~hc/RK356X_SDK_RELEASE.git

..	..	@@ -28,7 +28,10 @@
28	28	SCAN_SUCCEED,
29	29	SCAN_PMD_NULL,
30	30	SCAN_EXCEED_NONE_PTE,
	31	+ SCAN_EXCEED_SWAP_PTE,
	32	+ SCAN_EXCEED_SHARED_PTE,
31	33	SCAN_PTE_NON_PRESENT,
	34	+ SCAN_PTE_UFFD_WP,
32	35	SCAN_PAGE_RO,
33	36	SCAN_LACK_REFERENCED_PAGE,
34	37	SCAN_PAGE_NULL,
..	..	@@ -46,8 +49,8 @@
46	49	SCAN_DEL_PAGE_LRU,
47	50	SCAN_ALLOC_HUGE_PAGE_FAIL,
48	51	SCAN_CGROUP_CHARGE_FAIL,
49		- SCAN_EXCEED_SWAP_PTE,
50	52	SCAN_TRUNCATED,
	53	+ SCAN_PAGE_HAS_PRIVATE,
51	54	};
52	55
53	56	#define CREATE_TRACE_POINTS
..	..	@@ -73,11 +76,14 @@
73	76	*/
74	77	static unsigned int khugepaged_max_ptes_none __read_mostly;
75	78	static unsigned int khugepaged_max_ptes_swap __read_mostly;
	79	+static unsigned int khugepaged_max_ptes_shared __read_mostly;
76	80
77	81	#define MM_SLOTS_HASH_BITS 10
78	82	static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
79	83
80	84	static struct kmem_cache *mm_slot_cache __read_mostly;
	85	+
	86	+#define MAX_PTE_MAPPED_THP 8
81	87
82	88	/**
83	89	* struct mm_slot - hash lookup from mm to mm_slot
..	..	@@ -89,6 +95,10 @@
89	95	struct hlist_node hash;
90	96	struct list_head mm_node;
91	97	struct mm_struct *mm;
	98	+
	99	+ /* pte-mapped THP in this mm */
	100	+ int nr_pte_mapped_thp;
	101	+ unsigned long pte_mapped_thp[MAX_PTE_MAPPED_THP];
92	102	};
93	103
94	104	/**
..	..	@@ -286,15 +296,43 @@
286	296	__ATTR(max_ptes_swap, 0644, khugepaged_max_ptes_swap_show,
287	297	khugepaged_max_ptes_swap_store);
288	298
	299	+static ssize_t khugepaged_max_ptes_shared_show(struct kobject *kobj,
	300	+ struct kobj_attribute *attr,
	301	+ char *buf)
	302	+{
	303	+ return sprintf(buf, "%u\n", khugepaged_max_ptes_shared);
	304	+}
	305	+
	306	+static ssize_t khugepaged_max_ptes_shared_store(struct kobject *kobj,
	307	+ struct kobj_attribute *attr,
	308	+ const char *buf, size_t count)
	309	+{
	310	+ int err;
	311	+ unsigned long max_ptes_shared;
	312	+
	313	+ err = kstrtoul(buf, 10, &max_ptes_shared);
	314	+ if (err \|\| max_ptes_shared > HPAGE_PMD_NR-1)
	315	+ return -EINVAL;
	316	+
	317	+ khugepaged_max_ptes_shared = max_ptes_shared;
	318	+
	319	+ return count;
	320	+}
	321	+
	322	+static struct kobj_attribute khugepaged_max_ptes_shared_attr =
	323	+ __ATTR(max_ptes_shared, 0644, khugepaged_max_ptes_shared_show,
	324	+ khugepaged_max_ptes_shared_store);
	325	+
289	326	static struct attribute *khugepaged_attr[] = {
290	327	&khugepaged_defrag_attr.attr,
291	328	&khugepaged_max_ptes_none_attr.attr,
	329	+ &khugepaged_max_ptes_swap_attr.attr,
	330	+ &khugepaged_max_ptes_shared_attr.attr,
292	331	&pages_to_scan_attr.attr,
293	332	&pages_collapsed_attr.attr,
294	333	&full_scans_attr.attr,
295	334	&scan_sleep_millisecs_attr.attr,
296	335	&alloc_sleep_millisecs_attr.attr,
297		- &khugepaged_max_ptes_swap_attr.attr,
298	336	NULL,
299	337	};
300	338
..	..	@@ -303,8 +341,6 @@
303	341	.name = "khugepaged",
304	342	};
305	343	#endif /* CONFIG_SYSFS */
306		-
307		-#define VM_NO_KHUGEPAGED (VM_SPECIAL \| VM_HUGETLB)
308	344
309	345	int hugepage_madvise(struct vm_area_struct *vma,
310	346	unsigned long *vm_flags, int advice)
..	..	@@ -356,6 +392,7 @@
356	392	khugepaged_pages_to_scan = HPAGE_PMD_NR * 8;
357	393	khugepaged_max_ptes_none = HPAGE_PMD_NR - 1;
358	394	khugepaged_max_ptes_swap = HPAGE_PMD_NR / 8;
	395	+ khugepaged_max_ptes_shared = HPAGE_PMD_NR / 2;
359	396
360	397	return 0;
361	398	}
..	..	@@ -397,25 +434,39 @@
397	434
398	435	static inline int khugepaged_test_exit(struct mm_struct *mm)
399	436	{
400		- return atomic_read(&mm->mm_users) == 0 \|\| !mmget_still_valid(mm);
	437	+ return atomic_read(&mm->mm_users) == 0;
401	438	}
402	439
403	440	static bool hugepage_vma_check(struct vm_area_struct *vma,
404	441	unsigned long vm_flags)
405	442	{
406		- if ((!(vm_flags & VM_HUGEPAGE) && !khugepaged_always()) \|\|
407		- (vm_flags & VM_NOHUGEPAGE) \|\|
408		- test_bit(MMF_DISABLE_THP, &vma->vm_mm->flags))
	443	+ if (!transhuge_vma_enabled(vma, vm_flags))
409	444	return false;
410		- if (shmem_file(vma->vm_file)) {
411		- if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE))
412		- return false;
413		- return IS_ALIGNED((vma->vm_start >> PAGE_SHIFT) - vma->vm_pgoff,
414		- HPAGE_PMD_NR);
	445	+
	446	+ if (vma->vm_file && !IS_ALIGNED((vma->vm_start >> PAGE_SHIFT) -
	447	+ vma->vm_pgoff, HPAGE_PMD_NR))
	448	+ return false;
	449	+
	450	+ /* Enabled via shmem mount options or sysfs settings. */
	451	+ if (shmem_file(vma->vm_file))
	452	+ return shmem_huge_enabled(vma);
	453	+
	454	+ /* THP settings require madvise. */
	455	+ if (!(vm_flags & VM_HUGEPAGE) && !khugepaged_always())
	456	+ return false;
	457	+
	458	+ /* Only regular file is valid */
	459	+ if (IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) && vma->vm_file &&
	460	+ !inode_is_open_for_write(vma->vm_file->f_inode) &&
	461	+ (vm_flags & VM_EXEC)) {
	462	+ struct inode *inode = vma->vm_file->f_inode;
	463	+
	464	+ return S_ISREG(inode->i_mode);
415	465	}
	466	+
416	467	if (!vma->anon_vma \|\| vma->vm_ops)
417	468	return false;
418		- if (is_vma_temporary_stack(vma))
	469	+ if (vma_is_temporary_stack(vma))
419	470	return false;
420	471	return !(vm_flags & VM_NO_KHUGEPAGED);
421	472	}
..	..	@@ -459,8 +510,9 @@
459	510	unsigned long hstart, hend;
460	511
461	512	/*
462		- * khugepaged does not yet work on non-shmem files or special
463		- * mappings. And file-private shmem THP is not supported.
	513	+ * khugepaged only supports read-only files for non-shmem files.
	514	+ * khugepaged does not yet work on special mappings. And
	515	+ * file-private shmem THP is not supported.
464	516	*/
465	517	if (!hugepage_vma_check(vma, vm_flags))
466	518	return 0;
..	..	@@ -497,36 +549,61 @@
497	549	* under mmap sem read mode). Stop here (after we
498	550	* return all pagetables will be destroyed) until
499	551	* khugepaged has finished working on the pagetables
500		- * under the mmap_sem.
	552	+ * under the mmap_lock.
