The patch titled
uninline page_cache_get_speculative()
has been removed from the -mm tree. Its filename is
mm-speculative-get_page-uninlining.patch
This patch was dropped because it is obsolete
------------------------------------------------------
Subject: uninline page_cache_get_speculative()
From: Andrew Morton <akpm@xxxxxxxx>
Shrinks my SMP kernel by ~900 bytes. In the fastpath.
This is so obvious, I have a feeling I'm missing something.
Cc: Nick Piggin <nickpiggin@xxxxxxxxxxxx>
Cc: Hugh Dickins <hugh@xxxxxxxxxxx>
Signed-off-by: Andrew Morton <akpm@xxxxxxxx>
---
include/linux/pagemap.h | 102 --------------------------------------
mm/swap.c | 101 +++++++++++++++++++++++++++++++++++++
2 files changed, 102 insertions(+), 101 deletions(-)
diff -puN include/linux/pagemap.h~mm-speculative-get_page-uninlining include/linux/pagemap.h
--- a/include/linux/pagemap.h~mm-speculative-get_page-uninlining
+++ a/include/linux/pagemap.h
@@ -52,107 +52,7 @@ static inline void mapping_set_gfp_mask(
#define page_cache_get(page) get_page(page)
#define page_cache_release(page) put_page(page)
void release_pages(struct page **pages, int nr, int cold);
-
-/*
- * speculatively take a reference to a page.
- * If the page is free (_count == 0), then _count is untouched, and NULL
- * is returned. Otherwise, _count is incremented by 1 and page is returned.
- *
- * This function must be run in the same rcu_read_lock() section as has
- * been used to lookup the page in the pagecache radix-tree: this allows
- * allocators to use a synchronize_rcu() to stabilize _count.
- *
- * Unless an RCU grace period has passed, the count of all pages coming out
- * of the allocator must be considered unstable. page_count may return higher
- * than expected, and put_page must be able to do the right thing when the
- * page has been finished with (because put_page is what is used to drop an
- * invalid speculative reference).
- *
- * After incrementing the refcount, this function spins until PageNoNewRefs
- * is clear, then a read memory barrier is issued.
- *
- * This forms the core of the lockless pagecache locking protocol, where
- * the lookup-side (eg. find_get_page) has the following pattern:
- * 1. find page in radix tree
- * 2. conditionally increment refcount
- * 3. wait for PageNoNewRefs
- * 4. check the page is still in pagecache
- *
- * Remove-side (that cares about _count, eg. reclaim) has the following:
- * A. SetPageNoNewRefs
- * B. check refcount is correct
- * C. remove page
- * D. ClearPageNoNewRefs
- *
- * There are 2 critical interleavings that matter:
- * - 2 runs before B: in this case, B sees elevated refcount and bails out
- * - B runs before 2: in this case, 3 ensures 4 will not run until *after* C
- * (after D, even). In which case, 4 will notice C and lookup side can retry
- *
- * It is possible that between 1 and 2, the page is removed then the exact same
- * page is inserted into the same position in pagecache. That's OK: the
- * old find_get_page using tree_lock could equally have run before or after
- * the write-side, depending on timing.
- *
- * Pagecache insertion isn't a big problem: either 1 will find the page or
- * it will not. Likewise, the old find_get_page could run either before the
- * insertion or afterwards, depending on timing.
- */
-static inline struct page *page_cache_get_speculative(struct page *page)
-{
- VM_BUG_ON(in_interrupt());
-
-#ifndef CONFIG_SMP
- VM_BUG_ON(!in_atomic());
- /*
- * Preempt must be disabled here - we rely on rcu_read_lock doing
- * this for us.
- *
- * Pagecache won't be truncated from interrupt context, so if we have
- * found a page in the radix tree here, we have pinned its refcount by
- * disabling preempt, and hence no need for the "speculative get" that
- * SMP requires.
- */
- VM_BUG_ON(page_count(page) == 0);
- atomic_inc(&page->_count);
-
-#else
- if (unlikely(!get_page_unless_zero(page)))
- return NULL; /* page has been freed */
-
- /*
- * Note that get_page_unless_zero provides a memory barrier.
- * This is needed to ensure PageNoNewRefs is evaluated after the
- * page refcount has been raised. See below comment.
- */
-
- while (unlikely(PageNoNewRefs(page)))
- cpu_relax();
-
- /*
- * smp_rmb is to ensure the load of page->flags (for PageNoNewRefs())
- * is performed before a future load used to ensure the page is
- * the correct on (usually: page->mapping and page->index).
