Linux kernel mirror (for testing)
git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git
kernel
os
linux
Merge tag 'x86-urgent-2024-08-11' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip
Pull x86 fixes from Thomas Gleixner:
- Fix 32-bit PTI for real.
pti_clone_entry_text() is called twice, once before initcalls so that
initcalls can use the user-mode helper and then again after text is
set read only. Setting read only on 32-bit might break up the PMD
mapping, which makes the second invocation of pti_clone_entry_text()
find the mappings out of sync and failing.
Allow the second call to split the existing PMDs in the user mapping
and synchronize with the kernel mapping.
- Don't make acpi_mp_wake_mailbox read-only after init as the mail box
must be writable in the case that CPU hotplug operations happen after
boot. Otherwise the attempt to start a CPU crashes with a write to
read only memory.
- Add a missing sanity check in mtrr_save_state() to ensure that the
fixed MTRR MSRs are supported.
Otherwise mtrr_save_state() ends up in a #GP, which is fixed up, but
the WARN_ON() can bring systems down when panic on warn is set.
* tag 'x86-urgent-2024-08-11' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip:
x86/mtrr: Check if fixed MTRRs exist before saving them
x86/paravirt: Fix incorrect virt spinlock setting on bare metal
x86/acpi: Remove __ro_after_init from acpi_mp_wake_mailbox
x86/mm: Fix PTI for i386 some more
+41
-27
+7
-5
arch/x86/include/asm/qspinlock.h
+7
-5
arch/x86/include/asm/qspinlock.h
···
66
66
67
67
#ifdef CONFIG_PARAVIRT
68
68
/*
69
-
* virt_spin_lock_key - enables (by default) the virt_spin_lock() hijack.
69
+
* virt_spin_lock_key - disables by default the virt_spin_lock() hijack.
70
70
*
71
-
* Native (and PV wanting native due to vCPU pinning) should disable this key.
72
-
* It is done in this backwards fashion to only have a single direction change,
73
-
* which removes ordering between native_pv_spin_init() and HV setup.
71
+
* Native (and PV wanting native due to vCPU pinning) should keep this key
72
+
* disabled. Native does not touch the key.
73
+
*
74
+
* When in a guest then native_pv_lock_init() enables the key first and
75
+
* KVM/XEN might conditionally disable it later in the boot process again.
74
76
*/
75
-
DECLARE_STATIC_KEY_TRUE(virt_spin_lock_key);
77
+
DECLARE_STATIC_KEY_FALSE(virt_spin_lock_key);
76
78
77
79
/*
78
80
* Shortcut for the queued_spin_lock_slowpath() function that allows
+1
-1
arch/x86/kernel/acpi/madt_wakeup.c
+1
-1
arch/x86/kernel/acpi/madt_wakeup.c
···
19
19
static u64 acpi_mp_wake_mailbox_paddr __ro_after_init;
20
20
21
21
/* Virtual address of the Multiprocessor Wakeup Structure mailbox */
22
-
static struct acpi_madt_multiproc_wakeup_mailbox *acpi_mp_wake_mailbox __ro_after_init;
22
+
static struct acpi_madt_multiproc_wakeup_mailbox *acpi_mp_wake_mailbox;
23
23
24
24
static u64 acpi_mp_pgd __ro_after_init;
25
25
static u64 acpi_mp_reset_vector_paddr __ro_after_init;
+1
-1
arch/x86/kernel/cpu/mtrr/mtrr.c
+1
-1
arch/x86/kernel/cpu/mtrr/mtrr.c
+3
-4
arch/x86/kernel/paravirt.c
+3
-4
arch/x86/kernel/paravirt.c
···
51
51
DEFINE_ASM_FUNC(pv_native_read_cr2, "mov %cr2, %rax", .noinstr.text);
52
52
#endif
53
53
54
-
DEFINE_STATIC_KEY_TRUE(virt_spin_lock_key);
54
+
DEFINE_STATIC_KEY_FALSE(virt_spin_lock_key);
55
55
56
56
void __init native_pv_lock_init(void)
57
57
{
58
-
if (IS_ENABLED(CONFIG_PARAVIRT_SPINLOCKS) &&
59
-
!boot_cpu_has(X86_FEATURE_HYPERVISOR))
60
-
static_branch_disable(&virt_spin_lock_key);
58
+
if (boot_cpu_has(X86_FEATURE_HYPERVISOR))
59
+
static_branch_enable(&virt_spin_lock_key);
61
60
}
62
61
63
62
static void native_tlb_remove_table(struct mmu_gather *tlb, void *table)
+29
-16
arch/x86/mm/pti.c
+29
-16
arch/x86/mm/pti.c
···
241
241
*
242
242
* Returns a pointer to a PTE on success, or NULL on failure.