501	553	*/
502		- down_write(&mm->mmap_sem);
503		- up_write(&mm->mmap_sem);
	554	+ mmap_write_lock(mm);
	555	+ mmap_write_unlock(mm);
504	556	}
505	557	}
506	558
507	559	static void release_pte_page(struct page *page)
508	560	{
509		- dec_node_page_state(page, NR_ISOLATED_ANON + page_is_file_cache(page));
	561	+ mod_node_page_state(page_pgdat(page),
	562	+ NR_ISOLATED_ANON + page_is_file_lru(page),
	563	+ -compound_nr(page));
510	564	unlock_page(page);
511	565	putback_lru_page(page);
512	566	}
513	567
514		-static void release_pte_pages(pte_t pte, pte_t _pte)
	568	+static void release_pte_pages(pte_t pte, pte_t _pte,
	569	+ struct list_head *compound_pagelist)
515	570	{
	571	+ struct page page, tmp;
	572	+
516	573	while (--_pte >= pte) {
517	574	pte_t pteval = *_pte;
518		- if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval)))
519		- release_pte_page(pte_page(pteval));
	575	+
	576	+ page = pte_page(pteval);
	577	+ if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval)) &&
	578	+ !PageCompound(page))
	579	+ release_pte_page(page);
520	580	}
	581	+
	582	+ list_for_each_entry_safe(page, tmp, compound_pagelist, lru) {
	583	+ list_del(&page->lru);
	584	+ release_pte_page(page);
	585	+ }
	586	+}
	587	+
	588	+static bool is_refcount_suitable(struct page *page)
	589	+{
	590	+ int expected_refcount;
	591	+
	592	+ expected_refcount = total_mapcount(page);
	593	+ if (PageSwapCache(page))
	594	+ expected_refcount += compound_nr(page);
	595	+
	596	+ return page_count(page) == expected_refcount;
521	597	}
522	598
523	599	static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
524	600	unsigned long address,
525		- pte_t *pte)
	601	+ pte_t *pte,
	602	+ struct list_head *compound_pagelist)
526	603	{
527	604	struct page *page = NULL;
528	605	pte_t *_pte;
529		- int none_or_zero = 0, result = 0, referenced = 0;
	606	+ int none_or_zero = 0, shared = 0, result = 0, referenced = 0;
530	607	bool writable = false;
531	608
532	609	for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
..	..	@@ -552,13 +629,27 @@
552	629	goto out;
553	630	}
554	631
555		- /* TODO: teach khugepaged to collapse THP mapped with pte */
556		- if (PageCompound(page)) {
557		- result = SCAN_PAGE_COMPOUND;
	632	+ VM_BUG_ON_PAGE(!PageAnon(page), page);
	633	+
	634	+ if (page_mapcount(page) > 1 &&
	635	+ ++shared > khugepaged_max_ptes_shared) {
	636	+ result = SCAN_EXCEED_SHARED_PTE;
558	637	goto out;
559	638	}
560	639
561		- VM_BUG_ON_PAGE(!PageAnon(page), page);
	640	+ if (PageCompound(page)) {
	641	+ struct page *p;
	642	+ page = compound_head(page);
	643	+
	644	+ /*
	645	+ * Check if we have dealt with the compound page
	646	+ * already
	647	+ */
	648	+ list_for_each_entry(p, compound_pagelist, lru) {
	649	+ if (page == p)
	650	+ goto next;
	651	+ }
	652	+ }
562	653
563	654	/*
564	655	* We can do it before isolate_lru_page because the
..	..	@@ -572,28 +663,30 @@
572	663	}
573	664
574	665	/*
575		- * cannot use mapcount: can't collapse if there's a gup pin.
576		- * The page must only be referenced by the scanned process
577		- * and page swap cache.
	666	+ * Check if the page has any GUP (or other external) pins.
	667	+ *
	668	+ * The page table that maps the page has been already unlinked
	669	+ * from the page table tree and this process cannot get
	670	+ * an additinal pin on the page.
	671	+ *
	672	+ * New pins can come later if the page is shared across fork,
	673	+ * but not from this process. The other process cannot write to
	674	+ * the page, only trigger CoW.
578	675	*/
579		- if (page_count(page) != 1 + PageSwapCache(page)) {
	676	+ if (!is_refcount_suitable(page)) {
580	677	unlock_page(page);
581	678	result = SCAN_PAGE_COUNT;
582	679	goto out;
583	680	}
584		- if (pte_write(pteval)) {
585		- writable = true;
586		- } else {
587		- if (PageSwapCache(page) &&
588		- !reuse_swap_page(page, NULL)) {
589		- unlock_page(page);
590		- result = SCAN_SWAP_CACHE_PAGE;
591		- goto out;
592		- }
	681	+ if (!pte_write(pteval) && PageSwapCache(page) &&
	682	+ !reuse_swap_page(page, NULL)) {
593	683	/*
594		- * Page is not in the swap cache. It can be collapsed
595		- * into a THP.
	684	+ * Page is in the swap cache and cannot be re-used.
	685	+ * It cannot be collapsed into a THP.
596	686	*/
	687	+ unlock_page(page);
	688	+ result = SCAN_SWAP_CACHE_PAGE;
	689	+ goto out;
597	690	}
598	691
599	692	/*
..	..	@@ -605,16 +698,23 @@
605	698	result = SCAN_DEL_PAGE_LRU;
606	699	goto out;
607	700	}
608		- inc_node_page_state(page,
609		- NR_ISOLATED_ANON + page_is_file_cache(page));
	701	+ mod_node_page_state(page_pgdat(page),
	702	+ NR_ISOLATED_ANON + page_is_file_lru(page),
	703	+ compound_nr(page));
610	704	VM_BUG_ON_PAGE(!PageLocked(page), page);
611	705	VM_BUG_ON_PAGE(PageLRU(page), page);
612	706
	707	+ if (PageCompound(page))
	708	+ list_add_tail(&page->lru, compound_pagelist);
	709	+next:
613	710	/* There should be enough young pte to collapse the page */
614	711	if (pte_young(pteval) \|\|
615	712	page_is_young(page) \|\| PageReferenced(page) \|\|
616	713	mmu_notifier_test_young(vma->vm_mm, address))
617	714	referenced++;
	715	+
	716	+ if (pte_write(pteval))
	717	+ writable = true;
618	718	}
619	719
620	720	if (unlikely(!writable)) {
..	..	@@ -628,7 +728,7 @@
628	728	return 1;
629	729	}
630	730	out:
631		- release_pte_pages(pte, _pte);
	731	+ release_pte_pages(pte, _pte, compound_pagelist);
632	732	trace_mm_collapse_huge_page_isolate(page, none_or_zero,
633	733	referenced, writable, result);
634	734	return 0;
..	..	@@ -637,13 +737,14 @@
637	737	static void __collapse_huge_page_copy(pte_t pte, struct page page,
638	738	struct vm_area_struct *vma,
639	739	unsigned long address,
640		- spinlock_t *ptl)
	740	+ spinlock_t *ptl,
	741	+ struct list_head *compound_pagelist)
641	742	{
	743	+ struct page src_page, tmp;
642	744	pte_t *_pte;
643	745	for (_pte = pte; _pte < pte + HPAGE_PMD_NR;
644	746	_pte++, page++, address += PAGE_SIZE) {
645	747	pte_t pteval = *_pte;
646		- struct page *src_page;
647	748
648	749	if (pte_none(pteval) \|\| is_zero_pfn(pte_pfn(pteval))) {
649	750	clear_user_highpage(page, address);
..	..	@@ -663,8 +764,8 @@
663	764	} else {
664	765	src_page = pte_page(pteval);
665	766	copy_user_highpage(page, src_page, address, vma);
666		- VM_BUG_ON_PAGE(page_mapcount(src_page) != 1, src_page);
667		- release_pte_page(src_page);
	767	+ if (!PageCompound(src_page))
	768	+ release_pte_page(src_page);
668	769	/*
669	770	* ptl mostly unnecessary, but preempt has to
670	771	* be disabled to update the per-cpu stats
..	..	@@ -680,6 +781,11 @@
680	781	spin_unlock(ptl);
681	782	free_page_and_swap_cache(src_page);
682	783	}
	784	+ }
	785	+
	786	+ list_for_each_entry_safe(src_page, tmp, compound_pagelist, lru) {
	787	+ list_del(&src_page->lru);
	788	+ release_pte_page(src_page);
683	789	}
684	790	}
685	791
..	..	@@ -713,7 +819,7 @@
713	819	for (i = 0; i < MAX_NUMNODES; i++) {
714	820	if (!khugepaged_node_load[i])
715	821	continue;
716		- if (node_distance(nid, i) > RECLAIM_DISTANCE)
	822	+ if (node_distance(nid, i) > node_reclaim_distance)
717	823	return true;
718	824	}
719	825	return false;
..	..	@@ -854,8 +960,8 @@
854	960	#endif
855	961
856	962	/*
857		- * If mmap_sem temporarily dropped, revalidate vma
858		- * before taking mmap_sem.
	963	+ * If mmap_lock temporarily dropped, revalidate vma
	964	+ * before taking mmap_lock.
859	965	* Return 0 if succeeds, otherwise return none-zero
860	966	* value (scan code).
861	967	*/
..	..	@@ -879,6 +985,9 @@
879	985	return SCAN_ADDRESS_RANGE;
880	986	if (!hugepage_vma_check(vma, vma->vm_flags))
881	987	return SCAN_VMA_CHECK;
	988	+ /* Anon VMA expected */
	989	+ if (!vma->anon_vma \|\| vma->vm_ops)
	990	+ return SCAN_VMA_CHECK;
882	991	return 0;
883	992	}
884	993
..	..	@@ -887,48 +996,48 @@
887	996	* Only done if khugepaged_scan_pmd believes it is worthwhile.
888	997	*
889	998	* Called and returns without pte mapped or spinlocks held,
890		- * but with mmap_sem held to protect against vma changes.
	999	+ * but with mmap_lock held to protect against vma changes.