- *
- * Those places that set PageNoNewRefs have the following pattern:
- * SetPageNoNewRefs(page)
- * wmb();
- * if (page_count(page) == X)
- * remove page from pagecache
- * wmb();
- * ClearPageNoNewRefs(page)
- *
- * If the load was out of order, page->mapping might be loaded before
- * the page is removed from pagecache but PageNoNewRefs evaluated
- * after the ClearPageNoNewRefs().
- */
- smp_rmb();
-
-#endif
- VM_BUG_ON(PageCompound(page) && (struct page *)page_private(page) != page);
-
- return page;
-}
+struct page *page_cache_get_speculative(struct page *page);
#ifdef CONFIG_NUMA
extern struct page *page_cache_alloc(struct address_space *x);
diff -puN mm/swap.c~mm-speculative-get_page-uninlining mm/swap.c
--- a/mm/swap.c~mm-speculative-get_page-uninlining
+++ a/mm/swap.c
@@ -74,6 +74,107 @@ void put_page(struct page *page)
EXPORT_SYMBOL(put_page);
/*
+ * speculatively take a reference to a page.
+ * If the page is free (_count == 0), then _count is untouched, and NULL
+ * is returned. Otherwise, _count is incremented by 1 and page is returned.
+ *
+ * This function must be run in the same rcu_read_lock() section as has
+ * been used to lookup the page in the pagecache radix-tree: this allows
+ * allocators to use a synchronize_rcu() to stabilize _count.
+ *
+ * Unless an RCU grace period has passed, the count of all pages coming out
+ * of the allocator must be considered unstable. page_count may return higher
+ * than expected, and put_page must be able to do the right thing when the
+ * page has been finished with (because put_page is what is used to drop an
+ * invalid speculative reference).
+ *
+ * After incrementing the refcount, this function spins until PageNoNewRefs
+ * is clear, then a read memory barrier is issued.
+ *
+ * This forms the core of the lockless pagecache locking protocol, where
+ * the lookup-side (eg. find_get_page) has the following pattern:
+ * 1. find page in radix tree
+ * 2. conditionally increment refcount
+ * 3. wait for PageNoNewRefs
+ * 4. check the page is still in pagecache
+ *
+ * Remove-side (that cares about _count, eg. reclaim) has the following:
+ * A. SetPageNoNewRefs
+ * B. check refcount is correct
+ * C. remove page
+ * D. ClearPageNoNewRefs
+ *
+ * There are 2 critical interleavings that matter:
+ * - 2 runs before B: in this case, B sees elevated refcount and bails out
+ * - B runs before 2: in this case, 3 ensures 4 will not run until *after* C
+ * (after D, even). In which case, 4 will notice C and lookup side can retry
+ *
+ * It is possible that between 1 and 2, the page is removed then the exact same
+ * page is inserted into the same position in pagecache. That's OK: the
+ * old find_get_page using tree_lock could equally have run before or after
+ * the write-side, depending on timing.
+ *
+ * Pagecache insertion isn't a big problem: either 1 will find the page or
+ * it will not. Likewise, the old find_get_page could run either before the
+ * insertion or afterwards, depending on timing.
+ */
+struct page *page_cache_get_speculative(struct page *page)
+{
+ VM_BUG_ON(in_interrupt());
+
+#ifndef CONFIG_SMP
+ VM_BUG_ON(!in_atomic());
+ /*
+ * Preempt must be disabled here - we rely on rcu_read_lock doing
+ * this for us.
+ *
+ * Pagecache won't be truncated from interrupt context, so if we have
+ * found a page in the radix tree here, we have pinned its refcount by
+ * disabling preempt, and hence no need for the "speculative get" that
+ * SMP requires.
+ */
+ VM_BUG_ON(page_count(page) == 0);
+ atomic_inc(&page->_count);
+
+#else
+ if (unlikely(!get_page_unless_zero(page)))
+ return NULL; /* page has been freed */
+
+ /*
+ * Note that get_page_unless_zero provides a memory barrier.
+ * This is needed to ensure PageNoNewRefs is evaluated after the
+ * page refcount has been raised. See below comment.
+ */
+
+ while (unlikely(PageNoNewRefs(page)))
+ cpu_relax();
+
+ /*
+ * smp_rmb is to ensure the load of page->flags (for PageNoNewRefs())
+ * is performed before a future load used to ensure the page is
+ * the correct on (usually: page->mapping and page->index).
+ *
+ * Those places that set PageNoNewRefs have the following pattern:
+ * SetPageNoNewRefs(page)
+ * wmb();
+ * if (page_count(page) == X)
+ * remove page from pagecache
+ * wmb();
+ * ClearPageNoNewRefs(page)
+ *
+ * If the load was out of order, page->mapping might be loaded before
+ * the page is removed from pagecache but PageNoNewRefs evaluated
+ * after the ClearPageNoNewRefs().
