[PATCH v4 0/9] mm/vmalloc: fix up vmap error handling

[PATCH v4 0/9] mm/vmalloc: fix up vmap error handling

[Hao Ge]

__vmap_pages_range_noflush() and friends can install some PTEs before
failing and leave them mapped, and the kernel callers do not agree on
who cleans them up. pcpu_map_pages() and kmsan_ioremap_page_range()
roll back what they mapped before the failure, while
vm_module_tags_populate() and the __GFP_NOFAIL retry loop in
__vmalloc_area_node() rely on the mapping functions cleaning up and
do not call anything like vunmap_range() themselves. When the same
range is mapped again, the attempt hits the leftovers and fails with
a warning in vmap_pages_pte_range(). A leftover huge pmd is worse.
During the retry, vmap_try_huge_pmd() calls pmd_free_pte_page() on
it, and the x86 implementation cannot tell a leaf entry from a page
table and frees the mapped page.

After discussing with Suren and Ulad [1], we move the rollback into
the mapping functions. A failed vmap leaves nothing mapped, and the
callers no longer clean up after it.

Patch 1 is the KMSAN part. The shadow and the origin of a range are
mapped by two separate calls, so the rollback there has to cover
both. Patch 2 is its disabled-KMSAN counterpart. Patches 3 to 5 bug
out on an unaligned range at the entry of the three range walkers.
Patch 6 only flushes the cache once the mapping succeeded. Patch 7
adds the rollback to __vmap_pages_range_noflush() and
vmap_page_range(). Patches 8 and 9 drop the powerpc and percpu
cleanups that are not needed anymore.

Changes since v3:
- drop v3 patch 3. The clamp papered over the caller bug, a bad
  range now BUG_ONs at the unmap entry
- drop v3 patch 4. The reworked walk (24bbe1c56f63a) propagates
  the errors already
- drop v3 patch 6. The walk below checks the range already
- the map side checks moved from the pte level to the entries of
  vmap_range_noflush() and the walk, and became BUG_ON, in line
  with the unmap side
- rebase onto the walk rework, commit messages updated accordingly

[1] https://lore.kernel.org/all/20260915070001.113559-3-hao.ge@linux.dev/

v3 link: https://lore.kernel.org/all/20261026080329.140244-1-hao.ge@linux.dev/
v2 link: https://lore.kernel.org/all/20261019054531.87652-1-hao.ge@linux.dev/
v1 link: https://lore.kernel.org/all/20261009063619.112313-1-hao.ge@linux.dev/

Hao Ge (9):
  mm/kmsan: undo the shadow mapping when the origin mapping fails
  mm/kmsan: skip vunmap_range_noflush when KMSAN is disabled
  mm/vmalloc: BUG_ON() on an unaligned range in vmap_range_noflush()
  mm/vmalloc: BUG_ON() on an unaligned range in
    vmap_pages_range_noflush_walk()
  mm/vmalloc: BUG_ON() on an unaligned range in __vunmap_range_noflush()
  mm/vmalloc: only flush the cache after the vmap succeeded
  mm/vmalloc: undo partial mappings inside the mapping functions
  powerpc: drop redundant unmaps of failed vmap mappings
  mm/percpu: stop unmapping the CPU that failed to map

 arch/powerpc/kernel/isa-bridge.c |  5 +--
 arch/powerpc/kernel/pci_64.c     |  4 +-
 include/linux/kmsan.h            |  8 +++-
 mm/kmsan/hooks.c                 | 21 ++++++---
 mm/kmsan/shadow.c                |  9 +++-
 mm/percpu-vm.c                   |  5 ++-
 mm/vmalloc.c                     | 75 ++++++++++++++++++++++++--------
 mm/vmalloc.h                     |  6 ++-
 8 files changed, 94 insertions(+), 39 deletions(-)

-- 
2.25.1

[PATCH v4 1/9] mm/kmsan: undo the shadow mapping when the origin mapping fails

[Hao Ge]

kmsan_vmap_pages_range_noflush() first maps the shadow pages and then
the origin pages. If the second mapping fails, the first one is left
mapped, and the callers do not roll it back. The next vmap of the
same metadata range hits the stale PTEs again and fails with a
WARN_ON() and -EBUSY in vmap_pages_pte_range().

Undo the shadow mapping on that error path, the same way
kmsan_ioremap_page_range() cleans up after a partial failure.
__vunmap_range_noflush() only clears the PTEs; nothing has touched
the shadow mapping, so no TLB flush is needed.