891	1000	*/
892	1001
893	1002	static bool __collapse_huge_page_swapin(struct mm_struct *mm,
894	1003	struct vm_area_struct *vma,
895		- unsigned long address, pmd_t *pmd,
	1004	+ unsigned long haddr, pmd_t *pmd,
896	1005	int referenced)
897	1006	{
898	1007	int swapped_in = 0;
899	1008	vm_fault_t ret = 0;
900		- struct vm_fault vmf = {
901		- .vma = vma,
902		- .address = address,
903		- .flags = FAULT_FLAG_ALLOW_RETRY,
904		- .pmd = pmd,
905		- .pgoff = linear_page_index(vma, address),
906		- };
	1009	+ unsigned long address, end = haddr + (HPAGE_PMD_NR * PAGE_SIZE);
907	1010
908		- /* we only decide to swapin, if there is enough young ptes */
909		- if (referenced < HPAGE_PMD_NR/2) {
910		- trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
911		- return false;
912		- }
913		- vmf.pte = pte_offset_map(pmd, address);
914		- for (; vmf.address < address + HPAGE_PMD_NR*PAGE_SIZE;
915		- vmf.pte++, vmf.address += PAGE_SIZE) {
	1011	+ for (address = haddr; address < end; address += PAGE_SIZE) {
	1012	+ struct vm_fault vmf = {
	1013	+ .vma = vma,
	1014	+ .address = address,
	1015	+ .pgoff = linear_page_index(vma, haddr),
	1016	+ .flags = FAULT_FLAG_ALLOW_RETRY,
	1017	+ .pmd = pmd,
	1018	+ .vma_flags = vma->vm_flags,
	1019	+ .vma_page_prot = vma->vm_page_prot,
	1020	+ };
	1021	+
	1022	+ vmf.pte = pte_offset_map(pmd, address);
916	1023	vmf.orig_pte = *vmf.pte;
917		- if (!is_swap_pte(vmf.orig_pte))
	1024	+ if (!is_swap_pte(vmf.orig_pte)) {
	1025	+ pte_unmap(vmf.pte);
918	1026	continue;
	1027	+ }
919	1028	swapped_in++;
920	1029	ret = do_swap_page(&vmf);
921	1030
922		- /* do_swap_page returns VM_FAULT_RETRY with released mmap_sem */
	1031	+ /* do_swap_page returns VM_FAULT_RETRY with released mmap_lock */
923	1032	if (ret & VM_FAULT_RETRY) {
924		- down_read(&mm->mmap_sem);
925		- if (hugepage_vma_revalidate(mm, address, &vmf.vma)) {
	1033	+ mmap_read_lock(mm);
	1034	+ if (hugepage_vma_revalidate(mm, haddr, &vma)) {
926	1035	/* vma is no longer available, don't continue to swapin */
927	1036	trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
928	1037	return false;
929	1038	}
930	1039	/* check if the pmd is still valid */
931		- if (mm_find_pmd(mm, address) != pmd) {
	1040	+ if (mm_find_pmd(mm, haddr) != pmd) {
932	1041	trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
933	1042	return false;
934	1043	}
..	..	@@ -937,11 +1046,12 @@
937	1046	trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
938	1047	return false;
939	1048	}
940		- /* pte is unmapped now, we need to map it */
941		- vmf.pte = pte_offset_map(pmd, vmf.address);
942	1049	}
943		- vmf.pte--;
944		- pte_unmap(vmf.pte);
	1050	+
	1051	+ /* Drain LRU add pagevec to remove extra pin on the swapped in pages */
	1052	+ if (swapped_in)
	1053	+ lru_add_drain();
	1054	+
945	1055	trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 1);
946	1056	return true;
947	1057	}
..	..	@@ -949,18 +1059,17 @@
949	1059	static void collapse_huge_page(struct mm_struct *mm,
950	1060	unsigned long address,
951	1061	struct page **hpage,
952		- int node, int referenced)
	1062	+ int node, int referenced, int unmapped)
953	1063	{
	1064	+ LIST_HEAD(compound_pagelist);
954	1065	pmd_t *pmd, _pmd;
955	1066	pte_t *pte;
956	1067	pgtable_t pgtable;
957	1068	struct page *new_page;
958	1069	spinlock_t pmd_ptl, pte_ptl;
959	1070	int isolated = 0, result = 0;
960		- struct mem_cgroup *memcg;
961	1071	struct vm_area_struct *vma;
962		- unsigned long mmun_start; /* For mmu_notifiers */
963		- unsigned long mmun_end; /* For mmu_notifiers */
	1072	+ struct mmu_notifier_range range;
964	1073	gfp_t gfp;
965	1074
966	1075	VM_BUG_ON(address & ~HPAGE_PMD_MASK);
..	..	@@ -969,57 +1078,56 @@
969	1078	gfp = alloc_hugepage_khugepaged_gfpmask() \| __GFP_THISNODE;
970	1079
971	1080	/*
972		- * Before allocating the hugepage, release the mmap_sem read lock.
	1081	+ * Before allocating the hugepage, release the mmap_lock read lock.
973	1082	* The allocation can take potentially a long time if it involves
974		- * sync compaction, and we do not need to hold the mmap_sem during
	1083	+ * sync compaction, and we do not need to hold the mmap_lock during
975	1084	* that. We will recheck the vma after taking it again in write mode.
976	1085	*/
977		- up_read(&mm->mmap_sem);
	1086	+ mmap_read_unlock(mm);
978	1087	new_page = khugepaged_alloc_page(hpage, gfp, node);
979	1088	if (!new_page) {
980	1089	result = SCAN_ALLOC_HUGE_PAGE_FAIL;
981	1090	goto out_nolock;
982	1091	}
983	1092
984		- if (unlikely(mem_cgroup_try_charge(new_page, mm, gfp, &memcg, true))) {
	1093	+ if (unlikely(mem_cgroup_charge(new_page, mm, gfp))) {
985	1094	result = SCAN_CGROUP_CHARGE_FAIL;
986	1095	goto out_nolock;
987	1096	}
	1097	+ count_memcg_page_event(new_page, THP_COLLAPSE_ALLOC);
988	1098
989		- down_read(&mm->mmap_sem);
	1099	+ mmap_read_lock(mm);
990	1100	result = hugepage_vma_revalidate(mm, address, &vma);
991	1101	if (result) {
992		- mem_cgroup_cancel_charge(new_page, memcg, true);
993		- up_read(&mm->mmap_sem);
	1102	+ mmap_read_unlock(mm);
994	1103	goto out_nolock;
995	1104	}
996	1105
997	1106	pmd = mm_find_pmd(mm, address);
998	1107	if (!pmd) {
999	1108	result = SCAN_PMD_NULL;
1000		- mem_cgroup_cancel_charge(new_page, memcg, true);
1001		- up_read(&mm->mmap_sem);
	1109	+ mmap_read_unlock(mm);
1002	1110	goto out_nolock;
1003	1111	}
1004	1112
1005	1113	/*
1006		- * __collapse_huge_page_swapin always returns with mmap_sem locked.
1007		- * If it fails, we release mmap_sem and jump out_nolock.
	1114	+ * __collapse_huge_page_swapin always returns with mmap_lock locked.
	1115	+ * If it fails, we release mmap_lock and jump out_nolock.
1008	1116	* Continuing to collapse causes inconsistency.
1009	1117	*/
1010		- if (!__collapse_huge_page_swapin(mm, vma, address, pmd, referenced)) {
1011		- mem_cgroup_cancel_charge(new_page, memcg, true);
1012		- up_read(&mm->mmap_sem);
	1118	+ if (unmapped && !__collapse_huge_page_swapin(mm, vma, address,
	1119	+ pmd, referenced)) {
	1120	+ mmap_read_unlock(mm);
1013	1121	goto out_nolock;
1014	1122	}
1015	1123
1016		- up_read(&mm->mmap_sem);
	1124	+ mmap_read_unlock(mm);
1017	1125	/*
1018	1126	* Prevent all access to pagetables with the exception of
1019	1127	* gup_fast later handled by the ptep_clear_flush and the VM
1020	1128	* handled by the anon_vma lock + PG_lock.
1021	1129	*/
1022		- down_write(&mm->mmap_sem);
	1130	+ mmap_write_lock(mm);
1023	1131	result = hugepage_vma_revalidate(mm, address, &vma);
1024	1132	if (result)
1025	1133	goto out;
..	..	@@ -1027,27 +1135,33 @@
1027	1135	if (mm_find_pmd(mm, address) != pmd)
1028	1136	goto out;
1029	1137
	1138	+ vm_write_begin(vma);
1030	1139	anon_vma_lock_write(vma->anon_vma);
	1140	+
	1141	+ mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, NULL, mm,
	1142	+ address, address + HPAGE_PMD_SIZE);
	1143	+ mmu_notifier_invalidate_range_start(&range);
1031	1144
1032	1145	pte = pte_offset_map(pmd, address);
1033	1146	pte_ptl = pte_lockptr(mm, pmd);
1034	1147
1035		- mmun_start = address;
1036		- mmun_end = address + HPAGE_PMD_SIZE;
1037		- mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1038	1148	pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
1039	1149	/*
1040		- * After this gup_fast can't run anymore. This also removes
1041		- * any huge TLB entry from the CPU so we won't allow
1042		- * huge and small TLB entries for the same virtual address
1043		- * to avoid the risk of CPU bugs in that area.
	1150	+ * This removes any huge TLB entry from the CPU so we won't allow
	1151	+ * huge and small TLB entries for the same virtual address to
	1152	+ * avoid the risk of CPU bugs in that area.
	1153	+ *
	1154	+ * Parallel fast GUP is fine since fast GUP will back off when
	1155	+ * it detects PMD is changed.
