In remove_pagetable(), page tables may be freed before the TLB is
flushed. The upper page tables are zapped before freeing the lower
levels. However, without the flush the lower tables can still remain in
paging-structure caches and so data that is written to the re-allocated
page can control these mappings. For some reason there is only a flush
lower down in remove_pte_table(), however, this will not be hit in the
case of large pages on the direct map which is common.
Currently remove_pagetable() is called from a few places in the
hot unplug codepath and memremap unmapping operations.
To properly tear down these mappings, gather the page tables using a
simple linked list based in the table's struct page. Then flush the TLB
before actually freeing the pages.
Cc: [email protected]
Fixes: ae9aae9eda2d ("memory-hotplug: common APIs to support page tables hot-remove")
Acked-by: Dave Hansen <[email protected]>
Signed-off-by: Rick Edgecombe <[email protected]>
---
This wasn't observed causing any functional problem in normal runtime. AFAICT it
can't be triggered from unprivileged userspace.
arch/x86/mm/init_64.c | 60 ++++++++++++++++++++++++++++---------------
1 file changed, 39 insertions(+), 21 deletions(-)
diff --git a/arch/x86/mm/init_64.c b/arch/x86/mm/init_64.c
index ddeaba947eb3..3c0323ad99da 100644
--- a/arch/x86/mm/init_64.c
+++ b/arch/x86/mm/init_64.c
@@ -992,6 +992,23 @@ static void __meminit free_pagetable(struct page *page, int order)
free_pages((unsigned long)page_address(page), order);
}
+static void __meminit gather_table(struct page *page, struct list_head *tables)
+{
+ list_add(&page->lru, tables);
+}
+
+static void __meminit gather_table_finish(struct list_head *tables)
+{
+ struct page *page, *next;
+
+ flush_tlb_all();
+
+ list_for_each_entry_safe(page, next, tables, lru) {
+ list_del(&page->lru);
+ free_pagetable(page, 0);
+ }
+}
+
static void __meminit free_hugepage_table(struct page *page,
struct vmem_altmap *altmap)
{
@@ -1001,7 +1018,7 @@ static void __meminit free_hugepage_table(struct page *page,
free_pagetable(page, get_order(PMD_SIZE));
}
-static void __meminit free_pte_table(pte_t *pte_start, pmd_t *pmd)
+static void __meminit free_pte_table(pte_t *pte_start, pmd_t *pmd, struct list_head *tables)
{
pte_t *pte;
int i;
@@ -1012,14 +1029,14 @@ static void __meminit free_pte_table(pte_t *pte_start, pmd_t *pmd)
return;
}
- /* free a pte talbe */
- free_pagetable(pmd_page(*pmd), 0);
+ /* gather a pte table */
+ gather_table(pmd_page(*pmd), tables);
spin_lock(&init_mm.page_table_lock);
pmd_clear(pmd);
spin_unlock(&init_mm.page_table_lock);
}
-static void __meminit free_pmd_table(pmd_t *pmd_start, pud_t *pud)
+static void __meminit free_pmd_table(pmd_t *pmd_start, pud_t *pud, struct list_head *tables)
{
pmd_t *pmd;
int i;
@@ -1030,14 +1047,14 @@ static void __meminit free_pmd_table(pmd_t *pmd_start, pud_t *pud)
return;
}
- /* free a pmd talbe */
- free_pagetable(pud_page(*pud), 0);
+ /* gather a pmd table */
+ gather_table(pud_page(*pud), tables);
spin_lock(&init_mm.page_table_lock);
pud_clear(pud);
spin_unlock(&init_mm.page_table_lock);
}
-static void __meminit free_pud_table(pud_t *pud_start, p4d_t *p4d)
+static void __meminit free_pud_table(pud_t *pud_start, p4d_t *p4d, struct list_head *tables)
{
pud_t *pud;
int i;
@@ -1048,8 +1065,8 @@ static void __meminit free_pud_table(pud_t *pud_start, p4d_t *p4d)
return;
}
- /* free a pud talbe */
- free_pagetable(p4d_page(*p4d), 0);
+ /* gather a pud table */
+ gather_table(p4d_page(*p4d), tables);
spin_lock(&init_mm.page_table_lock);
p4d_clear(p4d);
spin_unlock(&init_mm.page_table_lock);
@@ -1057,7 +1074,7 @@ static void __meminit free_pud_table(pud_t *pud_start, p4d_t *p4d)
static void __meminit
remove_pte_table(pte_t *pte_start, unsigned long addr, unsigned long end,
- bool direct)
+ bool direct, struct list_head *tables)
{
unsigned long next, pages = 0;
pte_t *pte;
@@ -1100,7 +1117,7 @@ remove_pte_table(pte_t *pte_start, unsigned long addr, unsigned long end,
static void __meminit
remove_pmd_table(pmd_t *pmd_start, unsigned long addr, unsigned long end,
- bool direct, struct vmem_altmap *altmap)