Fixes: 47ebd0310e89 ("mm: kmsan: handle alloc failures in kmsan_vmap_pages_range_noflush()")
Cc: stable@vger.kernel.org
Reviewed-by: Suren Baghdasaryan <surenb@google.com>
Reviewed-by: Dev Jain <dev.jain@arm.com>
Signed-off-by: Hao Ge <hao.ge@linux.dev>
---
 mm/kmsan/shadow.c | 4 ++++
 1 file changed, 4 insertions(+)

diff --git a/mm/kmsan/shadow.c b/mm/kmsan/shadow.c
index 0c88d89bf0d6..2166086d3dc3 100644
--- a/mm/kmsan/shadow.c
+++ b/mm/kmsan/shadow.c
@@ -258,6 +258,10 @@ int kmsan_vmap_pages_range_noflush(unsigned long start, unsigned long end,
 					    o_pages, page_shift);
 	kmsan_leave_runtime();
 	if (mapped) {
+		/* Undo the shadow mapping set up above. */
+		kmsan_enter_runtime();
+		__vunmap_range_noflush(shadow_start, shadow_end);
+		kmsan_leave_runtime();
 		err = mapped;
 		goto ret;
 	}
-- 
2.25.1

[PATCH v4 2/9] mm/kmsan: skip vunmap_range_noflush when KMSAN is disabled

[Hao Ge]

kmsan_vmap_pages_range_noflush() returns 0 when kmsan_enabled is
false.  The unmap side does not, so it still walks the shadow and
origin page tables even though KMSAN never installed any metadata
there.

Add the matching early return.

Reported-by: Sashiko <sashiko-bot@kernel.org>
Reviewed-by: Dev Jain <dev.jain@arm.com>
Signed-off-by: Hao Ge <hao.ge@linux.dev>
---
 mm/kmsan/hooks.c | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/mm/kmsan/hooks.c b/mm/kmsan/hooks.c
index a706666db097..107f66841274 100644
--- a/mm/kmsan/hooks.c
+++ b/mm/kmsan/hooks.c
@@ -138,6 +138,9 @@ static unsigned long vmalloc_origin(unsigned long addr)
 
 void kmsan_vunmap_range_noflush(unsigned long start, unsigned long end)
 {
+	if (!kmsan_enabled)
+		return;
+
 	__vunmap_range_noflush(vmalloc_shadow(start), vmalloc_shadow(end));
 	__vunmap_range_noflush(vmalloc_origin(start), vmalloc_origin(end));
 	flush_cache_vmap(vmalloc_shadow(start), vmalloc_shadow(end));
-- 
2.25.1

[PATCH v4 3/9] mm/vmalloc: BUG_ON() on an unaligned range in vmap_range_noflush()

[Hao Ge]

commit a739e6b557af ("mm: vmalloc: WARN_ON if mapping size is not
PAGE_SIZE aligned") put the check in vmap_pte_range(), which is too
late, because the upper levels may already have installed page
tables for a range that is rejected anyway.

The pte loop maps whole pages and stops only when addr lands
exactly on end.  An unaligned end lets addr jump over it and the
loop keeps mapping past the range.  An unaligned start maps the
page below the range and usually misses end too.  Such a range is
never a legitimate request, and the corruption it leaves behind is
hard to attribute to the caller afterwards, so bug out on addr and
end at the entry of vmap_range_noflush() and drop the check from
vmap_pte_range().

Signed-off-by: Hao Ge <hao.ge@linux.dev>
---
 mm/vmalloc.c | 4 +---
 1 file changed, 1 insertion(+), 3 deletions(-)

diff --git a/mm/vmalloc.c b/mm/vmalloc.c
index 093b8d5f7b0e..27f2674a5556 100644
--- a/mm/vmalloc.c
+++ b/mm/vmalloc.c
@@ -125,9 +125,6 @@ static int vmap_pte_range(pmd_t *pmd, unsigned long addr, unsigned long end,
 	unsigned long size;
 	unsigned int nr_pages;
 
-	if (WARN_ON_ONCE(!PAGE_ALIGNED(end - addr)))
-		return -EINVAL;
-
 	pfn = phys_addr >> PAGE_SHIFT;
 	pte = pte_alloc_kernel_track(pmd, addr, mask);
 	if (!pte)
@@ -347,6 +344,7 @@ static int vmap_range_noflush(unsigned long addr, unsigned long end,
 	 */
 	might_sleep();
 	BUG_ON(addr >= end);
+	BUG_ON(!PAGE_ALIGNED(addr) || !PAGE_ALIGNED(end));
 
 	start = addr;
 	pgd = pgd_offset_k(addr);
-- 
2.25.1

Re: [PATCH v4 3/9] mm/vmalloc: BUG_ON() on an unaligned range in vmap_range_noflush()

[Hao Ge]

