* Peng Zhang <zhangpeng.00@xxxxxxxxxxxxx> [231009 05:04]:
In dup_mmap(), using __mt_dup() to duplicate the old maple tree and then
directly replacing the entries of VMAs in the new maple tree can result
in better performance. __mt_dup() uses DFS pre-order to duplicate the
maple tree, so it is efficient.
The average time complexity of __mt_dup() is O(n), where n is the number
of VMAs. The proof of the time complexity is provided in the commit log
that introduces __mt_dup(). After duplicating the maple tree, each element
is traversed and replaced (ignoring the cases of deletion, which are rare).
Since it is only a replacement operation for each element, this process is
also O(n).
Analyzing the exact time complexity of the previous algorithm is
challenging because each insertion can involve appending to a node, pushing
data to adjacent nodes, or even splitting nodes. The frequency of each
action is difficult to calculate. The worst-case scenario for a single
insertion is when the tree undergoes splitting at every level. If we
consider each insertion as the worst-case scenario, we can determine that
the upper bound of the time complexity is O(n*log(n)), although this is a
loose upper bound. However, based on the test data, it appears that the
actual time complexity is likely to be O(n).
As the entire maple tree is duplicated using __mt_dup(), if dup_mmap()
fails, there will be a portion of VMAs that have not been duplicated in
the maple tree. This makes it impossible to unmap all VMAs in exit_mmap().
To solve this problem, undo_dup_mmap() is introduced to handle the failure
of dup_mmap(). I have carefully tested the failure path and so far it
seems there are no issues.
There is a "spawn" in byte-unixbench[1], which can be used to test the
performance of fork(). I modified it slightly to make it work with
different number of VMAs.
Below are the test results. The first row shows the number of VMAs.
The second and third rows show the number of fork() calls per ten seconds,
corresponding to next-20231006 and the this patchset, respectively. The
test results were obtained with CPU binding to avoid scheduler load
balancing that could cause unstable results. There are still some
fluctuations in the test results, but at least they are better than the
original performance.
21 121 221 421 821 1621 3221 6421 12821 25621 51221
112100 76261 54227 34035 20195 11112 6017 3161 1606 802 393
114558 83067 65008 45824 28751 16072 8922 4747 2436 1233 599
2.19% 8.92% 19.88% 34.64% 42.37% 44.64% 48.28% 50.17% 51.68% 53.74% 52.42%
[1] https://github.com/kdlucas/byte-unixbench/tree/master
Signed-off-by: Peng Zhang <zhangpeng.00@xxxxxxxxxxxxx>
---
include/linux/mm.h | 1 +
kernel/fork.c | 34 +++++++++++++++++++++----------
mm/internal.h | 3 ++-
mm/memory.c | 7 ++++---
mm/mmap.c | 50 ++++++++++++++++++++++++++++++++++++++++++++--
5 files changed, 78 insertions(+), 17 deletions(-)
diff --git a/include/linux/mm.h b/include/linux/mm.h
index 14e50925b76d..d039f10d258e 100644
--- a/include/linux/mm.h
+++ b/include/linux/mm.h
@@ -3248,6 +3248,7 @@ extern void unlink_file_vma(struct vm_area_struct *);
extern struct vm_area_struct *copy_vma(struct vm_area_struct **,
unsigned long addr, unsigned long len, pgoff_t pgoff,
bool *need_rmap_locks);
+extern void undo_dup_mmap(struct mm_struct *mm, struct vm_area_struct *vma_end);
extern void exit_mmap(struct mm_struct *);
static inline int check_data_rlimit(unsigned long rlim,
diff --git a/kernel/fork.c b/kernel/fork.c
index 0ff2e0cd4109..5f24f6d68ea4 100644
--- a/kernel/fork.c
+++ b/kernel/fork.c
