On Wed, Jun 14, 2023 at 02:23:39PM +0100, Lee Jones wrote:
> Good afternoon,
>
> I'm looking for a little expert help with a out-of-bounds issue I've
> been working on. Please let me know if my present understanding does
> not quite match reality.
>
> - Report
>
> The following highly indeterministic kernel panic was reported to occur every
> few hundred boots:
>
> Kernel panic - not syncing: Fatal exception
> RIP: 0010:alloc_ucounts+0x68/0x280
> RSP: 0018:ffffb53ac13dfe98 EFLAGS: 00010006
> RAX: 0000000000000000 RBX: 000000015b804063 RCX: ffff9bb60d5aa500
> RDX: 0000000000000001 RSI: 00000000000003fb RDI: 0000000000000001
> RBP: ffffb53ac13dfec0 R08: 0000000000000000 R09: ffffd53abf600000
> R10: ffff9bb60b4bdd80 R11: ffffffffbcde7bb0 R12: ffffffffbf04e710
> R13: ffffffffbf405160 R14: 00000000000003fb R15: ffffffffbf1afc60
> FS: 00007f5be44d5000(0000) GS:ffff9bb6b9cc0000(0000) knlGS:0000000000000000
> CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033
> CR2: 000000015b80407b CR3: 00000001038ea000 CR4: 00000000000406a0
> DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
> DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400
> Call Trace:
> <TASK>
> __sys_setuid+0x1a0/0x290
> __x64_sys_setuid+0xc/0x10
> do_syscall_64+0x43/0x90
> ? asm_exc_page_fault+0x8/0x30
> entry_SYSCALL_64_after_hwframe+0x44/0xae
>
> - Problem
>
> The issue was eventually tracked down to the attempted dereference of a value
> located in a corrupted hash table. ucounts_hashtable is an array of 1024
> struct hlists. Each element is the head of its own linked list where previous
> ucount allocations are stored. The [20]th element of ucounts_hashtable was
> being consistently trashed on each and every boot. However the indeterminism
> comes from it being accessed only every few hundred boots.
>
> The issue disappeared, or was at least unidentifiable when !(LTO=full) or when
> memory base randomisation (a.k.a. KASLR) was disabled, rending GDB all but
> impossible to use effectively.
>
> The cause of the corruption was uncovered using a verity of different debugging
> techniques and was eventually tracked down to page table manipulation in early
> architecture setup.
>
> The following line in arch/x86/realmode/init.c [0] allocates a variable, just 8
> Bytes in size, to "hold the pgd entry used on booting additional CPUs":
>
> pgd_t trampoline_pgd_entry;
>
> The address of that variable is then passed from init_trampoline_kaslr() via a
> call to set_pgd() [1] to have a value (not too important here) assigned to it.
> Numerous abstractions take place, eventually leading to native_set_p4d(), an
> inline function [2] contained in arch/x86/include/asm/pgtable_64.h.
>
> From here, intentionally or otherwise, a call to pti_set_user_pgtbl() is made.
> This is where the out-of-bounds write eventually occurs. It is not known (by
> me) why this function is called. The returned result is subsequently used as a
> value to write using the WRITE_ONCE macro. Perhaps the premature write is not
> intended. This is what I hope to find out.
>
> A little way down in pti_set_user_pgtbl() [3] the following line occurs:
>
> kernel_to_user_pgdp(pgdp)->pgd = pgd.pgd
>
> The kernel_to_user_pgdp() part takes the address of pgdp (a.k.a.
> trampoline_pgd_entry) and ends up flipping the 12th bit, essentially adding 4k
> (0x1000) to the address. Then the part at the end assigns our value (still not
> important here) to it. However, if we remember that only 8 Bytes was allocated
> (globally) for trampoline_pgd_entry, then means we just stored the value into
> the outlands (what we now know to be allocated to another global storage user
> ucounts_hashtable).
>
> - Ask
>
> Hopefully I haven't messed anything up in the explanation here. Please let me
> know if this is the case. I'm keen to understand what the intention was and
> what we might do to fix it, if of course this hasn't already been remedied
> somewhere.