On 2026/11/10 19:25, Hao Ge wrote:
> commit a739e6b557af ("mm: vmalloc: WARN_ON if mapping size is not
> PAGE_SIZE aligned") put the check in vmap_pte_range(), which is too
> late, because the upper levels may already have installed page
> tables for a range that is rejected anyway.
> 
> The pte loop maps whole pages and stops only when addr lands
> exactly on end.  An unaligned end lets addr jump over it and the
> loop keeps mapping past the range.  An unaligned start maps the
> page below the range and usually misses end too.  Such a range is
> never a legitimate request, and the corruption it leaves behind is
> hard to attribute to the caller afterwards, so bug out on addr and
> end at the entry of vmap_range_noflush() and drop the check from
> vmap_pte_range().
> 
> Signed-off-by: Hao Ge <hao.ge@linux.dev>
> ---
>  mm/vmalloc.c | 4 +---
>  1 file changed, 1 insertion(+), 3 deletions(-)
> 
> diff --git a/mm/vmalloc.c b/mm/vmalloc.c
> index 093b8d5f7b0e..27f2674a5556 100644
> --- a/mm/vmalloc.c
> +++ b/mm/vmalloc.c
> @@ -125,9 +125,6 @@ static int vmap_pte_range(pmd_t *pmd, unsigned long addr, unsigned long end,
>  	unsigned long size;
>  	unsigned int nr_pages;
>  
> -	if (WARN_ON_ONCE(!PAGE_ALIGNED(end - addr)))
> -		return -EINVAL;
> -
>  	pfn = phys_addr >> PAGE_SHIFT;
>  	pte = pte_alloc_kernel_track(pmd, addr, mask);
>  	if (!pte)
> @@ -347,6 +344,7 @@ static int vmap_range_noflush(unsigned long addr, unsigned long end,
>  	 */
>  	might_sleep();
>  	BUG_ON(addr >= end);
> +	BUG_ON(!PAGE_ALIGNED(addr) || !PAGE_ALIGNED(end));

I know BUG_ON() is ugly and it was not my first choice either, but I
have not figured out a better way to handle __vunmap_range_noflush().

If an unaligned range is passed to it, the pte loop may clear PTEs
adjacent to the range. Say a caller passes [3, 12) while [0, 3)
belongs to somebody else. The loop operates on whole pages, so it
starts at the page containing 3 and clears [0, 3) with it. That
silently corrupts a live mapping, and the corruption is not detected
until the owner touches the page. Bugs like that are nearly
impossible to debug afterwards.

It is not feasible for __vunmap_range_noflush() to just return an
invalid-argument error upfront. Its job is to clear PTEs, and an
early return leaves the residual PTEs behind.

I also thought about warning and returning -EINVAL on the map side
and keeping BUG_ON() only in __vunmap_range_noflush(), but the
rollback kills that idea.

int __vmap_pages_range_noflush(unsigned long addr, unsigned long end,
                pgprot_t prot, struct page **pages, unsigned int page_shift,
                gfp_t gfp_mask)
{
        int err;

        if (WARN_ON_ONCE(page_shift < PAGE_SHIFT))
                return -EINVAL;

        if (!IS_ENABLED(CONFIG_HAVE_ARCH_HUGE_VMALLOC))
                page_shift = PAGE_SHIFT;

        err = vmap_pages_range_noflush_walk(addr, end, prot, pages, page_shift);

        /* Undo the PTEs installed before the failure. */
        if (err)
                __vunmap_range_noflush(addr, end,
                                       gfpflags_allow_blocking(gfp_mask));

        return err;
}

If the walk WARNs and returns -EINVAL for a bad range, the undo
above passes that same range to __vunmap_range_noflush(), and we
BUG right after the WARN. vmap_page_range() runs into the same
thing on its error path.

So the map side cannot warn while the unmap side BUGs, the rollback
drags the rejected range into the unmap anyway. And the unmap side
cannot warn and skip either, that leaves the stale PTEs.

Also, these checks are the last line of defense. All in-tree callers
pass page aligned, non-empty ranges, and I am not aware of anyone ever
hitting this line. So for correct code the BUG_ON never fires at all,
no user will ever see it. The only one who can trip it is the developer
who just passed the bad range, and for him a stack trace right at the
spot is much better than chasing a silent corruption afterwards.

I settled on BUG_ON here. It may not be an ideal solution, but if anyone
has a better idea or alternative approach, please let me know.
I would greatly appreciate it.

Thanks
Best Regards
Hao

>  
>  	start = addr;
>  	pgd = pgd_offset_k(addr);

[PATCH v4 4/9] mm/vmalloc: BUG_ON() on an unaligned range in vmap_pages_range_noflush_walk()

[Hao Ge]

The pte loop of vmap_pages_pte_range() maps whole pages and stops
only when addr lands exactly on end.  An unaligned end lets addr
jump over it and the loop keeps mapping into whatever follows the
range.  An unaligned start maps the page below the range with its
first pte.  That corruption is impossible to debug afterwards, so
bug out on addr and end at the entry of
vmap_pages_range_noflush_walk().