1044	1156	*/
1045	1157	_pmd = pmdp_collapse_flush(vma, address, pmd);
1046	1158	spin_unlock(pmd_ptl);
1047		- mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
	1159	+ mmu_notifier_invalidate_range_end(&range);
	1160	+ tlb_remove_table_sync_one();
1048	1161
1049	1162	spin_lock(pte_ptl);
1050		- isolated = __collapse_huge_page_isolate(vma, address, pte);
	1163	+ isolated = __collapse_huge_page_isolate(vma, address, pte,
	1164	+ &compound_pagelist);
1051	1165	spin_unlock(pte_ptl);
1052	1166
1053	1167	if (unlikely(!isolated)) {
..	..	@@ -1062,6 +1176,7 @@
1062	1176	pmd_populate(mm, pmd, pmd_pgtable(_pmd));
1063	1177	spin_unlock(pmd_ptl);
1064	1178	anon_vma_unlock_write(vma->anon_vma);
	1179	+ vm_write_end(vma);
1065	1180	result = SCAN_FAIL;
1066	1181	goto out;
1067	1182	}
..	..	@@ -1072,7 +1187,8 @@
1072	1187	*/
1073	1188	anon_vma_unlock_write(vma->anon_vma);
1074	1189
1075		- __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl);
	1190	+ __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl,
	1191	+ &compound_pagelist);
1076	1192	pte_unmap(pte);
1077	1193	__SetPageUptodate(new_page);
1078	1194	pgtable = pmd_pgtable(_pmd);
..	..	@@ -1090,24 +1206,25 @@
1090	1206	spin_lock(pmd_ptl);
1091	1207	BUG_ON(!pmd_none(*pmd));
1092	1208	page_add_new_anon_rmap(new_page, vma, address, true);
1093		- mem_cgroup_commit_charge(new_page, memcg, false, true);
1094		- lru_cache_add_active_or_unevictable(new_page, vma);
	1209	+ lru_cache_add_inactive_or_unevictable(new_page, vma);
1095	1210	pgtable_trans_huge_deposit(mm, pmd, pgtable);
1096	1211	set_pmd_at(mm, address, pmd, _pmd);
1097	1212	update_mmu_cache_pmd(vma, address, pmd);
1098	1213	spin_unlock(pmd_ptl);
	1214	+ vm_write_end(vma);
1099	1215
1100	1216	*hpage = NULL;
1101	1217
1102	1218	khugepaged_pages_collapsed++;
1103	1219	result = SCAN_SUCCEED;
1104	1220	out_up_write:
1105		- up_write(&mm->mmap_sem);
	1221	+ mmap_write_unlock(mm);
1106	1222	out_nolock:
	1223	+ if (!IS_ERR_OR_NULL(*hpage))
	1224	+ mem_cgroup_uncharge(*hpage);
1107	1225	trace_mm_collapse_huge_page(mm, isolated, result);
1108	1226	return;
1109	1227	out:
1110		- mem_cgroup_cancel_charge(new_page, memcg, true);
1111	1228	goto out_up_write;
1112	1229	}
1113	1230
..	..	@@ -1118,7 +1235,8 @@
1118	1235	{
1119	1236	pmd_t *pmd;
1120	1237	pte_t pte, _pte;
1121		- int ret = 0, none_or_zero = 0, result = 0, referenced = 0;
	1238	+ int ret = 0, result = 0, referenced = 0;
	1239	+ int none_or_zero = 0, shared = 0;
1122	1240	struct page *page = NULL;
1123	1241	unsigned long _address;
1124	1242	spinlock_t *ptl;
..	..	@@ -1140,6 +1258,15 @@
1140	1258	pte_t pteval = *_pte;
1141	1259	if (is_swap_pte(pteval)) {
1142	1260	if (++unmapped <= khugepaged_max_ptes_swap) {
	1261	+ /*
	1262	+ * Always be strict with uffd-wp
	1263	+ * enabled swap entries. Please see
	1264	+ * comment below for pte_uffd_wp().
	1265	+ */
	1266	+ if (pte_swp_uffd_wp(pteval)) {
	1267	+ result = SCAN_PTE_UFFD_WP;
	1268	+ goto out_unmap;
	1269	+ }
1143	1270	continue;
1144	1271	} else {
1145	1272	result = SCAN_EXCEED_SWAP_PTE;
..	..	@@ -1159,6 +1286,19 @@
1159	1286	result = SCAN_PTE_NON_PRESENT;
1160	1287	goto out_unmap;
1161	1288	}
	1289	+ if (pte_uffd_wp(pteval)) {
	1290	+ /*
	1291	+ * Don't collapse the page if any of the small
	1292	+ * PTEs are armed with uffd write protection.
	1293	+ * Here we can also mark the new huge pmd as
	1294	+ * write protected if any of the small ones is
	1295	+ * marked but that could bring uknown
	1296	+ * userfault messages that falls outside of
	1297	+ * the registered range. So, just be simple.
	1298	+ */
	1299	+ result = SCAN_PTE_UFFD_WP;
	1300	+ goto out_unmap;
	1301	+ }
1162	1302	if (pte_write(pteval))
1163	1303	writable = true;
1164	1304
..	..	@@ -1168,11 +1308,13 @@
1168	1308	goto out_unmap;
1169	1309	}
1170	1310
1171		- /* TODO: teach khugepaged to collapse THP mapped with pte */
1172		- if (PageCompound(page)) {
1173		- result = SCAN_PAGE_COMPOUND;
	1311	+ if (page_mapcount(page) > 1 &&
	1312	+ ++shared > khugepaged_max_ptes_shared) {
	1313	+ result = SCAN_EXCEED_SHARED_PTE;
1174	1314	goto out_unmap;
1175	1315	}
	1316	+
	1317	+ page = compound_head(page);
1176	1318
1177	1319	/*
1178	1320	* Record which node the original page is from and save this
..	..	@@ -1200,11 +1342,23 @@
1200	1342	}
1201	1343
1202	1344	/*
1203		- * cannot use mapcount: can't collapse if there's a gup pin.
1204		- * The page must only be referenced by the scanned process
1205		- * and page swap cache.
	1345	+ * Check if the page has any GUP (or other external) pins.
	1346	+ *
	1347	+ * Here the check is racy it may see totmal_mapcount > refcount
	1348	+ * in some cases.
	1349	+ * For example, one process with one forked child process.
	1350	+ * The parent has the PMD split due to MADV_DONTNEED, then
	1351	+ * the child is trying unmap the whole PMD, but khugepaged
	1352	+ * may be scanning the parent between the child has
	1353	+ * PageDoubleMap flag cleared and dec the mapcount. So
	1354	+ * khugepaged may see total_mapcount > refcount.
	1355	+ *
	1356	+ * But such case is ephemeral we could always retry collapse
	1357	+ * later. However it may report false positive if the page
	1358	+ * has excessive GUP pins (i.e. 512). Anyway the same check
	1359	+ * will be done again later the risk seems low.
1206	1360	*/
1207		- if (page_count(page) != 1 + PageSwapCache(page)) {
	1361	+ if (!is_refcount_suitable(page)) {
1208	1362	result = SCAN_PAGE_COUNT;
1209	1363	goto out_unmap;
1210	1364	}
..	..	@@ -1213,22 +1367,21 @@
1213	1367	mmu_notifier_test_young(vma->vm_mm, address))
1214	1368	referenced++;
1215	1369	}
1216		- if (writable) {
1217		- if (referenced) {
1218		- result = SCAN_SUCCEED;
1219		- ret = 1;
1220		- } else {
1221		- result = SCAN_LACK_REFERENCED_PAGE;
1222		- }
1223		- } else {
	1370	+ if (!writable) {
1224	1371	result = SCAN_PAGE_RO;
	1372	+ } else if (!referenced \|\| (unmapped && referenced < HPAGE_PMD_NR/2)) {
	1373	+ result = SCAN_LACK_REFERENCED_PAGE;
	1374	+ } else {
	1375	+ result = SCAN_SUCCEED;
	1376	+ ret = 1;
1225	1377	}
1226	1378	out_unmap:
1227	1379	pte_unmap_unlock(pte, ptl);
1228	1380	if (ret) {
1229	1381	node = khugepaged_find_target_node();
1230		- /* collapse_huge_page will return with the mmap_sem released */
1231		- collapse_huge_page(mm, address, hpage, node, referenced);
	1382	+ /* collapse_huge_page will return with the mmap_lock released */
	1383	+ collapse_huge_page(mm, address, hpage, node,
	1384	+ referenced, unmapped);
1232	1385	}
1233	1386	out:
1234	1387	trace_mm_khugepaged_scan_pmd(mm, page, writable, referenced,
..	..	@@ -1240,7 +1393,7 @@
1240	1393	{
1241	1394	struct mm_struct *mm = mm_slot->mm;
1242	1395
1243		- VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
	1396	+ lockdep_assert_held(&khugepaged_mm_lock);
1244	1397
1245	1398	if (khugepaged_test_exit(mm)) {
1246	1399	/* free mm_slot */
..	..	@@ -1259,7 +1412,186 @@
1259	1412	}
1260	1413	}
1261	1414
1262		-#if defined(CONFIG_SHMEM) && defined(CONFIG_TRANSPARENT_HUGE_PAGECACHE)
	1415	+#ifdef CONFIG_SHMEM
	1416	+/*
	1417	+ * Notify khugepaged that given addr of the mm is pte-mapped THP. Then
	1418	+ * khugepaged should try to collapse the page table.
	1419	+ */
	1420	+static int khugepaged_add_pte_mapped_thp(struct mm_struct *mm,
	1421	+ unsigned long addr)
	1422	+{
	1423	+ struct mm_slot *mm_slot;
	1424	+
	1425	+ VM_BUG_ON(addr & ~HPAGE_PMD_MASK);
	1426	+
	1427	+ spin_lock(&khugepaged_mm_lock);
	1428	+ mm_slot = get_mm_slot(mm);
	1429	+ if (likely(mm_slot && mm_slot->nr_pte_mapped_thp < MAX_PTE_MAPPED_THP))
	1430	+ mm_slot->pte_mapped_thp[mm_slot->nr_pte_mapped_thp++] = addr;
	1431	+ spin_unlock(&khugepaged_mm_lock);
	1432	+ return 0;
	1433	+}
	1434	+
	1435	+/**
	1436	+ * Try to collapse a pte-mapped THP for mm at address haddr.
