Skip to contentSkip to conversation

i386: fix assumed to be contiguous leaf page tables for kmap_atomic region

PATCH In review Jan Beulich 2 participants 2 messages Jan 14, 2009 – Jan 15, 2009

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 i386: fix assumed to be contiguous leaf page tables for kmap_atomic region

Debugging and original patch from Nick Piggin <[email protected]>

The early fixmap pmd entry inserted at the very top of the KVA is causing the subsequent fixmap mapping code to not provide physically linear pte pages over the kmap atomic portion of the fixmap (which relies on said property to calculate pte addresses).

This has caused weird boot failures in kmap_atomic much later in the boot process (initial userspace faults) on a 32-bit PAE system with a larger number of CPUs (smaller CPU counts tend not to run over into the next page so don't show up the problem).

Solve this by attempting to clear out the page table, and copy any of its entries to the new one. Also, add a bug if a nonlinear condition is encounted and can't be resolved, which might save some hours of debugging if this fragile scheme ever breaks again...

Once we have such logic, we can also use it to eliminate the early ioremap trickery around the page table setup for the fixmap area. This also fixes potential issues with FIX_* entries sharing the leaf page table with the early ioremap ones getting discarded by early_ioremap_clear() and not restored by early_ioremap_reset(). It at once eliminates the temporary (and configuration, namely NR_CPUS, dependent) unavailability of early fixed mappings during the time the fixmap area page tables get constructed.

Finally, also replace the hard coded calculation of the initial table space needed for the fixmap area with a proper one, allowing kernels configured for large CPU counts to actually boot.

---
 arch/x86/include/asm/io.h |    1 -
 arch/x86/mm/init_32.c     |   46 +++++++++++++++++++++++++++++++++++++++++++---
 arch/x86/mm/ioremap.c     |   25 -------------------------
 3 files changed, 43 insertions(+), 29 deletions(-)
arch/x86/include/asm/io.h +0 -1
@@ -99,7 +99,6 @@ extern void __iomem *ioremap_wc(unsigned
  * A boot-time mapping is currently limited to at most 16 pages.
  */
 extern void early_ioremap_init(void);
-extern void early_ioremap_clear(void);
 extern void early_ioremap_reset(void);
 extern void __iomem *early_ioremap(unsigned long offset, unsigned long size);
 extern void __iomem *early_memremap(unsigned long offset, unsigned long size);
arch/x86/mm/init_32.c +43 -3

70 lines not on this page · raw message

arch/x86/mm/ioremap.c +0 -25
@@ -557,34 +557,9 @@ void __init early_ioremap_init(void)
 	}
 }
 
-void __init early_ioremap_clear(void)
-{
-	pmd_t *pmd;
-
-	if (early_ioremap_debug)
-		printk(KERN_INFO "early_ioremap_clear()\n");
-
-	pmd = early_ioremap_pmd(fix_to_virt(FIX_BTMAP_BEGIN));
-	pmd_clear(pmd);
-	paravirt_release_pte(__pa(bm_pte) >> PAGE_SHIFT);
-	__flush_tlb_all();
-}
-
 void __init early_ioremap_reset(void)
 {
-	enum fixed_addresses idx;
-	unsigned long addr, phys;
-	pte_t *pte;
-
 	after_paging_init = 1;
-	for (idx = FIX_BTMAP_BEGIN; idx >= FIX_BTMAP_END; idx--) {
-		addr = fix_to_virt(idx);
-		pte = early_ioremap_pte(addr);
-		if (pte_present(*pte)) {
-			phys = pte_val(*pte) & PAGE_MASK;
-			set_fixmap(idx, phys);
-		}
-	}
 }
 
 static void __init __early_set_fixmap(enum fixed_addresses idx,
#1 lore raw
Quoted 31 lines · Jan Beulich

Debugging and original patch from Nick Piggin <[email protected]>

The early fixmap pmd entry inserted at the very top of the KVA is causing the subsequent fixmap mapping code to not provide physically linear pte pages over the kmap atomic portion of the fixmap (which relies on said property to calculate pte addresses).

This has caused weird boot failures in kmap_atomic much later in the boot process (initial userspace faults) on a 32-bit PAE system with a larger number of CPUs (smaller CPU counts tend not to run over into the next page so don't show up the problem).