Reported-by: Sashiko <sashiko-bot@kernel.org>
Signed-off-by: Hao Ge <hao.ge@linux.dev>
---
 mm/vmalloc.c | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/mm/vmalloc.c b/mm/vmalloc.c
index 27f2674a5556..86fbb1d2b27a 100644
--- a/mm/vmalloc.c
+++ b/mm/vmalloc.c
@@ -698,6 +698,8 @@ static int vmap_pages_range_noflush_walk(unsigned long addr, unsigned long end,
 	pgtbl_mod_mask mask = 0;
 
 	BUG_ON(addr >= end);
+	BUG_ON(!PAGE_ALIGNED(addr) || !PAGE_ALIGNED(end));
+
 	pgd = pgd_offset_k(addr);
 	do {
 		next = pgd_addr_end(addr, end);
-- 
2.25.1

[PATCH v4 5/9] mm/vmalloc: BUG_ON() on an unaligned range in __vunmap_range_noflush()

[Hao Ge]

The PTE loop of the unmap path stops only at addr == end.  With an
unaligned end addr never lands on end, so the loop walks past the
range and keeps clearing.  With an unaligned start the loop begins
in the page containing start, which extends below the range.  Either
way the unmap clears the mappings adjacent to the range, and that is
nearly impossible to debug afterwards.

vmap only installs page-aligned, page-granular mappings, so an
unaligned unmap request is never a legitimate teardown, and there is
no way to tell which pages the caller meant.  Check addr and end at
the entry of __vunmap_range_noflush() and bug out before any PTE is
cleared.

Signed-off-by: Hao Ge <hao.ge@linux.dev>
---
 mm/vmalloc.c | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/mm/vmalloc.c b/mm/vmalloc.c
index 86fbb1d2b27a..0565cbfe325c 100644
--- a/mm/vmalloc.c
+++ b/mm/vmalloc.c
@@ -518,6 +518,8 @@ void __vunmap_range_noflush(unsigned long start, unsigned long end)
 	pgtbl_mod_mask mask = 0;
 
 	BUG_ON(addr >= end);
+	BUG_ON(!PAGE_ALIGNED(addr) || !PAGE_ALIGNED(end));
+
 	pgd = pgd_offset_k(addr);
 	do {
 		next = pgd_addr_end(addr, end);
-- 
2.25.1

[PATCH v4 6/9] mm/vmalloc: only flush the cache after the vmap succeeded

[Hao Ge]

vmap_page_range() and __vmap_pages_range() call flush_cache_vmap()
before they even look at the error code. A failed mapping is never
used, so the flush is pointless at best.

According to the Sashiko review it is worse than pointless on UML:
flush_cache_vmap() there is flush_tlb_kernel_range(), which syncs the
PTEs into the host OS. Flushing a half-installed mapping therefore
leaves host mappings behind for a range that was just freed again.

Only flush when the mapping, and for vmap_page_range() also the KMSAN
metadata, is fully set up.

Reported-by: Sashiko <sashiko-bot@kernel.org>
Signed-off-by: Hao Ge <hao.ge@linux.dev>
---
 mm/vmalloc.c | 8 ++++++--
 1 file changed, 6 insertions(+), 2 deletions(-)

diff --git a/mm/vmalloc.c b/mm/vmalloc.c
index 0565cbfe325c..3f98056f8c7e 100644
--- a/mm/vmalloc.c
+++ b/mm/vmalloc.c
@@ -369,10 +369,12 @@ int vmap_page_range(unsigned long addr, unsigned long end,
 
 	err = vmap_range_noflush(addr, end, phys_addr, pgprot_nx(prot),
 				 ioremap_max_page_shift);
-	flush_cache_vmap(addr, end);
 	if (!err)
 		err = kmsan_ioremap_page_range(addr, end, phys_addr, prot,
 					       ioremap_max_page_shift);
+	if (!err)
+		flush_cache_vmap(addr, end);
+
 	return err;
 }
 
@@ -758,7 +760,9 @@ static int __vmap_pages_range(unsigned long addr, unsigned long end,
 	int err;
 
 	err = vmap_pages_range_noflush(addr, end, prot, pages, page_shift, gfp_mask);
-	flush_cache_vmap(addr, end);
+	if (!err)
+		flush_cache_vmap(addr, end);
+
 	return err;
 }
 
-- 
2.25.1

[PATCH v4 7/9] mm/vmalloc: undo partial mappings inside the mapping functions

[Hao Ge]

The vmap entry points may install some PTEs and then fail, leaving a
partial mapping behind.