	1437	+ *
	1438	+ * This function checks whether all the PTEs in the PMD are pointing to the
	1439	+ * right THP. If so, retract the page table so the THP can refault in with
	1440	+ * as pmd-mapped.
	1441	+ */
	1442	+void collapse_pte_mapped_thp(struct mm_struct *mm, unsigned long addr)
	1443	+{
	1444	+ unsigned long haddr = addr & HPAGE_PMD_MASK;
	1445	+ struct vm_area_struct *vma = find_vma(mm, haddr);
	1446	+ struct page *hpage;
	1447	+ pte_t start_pte, pte;
	1448	+ pmd_t *pmd, _pmd;
	1449	+ spinlock_t *ptl;
	1450	+ int count = 0;
	1451	+ int i;
	1452	+ struct mmu_notifier_range range;
	1453	+
	1454	+ if (!vma \|\| !vma->vm_file \|\|
	1455	+ vma->vm_start > haddr \|\| vma->vm_end < haddr + HPAGE_PMD_SIZE)
	1456	+ return;
	1457	+
	1458	+ /*
	1459	+ * This vm_flags may not have VM_HUGEPAGE if the page was not
	1460	+ * collapsed by this mm. But we can still collapse if the page is
	1461	+ * the valid THP. Add extra VM_HUGEPAGE so hugepage_vma_check()
	1462	+ * will not fail the vma for missing VM_HUGEPAGE
	1463	+ */
	1464	+ if (!hugepage_vma_check(vma, vma->vm_flags \| VM_HUGEPAGE))
	1465	+ return;
	1466	+
	1467	+ /*
	1468	+ * Symmetry with retract_page_tables(): Exclude MAP_PRIVATE mappings
	1469	+ * that got written to. Without this, we'd have to also lock the
	1470	+ * anon_vma if one exists.
	1471	+ */
	1472	+ if (vma->anon_vma)
	1473	+ return;
	1474	+
	1475	+ hpage = find_lock_page(vma->vm_file->f_mapping,
	1476	+ linear_page_index(vma, haddr));
	1477	+ if (!hpage)
	1478	+ return;
	1479	+
	1480	+ if (!PageHead(hpage))
	1481	+ goto drop_hpage;
	1482	+
	1483	+ pmd = mm_find_pmd(mm, haddr);
	1484	+ if (!pmd)
	1485	+ goto drop_hpage;
	1486	+
	1487	+ vm_write_begin(vma);
	1488	+
	1489	+ /*
	1490	+ * We need to lock the mapping so that from here on, only GUP-fast and
	1491	+ * hardware page walks can access the parts of the page tables that
	1492	+ * we're operating on.
	1493	+ */
	1494	+ i_mmap_lock_write(vma->vm_file->f_mapping);
	1495	+
	1496	+ /*
	1497	+ * This spinlock should be unnecessary: Nobody else should be accessing
	1498	+ * the page tables under spinlock protection here, only
	1499	+ * lockless_pages_from_mm() and the hardware page walker can access page
	1500	+ * tables while all the high-level locks are held in write mode.
	1501	+ */
	1502	+ start_pte = pte_offset_map_lock(mm, pmd, haddr, &ptl);
	1503	+
	1504	+ /* step 1: check all mapped PTEs are to the right huge page */
	1505	+ for (i = 0, addr = haddr, pte = start_pte;
	1506	+ i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) {
	1507	+ struct page *page;
	1508	+
	1509	+ /* empty pte, skip */
	1510	+ if (pte_none(*pte))
	1511	+ continue;
	1512	+
	1513	+ /* page swapped out, abort */
	1514	+ if (!pte_present(*pte))
	1515	+ goto abort;
	1516	+
	1517	+ page = vm_normal_page(vma, addr, *pte);
	1518	+
	1519	+ /*
	1520	+ * Note that uprobe, debugger, or MAP_PRIVATE may change the
	1521	+ * page table, but the new page will not be a subpage of hpage.
	1522	+ */
	1523	+ if (hpage + i != page)
	1524	+ goto abort;
	1525	+ count++;
	1526	+ }
	1527	+
	1528	+ /* step 2: adjust rmap */
	1529	+ for (i = 0, addr = haddr, pte = start_pte;
	1530	+ i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) {
	1531	+ struct page *page;
	1532	+
	1533	+ if (pte_none(*pte))
	1534	+ continue;
	1535	+ page = vm_normal_page(vma, addr, *pte);
	1536	+ page_remove_rmap(page, false);
	1537	+ }
	1538	+
	1539	+ pte_unmap_unlock(start_pte, ptl);
	1540	+
	1541	+ /* step 3: set proper refcount and mm_counters. */
	1542	+ if (count) {
	1543	+ page_ref_sub(hpage, count);
	1544	+ add_mm_counter(vma->vm_mm, mm_counter_file(hpage), -count);
	1545	+ }
	1546	+
	1547	+ /* step 4: collapse pmd */
	1548	+ mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, NULL, mm, haddr,
	1549	+ haddr + HPAGE_PMD_SIZE);
	1550	+ mmu_notifier_invalidate_range_start(&range);
	1551	+ _pmd = pmdp_collapse_flush(vma, haddr, pmd);
	1552	+ vm_write_end(vma);
	1553	+ mm_dec_nr_ptes(mm);
	1554	+ tlb_remove_table_sync_one();
	1555	+ mmu_notifier_invalidate_range_end(&range);
	1556	+ pte_free(mm, pmd_pgtable(_pmd));
	1557	+
	1558	+ i_mmap_unlock_write(vma->vm_file->f_mapping);
	1559	+
	1560	+drop_hpage:
	1561	+ unlock_page(hpage);
	1562	+ put_page(hpage);
	1563	+ return;
	1564	+
	1565	+abort:
	1566	+ pte_unmap_unlock(start_pte, ptl);
	1567	+ vm_write_end(vma);
	1568	+ i_mmap_unlock_write(vma->vm_file->f_mapping);
	1569	+ goto drop_hpage;
	1570	+}
	1571	+
	1572	+static int khugepaged_collapse_pte_mapped_thps(struct mm_slot *mm_slot)
	1573	+{
	1574	+ struct mm_struct *mm = mm_slot->mm;
	1575	+ int i;
	1576	+
	1577	+ if (likely(mm_slot->nr_pte_mapped_thp == 0))
	1578	+ return 0;
	1579	+
	1580	+ if (!mmap_write_trylock(mm))
	1581	+ return -EBUSY;
	1582	+
	1583	+ if (unlikely(khugepaged_test_exit(mm)))
	1584	+ goto out;
	1585	+
	1586	+ for (i = 0; i < mm_slot->nr_pte_mapped_thp; i++)
	1587	+ collapse_pte_mapped_thp(mm, mm_slot->pte_mapped_thp[i]);
	1588	+
	1589	+out:
	1590	+ mm_slot->nr_pte_mapped_thp = 0;
	1591	+ mmap_write_unlock(mm);
	1592	+ return 0;
	1593	+}
	1594	+
1263	1595	static void retract_page_tables(struct address_space *mapping, pgoff_t pgoff)
1264	1596	{
1265	1597	struct vm_area_struct *vma;
..	..	@@ -1269,7 +1601,23 @@
1269	1601
1270	1602	i_mmap_lock_write(mapping);
1271	1603	vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) {
1272		- /* probably overkill */
	1604	+ /*
	1605	+ * Check vma->anon_vma to exclude MAP_PRIVATE mappings that
	1606	+ * got written to. These VMAs are likely not worth investing
	1607	+ * mmap_write_lock(mm) as PMD-mapping is likely to be split
	1608	+ * later.
	1609	+ *
	1610	+ * Not that vma->anon_vma check is racy: it can be set up after
	1611	+ * the check but before we took mmap_lock by the fault path.
	1612	+ * But page lock would prevent establishing any new ptes of the
	1613	+ * page, so we are safe.
	1614	+ *
	1615	+ * An alternative would be drop the check, but check that page
	1616	+ * table is clear before calling pmdp_collapse_flush() under
	1617	+ * ptl. It has higher chance to recover THP for the VMA, but
	1618	+ * has higher cost too. It would also probably require locking
	1619	+ * the anon_vma.
	1620	+ */
1273	1621	if (vma->anon_vma)
1274	1622	continue;
1275	1623	addr = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
..	..	@@ -1282,57 +1630,71 @@
1282	1630	if (!pmd)
1283	1631	continue;
1284	1632	/*
1285		- * We need exclusive mmap_sem to retract page table.
1286		- * If trylock fails we would end up with pte-mapped THP after
1287		- * re-fault. Not ideal, but it's more important to not disturb
1288		- * the system too much.
	1633	+ * We need exclusive mmap_lock to retract page table.
	1634	+ *
	1635	+ * We use trylock due to lock inversion: we need to acquire
	1636	+ * mmap_lock while holding page lock. Fault path does it in
	1637	+ * reverse order. Trylock is a way to avoid deadlock.
