Hi Ingo,
On Tue, Nov 26, 2019 at 12:11:19PM +0100, Ingo Molnar wrote:
> The vmalloc_sync_all() also iterating over the LDT range is buggy,
> because for the LDT the mappings are *intentionally* and fundamentally
> different between processes, i.e. not synchronized.
Yes, you are right, your patch description is much better, thanks for
making it more clear and correct.
> Furthermore I'm not sure we need to iterate over the PKMAP range either:
> those are effectively permanent PMDs as well, and they are not part of
> the vmalloc.c lazy deallocation scheme in any case - they are handled
> entirely separately in mm/highmem.c et al.
I looked a bit at that, and I didn't find an explict place where the
PKMAP PMD gets established. It probably happens implicitly on the first
kmap() call, so we are safe as long as the first call to kmap happens
before the kernel starts the first userspace process.
But that is not an issue that should be handled by vmalloc_sync_all(),
as the name already implies that it only cares about the vmalloc range.
So your change to only iterate to VMALLOC_END makes sense and we should
establish the PKMAP PMD at a defined place to make sure it exists when
we start the first process.
> Note that this is *completely* untested - I might have wrecked PKMAP in
> my ignorance. Mind giving it a careful review and a test?
My testing environment for 32 bit is quite limited these days, but I
tested it in my PTI-x32 environment and the patch below works perfectly
fine there and still fixes the ldt_gdt selftest.
Regards,
Joerg
> ===========================>
> Subject: x86/mm/32: Sync only to VMALLOC_END in vmalloc_sync_all()
> From: Joerg Roedel <[email protected]>
> Date: Tue, 26 Nov 2019 11:09:42 +0100
>
> From: Joerg Roedel <[email protected]>
>
> The job of vmalloc_sync_all() is to help the lazy freeing of vmalloc()
> ranges: before such vmap ranges are reused we make sure that they are
> unmapped from every task's page tables.
>
> This is really easy on pagetable setups where the kernel page tables
> are shared between all tasks - this is the case on 32-bit kernels
> with SHARED_KERNEL_PMD = 1.
>
> But on !SHARED_KERNEL_PMD 32-bit kernels this involves iterating
> over the pgd_list and clearing all pmd entries in the pgds that
> are cleared in the init_mm.pgd, which is the reference pagetable
> that the vmalloc() code uses.
>
> In that context the current practice of vmalloc_sync_all() iterating
> until FIX_ADDR_TOP is buggy:
>
> for (address = VMALLOC_START & PMD_MASK;
> address >= TASK_SIZE_MAX && address < FIXADDR_TOP;
> address += PMD_SIZE) {
> struct page *page;
>
> Because iterating up to FIXADDR_TOP will involve a lot of non-vmalloc
> address ranges:
>
> VMALLOC -> PKMAP -> LDT -> CPU_ENTRY_AREA -> FIX_ADDR
>
> This is mostly harmless for the FIX_ADDR and CPU_ENTRY_AREA ranges
> that don't clear their pmds, but it's lethal for the LDT range,
> which relies on having different mappings in different processes,
> and 'synchronizing' them in the vmalloc sense corrupts those
> pagetable entries (clearing them).
>
> This got particularly prominent with PTI, which turns SHARED_KERNEL_PMD
> off and makes this the dominant mapping mode on 32-bit.
>
> To make LDT working again vmalloc_sync_all() must only iterate over
> the volatile parts of the kernel address range that are identical
> between all processes.
>
> So the correct check in vmalloc_sync_all() is "address < VMALLOC_END"
> to make sure the VMALLOC areas are synchronized and the LDT
> mapping is not falsely overwritten.
>
> The CPU_ENTRY_AREA and the FIXMAP area are no longer synced either,
> but this is not really a proplem since their PMDs get established
> during bootup and never change.
>
> This change fixes the ldt_gdt selftest in my setup.