The callers do not agree on who cleans it up. pcpu_map_pages() and
kmsan_ioremap_page_range() roll back what they mapped, while
vm_module_tags_populate() and the __GFP_NOFAIL retry loop in
__vmalloc_area_node() expect a clean range from the mapping function.
For them the leftovers stay mapped, and the next attempt to map the
same range trips the WARN_ON(!pte_none()) in vmap_pages_pte_range().
A leftover huge pmd is even worse. During the retry, vmap_try_huge_pmd()
calls pmd_free_pte_page() on it, and the x86 implementation cannot
tell a leaf entry from a page table and frees the mapped page.

After discussing with Suren and Ulad, we decided to put the rollback
into the entry points of the vmap API rather than into every
low-level helper. __vmap_pages_range_noflush() undoes the whole range
it was asked to map when it fails, and vmap_page_range() undoes its
range for the ioremap-style mappings.

The rollback is __vunmap_range_noflush(), which only clears the PTEs.
No TLB flush is needed, nothing has touched the mapping yet.
vmap_pages_range_noflush() also drops the KMSAN metadata of the
failed mapping.

The undo walk skips its cond_resched() when the caller's gfp mask
cannot block. __vmalloc() with GFP_ATOMIC or GFP_NOWAIT can run in
atomic context, where cond_resched() cannot schedule.
__vunmap_range_noflush() takes a can_resched argument for that, which
the sleepable callers pass as true.

Suggested-by: Uladyslau Rezki <urezki@gmail.com>
Signed-off-by: Hao Ge <hao.ge@linux.dev>
---
 include/linux/kmsan.h |  8 ++++--
 mm/kmsan/hooks.c      | 18 ++++++++-----
 mm/kmsan/shadow.c     |  7 ++---
 mm/vmalloc.c          | 61 +++++++++++++++++++++++++++++++++----------
 mm/vmalloc.h          |  6 +++--
 5 files changed, 72 insertions(+), 28 deletions(-)

diff --git a/include/linux/kmsan.h b/include/linux/kmsan.h
index 7da9fd506b39..0016c032d1e7 100644
--- a/include/linux/kmsan.h
+++ b/include/linux/kmsan.h
@@ -150,11 +150,14 @@ int __must_check kmsan_vmap_pages_range_noflush(unsigned long start,
  * kmsan_vunmap_kernel_range_noflush() - Notify KMSAN about a vunmap.
  * @start: start of vunmapped range.
  * @end:   end of vunmapped range.
+ * @can_resched: whether the unmap walk may call cond_resched(); pass
+ * false from atomic context.
  *
  * KMSAN unmaps the contiguous metadata ranges created by
  * kmsan_map_kernel_range_noflush().
  */
-void kmsan_vunmap_range_noflush(unsigned long start, unsigned long end);
+void kmsan_vunmap_range_noflush(unsigned long start, unsigned long end,
+				bool can_resched);
 
 /**
  * kmsan_ioremap_page_range() - Notify KMSAN about a ioremap_page_range() call.
@@ -355,7 +358,8 @@ static inline int __must_check kmsan_vmap_pages_range_noflush(
 }
 
 static inline void kmsan_vunmap_range_noflush(unsigned long start,
-					      unsigned long end)
+					      unsigned long end,
+					      bool can_resched)
 {
 }
 
diff --git a/mm/kmsan/hooks.c b/mm/kmsan/hooks.c
index 107f66841274..0cd1edaa1120 100644
--- a/mm/kmsan/hooks.c
+++ b/mm/kmsan/hooks.c
@@ -136,13 +136,16 @@ static unsigned long vmalloc_origin(unsigned long addr)
 						 KMSAN_META_ORIGIN);
 }
 
-void kmsan_vunmap_range_noflush(unsigned long start, unsigned long end)
+void kmsan_vunmap_range_noflush(unsigned long start, unsigned long end,
+				bool can_resched)
 {
 	if (!kmsan_enabled)
 		return;
 