1289	1638	*/
1290		- if (down_write_trylock(&mm->mmap_sem)) {
	1639	+ if (mmap_write_trylock(mm)) {
1291	1640	if (!khugepaged_test_exit(mm)) {
1292		- spinlock_t *ptl = pmd_lock(mm, pmd);
	1641	+ struct mmu_notifier_range range;
	1642	+
	1643	+ vm_write_begin(vma);
	1644	+ mmu_notifier_range_init(&range,
	1645	+ MMU_NOTIFY_CLEAR, 0,
	1646	+ NULL, mm, addr,
	1647	+ addr + HPAGE_PMD_SIZE);
	1648	+ mmu_notifier_invalidate_range_start(&range);
1293	1649	/* assume page table is clear */
1294	1650	_pmd = pmdp_collapse_flush(vma, addr, pmd);
1295		- spin_unlock(ptl);
	1651	+ vm_write_end(vma);
1296	1652	mm_dec_nr_ptes(mm);
	1653	+ tlb_remove_table_sync_one();
1297	1654	pte_free(mm, pmd_pgtable(_pmd));
	1655	+ mmu_notifier_invalidate_range_end(&range);
1298	1656	}
1299		- up_write(&mm->mmap_sem);
	1657	+ mmap_write_unlock(mm);
	1658	+ } else {
	1659	+ /* Try again later */
	1660	+ khugepaged_add_pte_mapped_thp(mm, addr);
1300	1661	}
1301	1662	}
1302	1663	i_mmap_unlock_write(mapping);
1303	1664	}
1304	1665
1305	1666	/**
1306		- * collapse_shmem - collapse small tmpfs/shmem pages into huge one.
	1667	+ * collapse_file - collapse filemap/tmpfs/shmem pages into huge one.
1307	1668	*
1308	1669	* Basic scheme is simple, details are more complex:
1309	1670	* - allocate and lock a new huge page;
1310		- * - scan over radix tree replacing old pages the new one
1311		- * + swap in pages if necessary;
	1671	+ * - scan page cache replacing old pages with the new one
	1672	+ * + swap/gup in pages if necessary;
1312	1673	* + fill in gaps;
1313		- * + keep old pages around in case if rollback is required;
1314		- * - if replacing succeed:
	1674	+ * + keep old pages around in case rollback is required;
	1675	+ * - if replacing succeeds:
1315	1676	* + copy data over;
1316	1677	* + free old pages;
1317	1678	* + unlock huge page;
1318	1679	* - if replacing failed;
1319	1680	* + put all pages back and unfreeze them;
1320		- * + restore gaps in the radix-tree;
	1681	+ * + restore gaps in the page cache;
1321	1682	* + unlock and free huge page;
1322	1683	*/
1323		-static void collapse_shmem(struct mm_struct *mm,
1324		- struct address_space *mapping, pgoff_t start,
	1684	+static void collapse_file(struct mm_struct *mm,
	1685	+ struct file *file, pgoff_t start,
1325	1686	struct page **hpage, int node)
1326	1687	{
	1688	+ struct address_space *mapping = file->f_mapping;
1327	1689	gfp_t gfp;
1328		- struct page page, new_page, *tmp;
1329		- struct mem_cgroup *memcg;
	1690	+ struct page *new_page;
1330	1691	pgoff_t index, end = start + HPAGE_PMD_NR;
1331	1692	LIST_HEAD(pagelist);
1332		- struct radix_tree_iter iter;
1333		- void **slot;
	1693	+ XA_STATE_ORDER(xas, &mapping->i_pages, start, HPAGE_PMD_ORDER);
1334	1694	int nr_none = 0, result = SCAN_SUCCEED;
	1695	+ bool is_shmem = shmem_file(file);
1335	1696
	1697	+ VM_BUG_ON(!IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) && !is_shmem);
1336	1698	VM_BUG_ON(start & (HPAGE_PMD_NR - 1));
1337	1699
1338	1700	/* Only allocate from the target node */
..	..	@@ -1344,13 +1706,28 @@
1344	1706	goto out;
1345	1707	}
1346	1708
1347		- if (unlikely(mem_cgroup_try_charge(new_page, mm, gfp, &memcg, true))) {
	1709	+ if (unlikely(mem_cgroup_charge(new_page, mm, gfp))) {
1348	1710	result = SCAN_CGROUP_CHARGE_FAIL;
1349	1711	goto out;
1350	1712	}
	1713	+ count_memcg_page_event(new_page, THP_COLLAPSE_ALLOC);
	1714	+
	1715	+ /* This will be less messy when we use multi-index entries */
	1716	+ do {
	1717	+ xas_lock_irq(&xas);
	1718	+ xas_create_range(&xas);
	1719	+ if (!xas_error(&xas))
	1720	+ break;
	1721	+ xas_unlock_irq(&xas);
	1722	+ if (!xas_nomem(&xas, GFP_KERNEL)) {
	1723	+ result = SCAN_FAIL;
	1724	+ goto out;
	1725	+ }
	1726	+ } while (1);
1351	1727
1352	1728	__SetPageLocked(new_page);
1353		- __SetPageSwapBacked(new_page);
	1729	+ if (is_shmem)
	1730	+ __SetPageSwapBacked(new_page);
1354	1731	new_page->index = start;
1355	1732	new_page->mapping = mapping;
1356	1733
..	..	@@ -1360,55 +1737,91 @@
1360	1737	* be able to map it or use it in another way until we unlock it.
1361	1738	*/
1362	1739
1363		- index = start;
1364		- xa_lock_irq(&mapping->i_pages);
1365		- radix_tree_for_each_slot(slot, &mapping->i_pages, &iter, start) {
1366		- int n = min(iter.index, end) - index;
	1740	+ xas_set(&xas, start);
	1741	+ for (index = start; index < end; index++) {
	1742	+ struct page *page = xas_next(&xas);
1367	1743
1368		- /*
1369		- * Stop if extent has been hole-punched, and is now completely
1370		- * empty (the more obvious i_size_read() check would take an
1371		- * irq-unsafe seqlock on 32-bit).
1372		- */
1373		- if (n >= HPAGE_PMD_NR) {
1374		- result = SCAN_TRUNCATED;
1375		- goto tree_locked;
1376		- }
1377		-
1378		- /*
1379		- * Handle holes in the radix tree: charge it from shmem and
1380		- * insert relevant subpage of new_page into the radix-tree.
1381		- */
1382		- if (n && !shmem_charge(mapping->host, n)) {
1383		- result = SCAN_FAIL;
1384		- goto tree_locked;
1385		- }
1386		- for (; index < min(iter.index, end); index++) {
1387		- radix_tree_insert(&mapping->i_pages, index,
1388		- new_page + (index % HPAGE_PMD_NR));
1389		- }
1390		- nr_none += n;
1391		-
1392		- /* We are done. */
1393		- if (index >= end)
1394		- break;
1395		-
1396		- page = radix_tree_deref_slot_protected(slot,
1397		- &mapping->i_pages.xa_lock);
1398		- if (radix_tree_exceptional_entry(page) \|\| !PageUptodate(page)) {
1399		- xa_unlock_irq(&mapping->i_pages);
1400		- /* swap in or instantiate fallocated page */
1401		- if (shmem_getpage(mapping->host, index, &page,
1402		- SGP_NOHUGE)) {
1403		- result = SCAN_FAIL;
1404		- goto tree_unlocked;
	1744	+ VM_BUG_ON(index != xas.xa_index);
	1745	+ if (is_shmem) {
	1746	+ if (!page) {
	1747	+ /*
	1748	+ * Stop if extent has been truncated or
	1749	+ * hole-punched, and is now completely
	1750	+ * empty.
	1751	+ */
	1752	+ if (index == start) {
	1753	+ if (!xas_next_entry(&xas, end - 1)) {
	1754	+ result = SCAN_TRUNCATED;
	1755	+ goto xa_locked;
	1756	+ }
	1757	+ xas_set(&xas, index);
	1758	+ }
	1759	+ if (!shmem_charge(mapping->host, 1)) {
	1760	+ result = SCAN_FAIL;
	1761	+ goto xa_locked;
	1762	+ }
	1763	+ xas_store(&xas, new_page);
	1764	+ nr_none++;
	1765	+ continue;
1405	1766	}
1406		- } else if (trylock_page(page)) {
1407		- get_page(page);
1408		- xa_unlock_irq(&mapping->i_pages);
1409		- } else {
1410		- result = SCAN_PAGE_LOCK;
1411		- goto tree_locked;
	1767	+
	1768	+ if (xa_is_value(page) \|\| !PageUptodate(page)) {
	1769	+ xas_unlock_irq(&xas);
	1770	+ /* swap in or instantiate fallocated page */
	1771	+ if (shmem_getpage(mapping->host, index, &page,
	1772	+ SGP_NOHUGE)) {
	1773	+ result = SCAN_FAIL;
	1774	+ goto xa_unlocked;
	1775	+ }
	1776	+ } else if (trylock_page(page)) {
	1777	+ get_page(page);
	1778	+ xas_unlock_irq(&xas);
	1779	+ } else {
	1780	+ result = SCAN_PAGE_LOCK;
	1781	+ goto xa_locked;
	1782	+ }
	1783	+ } else { /* !is_shmem */
	1784	+ if (!page \|\| xa_is_value(page)) {
	1785	+ xas_unlock_irq(&xas);
	1786	+ page_cache_sync_readahead(mapping, &file->f_ra,
	1787	+ file, index,
	1788	+ end - index);
	1789	+ /* drain pagevecs to help isolate_lru_page() */
	1790	+ lru_add_drain();
	1791	+ page = find_lock_page(mapping, index);
	1792	+ if (unlikely(page == NULL)) {
	1793	+ result = SCAN_FAIL;
	1794	+ goto xa_unlocked;
	1795	+ }
	1796	+ } else if (PageDirty(page)) {
	1797	+ /*
	1798	+ * khugepaged only works on read-only fd,
	1799	+ * so this page is dirty because it hasn't
	1800	+ * been flushed since first write. There
	1801	+ * won't be new dirty pages.
