- 30 Oct, 2005 40 commits
-
-
Andrea Arcangeli authored
We had a problem on ppc64 where with more than 4 threads a large system wouldn't scale well while faulting in the .text (most of the time was spent in the kernel despite it was an userland compute intensive app). The reason is the useless overwrite of the same pte from all cpu. I fixed it this way (verified on an older kernel but the forward port is almost identical). This will benefit all archs not just ppc64. Signed-off-by:
Andrea Arcangeli <andrea@suse.de> Cc: Hugh Dickins <hugh@veritas.com> Signed-off-by:
Andrew Morton <akpm@osdl.org> Signed-off-by:
Linus Torvalds <torvalds@osdl.org>
-
Adam Litke authored
Basic overcommit checking for hugetlb_file_map() based on an implementation used with demand faulting in SLES9. Since demand faulting can't guarantee the availability of pages at mmap time, this patch implements a basic sanity check to ensure that the number of huge pages required to satisfy the mmap are currently available. Despite the obvious race, I think it is a good start on doing proper accounting. I'd like to work towards an accounting system that mimics the semantics of normal pages (especially for the MAP_PRIVATE/COW case). That work is underway and builds on what this patch starts. Huge page shared memory segments are simpler and still maintain their commit on shmget semantics. Signed-off-by:
Adam Litke <agl@us.ibm.com> Signed-off-by:
-
Adam Litke authored
Below is a patch to implement demand faulting for huge pages. The main motivation for changing from prefaulting to demand faulting is so that huge page memory areas can be allocated according to NUMA policy. Thanks to consolidated hugetlb code, switching the behavior requires changing only one fault handler. The bulk of the patch just moves the logic from hugelb_prefault() to hugetlb_pte_fault() and find_get_huge_page(). Signed-off-by:
-
Krishnakumar R authored
Clean up some repeated code related to HugeTLB. hugetlb_zero_setup would have already allocated the file->f_op. Signed-off-by:
Krishnakumar. R <rkrishnakumar@gmail.com> Signed-off-by:
-
Christoph Hellwig authored
Reformat hugelbfs_forget_inode and add the missing but harmless write_inode_now call. It looks the same as generic_forget_inode now except for the call to truncate_hugepages instead of truncate_inode_pages. Signed-off-by:
Christoph Hellwig <hch@lst.de> Signed-off-by:
-
Christoph Hellwig authored
hugetlbfs_do_delete_inode is the same as generic_delete_inode now, so remove it in favour of the latter. Signed-off-by:
-
Christoph Hellwig authored
Make hugetlbfs looks the same as generic_detelte_inode, fixing a bunch of missing updates to it at the same time. Rename it to hugetlbfs_do_delete_inode and add a real hugetlbfs_delete_inode that implements ->delete_inode. Signed-off-by:
-
Christoph Hellwig authored
Move hugetlbfs accounting into ->alloc_inode / ->destroy_inode. This keeps the code simpler, fixes a loeak where a failing inode allocation wouldn't decrement the counter and moves hugetlbfs_delete_inode and hugetlbfs_forget_inode closer to their generic counterparts. Signed-off-by:
-
Hugh Dickins authored
Updated several references to page_table_lock in common code comments. Signed-off-by:
Hugh Dickins <hugh@veritas.com> Signed-off-by:
-
Hugh Dickins authored
A couple of oddities were guarded by page_table_lock, no longer properly guarded when that is split. The mm_counters of file_rss and anon_rss: make those an atomic_t, or an atomic64_t if the architecture supports it, in such a case. Definitions by courtesy of Christoph Lameter: who spent considerable effort on more scalable ways of counting, but found insufficient benefit in practice. And adding an mm with swap to the mmlist for swapoff: the list is well- guarded by its own lock, but the list_empty check now has to be repeated inside it. Signed-off-by:
-
Hugh Dickins authored