-	__vunmap_range_noflush(vmalloc_shadow(start), vmalloc_shadow(end));
-	__vunmap_range_noflush(vmalloc_origin(start), vmalloc_origin(end));
+	__vunmap_range_noflush(vmalloc_shadow(start), vmalloc_shadow(end),
+			       can_resched);
+	__vunmap_range_noflush(vmalloc_origin(start), vmalloc_origin(end),
+			       can_resched);
 	flush_cache_vmap(vmalloc_shadow(start), vmalloc_shadow(end));
 	flush_cache_vmap(vmalloc_origin(start), vmalloc_origin(end));
 }
@@ -171,8 +174,9 @@ static void kmsan_iounmap_pages(unsigned long start, unsigned long end)
 		if (origin)
 			__free_pages(origin, KMSAN_IOREMAP_META_ORDER);
 	}
-	__vunmap_range_noflush(shadow_start, shadow_end);
-	__vunmap_range_noflush(origin_start, origin_end);
+	/* ioremap metadata mappings are always GFP_KERNEL, so resched is safe. */
+	__vunmap_range_noflush(shadow_start, shadow_end, true);
+	__vunmap_range_noflush(origin_start, origin_end, true);
 	flush_tlb_kernel_range(shadow_start, shadow_end);
 	flush_tlb_kernel_range(origin_start, origin_end);
 }
@@ -207,7 +211,7 @@ int kmsan_ioremap_page_range(unsigned long start, unsigned long end,
 		mapped = __vmap_pages_range_noflush(
 			vmalloc_shadow(start + off),
 			vmalloc_shadow(start + off + PAGE_SIZE), prot, &shadow,
-			PAGE_SHIFT);
+			PAGE_SHIFT, gfp_mask);
 		if (mapped) {
 			err = mapped;
 			goto ret;
@@ -217,7 +221,7 @@ int kmsan_ioremap_page_range(unsigned long start, unsigned long end,
 		mapped = __vmap_pages_range_noflush(
 			vmalloc_origin(start + off),
 			vmalloc_origin(start + off + PAGE_SIZE), prot, &origin,
-			PAGE_SHIFT);
+			PAGE_SHIFT, gfp_mask);
 		if (mapped) {
 			err = mapped;
 			goto ret;
diff --git a/mm/kmsan/shadow.c b/mm/kmsan/shadow.c
index 2166086d3dc3..452760a333e4 100644
--- a/mm/kmsan/shadow.c
+++ b/mm/kmsan/shadow.c
@@ -247,7 +247,7 @@ int kmsan_vmap_pages_range_noflush(unsigned long start, unsigned long end,
 	origin_end = vmalloc_meta((void *)end, KMSAN_META_ORIGIN);
 	kmsan_enter_runtime();
 	mapped = __vmap_pages_range_noflush(shadow_start, shadow_end, prot,
-					    s_pages, page_shift);
+					    s_pages, page_shift, gfp_mask);
 	kmsan_leave_runtime();
 	if (mapped) {
 		err = mapped;
@@ -255,12 +255,13 @@ int kmsan_vmap_pages_range_noflush(unsigned long start, unsigned long end,
 	}
 	kmsan_enter_runtime();
 	mapped = __vmap_pages_range_noflush(origin_start, origin_end, prot,
-					    o_pages, page_shift);
+					    o_pages, page_shift, gfp_mask);
 	kmsan_leave_runtime();
 	if (mapped) {
 		/* Undo the shadow mapping set up above. */
 		kmsan_enter_runtime();
-		__vunmap_range_noflush(shadow_start, shadow_end);
+		__vunmap_range_noflush(shadow_start, shadow_end,
+				       gfpflags_allow_blocking(gfp_mask));
 		kmsan_leave_runtime();
 		err = mapped;
 		goto ret;
diff --git a/mm/vmalloc.c b/mm/vmalloc.c
index 3f98056f8c7e..d1986f68eab4 100644
--- a/mm/vmalloc.c
+++ b/mm/vmalloc.c
@@ -372,8 +372,12 @@ int vmap_page_range(unsigned long addr, unsigned long end,
 	if (!err)
 		err = kmsan_ioremap_page_range(addr, end, phys_addr, prot,
 					       ioremap_max_page_shift);
-	if (!err)
+	if (err) {
+		/* Callers are ioremap and arch setup code, always sleepable. */
+		__vunmap_range_noflush(addr, end, true);
+	} else {
 		flush_cache_vmap(addr, end);
+	}
 
 	return err;
 }
@@ -429,7 +433,7 @@ static void vunmap_pte_range(pmd_t *pmd, unsigned long addr, unsigned long end,
 }
 
 static void vunmap_pmd_range(pud_t *pud, unsigned long addr, unsigned long end,
-			     pgtbl_mod_mask *mask)
+			     pgtbl_mod_mask *mask, bool can_resched)
 {
 	pmd_t *pmd;
 	unsigned long next;
@@ -451,12 +455,14 @@ static void vunmap_pmd_range(pud_t *pud, unsigned long addr, unsigned long end,
 			continue;
 		vunmap_pte_range(pmd, addr, next, mask);
 