+ bool direct, struct vmem_altmap *altmap, struct list_head *tables)
{
unsigned long next, pages = 0;
pte_t *pte_base;
@@ -1138,8 +1155,8 @@ remove_pmd_table(pmd_t *pmd_start, unsigned long addr, unsigned long end,
}
pte_base = (pte_t *)pmd_page_vaddr(*pmd);
- remove_pte_table(pte_base, addr, next, direct);
- free_pte_table(pte_base, pmd);
+ remove_pte_table(pte_base, addr, next, direct, tables);
+ free_pte_table(pte_base, pmd, tables);
}
/* Call free_pmd_table() in remove_pud_table(). */
@@ -1149,7 +1166,7 @@ remove_pmd_table(pmd_t *pmd_start, unsigned long addr, unsigned long end,
static void __meminit
remove_pud_table(pud_t *pud_start, unsigned long addr, unsigned long end,
- struct vmem_altmap *altmap, bool direct)
+ struct vmem_altmap *altmap, bool direct, struct list_head *tables)
{
unsigned long next, pages = 0;
pmd_t *pmd_base;
@@ -1173,8 +1190,8 @@ remove_pud_table(pud_t *pud_start, unsigned long addr, unsigned long end,
}
pmd_base = pmd_offset(pud, 0);
- remove_pmd_table(pmd_base, addr, next, direct, altmap);
- free_pmd_table(pmd_base, pud);
+ remove_pmd_table(pmd_base, addr, next, direct, altmap, tables);
+ free_pmd_table(pmd_base, pud, tables);
}
if (direct)
@@ -1183,7 +1200,7 @@ remove_pud_table(pud_t *pud_start, unsigned long addr, unsigned long end,
static void __meminit
remove_p4d_table(p4d_t *p4d_start, unsigned long addr, unsigned long end,
- struct vmem_altmap *altmap, bool direct)
+ struct vmem_altmap *altmap, bool direct, struct list_head *tables)
{
unsigned long next, pages = 0;
pud_t *pud_base;
@@ -1199,14 +1216,14 @@ remove_p4d_table(p4d_t *p4d_start, unsigned long addr, unsigned long end,
BUILD_BUG_ON(p4d_large(*p4d));
pud_base = pud_offset(p4d, 0);
- remove_pud_table(pud_base, addr, next, altmap, direct);
+ remove_pud_table(pud_base, addr, next, altmap, direct, tables);
/*
* For 4-level page tables we do not want to free PUDs, but in the
* 5-level case we should free them. This code will have to change
* to adapt for boot-time switching between 4 and 5 level page tables.
*/
if (pgtable_l5_enabled())
- free_pud_table(pud_base, p4d);
+ free_pud_table(pud_base, p4d, tables);
}
if (direct)
@@ -1220,6 +1237,7 @@ remove_pagetable(unsigned long start, unsigned long end, bool direct,
{
unsigned long next;
unsigned long addr;
+ LIST_HEAD(tables);
pgd_t *pgd;
p4d_t *p4d;
@@ -1231,10 +1249,10 @@ remove_pagetable(unsigned long start, unsigned long end, bool direct,
continue;
p4d = p4d_offset(pgd, 0);
- remove_p4d_table(p4d, addr, next, altmap, direct);
+ remove_p4d_table(p4d, addr, next, altmap, direct, &tables);
}
- flush_tlb_all();
+ gather_table_finish(&tables);
}
void __ref vmemmap_free(unsigned long start, unsigned long end,
--
2.17.1
On Wed, 2021-08-18 at 15:10 -0700, Rick Edgecombe wrote:
>
> In remove_pagetable(), page tables may be freed before the TLB is
> flushed. The upper page tables are zapped before freeing the lower
> levels. However, without the flush the lower tables can still remain in
> paging-structure caches and so data that is written to the re-allocated
> page can control these mappings. For some reason there is only a flush
> lower down in remove_pte_table(), however, this will not be hit in the
> case of large pages on the direct map which is common.
It's also common for device-dax reconfiguration which we are in the
process of adding udev automation to replug devices from ZONE_DEVICE to
ZONE_{NORMAL,MOVABLE} automatically depending on what setup was
estabished on the previous boot. So even if unpriveleged userspace
can't force this, there may be more opportunities to find this gap in
the future.
>
> Currently remove_pagetable() is called from a few places in the
> hot unplug codepath and memremap unmapping operations.
>
> To properly tear down these mappings, gather the page tables using a
> simple linked list based in the table's struct page. Then flush the TLB
> before actually freeing the pages.
>
> Cc: [email protected]
> Fixes: ae9aae9eda2d ("memory-hotplug: common APIs to support page tables hot-remove")
> Acked-by: Dave Hansen <[email protected]>
> Signed-off-by: Rick Edgecombe <[email protected]>
Acked-by: Dan Williams <[email protected]>