Christoph Lameter demonstrated very poor scalability on the SGI 512-way, with a many-threaded application which concurrently initializes different parts of a large anonymous area. This patch corrects that, by using a separate spinlock per page table page, to guard the page table entries in that page, instead of using the mm's single page_table_lock. (But even then, page_table_lock is still used to guard page table allocation, and anon_vma allocation.) In this implementation, the spinlock is tucked inside the struct page of the page table page: with a BUILD_BUG_ON in case it overflows - which it would in the case of 32-bit PA-RISC with spinlock debugging enabled. Splitting the lock is not quite for free: another cacheline access. Ideally, I suppose we would use split ptlock only for multi-threaded processes on multi-cpu machines; but deciding that dynamically would have its own costs. So for now enable it by config, at some number of cpus - since the Kconfig language doesn't support inequalities, let preprocessor compare that with NR_CPUS. But I don't think it's worth being user-configurable: for good testing of both split and unsplit configs, split now at 4 cpus, and perhaps change that to 8 later. There is a benefit even for singly threaded processes: kswapd can be attacking one part of the mm while another part is busy faulting. Signed-off-by:
-
Hugh Dickins authored
In worrying over the various pte operations in different architectures, I came across some unused functions in UML: remove mprotect_kernel_vm, protect_vm_page and addr_pte. Signed-off-by:
-
Hugh Dickins authored
There's usually a good reason when a pte is examined without the lock; but it makes me nervous when the pointer is dereferenced more than once. Signed-off-by:
-
Hugh Dickins authored
The cris v32 switch_mm guards get_mmu_context with next->page_table_lock: good it's not really SMP yet, since get_mmu_context messes with global variables affecting other mms. Replace by global mmu_context_lock. Signed-off-by:
-
Hugh Dickins authored
There's a worrying function translation_exists in parisc cacheflush.h, unaffected by split ptlock since flush_dcache_page is using it on some other mm, without any relevant lock. Oh well, make it a slightly more robust by factoring the pfn check within it. And it looked liable to confuse a camouflaged swap or file entry with a good pte: fix that too. Signed-off-by:
-
Hugh Dickins authored
Prepare arm for the split page_table_lock: three issues. Signal handling's preserve and restore of iwmmxt context currently involves reading and writing that context to and from user space, while holding page_table_lock to secure the user page(s) against kswapd. If we split the lock, then the structure might span two pages, secured by to read into and write from a kernel stack buffer, copying that out and in without locking (the structure is 160 bytes in size, and here we're near the top of the kernel stack). Or would the overhead be noticeable? arm_syscall's cmpxchg emulation use pte_offset_map_lock, instead of pte_offset_map and mm-wide page_table_lock; and strictly, it should now also take mmap_sem before descending to pmd, to guard against another thread munmapping, and the page table pulled out beneath this thread. Updated two comments in fault-armv.c. adjust_pte is interesting, since its modification of a pte in one part of the mm depends on the lock held when calling update_mmu_cache for a pte in some other part of that mm. This can't be done with a split page_table_lock (and we've already taken the lowest lock in the hierarchy here): so we'll have to disable split on arm, unless CONFIG_CPU_CACHE_VIPT to ensures adjust_pte never used. Signed-off-by:
-
Hugh Dickins authored
Use pte_offset_map_lock, instead of pte_offset_map (or inappropriate pte_offset_kernel) and mm-wide page_table_lock, in sundry arch places. The i386 vm86 mark_screen_rdonly: yes, there was and is an assumption that the screen fits inside the one page table, as indeed it does. The sh __do_page_fault: which handles both kernel faults (without lock) and user mm faults (locked - though it set_pte without locking before). The sh64 flush_cache_range and helpers: which wrongly thought callers held page_table_lock before (only its tlb_start_vma did, and no longer does so); moved the flush loop down, and adjusted the large versus small range decision to consider a range which spans page tables as large. Signed-off-by:
Paul Mundt <lethal@linux-sh.org> Signed-off-by:
-
Hugh Dickins authored
Final step in pushing down common core's page_table_lock. follow_page no longer wants caller to hold page_table_lock, uses pte_offset_map_lock itself; and so no page_table_lock is taken in get_user_pages itself. But get_user_pages (and get_futex_key) do then need follow_page to pin the page for them: take Daniel's suggestion of bitflags to follow_page. Need one for WRITE, another for TOUCH (it was the accessed flag before: vanished along with check_user_page_readable, but surely get_numa_maps is wrong to mark every page it finds as accessed), another for GET. And another, ANON to dispose of untouched_anonymous_page: it seems silly for that to descend a second time, let follow_page observe if there was no page table and return ZERO_PAGE if so. Fix minor bug in that: check VM_LOCKED - make_pages_present ought to make readonly anonymous present. Give get_numa_maps a cond_resched while we're there. Signed-off-by:
-
Hugh Dickins authored
check_user_page_readable is a problematic variant of follow_page. It's used only by oprofile's i386 and arm backtrace code, at interrupt time, to establish whether a userspace stackframe is currently readable. This is problematic, because we want to push the page_table_lock down inside follow_page, and later split it; whereas oprofile is doing a spin_trylock on it (in the i386 case, forgotten in the arm case), and needs that to pin perhaps two pages spanned by the stackframe (which might be covered by different locks when we split). I think oprofile is going about this in the wrong way: it doesn't need to know the area is readable (neither i386 nor arm uses read protection of user pages), it doesn't need to pin the memory, it should simply __copy_from_user_inatomic, and see if that succeeds or not. Sorry, but I've not got around to devising the sparse __user annotations for this. Then we can eliminate check_user_page_readable, and return to a single follow_page without the __follow_page variants. Signed-off-by:
-
Hugh Dickins authored
rmap's page_check_address descend without page_table_lock. First just pte_offset_map in case there's no pte present worth locking for, then take page_table_lock for the full check, and pass ptl back to caller in the same style as pte_offset_map_lock. __xip_unmap, page_referenced_one and try_to_unmap_one use pte_unmap_unlock. try_to_unmap_cluster also. page_check_address reformatted to avoid progressive indentation. No use is made of its one error code, return NULL when it fails. Signed-off-by:
-
Hugh Dickins authored
Small fix to the PageReserved patch: the mips ZERO_PAGE(address) depends on address, so __xip_unmap is wrong to initialize page with that before address is initialized; and in fact must re-evaluate it each iteration. Signed-off-by:
-
Hugh Dickins authored
Remove the page_table_lock from around the calls to unmap_vmas, and replace the pte_offset_map in zap_pte_range by pte_offset_map_lock: all callers are now safe to descend without page_table_lock. Don't attempt fancy locking for hugepages, just take page_table_lock in unmap_hugepage_range. Which makes zap_hugepage_range, and the hugetlb test in zap_page_range, redundant: unmap_vmas calls unmap_hugepage_range anyway. Nor does unmap_vmas have much use for its mm arg now. The tlb_start_vma and tlb_end_vma in unmap_page_range are now called without page_table_lock: if they're implemented at all, they typically come down to flush_cache_range (usually done outside page_table_lock) and flush_tlb_range (which we already audited for the mprotect case). Signed-off-by:
-
Hugh Dickins authored
In most places the descent from pgd to pud to pmd to pte holds mmap_sem (exclusively or not), which ensures that free_pgtables cannot be freeing page tables from any level at the same time. But truncation and reverse mapping descend without mmap_sem. No problem: just make sure that a vma is unlinked from its prio_tree (or nonlinear list) and from its anon_vma list, after zapping the vma, but before freeing its page tables. Then neither vmtruncate nor rmap can reach that vma whose page tables are now volatile (nor do they need to reach it, since all its page entries have been zapped by this stage). The i_mmap_lock and anon_vma->lock already serialize this correctly; but the locking hierarchy is such that we cannot take them while holding page_table_lock. Well, we're trying to push that down anyway. So in this patch, move anon_vma_unlink and unlink_file_vma into free_pgtables, at the same time as moving page_table_lock around calls to unmap_vmas. tlb_gather_mmu and tlb_finish_mmu then fall outside the page_table_lock, but we made them preempt_disable and preempt_enable earlier; and a long source audit of all the architectures has shown no problem with removing page_table_lock from them. free_pgtables doesn't need page_table_lock for itself, nor for what it calls; tlb->mm->nr_ptes is usually protected by page_table_lock, but partly by non-exclusive mmap_sem - here it's decremented with exclusive mmap_sem, or mm_users 0. update_hiwater_rss and vm_unacct_memory don't need page_table_lock either. Signed-off-by:
-
Hugh Dickins authored
There was one small but very significant change in the previous patch: mprotect's flush_tlb_range fell outside the page_table_lock: as it is in 2.4, but that doesn't prove it safe in 2.6. On some architectures flush_tlb_range comes to the same as flush_tlb_mm, which has always been called from outside page_table_lock in dup_mmap, and is so proved safe. Others required a deeper audit: I could find no reliance on page_table_lock in any; but in ia64 and parisc found some code which looks a bit as if it might want preemption disabled. That won't do any actual harm, so pending a decision from the maintainers, disable preemption there. Remove comments on page_table_lock from flush_tlb_mm, flush_tlb_range and flush_tlb_page entries in cachetlb.txt: they were rather misleading (what generic code does is different from what usually happens), the rules are now changing, and it's not yet clear where we'll end up (will the generic tlb_flush_mmu happen always under lock? never under lock? or sometimes under and sometimes not?). Signed-off-by:
-
Hugh Dickins authored
Convert those common loops using page_table_lock on the outside and pte_offset_map within to use just pte_offset_map_lock within instead. These all hold mmap_sem (some exclusively, some not), so at no level can a page table be whipped away from beneath them. But whereas pte_alloc loops tested with the "atomic" pmd_present, these loops are testing with pmd_none, which on i386 PAE tests both lower and upper halves. That's now unsafe, so add a cast into pmd_none to test only the vital lower half: we lose a little sensitivity to a corrupt middle directory, but not enough to worry about. It appears that i386 and UML were the only architectures vulnerable in this way, and pgd and pud no problem. Signed-off-by:
-
Hugh Dickins authored
On the page fault path, the patch before last pushed acquiring the page_table_lock down to the head of handle_pte_fault (though it's also taken and dropped earlier when a new page table has to be allocated). Now delete that line, read "entry = *pte" without it, and go off to this or that page fault handler on the basis of this unlocked peek. Usually the handler can proceed without the lock, relying on the subsequent locked pte_same or pte_none test to back out when necessary; though do_wp_page needs the lock immediately, and do_file_page doesn't check (if there's a race, install_page just zaps the entry and reinstalls it). But on those architectures (notably i386 with PAE) whose pte is too big to be read atomically, if SMP or preemption is enabled, do_swap_page and do_file_page might cause irretrievable damage if passed a Frankenstein entry stitched together from unrelated parts. In those configs, "pte_unmap_same" has to take page_table_lock, validate orig_pte still the same, and drop page_table_lock before unmapping, before proceeding. Use pte_offset_map_lock and pte_unmap_unlock throughout the handlers; but lock avoidance leaves more lone maps and unmaps than elsewhere. Signed-off-by:
-
Hugh Dickins authored
Convert those few architectures which are calling pud_alloc, pmd_alloc, pte_alloc_map on a user mm, not to take the page_table_lock first, nor drop it after. Each of these can continue to use pte_alloc_map, no need to change over to pte_alloc_map_lock, they're neither racy nor swappable. In the sparc64 io_remap_pfn_range, flush_tlb_range then falls outside of the page_table_lock: that's okay, on sparc64 it's like flush_tlb_mm, and that has always been called from outside of page_table_lock in dup_mmap. Signed-off-by:
-
Hugh Dickins authored
Second step in pushing down the page_table_lock. Remove the temporary bridging hack from __pud_alloc, __pmd_alloc, __pte_alloc: expect callers not to hold page_table_lock, whether it's on init_mm or a user mm; take page_table_lock internally to check if a racing task already allocated. Convert their callers from common code. But avoid coming back to change them again later: instead of moving the spin_lock(&mm->page_table_lock) down, switch over to new macros pte_alloc_map_lock and pte_unmap_unlock, which encapsulate the mapping+locking and unlocking+unmapping together, and in the end may use alternatives to the mm page_table_lock itself. These callers all hold mmap_sem (some exclusively, some not), so at no level can a page table be whipped away from beneath them; and pte_alloc uses the "atomic" pmd_present to test whether it needs to allocate. It appears that on all arches we can safely descend without page_table_lock. Signed-off-by:
-
Hugh Dickins authored
It seems odd to me that, whereas pud_alloc and pmd_alloc test inline, only calling out-of-line __pud_alloc __pmd_alloc if allocation needed, pte_alloc_map and pte_alloc_kernel are entirely out-of-line. Though it does add a little to kernel size, change them to macros testing inline, calling __pte_alloc or __pte_alloc_kernel to allocate out-of-line. Mark none of them as fastcalls, leave that to CONFIG_REGPARM or not. It also seems more natural for the out-of-line functions to leave the offset calculation and map to the inline, which has to do it anyway for the common case. At least mremap move wants __pte_alloc without _map. Macros rather than inline functions, certainly to avoid the header file issues which arise from CONFIG_HIGHPTE needing kmap_types.h, but also in case any architectures I haven't built would have other such problems. Signed-off-by:
-
Hugh Dickins authored
First step in pushing down the page_table_lock. init_mm.page_table_lock has been used throughout the architectures (usually for ioremap): not to serialize kernel address space allocation (that's usually vmlist_lock), but because pud_alloc,pmd_alloc,pte_alloc_kernel expect caller holds it. Reverse that: don't lock or unlock init_mm.page_table_lock in any of the architectures; instead rely on pud_alloc,pmd_alloc,pte_alloc_kernel to take and drop it when allocating a new one, to check lest a racing task already did. Similarly no page_table_lock in vmalloc's map_vm_area. Some temporary ugliness in __pud_alloc and __pmd_alloc: since they also handle user mms, which are converted only by a later patch, for now they have to lock differently according to whether or not it's init_mm. If sources get muddled, there's a danger that an arch source taking init_mm.page_table_lock will be mixed with common source also taking it (or neither take it). So break the rules and make another change, which should break the build for such a mismatch: remove the redundant mm arg from pte_alloc_kernel (ppc64 scrapped its distinct ioremap_mm in 2.6.13). Exceptions: arm26 used pte_alloc_kernel on user mm, now pte_alloc_map; ia64 used pte_alloc_map on init_mm, now pte_alloc_kernel; parisc had bad args to pmd_alloc and pte_alloc_kernel in unused USE_HPPA_IOREMAP code; ppc64 map_io_page forgot to unlock on failure; ppc mmu_mapin_ram and ppc64 im_free took page_table_lock for no good reason. Signed-off-by:
-
Hugh Dickins authored
ia64 has expand_backing_store function for growing its Register Backing Store vma upwards. But more complete code for this purpose is found in the CONFIG_STACK_GROWSUP part of mm/mmap.c. Uglify its #ifdefs further to provide expand_upwards for ia64 as well as expand_stack for parisc. The Register Backing Store vma should be marked VM_ACCOUNT. Implement the intention of growing it only a page at a time, instead of passing an address outside of the vma to handle_mm_fault, with unknown consequences. Signed-off-by:
-
Hugh Dickins authored
Slight and timid rearrangement of mm_struct: hiwater_rss and hiwater_vm were tacked on the end, but it seems better to keep them near _file_rss, _anon_rss and total_vm, in the same cacheline on those arches verified. There are likely to be more profitable rearrangements, but less obvious (is it good or bad that saved_auxv[AT_VECTOR_SIZE] isolates cpu_vm_mask and context from many others?), needing serious instrumentation. Signed-off-by:
-
Hugh Dickins authored
update_mem_hiwater has attracted various criticisms, in particular from those concerned with mm scalability. Originally it was called whenever rss or total_vm got raised. Then many of those callsites were replaced by a timer tick call from account_system_time. Now Frank van Maarseveen reports that to be found inadequate. How about this? Works for Frank. Replace update_mem_hiwater, a poor combination of two unrelated ops, by macros update_hiwater_rss and update_hiwater_vm. Don't attempt to keep mm->hiwater_rss up to date at timer tick, nor every time we raise rss (usually by 1): those are hot paths. Do the opposite, update only when about to lower rss (usually by many), or just before final accounting in do_exit. Handle mm->hiwater_vm