-		cond_resched();
+		/* The undo path may run in atomic context. */
+		if (can_resched)
+			cond_resched();
 	} while (pmd++, addr = next, addr != end);
 }
 
 static void vunmap_pud_range(p4d_t *p4d, unsigned long addr, unsigned long end,
-			     pgtbl_mod_mask *mask)
+			     pgtbl_mod_mask *mask, bool can_resched)
 {
 	pud_t *pud;
 	unsigned long next;
@@ -476,12 +482,12 @@ static void vunmap_pud_range(p4d_t *p4d, unsigned long addr, unsigned long end,
 		}
 		if (pud_none_or_clear_bad(pud))
 			continue;
-		vunmap_pmd_range(pud, addr, next, mask);
+		vunmap_pmd_range(pud, addr, next, mask, can_resched);
 	} while (pud++, addr = next, addr != end);
 }
 
 static void vunmap_p4d_range(pgd_t *pgd, unsigned long addr, unsigned long end,
-			     pgtbl_mod_mask *mask)
+			     pgtbl_mod_mask *mask, bool can_resched)
 {
 	p4d_t *p4d;
 	unsigned long next;
@@ -496,7 +502,7 @@ static void vunmap_p4d_range(pgd_t *pgd, unsigned long addr, unsigned long end,
 
 		if (p4d_none_or_clear_bad(p4d))
 			continue;
-		vunmap_pud_range(p4d, addr, next, mask);
+		vunmap_pud_range(p4d, addr, next, mask, can_resched);
 	} while (p4d++, addr = next, addr != end);
 }
 
@@ -512,7 +518,8 @@ static void vunmap_p4d_range(pgd_t *pgd, unsigned long addr, unsigned long end,
  *
  * This is an internal function only. Do not use outside mm/.
  */
-void __vunmap_range_noflush(unsigned long start, unsigned long end)
+void __vunmap_range_noflush(unsigned long start, unsigned long end,
+			    bool can_resched)
 {
 	unsigned long next;
 	pgd_t *pgd;
@@ -529,7 +536,7 @@ void __vunmap_range_noflush(unsigned long start, unsigned long end)
 			mask |= PGTBL_PGD_MODIFIED;
 		if (pgd_none_or_clear_bad(pgd))
 			continue;
-		vunmap_p4d_range(pgd, addr, next, &mask);
+		vunmap_p4d_range(pgd, addr, next, &mask, can_resched);
 	} while (pgd++, addr = next, addr != end);
 
 	if (mask & ARCH_PAGE_TABLE_SYNC_MASK)
@@ -538,8 +545,9 @@ void __vunmap_range_noflush(unsigned long start, unsigned long end)
 
 void vunmap_range_noflush(unsigned long start, unsigned long end)
 {
-	kmsan_vunmap_range_noflush(start, end);
-	__vunmap_range_noflush(start, end);
+	/* The unmap paths are sleepable, keep the resched point. */
+	kmsan_vunmap_range_noflush(start, end, true);
+	__vunmap_range_noflush(start, end, true);
 }
 
 /**
@@ -727,18 +735,32 @@ static int vmap_pages_range_noflush_walk(unsigned long addr, unsigned long end,
  * The caller is responsible for calling flush_cache_vmap() after this
  * function returns successfully and before the addresses are accessed.
  *
+ * On failure the mapping is torn down before returning, so a retry
+ * sees a clean range. When @gfp_mask cannot block, the teardown
+ * skips cond_resched(), it may run in atomic context.
+ *
  * This is an internal function only. Do not use outside mm/.
  */
 int __vmap_pages_range_noflush(unsigned long addr, unsigned long end,
-		pgprot_t prot, struct page **pages, unsigned int page_shift)
+		pgprot_t prot, struct page **pages, unsigned int page_shift,
+		gfp_t gfp_mask)
 {
+	int err;
+
 	if (WARN_ON_ONCE(page_shift < PAGE_SHIFT))
 		return -EINVAL;
 
 	if (!IS_ENABLED(CONFIG_HAVE_ARCH_HUGE_VMALLOC))
 		page_shift = PAGE_SHIFT;
 
-	return vmap_pages_range_noflush_walk(addr, end, prot, pages, page_shift);
+	err = vmap_pages_range_noflush_walk(addr, end, prot, pages, page_shift);
+
+	/* Undo the PTEs installed before the failure. */
+	if (err)
+		__vunmap_range_noflush(addr, end,
+				       gfpflags_allow_blocking(gfp_mask));
+
+	return err;
 }
 
 int vmap_pages_range_noflush(unsigned long addr, unsigned long end,
@@ -750,7 +772,18 @@ int vmap_pages_range_noflush(unsigned long addr, unsigned long end,
 
 	if (ret)
 		return ret;
-	return __vmap_pages_range_noflush(addr, end, prot, pages, page_shift);
+
+	ret = __vmap_pages_range_noflush(addr, end, prot, pages, page_shift,
+					 gfp_mask);
+	/*
+	 * The page tables undo themselves on failure. Tear down the
+	 * metadata that was fully set up before the mapping failed.
+	 */
+	if (ret)
+		kmsan_vunmap_range_noflush(addr, end,
+					   gfpflags_allow_blocking(gfp_mask));
+
+	return ret;
 }
 
 static int __vmap_pages_range(unsigned long addr, unsigned long end,
diff --git a/mm/vmalloc.h b/mm/vmalloc.h
index d8372a4f65c4..9ea28845f713 100644
--- a/mm/vmalloc.h
+++ b/mm/vmalloc.h
@@ -40,10 +40,12 @@ void clear_vm_uninitialized_flag(struct vm_struct *vm);
 
 int __must_check __vmap_pages_range_noflush(unsigned long addr,
 		unsigned long end, pgprot_t prot,
-		struct page **pages, unsigned int page_shift);
+		struct page **pages, unsigned int page_shift,
+		gfp_t gfp_mask);
 
 void vunmap_range_noflush(unsigned long start, unsigned long end);
 