>
> Reported-by: Borislav Petkov <[email protected]>
> Tested-by: Borislav Petkov <[email protected]>
> Signed-off-by: Joerg Roedel <[email protected]>
> Cc: <[email protected]>
> Cc: Andy Lutomirski <[email protected]>
> Cc: Borislav Petkov <[email protected]>
> Cc: Brian Gerst <[email protected]>
> Cc: Dave Hansen <[email protected]>
> Cc: H. Peter Anvin <[email protected]>
> Cc: Linus Torvalds <[email protected]>
> Cc: Peter Zijlstra <[email protected]>
> Cc: Thomas Gleixner <[email protected]>
> Fixes: 7757d607c6b3: ("x86/pti: Allow CONFIG_PAGE_TABLE_ISOLATION for x86_32")
> Link: https://lkml.kernel.org/r/[email protected]
> Signed-off-by: Ingo Molnar <[email protected]>
> ---
> arch/x86/mm/fault.c | 2 +-
> 1 file changed, 1 insertion(+), 1 deletion(-)
>
> Index: tip/arch/x86/mm/fault.c
> ===================================================================
> --- tip.orig/arch/x86/mm/fault.c
> +++ tip/arch/x86/mm/fault.c
> @@ -197,7 +197,7 @@ void vmalloc_sync_all(void)
> return;
>
> for (address = VMALLOC_START & PMD_MASK;
> - address >= TASK_SIZE_MAX && address < FIXADDR_TOP;
> + address >= TASK_SIZE_MAX && address < VMALLOC_END;
> address += PMD_SIZE) {
> struct page *page;
>
* Joerg Roedel <[email protected]> wrote:
> Hi Ingo,
>
> On Tue, Nov 26, 2019 at 12:11:19PM +0100, Ingo Molnar wrote:
> > The vmalloc_sync_all() also iterating over the LDT range is buggy,
> > because for the LDT the mappings are *intentionally* and fundamentally
> > different between processes, i.e. not synchronized.
>
> Yes, you are right, your patch description is much better, thanks for
> making it more clear and correct.
>
> > Furthermore I'm not sure we need to iterate over the PKMAP range either:
> > those are effectively permanent PMDs as well, and they are not part of
> > the vmalloc.c lazy deallocation scheme in any case - they are handled
> > entirely separately in mm/highmem.c et al.
>
> I looked a bit at that, and I didn't find an explict place where the
> PKMAP PMD gets established. It probably happens implicitly on the first
> kmap() call, so we are safe as long as the first call to kmap happens
> before the kernel starts the first userspace process.
No, it happens during early boot, in permanent_kmaps_init():
vaddr = PKMAP_BASE;
page_table_range_init(vaddr, vaddr + PAGE_SIZE*LAST_PKMAP, pgd_base);
That page_table_range_init() will go from PKMAP_BASE to the last PKMAP,
which on PAE kernels is typically 0xff600000...0xff800000, 2MB in size,
taking up exactly one PMD entry.
This single pagetable page, covering 2MB of virtual memory via 4K
entries, gets passed on to the mm/highmem.c code via:
pkmap_page_table = pte;
The pkmap_page_table is mapped early on into init_mm, every task started
after that with a new pgd inherits it, and the pmd entry never changes,
so there's nothing to synchronize.
The pte entries within this single pagetable page do change frequently
according to the kmap() code, but since the pagetable page is shared
between all tasks and the TLB flushes are SMP safe, it's all synchronized
by only modifying pkmap_page_table, as it should.
> But that is not an issue that should be handled by vmalloc_sync_all(),
> as the name already implies that it only cares about the vmalloc range.
Well, hypothetically it could *accidentally* have some essentially effect
on bootstrapping the PKMAP pagetables - I don't think that's so, based on
the reading of the code, but only testing will tell for sure.
> So your change to only iterate to VMALLOC_END makes sense and we should
> establish the PKMAP PMD at a defined place to make sure it exists when
> we start the first process.
I believe that's done in permanent_kmaps_init().
> > Note that this is *completely* untested - I might have wrecked PKMAP in
> > my ignorance. Mind giving it a careful review and a test?
>
> My testing environment for 32 bit is quite limited these days, but I
> tested it in my PTI-x32 environment and the patch below works perfectly
> fine there and still fixes the ldt_gdt selftest.
Cool, thanks! I'll apply it with your Tested-by.
Ingo