243
243
*/
244
-
static pte_t *pti_user_pagetable_walk_pte(unsigned long address)
244
+
static pte_t *pti_user_pagetable_walk_pte(unsigned long address, bool late_text)
245
245
{
246
246
gfp_t gfp = (GFP_KERNEL | __GFP_NOTRACK | __GFP_ZERO);
247
247
pmd_t *pmd;
···
251
251
if (!pmd)
252
252
return NULL;
253
253
254
-
/* We can't do anything sensible if we hit a large mapping. */
254
+
/* Large PMD mapping found */
255
255
if (pmd_leaf(*pmd)) {
256
-
WARN_ON(1);
257
-
return NULL;
256
+
/* Clear the PMD if we hit a large mapping from the first round */
257
+
if (late_text) {
258
+
set_pmd(pmd, __pmd(0));
259
+
} else {
260
+
WARN_ON_ONCE(1);
261
+
return NULL;
262
+
}
258
263
}
259
264
260
265
if (pmd_none(*pmd)) {
···
288
283
if (!pte || WARN_ON(level != PG_LEVEL_4K) || pte_none(*pte))
289
284
return;
290
285
291
-
target_pte = pti_user_pagetable_walk_pte(VSYSCALL_ADDR);
286
+
target_pte = pti_user_pagetable_walk_pte(VSYSCALL_ADDR, false);
292
287
if (WARN_ON(!target_pte))
293
288
return;
294
289
···
306
301
307
302
static void
308
303
pti_clone_pgtable(unsigned long start, unsigned long end,
309
-
enum pti_clone_level level)
304
+
enum pti_clone_level level, bool late_text)
310
305
{
311
306
unsigned long addr;
312
307
···
395
390
return;
396
391
397
392
/* Allocate PTE in the user page-table */
398
-
target_pte = pti_user_pagetable_walk_pte(addr);
393
+
target_pte = pti_user_pagetable_walk_pte(addr, late_text);
399
394
if (WARN_ON(!target_pte))
400
395
return;
401
396
···
457
452
phys_addr_t pa = per_cpu_ptr_to_phys((void *)va);
458
453
pte_t *target_pte;
459
454
460
-
target_pte = pti_user_pagetable_walk_pte(va);
455
+
target_pte = pti_user_pagetable_walk_pte(va, false);
461
456
if (WARN_ON(!target_pte))
462
457
return;
463
458
···
480
475
start = CPU_ENTRY_AREA_BASE;
481
476
end = start + (PAGE_SIZE * CPU_ENTRY_AREA_PAGES);
482
477
483
-
pti_clone_pgtable(start, end, PTI_CLONE_PMD);
478
+
pti_clone_pgtable(start, end, PTI_CLONE_PMD, false);
484
479
}
485
480
#endif /* CONFIG_X86_64 */
486
481
···
497
492
/*
498
493
* Clone the populated PMDs of the entry text and force it RO.
499
494
*/
500
-
static void pti_clone_entry_text(void)
495
+
static void pti_clone_entry_text(bool late)
501
496
{
502
497
pti_clone_pgtable((unsigned long) __entry_text_start,
503
498
(unsigned long) __entry_text_end,
504
-
PTI_LEVEL_KERNEL_IMAGE);
499
+
PTI_LEVEL_KERNEL_IMAGE, late);
505
500
}
506
501
507
502
/*
···
576
571
* pti_set_kernel_image_nonglobal() did to clear the
577
572
* global bit.
578
573
*/
579
-
pti_clone_pgtable(start, end_clone, PTI_LEVEL_KERNEL_IMAGE);
574
+
pti_clone_pgtable(start, end_clone, PTI_LEVEL_KERNEL_IMAGE, false);
580
575
581
576
/*
582
577
* pti_clone_pgtable() will set the global bit in any PMDs
···
643
638
644
639
/* Undo all global bits from the init pagetables in head_64.S: */
645
640
pti_set_kernel_image_nonglobal();
641
+
646
642
/* Replace some of the global bits just for shared entry text: */
647
-
pti_clone_entry_text();
643
+
/*
644
+
* This is very early in boot. Device and Late initcalls can do
645
+
* modprobe before free_initmem() and mark_readonly(). This
646
+
* pti_clone_entry_text() allows those user-mode-helpers to function,
647
+
* but notably the text is still RW.
648
+
*/
649
+
pti_clone_entry_text(false);
648
650
pti_setup_espfix64();
649
651
pti_setup_vsyscall();
650
652
}
···
668
656
if (!boot_cpu_has(X86_FEATURE_PTI))
669
657
return;
670
658
/*
671
-
* We need to clone everything (again) that maps parts of the
672
-
* kernel image.
659
+
* This is after free_initmem() (all initcalls are done) and we've done
660
+
* mark_readonly(). Text is now NX which might've split some PMDs
661
+
* relative to the early clone.
673
662
*/
674
-
pti_clone_entry_text();
663
+
pti_clone_entry_text(true);
675
664
pti_clone_kernel_text();
676
665
677
666
debug_checkwx_user();