-void __vunmap_range_noflush(unsigned long start, unsigned long end);
+void __vunmap_range_noflush(unsigned long start, unsigned long end,
+			    bool can_resched);
 
 #endif /* __MM_VMALLOC_H */
-- 
2.25.1

[PATCH v4 8/9] powerpc: drop redundant unmaps of failed vmap mappings

[Hao Ge]

vmap_page_range() and ioremap_page_range() clear the PTEs they
installed when they fail, so the vunmap_range() in remap_isa_base()
and ioremap_phb() has nothing left to remove.

Reviewed-by: Suren Baghdasaryan <surenb@google.com>
Reviewed-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Signed-off-by: Hao Ge <hao.ge@linux.dev>
---
 arch/powerpc/kernel/isa-bridge.c | 5 ++---
 arch/powerpc/kernel/pci_64.c     | 4 +---
 2 files changed, 3 insertions(+), 6 deletions(-)

diff --git a/arch/powerpc/kernel/isa-bridge.c b/arch/powerpc/kernel/isa-bridge.c
index 5c064485197a..93029d85e6ea 100644
--- a/arch/powerpc/kernel/isa-bridge.c
+++ b/arch/powerpc/kernel/isa-bridge.c
@@ -46,9 +46,8 @@ static void remap_isa_base(phys_addr_t pa, unsigned long size)
 	WARN_ON_ONCE(size & ~PAGE_MASK);
 
 	if (slab_is_available()) {
-		if (vmap_page_range(ISA_IO_BASE, ISA_IO_BASE + size, pa,
-				    pgprot_noncached(PAGE_KERNEL)))
-			vunmap_range(ISA_IO_BASE, ISA_IO_BASE + size);
+		vmap_page_range(ISA_IO_BASE, ISA_IO_BASE + size, pa,
+				pgprot_noncached(PAGE_KERNEL));
 	} else {
 		early_ioremap_range(ISA_IO_BASE, pa, size,
 				pgprot_noncached(PAGE_KERNEL));
diff --git a/arch/powerpc/kernel/pci_64.c b/arch/powerpc/kernel/pci_64.c
index e27342ef128b..f74269636d5e 100644
--- a/arch/powerpc/kernel/pci_64.c
+++ b/arch/powerpc/kernel/pci_64.c
@@ -139,10 +139,8 @@ void __iomem *ioremap_phb(phys_addr_t paddr, unsigned long size)
 
 	addr = (unsigned long)area->addr;
 	if (ioremap_page_range(addr, addr + size, paddr,
-			pgprot_noncached(PAGE_KERNEL))) {
-		vunmap_range(addr, addr + size);
+			pgprot_noncached(PAGE_KERNEL)))
 		return NULL;
-	}
 
 	return (void __iomem *)addr;
 }
-- 
2.25.1

[PATCH v4 9/9] mm/percpu: stop unmapping the CPU that failed to map

[Hao Ge]

__pcpu_map_pages() clears the PTEs it installed when it fails, so
pcpu_map_pages() only has to tear down the CPUs it mapped before
the failing one.

Reviewed-by: Suren Baghdasaryan <surenb@google.com>
Reviewed-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Signed-off-by: Hao Ge <hao.ge@linux.dev>
---
 mm/percpu-vm.c | 5 +++--
 1 file changed, 3 insertions(+), 2 deletions(-)

diff --git a/mm/percpu-vm.c b/mm/percpu-vm.c
index 509d8901835c..59d3a336a6bd 100644
--- a/mm/percpu-vm.c
+++ b/mm/percpu-vm.c
@@ -252,10 +252,11 @@ static int pcpu_map_pages(struct pcpu_chunk *chunk, struct page **pages,
 	return 0;
 err:
 	for_each_possible_cpu(tcpu) {
-		__pcpu_unmap_pages(pcpu_chunk_addr(chunk, tcpu, page_start),
-				   page_end - page_start);
+		/* The failing CPU's partial mapping undoes itself */
 		if (tcpu == cpu)
 			break;
+		__pcpu_unmap_pages(pcpu_chunk_addr(chunk, tcpu, page_start),
+				   page_end - page_start);
 	}
 	pcpu_post_unmap_tlb_flush(chunk, page_start, page_end);
 	return err;
-- 
2.25.1