[PATCH v2] powerpc/mm: Fix set_memory_*() against concurrent accesses

[Christophe Leroy]

Le 18/08/2021 à 14:05, Michael Ellerman a écrit :
> Laurent reported that STRICT_MODULE_RWX was causing intermittent crashes
> on one of his systems:
> 
>    kernel tried to execute exec-protected page (c008000004073278) - exploit attempt? (uid: 0)
>    BUG: Unable to handle kernel instruction fetch
>    Faulting instruction address: 0xc008000004073278
>    Oops: Kernel access of bad area, sig: 11 [#1]
>    LE PAGE_SIZE=64K MMU=Radix SMP NR_CPUS=2048 NUMA pSeries
>    Modules linked in: drm virtio_console fuse drm_panel_orientation_quirks ...
>    CPU: 3 PID: 44 Comm: kworker/3:1 Not tainted 5.14.0-rc4+ #12
>    Workqueue: events control_work_handler [virtio_console]
>    NIP:  c008000004073278 LR: c008000004073278 CTR: c0000000001e9de0
>    REGS: c00000002e4ef7e0 TRAP: 0400   Not tainted  (5.14.0-rc4+)
>    MSR:  800000004280b033 <SF,VEC,VSX,EE,FP,ME,IR,DR,RI,LE>  CR: 24002822 XER: 200400cf
>    ...
>    NIP fill_queue+0xf0/0x210 [virtio_console]
>    LR  fill_queue+0xf0/0x210 [virtio_console]
>    Call Trace:
>      fill_queue+0xb4/0x210 [virtio_console] (unreliable)
>      add_port+0x1a8/0x470 [virtio_console]
>      control_work_handler+0xbc/0x1e8 [virtio_console]
>      process_one_work+0x290/0x590
>      worker_thread+0x88/0x620
>      kthread+0x194/0x1a0
>      ret_from_kernel_thread+0x5c/0x64
> 
> Jordan, Fabiano & Murilo were able to reproduce and identify that the
> problem is caused by the call to module_enable_ro() in do_init_module(),
> which happens after the module's init function has already been called.
> 
> Our current implementation of change_page_attr() is not safe against
> concurrent accesses, because it invalidates the PTE before flushing the
> TLB and then installing the new PTE. That leaves a window in time where
> there is no valid PTE for the page, if another CPU tries to access the
> page at that time we see something like the fault above.
> 
> We can't simply switch to set_pte_at()/flush TLB, because our hash MMU
> code doesn't handle a set_pte_at() of a valid PTE. See [1].
> 
> But we do have pte_update(), which replaces the old PTE with the new,
> meaning there's no window where the PTE is invalid. And the hash MMU
> version hash__pte_update() deals with synchronising the hash page table
> correctly.
> 
> [1]: https://lore.kernel.org/linuxppc-dev/87y318wp9r.fsf@linux.ibm.com/
> 
> Fixes: 1f9ad21c3b38 ("powerpc/mm: Implement set_memory() routines")
> Reported-by: Laurent Vivier <lvivier at redhat.com>
> Signed-off-by: Fabiano Rosas <farosas at linux.ibm.com>
> Signed-off-by: Michael Ellerman <mpe at ellerman.id.au>

Reviewed-by: Christophe Leroy <christophe.leroy at csgroup.eu>

> ---
>   arch/powerpc/mm/pageattr.c | 23 ++++++++++-------------
>   1 file changed, 10 insertions(+), 13 deletions(-)
> 
> v2: Use pte_update(..., ~0, pte_val(pte), ...) as suggested by Fabiano,
>      and ptep_get() as suggested by Christophe.
> 
> diff --git a/arch/powerpc/mm/pageattr.c b/arch/powerpc/mm/pageattr.c
> index 0876216ceee6..edea388e9d3f 100644
> --- a/arch/powerpc/mm/pageattr.c
> +++ b/arch/powerpc/mm/pageattr.c
> @@ -18,16 +18,12 @@
>   /*
>    * Updates the attributes of a page in three steps:
>    *
> - * 1. invalidate the page table entry
> - * 2. flush the TLB
> - * 3. install the new entry with the updated attributes
> - *
> - * Invalidating the pte means there are situations where this will not work
> - * when in theory it should.
> - * For example:
> - * - removing write from page whilst it is being executed
> - * - setting a page read-only whilst it is being read by another CPU
> + * 1. take the page_table_lock
> + * 2. install the new entry with the updated attributes
> + * 3. flush the TLB
>    *
> + * This sequence is safe against concurrent updates, and also allows updating the
> + * attributes of a page currently being executed or accessed.
>    */
>   static int change_page_attr(pte_t *ptep, unsigned long addr, void *data)
>   {
> @@ -36,9 +32,7 @@ static int change_page_attr(pte_t *ptep, unsigned long addr, void *data)
>   
>   	spin_lock(&init_mm.page_table_lock);
>   
> -	/* invalidate the PTE so it's safe to modify */
> -	pte = ptep_get_and_clear(&init_mm, addr, ptep);
> -	flush_tlb_kernel_range(addr, addr + PAGE_SIZE);
> +	pte = ptep_get(ptep);
>   
>   	/* modify the PTE bits as desired, then apply */
>   	switch (action) {
> @@ -59,11 +53,14 @@ static int change_page_attr(pte_t *ptep, unsigned long addr, void *data)
>   		break;
>   	}
>   
> -	set_pte_at(&init_mm, addr, ptep, pte);
> +	pte_update(&init_mm, addr, ptep, ~0UL, pte_val(pte), 0);
>   
>   	/* See ptesync comment in radix__set_pte_at() */
>   	if (radix_enabled())
>   		asm volatile("ptesync": : :"memory");
> +
> +	flush_tlb_kernel_range(addr, addr + PAGE_SIZE);
> +
>   	spin_unlock(&init_mm.page_table_lock);
>   
>   	return 0;
> 
> base-commit: cbc06f051c524dcfe52ef0d1f30647828e226d30
>