>
> As ever, any help / guidance would be gratefully received.
This is the PTI, or Page-Table-Isolation muck that is the Meltdown
mitigation. Basically we end up running with 2 sets of page-tables, a
kernel and user pagetable, where the user pagetable only contains the
bare minimum of the kernel to make the transition (and consequently
userspace cannot access the kernel data).
The way this is done is by making the PGD two consecutive pages aligned
to 2*PAGE_SIZE, and a switch between the two CR3 values is a simple
bitop. The kernel is the first (or lower address) while the user is the
second (or higher address).
This is what kernel_to_user_pgdp() is doing, and what you'll find in
things like SWITCH_TO_{KERNEL,USER}_CR3 from arch/x86/entry/calling.h.
Now the problem is that:
1) this kaslr trampoline stuff does not play by the same rules, but
triggers it because it's about the first 1M of memory -- aka userspace
range.
2) we've already set X86_FEATURE_PTI by the time this happens
I've not yet remebered enough of this crap to figure out the best way to
cure this. Clearly we don't need PTI to perform kalsr, but all of this
early setup is a giant maze so perhaps PTI is needed for something else
this early.
IFF not, the solution might be to set PTI later.
On Wed, 14 Jun 2023, Peter Zijlstra wrote:
> On Wed, Jun 14, 2023 at 02:23:39PM +0100, Lee Jones wrote:
> > Good afternoon,
> >
> > I'm looking for a little expert help with a out-of-bounds issue I've
> > been working on. Please let me know if my present understanding does
> > not quite match reality.
> >
> > - Report
> >
> > The following highly indeterministic kernel panic was reported to occur every
> > few hundred boots:
> >
> > Kernel panic - not syncing: Fatal exception
> > RIP: 0010:alloc_ucounts+0x68/0x280
> > RSP: 0018:ffffb53ac13dfe98 EFLAGS: 00010006
> > RAX: 0000000000000000 RBX: 000000015b804063 RCX: ffff9bb60d5aa500
> > RDX: 0000000000000001 RSI: 00000000000003fb RDI: 0000000000000001
> > RBP: ffffb53ac13dfec0 R08: 0000000000000000 R09: ffffd53abf600000
> > R10: ffff9bb60b4bdd80 R11: ffffffffbcde7bb0 R12: ffffffffbf04e710
> > R13: ffffffffbf405160 R14: 00000000000003fb R15: ffffffffbf1afc60
> > FS: 00007f5be44d5000(0000) GS:ffff9bb6b9cc0000(0000) knlGS:0000000000000000
> > CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033
> > CR2: 000000015b80407b CR3: 00000001038ea000 CR4: 00000000000406a0
> > DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
> > DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400
> > Call Trace:
> > <TASK>
> > __sys_setuid+0x1a0/0x290
> > __x64_sys_setuid+0xc/0x10
> > do_syscall_64+0x43/0x90
> > ? asm_exc_page_fault+0x8/0x30
> > entry_SYSCALL_64_after_hwframe+0x44/0xae
> >
> > - Problem
> >
> > The issue was eventually tracked down to the attempted dereference of a value
> > located in a corrupted hash table. ucounts_hashtable is an array of 1024
> > struct hlists. Each element is the head of its own linked list where previous
> > ucount allocations are stored. The [20]th element of ucounts_hashtable was
> > being consistently trashed on each and every boot. However the indeterminism
> > comes from it being accessed only every few hundred boots.
> >
> > The issue disappeared, or was at least unidentifiable when !(LTO=full) or when
> > memory base randomisation (a.k.a. KASLR) was disabled, rending GDB all but
> > impossible to use effectively.
> >
> > The cause of the corruption was uncovered using a verity of different debugging
> > techniques and was eventually tracked down to page table manipulation in early
> > architecture setup.
> >
> > The following line in arch/x86/realmode/init.c [0] allocates a variable, just 8
> > Bytes in size, to "hold the pgd entry used on booting additional CPUs":
> >
> > pgd_t trampoline_pgd_entry;
> >
> > The address of that variable is then passed from init_trampoline_kaslr() via a
> > call to set_pgd() [1] to have a value (not too important here) assigned to it.