in the same way, though it's much less of an issue. Demand that whoever collects these hiwater statistics do the work of taking the maximum with rss or total_vm. And there has been no collector of these hiwater statistics in the tree. The new convention needs an example, so match Frank's usage by adding a VmPeak line above VmSize to /proc/<pid>/status, and also a VmHWM line above VmRSS (High-Water-Mark or High-Water-Memory). There was a particular anomaly during mremap move, that hiwater_vm might be captured too high. A fleeting such anomaly remains, but it's quickly corrected now, whereas before it would stick. What locking? None: if the app is racy then these statistics will be racy, it's not worth any overhead to make them exact. But whenever it suits, hiwater_vm is updated under exclusive mmap_sem, and hiwater_rss under page_table_lock (for now) or with preemption disabled (later on): without going to any trouble, minimize the time between reading current values and updating, to minimize those occasions when a racing thread bumps a count up and back down in between. Signed-off-by:
-
Hugh Dickins authored
There used to be just one call to zap_pte, but it shouldn't be inline now there are two. Check for the common case pte_none before calling, and move its rss accounting up into install_page or install_file_pte - which helps the next patch. Signed-off-by:
-
Hugh Dickins authored
Cleanup: relieve do_mremap from its surfeit of current->mms. Signed-off-by:
-
Hugh Dickins authored
Small adjustment: do_swap_page should report its !pte_same race as a major fault if it had to read into swap cache, because whatever raced with it will have found page already in cache and reported minor fault. Signed-off-by:
-
Hugh Dickins authored
Small adjustment: zap_pte_range decrement its rss counts from 0 then finally add, avoiding negations - we don't have or need a sub_mm_rss. Signed-off-by:
-
Hugh Dickins authored
Small adjustment, following Nick's suggestion: it's more straightforward for copy_pte_range to let copy_one_pte do the rss incrementation, than use an index it passed back. Saves a #define, and 16 bytes of .text. Signed-off-by:
-
Nick Piggin authored
Remove PageReserved() calls from core code by tightening VM_RESERVED handling in mm/ to cover PageReserved functionality. PageReserved special casing is removed from get_page and put_page. All setting and clearing of PageReserved is retained, and it is now flagged in the page_alloc checks to help ensure we don't introduce any refcount based freeing of Reserved pages. MAP_PRIVATE, PROT_WRITE of VM_RESERVED regions is tentatively being deprecated. We never completely handled it correctly anyway, and is be reintroduced in future if required (Hugh has a proof of concept). Once PageReserved() calls are removed from kernel/power/swsusp.c, and all arch/ and driver code, the Set and Clear calls, and the PG_reserved bit can be trivially removed. Last real user of PageReserved is swsusp, which uses PageReserved to determine whether a struct page points to valid memory or not. This still needs to be addressed (a generic page_is_ram() should work). A last caveat: the ZERO_PAGE is now refcounted and managed with rmap (and thus mapcounted and count towards shared rss). These writes to the struct page could cause excessive cacheline bouncing on big systems. There are a number of ways this could be addressed if it is an issue. Signed-off-by:
Nick Piggin <npiggin@suse.de> Refcount bug fix for filemap_xip.c Signed-off-by:
Carsten Otte <cotte@de.ibm.com> Signed-off-by:
-
Hugh Dickins authored
Please, please now delete the Atari CONFIG_STRAM_SWAP code. It may be excellent and ingenious code, but its reference to swap_vfsmnt betrays that it hasn't been built since 2.5.1 (four years old come December), it's delving deep into matters which are the preserve of core mm code, its only purpose is to give the more conscientious mm guys an anxiety attack from time to time; yet we keep on breaking it more and more. If you want to use RAM for swap, then if the MTD driver does not already provide just what you need, I'm sure David could be persuaded to add the extra. But you'd also like to be able to allocate extents of that swap for other use: we can give you a core interface for that if you need. But unbuilt for four years suggests to me that there's no need at all. I cannot swear the patch below won't break your build, but believe so. Signed-off-by:
-