Solve this by attempting to clear out the page table, and copy any of its entries to the new one. Also, add a bug if a nonlinear condition is encounted and can't be resolved, which might save some hours of debugging if this fragile scheme ever breaks again...

Once we have such logic, we can also use it to eliminate the early ioremap trickery around the page table setup for the fixmap area. This also fixes potential issues with FIX_* entries sharing the leaf page table with the early ioremap ones getting discarded by early_ioremap_clear() and not restored by early_ioremap_reset(). It at once eliminates the temporary (and configuration, namely NR_CPUS, dependent) unavailability of early fixed mappings during the time the fixmap area page tables get constructed.

Finally, also replace the hard coded calculation of the initial table space needed for the fixmap area with a proper one, allowing kernels configured for large CPU counts to actually boot.

Quoted 7 lines
---
 arch/x86/include/asm/io.h |    1 -
 arch/x86/mm/init_32.c     |   46 +++++++++++++++++++++++++++++++++++++++++++---
 arch/x86/mm/ioremap.c     |   25 -------------------------
 3 files changed, 43 insertions(+), 29 deletions(-)

looks good, but could you please do the following cleanup:

Quoted 44 lines
arch/x86/mm/init_32.c +32 -1
@@ -154,6 +154,11 @@ page_table_range_init(unsigned long star
 	unsigned long vaddr;
 	pgd_t *pgd;
 	pmd_t *pmd;
+#ifdef CONFIG_HIGHMEM
+	pte_t *lastpte = NULL;
+	int pmd_idx_kmap_begin = fix_to_virt(FIX_KMAP_END) >> PMD_SHIFT;
+	int pmd_idx_kmap_end = fix_to_virt(FIX_KMAP_BEGIN) >> PMD_SHIFT;
+#endif
 
 	vaddr = start;
 	pgd_idx = pgd_index(vaddr);
@@ -165,7 +170,43 @@ page_table_range_init(unsigned long star
 		pmd = pmd + pmd_index(vaddr);
 		for (; (pmd_idx < PTRS_PER_PMD) && (vaddr != end);
 							pmd++, pmd_idx++) {
-			one_page_table_init(pmd);
+			pte_t *pte;
+
+			pte = one_page_table_init(pmd);
+#ifdef CONFIG_HIGHMEM
+			/*
+			 * Something (early fixmap) may already have put a pte
+			 * page here, which causes the page table allocation
+			 * to become nonlinear. Attempt to fix it, and if it
+			 * is still nonlinear then we have to bug.
+			 */
+			if (pmd_idx_kmap_begin != pmd_idx_kmap_end
+			    && (vaddr >> PMD_SHIFT) >= pmd_idx_kmap_begin
+			    && (vaddr >> PMD_SHIFT) <= pmd_idx_kmap_end
+			    && ((__pa(pte) >> PAGE_SHIFT) < table_start
+				|| (__pa(pte) >> PAGE_SHIFT) >= table_end)) {
+				pte_t *newpte;
+				int i;
+
+				BUG_ON(after_init_bootmem);
+				newpte = alloc_low_page();
+				for (i = 0; i < PTRS_PER_PTE; i++)
+					set_pte(newpte + i, pte[i]);
+
+				paravirt_alloc_pte(&init_mm,
+						   __pa(newpte) >> PAGE_SHIFT);
+				set_pmd(pmd, __pmd(__pa(newpte)|_PAGE_TABLE));
+				BUG_ON(newpte != pte_offset_kernel(pmd, 0));

arch/x86/mm/init_32.c page_table_range_init(unsigned long star · line 209

Quoted 10 lines
arch/x86/mm/init_32.c +10 -0
+				__flush_tlb_all();
+
+				paravirt_release_pte(__pa(pte) >> PAGE_SHIFT);
+				pte = newpte;
+			}
+			BUG_ON(vaddr < fix_to_virt(FIX_KMAP_BEGIN - 1)
+			       && vaddr > fix_to_virt(FIX_KMAP_END)
+			       && lastpte && lastpte + PTRS_PER_PTE != pte);
+			lastpte = pte;
+#endif

please introduce a helper function here. That way both the ugly #ifdefs move out of this function (to a more logical place), and the very visible 80-line length artifacts from the code vanish as well.

Thanks,

Ingo