> > Numerous abstractions take place, eventually leading to native_set_p4d(), an
> > inline function [2] contained in arch/x86/include/asm/pgtable_64.h.
> >
> > From here, intentionally or otherwise, a call to pti_set_user_pgtbl() is made.
> > This is where the out-of-bounds write eventually occurs. It is not known (by
> > me) why this function is called. The returned result is subsequently used as a
> > value to write using the WRITE_ONCE macro. Perhaps the premature write is not
> > intended. This is what I hope to find out.
> >
> > A little way down in pti_set_user_pgtbl() [3] the following line occurs:
> >
> > kernel_to_user_pgdp(pgdp)->pgd = pgd.pgd
> >
> > The kernel_to_user_pgdp() part takes the address of pgdp (a.k.a.
> > trampoline_pgd_entry) and ends up flipping the 12th bit, essentially adding 4k
> > (0x1000) to the address. Then the part at the end assigns our value (still not
> > important here) to it. However, if we remember that only 8 Bytes was allocated
> > (globally) for trampoline_pgd_entry, then means we just stored the value into
> > the outlands (what we now know to be allocated to another global storage user
> > ucounts_hashtable).
> >
> > - Ask
> >
> > Hopefully I haven't messed anything up in the explanation here. Please let me
> > know if this is the case. I'm keen to understand what the intention was and
> > what we might do to fix it, if of course this hasn't already been remedied
> > somewhere.
> >
> > As ever, any help / guidance would be gratefully received.
>
> This is the PTI, or Page-Table-Isolation muck that is the Meltdown
> mitigation. Basically we end up running with 2 sets of page-tables, a
> kernel and user pagetable, where the user pagetable only contains the
> bare minimum of the kernel to make the transition (and consequently
> userspace cannot access the kernel data).
>
> The way this is done is by making the PGD two consecutive pages aligned
> to 2*PAGE_SIZE, and a switch between the two CR3 values is a simple
> bitop. The kernel is the first (or lower address) while the user is the
> second (or higher address).
>
> This is what kernel_to_user_pgdp() is doing, and what you'll find in
> things like SWITCH_TO_{KERNEL,USER}_CR3 from arch/x86/entry/calling.h.
>
> Now the problem is that:
>
> 1) this kaslr trampoline stuff does not play by the same rules, but
> triggers it because it's about the first 1M of memory -- aka userspace
> range.
>
> 2) we've already set X86_FEATURE_PTI by the time this happens
>
> I've not yet remebered enough of this crap to figure out the best way to
> cure this. Clearly we don't need PTI to perform kalsr, but all of this
> early setup is a giant maze so perhaps PTI is needed for something else
> this early.
>
> IFF not, the solution might be to set PTI later.
I see.
Still unsure how we (the kernel) can/should write to an area of memory
that does not belong to it. Should we allocate enough memory
(2*PAGE_SIZE? rather than 8-Bytes) for trampoline_pgd_entry to consume
in a more sane way?
--
Lee Jones [李琼斯]
On 6/14/23 07:37, Lee Jones wrote:
> Still unsure how we (the kernel) can/should write to an area of memory
> that does not belong to it. Should we allocate enough memory
> (2*PAGE_SIZE? rather than 8-Bytes) for trampoline_pgd_entry to consume
> in a more sane way?
No.
I think this:
set_pgd(&trampoline_pgd_entry,
__pgd(_KERNPG_TABLE | __pa(p4d_page_tramp)));
is bogus-ish. set_pgd() wants to operate on a pgd_t inside a pgd
*PAGE*. But it's just being pointed at a single _entry_. The address
of 'trampoline_pgd_entry' in your case also just (unfortunately)
happens to pass the:
__pti_set_user_pgtbl -> pgdp_maps_userspace()
test. I _think_ we want these to just be something like:
trampoline_pgd_entry = __pgd(_KERNPG_TABLE |
__pa(p4d_page_tramp);
That'll keep us away from all of the set_pgd()-induced nastiness.