x86/mm: Drop unnecessary PMD page copy when freeing

On a box with a discrete GPU, lockdep reports a possible deadlock as
soon as kswapd shrinks the TTM page pool. The immediate cause is an
x86 commit that added an mmap_read_lock() to kernel page protection
munging code.

The huge vmap code holds the same lock over a GFP_KERNEL allocation,
which is a no-no now that reclaim can take it. That allocation is in a
page table *free* path and ends up being for dubious purposes[1].
Basically, it tries to avoid hardware setting Accessed=1 in page table
entries that are unreachable by the hardware, a non-issue.

Remove the PMD copy. Detach the original PMD page at the PUD, flush
the mid-level caches, and free the PTE tables straight from the
detached PMD page. With no allocation left, the locking issue is gone.

Lockdep splat/analysis:

  WARNING: possible circular locking dependency detected
  7.3.0-rc3-f6e7b42bf05b+ #183 Tainted: G     U
  ------------------------------------------------------
  kswapd0/269 is trying to acquire lock:
  ((init_mm).mmap_lock){++++}-{4:4}, at: change_page_attr_set_clr+0x29a/0x4a0
  but task is already holding lock:
  (pool_shrink_rwsem){.+.+}-{4:4}, at: ttm_pool_shrink+0xb2/0x330 [ttm]
  Chain exists of:
    (init_mm).mmap_lock --> fs_reclaim --> pool_shrink_rwsem

The cycle is built from three edges:

  1) pool_shrink_rwsem -> (init_mm).mmap_lock

     The TTM shrinker restores the caching attribute of every page it
     frees, while holding pool_shrink_rwsem:

       ttm_pool_shrink()
       -> ttm_pool_dispose_list()
       -> ttm_pool_free_page()
       -> set_pages_wb()
       -> change_page_attr_set_clr()   [ init_mm mmap read lock ]

  2) fs_reclaim -> pool_shrink_rwsem

     The same shrinker, called from reclaim.

  3) (init_mm).mmap_lock -> fs_reclaim

     ioremap() installing a huge PUD mapping over an existing PMD table:

       ioremap_page_range()
       -> vmap_range_noflush()
       -> vmap_try_huge_pud()          [ init_mm mmap read lock ]
       -> pud_free_pmd_page()
       -> __get_free_page(GFP_KERNEL)  [ enters reclaim ]

[ dhansen: Lots of changelog munging/trimming and merged comments from my
	   version of the fix. ]

Fixes: d5d8b8662e6e ("x86/mm/pat: Acquire init_mm read lock on attribute changes to avoid UAF")
Suggested-by: Pedro Falcato <pfalcato@suse.de>
Signed-off-by: Mikhail Gavrilov <mikhail.v.gavrilov@gmail.com>
Signed-off-by: Dave Hansen <dave.hansen@linux.intel.com>
Reviewed-by: Pedro Falcato <pfalcato@suse.de>
Link: https://lore.kernel.org/20260916062222.27347-1-mikhail.v.gavrilov@gmail.com
Link: https://lore.kernel.org/all/e11449f0-d9ad-4d1b-ab21-2be7d71fe335@intel.com/ [1]
Link: https://patch.msgid.link/20260923223116.20090-1-mikhail.v.gavrilov@gmail.com
Cc: stable@vger.kernel.org

1 files changed, 14 insertions, 24 deletions

diff --git a/arch/x86/mm/pgtable.c b/arch/x86/mm/pgtable.c
index cb03f5a2b..4a105f283 100644
--- a/arch/x86/mm/pgtable.c
+++ b/arch/x86/mm/pgtable.c
@@ -705,47 +705,37 @@ int pmd_clear_huge(pmd_t *pmd)
 }
 
 #ifdef CONFIG_X86_64
-/**
- * pud_free_pmd_page - Clear PUD entry and free PMD page
- * @pud: Pointer to a PUD
- * @addr: Virtual address associated with PUD
- *
- * Context: The PUD range has been unmapped and TLB purged.
- * Return: 1 if clearing the entry succeeded. 0 otherwise.
- *
- * NOTE: Callers must allow a single page allocation.
+/*
+ * Given a PUD poitner, detach and free the pointed-to
+ * PMD page and any PTE page children. The entire range
+ * under the PUD must not have any valid translations
+ * and the TLB must have already been flushed.
  */
 int pud_free_pmd_page(pud_t *pud, unsigned long addr)
 {
-	pmd_t *pmd, *pmd_sv;
 	struct ptdesc *pt;
+	pmd_t *pmd;
 	int i;
 
 	pmd = pud_pgtable(*pud);
-	pmd_sv = (pmd_t *)__get_free_page(GFP_KERNEL);
-	if (!pmd_sv)
-		return 0;
-
-	for (i = 0; i < PTRS_PER_PMD; i++) {
-		pmd_sv[i] = pmd[i];
-		if (!pmd_none(pmd[i]))
-			pmd_clear(&pmd[i]);
-	}
 
+	/* Detach the PMD page: */
 	pud_clear(pud);
 
-	/* INVLPG to clear all paging-structure caches */
+	/*
+	 * PMD and all its descendents are unreachable
+	 * via normal page walks. Make them unreachable
+	 * in cached mid-level walks too:
+	 */
 	flush_tlb_kernel_range(addr, addr + PAGE_SIZE-1);
 
 	for (i = 0; i < PTRS_PER_PMD; i++) {
-		if (!pmd_none(pmd_sv[i])) {
-			pt = page_ptdesc(pmd_page(pmd_sv[i]));
+		if (!pmd_none(pmd[i])) {
+			pt = page_ptdesc(pmd_page(pmd[i]));
 			pagetable_dtor_free(pt);
 		}
 	}
 
-	free_page((unsigned long)pmd_sv);
-
 	pmd_free(&init_mm, pmd);
 
 	return 1;