@@ -650,7 +650,6 @@ static __latent_entropy int dup_mmap(struct mm_struct *mm,
int retval;
unsigned long charge = 0;
LIST_HEAD(uf);
- VMA_ITERATOR(old_vmi, oldmm, 0);
VMA_ITERATOR(vmi, mm, 0);
uprobe_start_dup_mmap();
@@ -678,16 +677,25 @@ static __latent_entropy int dup_mmap(struct mm_struct *mm,
goto out;
khugepaged_fork(mm, oldmm);
- retval = vma_iter_bulk_alloc(&vmi, oldmm->map_count);
- if (retval)
+ /* Use __mt_dup() to efficiently build an identical maple tree. */
+ retval = __mt_dup(&oldmm->mm_mt, &mm->mm_mt, GFP_KERNEL);
+ if (unlikely(retval))
goto out;
mt_clear_in_rcu(vmi.mas.tree);
- for_each_vma(old_vmi, mpnt) {
+ for_each_vma(vmi, mpnt) {
struct file *file;
vma_start_write(mpnt);
if (mpnt->vm_flags & VM_DONTCOPY) {
+ mas_store_gfp(&vmi.mas, NULL, GFP_KERNEL);
+
+ /* If failed, undo all completed duplications. */
+ if (unlikely(mas_is_err(&vmi.mas))) {
+ retval = xa_err(vmi.mas.node);
+ goto loop_out;
+ }
+
vm_stat_account(mm, mpnt->vm_flags, -vma_pages(mpnt));
I am not sure how we are getting away with this, but the mm stats are
copied before we enter this loop, so I'm surprised that we aren't
getting complaints about the VMAs that are later than the failure. I
don't think this needs to be fixed, it's just odd and it existed before
this change as well.
continue;
}
@@ -749,9 +757,11 @@ static __latent_entropy int dup_mmap(struct mm_struct *mm,
if (is_vm_hugetlb_page(tmp))
hugetlb_dup_vma_private(tmp);
- /* Link the vma into the MT */
- if (vma_iter_bulk_store(&vmi, tmp))
- goto fail_nomem_vmi_store;
+ /*
+ * Link the vma into the MT. After using __mt_dup(), memory
+ * allocation is not necessary here, so it cannot fail.
+ */
+ mas_store(&vmi.mas, tmp);
mm->map_count++;
if (!(tmp->vm_flags & VM_WIPEONFORK))
@@ -760,15 +770,19 @@ static __latent_entropy int dup_mmap(struct mm_struct *mm,
if (tmp->vm_ops && tmp->vm_ops->open)
tmp->vm_ops->open(tmp);
- if (retval)
+ if (retval) {
+ mpnt = vma_next(&vmi);
goto loop_out;
+ }
}
/* a new mm has just been created */
retval = arch_dup_mmap(oldmm, mm);
loop_out:
vma_iter_free(&vmi);
- if (!retval)
+ if (likely(!retval))
mt_set_in_rcu(vmi.mas.tree);
+ else
+ undo_dup_mmap(mm, mpnt);
out:
mmap_write_unlock(mm);
flush_tlb_mm(oldmm);
@@ -778,8 +792,6 @@ static __latent_entropy int dup_mmap(struct mm_struct *mm,
uprobe_end_dup_mmap();
return retval;
-fail_nomem_vmi_store:
- unlink_anon_vmas(tmp);
fail_nomem_anon_vma_fork:
mpol_put(vma_policy(tmp));
fail_nomem_policy:
diff --git a/mm/internal.h b/mm/internal.h
index 18e360fa53bc..bcd92a5b5474 100644
--- a/mm/internal.h
+++ b/mm/internal.h
@@ -111,7 +111,8 @@ void folio_activate(struct folio *folio);
void free_pgtables(struct mmu_gather *tlb, struct ma_state *mas,
struct vm_area_struct *start_vma, unsigned long floor,
- unsigned long ceiling, bool mm_wr_locked);
+ unsigned long ceiling, unsigned long tree_end,
+ bool mm_wr_locked);
void pmd_install(struct mm_struct *mm, pmd_t *pmd, pgtable_t *pte);
struct zap_details;
diff --git a/mm/memory.c b/mm/memory.c
index b320af6466cc..51bb1d16a54e 100644
--- a/mm/memory.c
+++ b/mm/memory.c
@@ -363,7 +363,8 @@ void free_pgd_range(struct mmu_gather *tlb,
void free_pgtables(struct mmu_gather *tlb, struct ma_state *mas,
struct vm_area_struct *vma, unsigned long floor,
- unsigned long ceiling, bool mm_wr_locked)
+ unsigned long ceiling, unsigned long tree_end,
+ bool mm_wr_locked)
{
do {
unsigned long addr = vma->vm_start;
@@ -373,7 +374,7 @@ void free_pgtables(struct mmu_gather *tlb, struct ma_state *mas,
* Note: USER_PGTABLES_CEILING may be passed as ceiling and may
* be 0. This will underflow and is okay.
