged out by the
+kernel. After this call the IOPROC must not access the physical memory
+again unless a new translation is loaded.
+
+Called holding the mm->page_table_lock.
+
+ioproc_update_[range|page]
+==========================
+The update hooks are called whenever a valid PTE is loaded
+e.g. mmaping memory, moving the brk up, when breaking COW or faulting
+in an anonymous page of memory. These give the IOPROC device the
+opportunity to load translations speculatively, which can improve
+performance by avoiding device translation faults.
+
+Called holding the mm->page_table_lock.
+
+ioproc_change_protection
+========================
+This hook is called when the protection on a region of memory is
+changed i.e. when the mprotect() syscall is invoked.
+
+The IOPROC must not be able to write to a read-only page, so if the
+permissions are downgraded then it must honour them. If they are
+upgraded it can treat this in the same way as the
+ioproc_update_[range|page]() calls.
+
+Called holding the mm->page_table_lock.
+
+
+Linux 2.6 IOPROC patch details
+==============================
+
+Here are the specific details of each ioproc hook added to the Linux
+2.6 VM system and the reasons for doing so:
+
+++++ FILE
+	mm/fremap.c
+
+==== FUNCTION
+	zap_pte
+
+CALLED FROM
+	install_page
+	install_file_pte
+
+PTE MODIFICATION
+	ptep_clear_flush
+
+ADDED HOOKS
+	ioproc_invalidate_page
+
+==== FUNCTION
+	install_page
+
+CALLED FROM
+	filemap_populate, shmem_populate
+
+PTE MODIFICATION
+	set_pte
+
+ADDED HOOKS
+	ioproc_update_page
+
+==== FUNCTION
+	install_file_pte
+
+CALLED FROM
+	filemap_populate, shmem_populate
+
+PTE MODIFICATION
+	set_pte
+
+ADDED HOOKS
+	ioproc_update_page
+
+
+++++ FILE
+	mm/memory.c
+
+==== FUNCTION
+	zap_page_range
+
+CALLED FROM
+	read_zero_pagealigned, madvise_dontneed, unmap_mapping_range,
+	unmap_mapping_range_list, do_mmap_pgoff
+
+PTE MODIFICATION
+	set_pte (unmap_vmas)
+
+ADDED HOOKS
+	ioproc_invalidate_range
+
+
+==== FUNCTION
+	zeromap_page_range
+
+CALLED FROM
+	read_zero_pagealigned, mmap_zero
+
+PTE MODIFICATION
+	set_pte (zeromap_pte_range)
+
+ADDED HOOKS
+	ioproc_invalidate_range
+	ioproc_update_range
+
+
+==== FUNCTION
+	remap_page_range
+
+CALLED FROM
+	many device drivers
+
+PTE MODIFICATION
+	set_pte (remap_pte_range)
+
+ADDED HOOKS
+	ioproc_invalidate_range
+	ioproc_update_range
+
+
+==== FUNCTION
+	break_cow
+
+CALLED FROM
+	do_wp_page
+
+PTE MODIFICATION
+	ptep_establish
+
+ADDED HOOKS
+	ioproc_invalidate_page
+	ioproc_update_page
+
+
+==== FUNCTION
+	do_wp_page
+
+CALLED FROM
+       do_swap_page, handle_pte_fault
+
+PTE MODIFICATION
+	ptep_set_access_flags
+
+ADDED HOOKS
+	ioproc_update_page
+
+
+==== FUNCTION
+	do_swap_page
+
+CALLED FROM
+	handle_pte_fault
+
+PTE MODIFICATION
+	set_pte
+
+ADDED HOOKS
+	ioproc_update_page
+
+
+==== FUNCTION
+	do_anonymous_page
+
+CALLED FROM
+	do_no_page
+
+PTE MODIFICATION
+	set_pte
+
+ADDED HOOKS
+	ioproc_update_page
+
+
+==== FUNCTION
+	do_no_page
+
+CALLED FROM
+	do_file_page, handle_pte_fault
+
+PTE MODIFICATION
+	set_pte
+
+ADDED HOOKS
+	ioproc_update_page
+
+
+++++ FILE
+	mm/mmap.c
+
+==== FUNCTION
+	unmap_region
+
+CALLED FROM
+	do_munmap
+
+PTE MODIFICATION
+	set_pte (unmap_vmas)
+
+ADDED HOOKS
+	ioproc_invalidate_range
+
+
+==== FUNCTION
+	exit_mmap
+
+CALLED FROM
+	mmput
+
+PTE MODIFICATION
+	set_pte (unmap_vmas)
+
+ADDED HOOKS
+	ioproc_release
+
+
+++++ FILE
+	mm/mprotect.c
+
+==== FUNCTION
+	change_protection
+
+CALLED FROM
+	mprotect_fixup
+
+PTE MODIFICATION
+	set_pte (change_pte_range)
+
+ADDED HOOKS
+	ioproc_change_protection
+
+
+++++ FILE
+	mm/mremap.c
+
+==== FUNCTION
+	move_page_tables
+
+CALLED FROM
+	move_vma
+
+PTE MODIFICATION
+	ptep_clear_flush (move_one_page)
+
+ADDED HOOKS
+	ioproc_invalidate_range
+	ioproc_invalidate_range
+
+
+++++ FILE
+	mm/rmap.c
+
+==== FUNCTION
+	try_to_unmap_one
+
+CALLED FROM
+	try_to_unmap_anon, try_to_unmap_file
+
+PTE MODIFICATION
+	ptep_clear_flush
+
+ADDED HOOKS
+	ioproc_invalidate_page
+
+
+==== FUNCTION
+	try_to_unmap_cluster
+
+CALLED FROM
+	try_to_unmap_file
+
+PTE MODIFICATION
+	ptep_clear_flush
+
+ADDED HOOKS
+	ioproc_invalidate_page
+
+
+
+++++ FILE 
+	mm/msync.c
+
+==== FUNCTION
+	filemap_sync
+
+CALLED FROM
+	msync_interval
+
+PTE MODIFICATION
+	ptep_clear_flush_dirty (filemap_sync_pte)
+
+ADDED HOOKS
+	ioproc_sync_range
+
+
+++++ FILE
+	mm/hugetlb.c
+
+==== FUNCTION
+	zap_hugepage_range
+
+CALLED FROM
+	hugetlb_vmtruncate_list
+
+PTE MODIFICATION
+	ptep_get_and_clear (unmap_hugepage_range)
+
+ADDED HOOK
+	ioproc_invalidate_range
+