+ */
+ smp_rmb();
+
+#endif
+ VM_BUG_ON(PageCompound(page) && (struct page *)page_private(page) != page);
+
+ return page;
+}
+
+/*
* Writeback is about to end against a page which has been marked for immediate
* reclaim. If it still appears to be reclaimable, move it to the tail of the
* inactive list. The page still has PageWriteback set, which will pin it.
_
Patches currently in -mm which might be from akpm@xxxxxxxx are
origin.patch
convert-i386-numa-kva-space-to-bootmem.patch
reduce-max_nr_zones-make-display-of-highmem-counters-conditional-on-config_highmem.patch
reduce-max_nr_zones-page-allocator-zone_highmem-cleanup.patch
reduce-max_nr_zones-use-enum-to-define-zones-reformat-and-comment.patch
reduce-max_nr_zones-remove-display-of-counters-for-unconfigured-zones.patch
out-of-memory-notifier.patch
mm-swap-write-failure-fixup.patch
slab-optimize-kmalloc_node-the-same-way-as-kmalloc-fix.patch
slab-fix-lockdep-warnings-fix-2.patch
have-x86-use-add_active_range-and-free_area_init_nodes-fix.patch
zone_reclaim-dynamic-slab-reclaim-tidy.patch
vm-add-per-zone-writeout-counter.patch
convert-s390-page-handling-macros-to-functions-fix.patch
slab-fix-kmalloc_node-applying-memory-policies-if-nodeid-==-numa_node_id-fix.patch
gfp_thisnode-for-the-slab-allocator-v2-fix.patch
get-rid-of-zone_table-fix.patch
do_no_pfn-tweaks.patch
shared-page-table-for-hugetlb-page-v2-tidy.patch
radix-tree-rcu-lockless-readside.patch
acx1xx-wireless-driver.patch
tiacx-pci-build-fix.patch
tiacx-ia64-fix.patch
tiacx-build-fix.patch
binfmt_elf-consistently-use-loff_t.patch
nommu-set-bdi-capabilities-for-dev-mem-and-dev-kmem-tidy.patch
nommu-move-the-fallback-arch_vma_name-to-a-sensible-place-fix.patch
convert-i386-summit-subarch-to-use-srat-info-for-apicid_to_node-calls-tidy.patch
i386-adds-smp_call_function_single-fix.patch
swsusp-write-timer.patch
swsusp-write-speedup.patch
swsusp-read-timer.patch
swsusp-read-speedup.patch
swsusp-read-speedup-fix.patch
swsusp-read-speedup-cleanup.patch
swsusp-read-speedup-cleanup-2.patch
swsusp-read-speedup-fix-fix-2.patch
deprecate-smbfs-in-favour-of-cifs.patch
edac-new-opteron-athlon64-memory-controller-driver-tidy.patch
inode_diet-replace-inodeugeneric_ip-with-inodei_private-gfs-fix.patch
inode-diet-eliminate-i_blksize-and-use-a-per-superblock-default-fix-fix.patch
x86-microcode-microcode-driver-cleanup-tidy.patch
x86-microcode-add-sysfs-and-hotplug-support-fix.patch
eisa-bus-modalias-attributes-support-1-fix-git-kbuild-fix.patch
add-address_space_operationsbatch_write-fix.patch
alloc_fdtable-cleanup.patch
sysctl-allow-proc-sys-without-sys_sysctl-fix.patch
add-probe_kernel_address.patch
x86-use-probe_kernel_address-in-handle_bug.patch
blockdevc-check-errors.patch
let-warn_on-warn_on_once-return-the-condition-fix.patch
let-warn_on-warn_on_once-return-the-condition-fix-2.patch
omap-add-watchdog-driver-support-tweaks.patch
move-valid_dma_direction-from-x86_64-to-generic-code-fix.patch
single-bit-flip-detector-tidy.patch
fix-unserialized-task-files-changing-fix.patch
tty-make-termios_sem-a-mutex-fix.patch
solaris-emulation-incorrect-tty-locking-fix.patch
solaris-emulation-incorrect-tty-locking-fix-2.patch
cpuset-top_cpuset-tracks-hotplug-changes-to-node_online_map-fix.patch
remove-sound-oss-copying.patch
maximum-latency-tracking-infrastructure-tidy.patch
fs-nameic-replace-multiple-current-fs-by-shortcut-variable-tidy.patch
call-mm-page-writebackcset_ratelimit-when-new-pages-tidy.patch
sys-modules-patch-allow-full-length-section-names.patch
expand_fdtable-remove-pointless-unlocklock.patch
add-config_headers_check-option-to-automatically-run-make-headers_check-nobble.patch