*/
- next = mas_find(mas, ceiling - 1);
+ next = mas_find(mas, tree_end - 1);
/*
* Hide vma from rmap and truncate_pagecache before freeing
@@ -394,7 +395,7 @@ void free_pgtables(struct mmu_gather *tlb, struct ma_state *mas,
while (next && next->vm_start <= vma->vm_end + PMD_SIZE
&& !is_vm_hugetlb_page(next)) {
vma = next;
- next = mas_find(mas, ceiling - 1);
+ next = mas_find(mas, tree_end - 1);
if (mm_wr_locked)
vma_start_write(vma);
unlink_anon_vmas(vma);
diff --git a/mm/mmap.c b/mm/mmap.c
index 1855a2d84200..d044d68d1361 100644
--- a/mm/mmap.c
+++ b/mm/mmap.c
@@ -2337,7 +2337,7 @@ static void unmap_region(struct mm_struct *mm, struct ma_state *mas,
mas_set(mas, mt_start);
free_pgtables(&tlb, mas, vma, prev ? prev->vm_end : FIRST_USER_ADDRESS,
next ? next->vm_start : USER_PGTABLES_CEILING,
- mm_wr_locked);
+ tree_end, mm_wr_locked);
tlb_finish_mmu(&tlb);
}
@@ -3197,6 +3197,52 @@ int vm_brk_flags(unsigned long addr, unsigned long request, unsigned long flags)
}
EXPORT_SYMBOL(vm_brk_flags);
+void undo_dup_mmap(struct mm_struct *mm, struct vm_area_struct *vma_end)
+{
+ unsigned long tree_end = USER_PGTABLES_CEILING;
+ VMA_ITERATOR(vmi, mm, 0);
+ struct vm_area_struct *vma;
+ unsigned long nr_accounted = 0;
+ int count = 0;
+
+ /*
+ * vma_end points to the first VMA that has not been duplicated. We need
+ * to unmap all VMAs before it.
+ * If vma_end is NULL, it means that all VMAs in the maple tree have
+ * been duplicated, so setting tree_end to USER_PGTABLES_CEILING will
+ * unmap all VMAs in the maple tree.
+ */
+ if (vma_end) {
+ tree_end = vma_end->vm_start;
+ if (tree_end == 0)
+ goto destroy;
+ }
+
+ vma = vma_find(&vmi, tree_end);
+ if (!vma)
+ goto destroy;
+
+ arch_unmap(mm, vma->vm_start, tree_end);
+
+ vma_iter_set(&vmi, vma->vm_end);
FYI, This missing vma_iter_set() in v3 was not caught by your testing
because it would not cause an issue, just avoid the optimisation in the
gathering of page tables.
+ unmap_region(mm, &vmi.mas, vma, NULL, NULL, 0, tree_end, tree_end, true);
+
I really don't like having to modify unmap_region() and free_pgtables()
for a rare error case. Looking into the issue, you are correct in the
rounding that is happening in free_pgd_range() and this alignment to
avoid "unnecessary work" is causing us issues. However, if we open code
it a lot like what exit_mmap() does, we can avoid changing these
functions:
+ lru_add_drain();
+ tlb_gather_mmu(&tlb, mm);
+ update_hiwater_rss(mm);
+ unmap_vmas(&tlb, &vmi.mas, vma, 0, tree_end, tree_end, true);
+ vma_iter_set(&vmi, vma->vm_end);
+ free_pgtables(&tlb, &vmi.mas, vma, FIRST_USER_ADDRESS, vma_end->vm_start,
+ true);
+ free_pgd_range(&tlb, vma->vm_start, vma_end->vm_start,
+ FIRST_USER_ADDRESS, USER_PGTABLES_CEILING);