	1802	+ *
	1803	+ * Trigger async flush here and hope the
	1804	+ * writeback is done when khugepaged
	1805	+ * revisits this page.
	1806	+ *
	1807	+ * This is a one-off situation. We are not
	1808	+ * forcing writeback in loop.
	1809	+ */
	1810	+ xas_unlock_irq(&xas);
	1811	+ filemap_flush(mapping);
	1812	+ result = SCAN_FAIL;
	1813	+ goto xa_unlocked;
	1814	+ } else if (PageWriteback(page)) {
	1815	+ xas_unlock_irq(&xas);
	1816	+ result = SCAN_FAIL;
	1817	+ goto xa_unlocked;
	1818	+ } else if (trylock_page(page)) {
	1819	+ get_page(page);
	1820	+ xas_unlock_irq(&xas);
	1821	+ } else {
	1822	+ result = SCAN_PAGE_LOCK;
	1823	+ goto xa_locked;
	1824	+ }
1412	1825	}
1413	1826
1414	1827	/*
..	..	@@ -1416,7 +1829,12 @@
1416	1829	* without racing with truncate.
1417	1830	*/
1418	1831	VM_BUG_ON_PAGE(!PageLocked(page), page);
1419		- VM_BUG_ON_PAGE(!PageUptodate(page), page);
	1832	+
	1833	+ /* make sure the page is up to date */
	1834	+ if (unlikely(!PageUptodate(page))) {
	1835	+ result = SCAN_FAIL;
	1836	+ goto out_unlock;
	1837	+ }
1420	1838
1421	1839	/*
1422	1840	* If file was truncated then extended, or hole-punched, before
..	..	@@ -1432,30 +1850,47 @@
1432	1850	goto out_unlock;
1433	1851	}
1434	1852
	1853	+ if (!is_shmem && (PageDirty(page) \|\|
	1854	+ PageWriteback(page))) {
	1855	+ /*
	1856	+ * khugepaged only works on read-only fd, so this
	1857	+ * page is dirty because it hasn't been flushed
	1858	+ * since first write.
	1859	+ */
	1860	+ result = SCAN_FAIL;
	1861	+ goto out_unlock;
	1862	+ }
	1863	+
1435	1864	if (isolate_lru_page(page)) {
1436	1865	result = SCAN_DEL_PAGE_LRU;
	1866	+ goto out_unlock;
	1867	+ }
	1868	+
	1869	+ if (page_has_private(page) &&
	1870	+ !try_to_release_page(page, GFP_KERNEL)) {
	1871	+ result = SCAN_PAGE_HAS_PRIVATE;
	1872	+ putback_lru_page(page);
1437	1873	goto out_unlock;
1438	1874	}
1439	1875
1440	1876	if (page_mapped(page))
1441	1877	unmap_mapping_pages(mapping, index, 1, false);
1442	1878
1443		- xa_lock_irq(&mapping->i_pages);
	1879	+ xas_lock_irq(&xas);
	1880	+ xas_set(&xas, index);
1444	1881
1445		- slot = radix_tree_lookup_slot(&mapping->i_pages, index);
1446		- VM_BUG_ON_PAGE(page != radix_tree_deref_slot_protected(slot,
1447		- &mapping->i_pages.xa_lock), page);
	1882	+ VM_BUG_ON_PAGE(page != xas_load(&xas), page);
1448	1883	VM_BUG_ON_PAGE(page_mapped(page), page);
1449	1884
1450	1885	/*
1451	1886	* The page is expected to have page_count() == 3:
1452	1887	* - we hold a pin on it;
1453		- * - one reference from radix tree;
	1888	+ * - one reference from page cache;
1454	1889	* - one from isolate_lru_page;
1455	1890	*/
1456	1891	if (!page_ref_freeze(page, 3)) {
1457	1892	result = SCAN_PAGE_COUNT;
1458		- xa_unlock_irq(&mapping->i_pages);
	1893	+ xas_unlock_irq(&xas);
1459	1894	putback_lru_page(page);
1460	1895	goto out_unlock;
1461	1896	}
..	..	@@ -1467,58 +1902,50 @@
1467	1902	list_add_tail(&page->lru, &pagelist);
1468	1903
1469	1904	/* Finally, replace with the new page. */
1470		- radix_tree_replace_slot(&mapping->i_pages, slot,
1471		- new_page + (index % HPAGE_PMD_NR));
1472		-
1473		- slot = radix_tree_iter_resume(slot, &iter);
1474		- index++;
	1905	+ xas_store(&xas, new_page);
1475	1906	continue;
1476	1907	out_unlock:
1477	1908	unlock_page(page);
1478	1909	put_page(page);
1479		- goto tree_unlocked;
	1910	+ goto xa_unlocked;
1480	1911	}
1481	1912
1482		- /*
1483		- * Handle hole in radix tree at the end of the range.
1484		- * This code only triggers if there's nothing in radix tree
1485		- * beyond 'end'.
1486		- */
1487		- if (index < end) {
1488		- int n = end - index;
1489		-
1490		- /* Stop if extent has been truncated, and is now empty */
1491		- if (n >= HPAGE_PMD_NR) {
1492		- result = SCAN_TRUNCATED;
1493		- goto tree_locked;
1494		- }
1495		- if (!shmem_charge(mapping->host, n)) {
	1913	+ if (is_shmem)
	1914	+ __inc_node_page_state(new_page, NR_SHMEM_THPS);
	1915	+ else {
	1916	+ __inc_node_page_state(new_page, NR_FILE_THPS);
	1917	+ filemap_nr_thps_inc(mapping);
	1918	+ /*
	1919	+ * Paired with smp_mb() in do_dentry_open() to ensure
	1920	+ * i_writecount is up to date and the update to nr_thps is
	1921	+ * visible. Ensures the page cache will be truncated if the
	1922	+ * file is opened writable.
	1923	+ */
	1924	+ smp_mb();
	1925	+ if (inode_is_open_for_write(mapping->host)) {
1496	1926	result = SCAN_FAIL;
1497		- goto tree_locked;
	1927	+ __dec_node_page_state(new_page, NR_FILE_THPS);
	1928	+ filemap_nr_thps_dec(mapping);
	1929	+ goto xa_locked;
1498	1930	}
1499		- for (; index < end; index++) {
1500		- radix_tree_insert(&mapping->i_pages, index,
1501		- new_page + (index % HPAGE_PMD_NR));
1502		- }
1503		- nr_none += n;
1504	1931	}
1505	1932
1506		- __inc_node_page_state(new_page, NR_SHMEM_THPS);
1507	1933	if (nr_none) {
1508		- struct zone *zone = page_zone(new_page);
1509		-
1510		- __mod_node_page_state(zone->zone_pgdat, NR_FILE_PAGES, nr_none);
1511		- __mod_node_page_state(zone->zone_pgdat, NR_SHMEM, nr_none);
	1934	+ __mod_lruvec_page_state(new_page, NR_FILE_PAGES, nr_none);
	1935	+ if (is_shmem)
	1936	+ __mod_lruvec_page_state(new_page, NR_SHMEM, nr_none);
1512	1937	}
1513	1938
1514		-tree_locked:
1515		- xa_unlock_irq(&mapping->i_pages);
1516		-tree_unlocked:
	1939	+xa_locked:
	1940	+ xas_unlock_irq(&xas);
	1941	+xa_unlocked:
1517	1942
1518	1943	if (result == SCAN_SUCCEED) {
	1944	+ struct page page, tmp;
	1945	+
1519	1946	/*
1520		- * Replacing old pages with new one has succeed, now we need to
1521		- * copy the content and free old pages.
	1947	+ * Replacing old pages with new one has succeeded, now we
	1948	+ * need to copy the content and free the old pages.
1522	1949	*/
1523	1950	index = start;
1524	1951	list_for_each_entry_safe(page, tmp, &pagelist, lru) {
..	..	@@ -1544,9 +1971,9 @@
1544	1971
1545	1972	SetPageUptodate(new_page);
1546	1973	page_ref_add(new_page, HPAGE_PMD_NR - 1);
1547		- set_page_dirty(new_page);
1548		- mem_cgroup_commit_charge(new_page, memcg, false, true);
1549		- lru_cache_add_anon(new_page);
	1974	+ if (is_shmem)
	1975	+ set_page_dirty(new_page);
	1976	+ lru_cache_add(new_page);
1550	1977
1551	1978	/*
1552	1979	* Remove pte page tables, so we can re-fault the page as huge.