submit-checklist-mention-headers_check.patch
ntp-move-all-the-ntp-related-code-to-ntpc-fix.patch
reiserfs-on-demand-bitmap-loading.patch
streamline-generic_file_-interfaces-and-filemap-gfs-fix.patch
add-vector-aio-support-fix.patch
csa-basic-accounting-over-taskstats-fix.patch
fs-cache-make-kafs-use-fs-cache-fix.patch
fs-cache-make-kafs-use-fs-cache-vs-streamline-generic_file_-interfaces-and-filemap.patch
nfs-use-local-caching-12-fix.patch
stack-overflow-safe-kdump-crash_use_safe_smp_processor_id-fix.patch
generic-ioremap_page_range-x86_64-conversion-fix.patch
vfs-make-filldir_t-and-struct-kstat-deal-in-64-bit-inode-numbers-alpha-fix.patch
some-cleanup-in-the-pipe-code-tidy.patch
support-piping-into-commands-in-proc-sys-kernel-core_pattern-fix.patch
move-pidmap-to-pspaceh-fix.patch
kprobes-handle-symbol-resolution-when-modulesymbol-is-specified-tidy.patch
isdn-work-around-excessive-udelay.patch
knfsd-add-a-callback-for-when-last-rpc-thread-finishes-tidy.patch
knfsd-add-a-callback-for-when-last-rpc-thread-finishes-fix.patch
knfsd-separate-out-some-parts-of-nfsd_svc-which-start-nfs-servers-tweaks.patch
knfsd-define-new-nfsdfs-file-portlist-contains-list-of-ports-tidy.patch
knfsd-define-new-nfsdfs-file-portlist-contains-list-of-ports-fix.patch
knfsd-drop-serv-option-to-svc_recv-and-svc_process-nfs-callback-fix-nfs-callback-fix.patch
knfsd-move-tempsock-aging-to-a-timer-tidy.patch
sched-remove-unnecessary-sched-group-allocations-fix.patch
swap_prefetch-vs-zoned-counters.patch
ecryptfs-mmap-operations.patch
ecryptfs-alpha-build-fix.patch
ecryptfs-more-elegant-aes-key-size-manipulation.patch
ecryptfs-get_sb_dev-fix.patch
make-kmem_cache_destroy-return-void-ecryptfs.patch
ecryptfs-versioning-fixes-tidy.patch
namespaces-add-nsproxy.patch
namespaces-utsname-switch-to-using-uts-namespaces.patch
namespaces-utsname-switch-to-using-uts-namespaces-klibc-bit-sparc.patch
namespaces-utsname-use-init_utsname-when-appropriate.patch
namespaces-utsname-implement-utsname-namespaces.patch
namespaces-utsname-sysctl-hack.patch
ipc-namespace-core.patch
rename-the-provided-execve-functions-to-kernel_execve-headers-fix.patch
replace-cad_pid-by-a-struct-pid-fixes.patch
readahead-sysctl-parameters-fix.patch
make-copy_from_user_inatomic-not-zero-the-tail-on-i386-vs-reiser4.patch
make-kmem_cache_destroy-return-void-reiser4.patch
reiser4-hardirq-include-fix.patch
reiser4-run-truncate_inode_pages-in-reiser4_delete_inode.patch
reiser4-get_sb_dev-fix.patch
reiser4-vs-zoned-allocator.patch
reiser4-rename-generic_sounding_globalspatch-fix.patch
hpt3xx-rework-rate-filtering-tidy.patch
genirq-convert-the-i386-architecture-to-irq-chips.patch
genirq-x86_64-irq-reenable-migrating-irqs-to-other-cpus.patch
genirq-msi-simplify-msi-enable-and-disable.patch
genirq-ia64-irq-dynamic-irq-support.patch
genirq-msi-only-build-msi-apicc-on-ia64-fix.patch
genirq-i386-irq-remove-the-msi-assumption-that-irq-==-vector.patch
genirq-x86_64-irq-make-vector_irq-per-cpu-fix.patch
genirq-x86_64-irq-make-vector_irq-per-cpu-warning-fix.patch
add-hypertransport-capability-defines-fix.patch
initial-generic-hypertransport-interrupt-support-Kconfig-fix.patch
srcu-report-out-of-memory-errors-fixlet.patch
isdn-debug-build-fix.patch
isdn-more-pr_debug-fixes.patch
nr_blockdev_pages-in_interrupt-warning.patch
device-suspend-debug.patch
slab-leaks3-default-y.patch
x86-kmap_atomic-debugging.patch
restore-rogue-readahead-printk.patch
put_bh-debug.patch
acpi_format_exception-debug.patch
jmicron-warning-fix.patch
-
To unsubscribe from this list: send the line "unsubscribe mm-commits" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