..	..	@@ -1556,57 +1983,60 @@
1556	1983
1557	1984	khugepaged_pages_collapsed++;
1558	1985	} else {
1559		- /* Something went wrong: rollback changes to the radix-tree */
1560		- xa_lock_irq(&mapping->i_pages);
1561		- mapping->nrpages -= nr_none;
1562		- shmem_uncharge(mapping->host, nr_none);
	1986	+ struct page *page;
1563	1987
1564		- radix_tree_for_each_slot(slot, &mapping->i_pages, &iter, start) {
1565		- if (iter.index >= end)
1566		- break;
	1988	+ /* Something went wrong: roll back page cache changes */
	1989	+ xas_lock_irq(&xas);
	1990	+ mapping->nrpages -= nr_none;
	1991	+
	1992	+ if (is_shmem)
	1993	+ shmem_uncharge(mapping->host, nr_none);
	1994	+
	1995	+ xas_set(&xas, start);
	1996	+ xas_for_each(&xas, page, end - 1) {
1567	1997	page = list_first_entry_or_null(&pagelist,
1568	1998	struct page, lru);
1569		- if (!page \|\| iter.index < page->index) {
	1999	+ if (!page \|\| xas.xa_index < page->index) {
1570	2000	if (!nr_none)
1571	2001	break;
1572	2002	nr_none--;
1573	2003	/* Put holes back where they were */
1574		- radix_tree_delete(&mapping->i_pages, iter.index);
	2004	+ xas_store(&xas, NULL);
1575	2005	continue;
1576	2006	}
1577	2007
1578		- VM_BUG_ON_PAGE(page->index != iter.index, page);
	2008	+ VM_BUG_ON_PAGE(page->index != xas.xa_index, page);
1579	2009
1580	2010	/* Unfreeze the page. */
1581	2011	list_del(&page->lru);
1582	2012	page_ref_unfreeze(page, 2);
1583		- radix_tree_replace_slot(&mapping->i_pages, slot, page);
1584		- slot = radix_tree_iter_resume(slot, &iter);
1585		- xa_unlock_irq(&mapping->i_pages);
	2013	+ xas_store(&xas, page);
	2014	+ xas_pause(&xas);
	2015	+ xas_unlock_irq(&xas);
1586	2016	unlock_page(page);
1587	2017	putback_lru_page(page);
1588		- xa_lock_irq(&mapping->i_pages);
	2018	+ xas_lock_irq(&xas);
1589	2019	}
1590	2020	VM_BUG_ON(nr_none);
1591		- xa_unlock_irq(&mapping->i_pages);
	2021	+ xas_unlock_irq(&xas);
1592	2022
1593		- mem_cgroup_cancel_charge(new_page, memcg, true);
1594	2023	new_page->mapping = NULL;
1595	2024	}
1596	2025
1597	2026	unlock_page(new_page);
1598	2027	out:
1599	2028	VM_BUG_ON(!list_empty(&pagelist));
	2029	+ if (!IS_ERR_OR_NULL(*hpage))
	2030	+ mem_cgroup_uncharge(*hpage);
1600	2031	/* TODO: tracepoints */
1601	2032	}
1602	2033
1603		-static void khugepaged_scan_shmem(struct mm_struct *mm,
1604		- struct address_space *mapping,
1605		- pgoff_t start, struct page **hpage)
	2034	+static void khugepaged_scan_file(struct mm_struct *mm,
	2035	+ struct file file, pgoff_t start, struct page *hpage)
1606	2036	{
1607	2037	struct page *page = NULL;
1608		- struct radix_tree_iter iter;
1609		- void **slot;
	2038	+ struct address_space *mapping = file->f_mapping;
	2039	+ XA_STATE(xas, &mapping->i_pages, start);
1610	2040	int present, swap;
1611	2041	int node = NUMA_NO_NODE;
1612	2042	int result = SCAN_SUCCEED;
..	..	@@ -1615,17 +2045,11 @@
1615	2045	swap = 0;
1616	2046	memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
1617	2047	rcu_read_lock();
1618		- radix_tree_for_each_slot(slot, &mapping->i_pages, &iter, start) {
1619		- if (iter.index >= start + HPAGE_PMD_NR)
1620		- break;
1621		-
1622		- page = radix_tree_deref_slot(slot);
1623		- if (radix_tree_deref_retry(page)) {
1624		- slot = radix_tree_iter_retry(&iter);
	2048	+ xas_for_each(&xas, page, start + HPAGE_PMD_NR - 1) {
	2049	+ if (xas_retry(&xas, page))
1625	2050	continue;
1626		- }
1627	2051
1628		- if (radix_tree_exception(page)) {
	2052	+ if (xa_is_value(page)) {
1629	2053	if (++swap > khugepaged_max_ptes_swap) {
1630	2054	result = SCAN_EXCEED_SWAP_PTE;
1631	2055	break;
..	..	@@ -1650,7 +2074,8 @@
1650	2074	break;
1651	2075	}
1652	2076
1653		- if (page_count(page) != 1 + page_mapcount(page)) {
	2077	+ if (page_count(page) !=
	2078	+ 1 + page_mapcount(page) + page_has_private(page)) {
1654	2079	result = SCAN_PAGE_COUNT;
1655	2080	break;
1656	2081	}
..	..	@@ -1664,7 +2089,7 @@
1664	2089	present++;
1665	2090
1666	2091	if (need_resched()) {
1667		- slot = radix_tree_iter_resume(slot, &iter);
	2092	+ xas_pause(&xas);
1668	2093	cond_resched_rcu();
1669	2094	}
1670	2095	}
..	..	@@ -1675,18 +2100,22 @@
1675	2100	result = SCAN_EXCEED_NONE_PTE;
1676	2101	} else {
1677	2102	node = khugepaged_find_target_node();
1678		- collapse_shmem(mm, mapping, start, hpage, node);
	2103	+ collapse_file(mm, file, start, hpage, node);
1679	2104	}
1680	2105	}
1681	2106
1682	2107	/* TODO: tracepoints */
1683	2108	}
1684	2109	#else
1685		-static void khugepaged_scan_shmem(struct mm_struct *mm,
1686		- struct address_space *mapping,
1687		- pgoff_t start, struct page **hpage)
	2110	+static void khugepaged_scan_file(struct mm_struct *mm,
	2111	+ struct file file, pgoff_t start, struct page *hpage)
1688	2112	{
1689	2113	BUILD_BUG();
	2114	+}
	2115	+
	2116	+static int khugepaged_collapse_pte_mapped_thps(struct mm_slot *mm_slot)
	2117	+{
	2118	+ return 0;
1690	2119	}
1691	2120	#endif
1692	2121
..	..	@@ -1701,7 +2130,7 @@
1701	2130	int progress = 0;
1702	2131
1703	2132	VM_BUG_ON(!pages);
1704		- VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
	2133	+ lockdep_assert_held(&khugepaged_mm_lock);
1705	2134
1706	2135	if (khugepaged_scan.mm_slot)
1707	2136	mm_slot = khugepaged_scan.mm_slot;
..	..	@@ -1712,6 +2141,7 @@
1712	2141	khugepaged_scan.mm_slot = mm_slot;
1713	2142	}
1714	2143	spin_unlock(&khugepaged_mm_lock);
	2144	+ khugepaged_collapse_pte_mapped_thps(mm_slot);
1715	2145
1716	2146	mm = mm_slot->mm;
1717	2147	/*
..	..	@@ -1719,8 +2149,8 @@
1719	2149	* the next mm on the list.
1720	2150	*/
1721	2151	vma = NULL;
1722		- if (unlikely(!down_read_trylock(&mm->mmap_sem)))
1723		- goto breakouterloop_mmap_sem;
	2152	+ if (unlikely(!mmap_read_trylock(mm)))
	2153	+ goto breakouterloop_mmap_lock;
1724	2154	if (likely(!khugepaged_test_exit(mm)))
1725	2155	vma = find_vma(mm, khugepaged_scan.address);
1726	2156
..	..	@@ -1747,6 +2177,8 @@
1747	2177	if (khugepaged_scan.address < hstart)
1748	2178	khugepaged_scan.address = hstart;
1749	2179	VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
	2180	+ if (shmem_file(vma->vm_file) && !shmem_huge_enabled(vma))
	2181	+ goto skip;
1750	2182
1751	2183	while (khugepaged_scan.address < hend) {
1752	2184	int ret;
..	..	@@ -1757,17 +2189,14 @@
1757	2189	VM_BUG_ON(khugepaged_scan.address < hstart \|\|
1758	2190	khugepaged_scan.address + HPAGE_PMD_SIZE >
1759	2191	hend);
1760		- if (shmem_file(vma->vm_file)) {
1761		- struct file *file;
	2192	+ if (IS_ENABLED(CONFIG_SHMEM) && vma->vm_file) {
	2193	+ struct file *file = get_file(vma->vm_file);
1762	2194	pgoff_t pgoff = linear_page_index(vma,
1763	2195	khugepaged_scan.address);
1764		- if (!shmem_huge_enabled(vma))
1765		- goto skip;
1766		- file = get_file(vma->vm_file);
1767		- up_read(&mm->mmap_sem);
	2196	+
	2197	+ mmap_read_unlock(mm);
1768	2198	ret = 1;
1769		- khugepaged_scan_shmem(mm, file->f_mapping,
1770		- pgoff, hpage);
	2199	+ khugepaged_scan_file(mm, file, pgoff, hpage);
1771	2200	fput(file);
1772	2201	} else {
1773	2202	ret = khugepaged_scan_pmd(mm, vma,
..	..	@@ -1778,15 +2207,15 @@
1778	2207	khugepaged_scan.address += HPAGE_PMD_SIZE;
1779	2208	progress += HPAGE_PMD_NR;
1780	2209	if (ret)
1781		- /* we released mmap_sem so break loop */
1782		- goto breakouterloop_mmap_sem;
	2210	+ /* we released mmap_lock so break loop */
	2211	+ goto breakouterloop_mmap_lock;
1783	2212	if (progress >= pages)
1784	2213	goto breakouterloop;
1785	2214	}
1786	2215	}
1787	2216	breakouterloop:
1788		- up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
1789		-breakouterloop_mmap_sem:
	2217	+ mmap_read_unlock(mm); /* exit_mmap will destroy ptes after this */
	2218	+breakouterloop_mmap_lock:
1790	2219
1791	2220	spin_lock(&khugepaged_mm_lock);
1792	2221	VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
..	..	@@ -1837,6 +2266,8 @@
1837	2266
1838	2267	barrier(); /* write khugepaged_pages_to_scan to local stack */
1839	2268
	2269	+ lru_add_drain_all();
	2270	+
1840	2271	while (progress < pages) {
1841	2272	if (!khugepaged_prealloc_page(&hpage, &wait))
1842	2273	break;