commit 5e1d824f9a283cbf90f25241b66d1f69adb3835b upstream.
During floating point and vector save to thread data f0/vs0 are
clobbered by the FPSCR/VSCR store routine. This has been obvserved to
lead to userspace register corruption and application data corruption
with io-uring.
Fix it by restoring f0/vs0 after FPSCR/VSCR store has completed for
all the FP, altivec, VMX register save paths.
Tested under QEMU in kvm mode, running on a Talos II workstation with
dual POWER9 DD2.2 CPUs.
Additional detail (mpe):
Typically save_fpu() is called from __giveup_fpu() which saves the FP
regs and also *turns off FP* in the tasks MSR, meaning the kernel will
reload the FP regs from the thread struct before letting the task use FP
again. So in that case save_fpu() is free to clobber f0 because the FP
regs no longer hold live values for the task.
There is another case though, which is the path via:
sys_clone()
...
copy_process()
dup_task_struct()
arch_dup_task_struct()
flush_all_to_thread()
save_all()
That path saves the FP regs but leaves them live. That's meant as an
optimisation for a process that's using FP/VSX and then calls fork(),
leaving the regs live means the parent process doesn't have to take a
fault after the fork to get its FP regs back. The optimisation was added
in commit 8792468da5e1 ("powerpc: Add the ability to save FPU without
giving it up").
That path does clobber f0, but f0 is volatile across function calls,
and typically programs reach copy_process() from userspace via a syscall
wrapper function. So in normal usage f0 being clobbered across a
syscall doesn't cause visible data corruption.
But there is now a new path, because io-uring can call copy_process()
via create_io_thread() from the signal handling path. That's OK if the
signal is handled as part of syscall return, but it's not OK if the
signal is handled due to some other interrupt.
That path is:
interrupt_return_srr_user()
interrupt_exit_user_prepare()
interrupt_exit_user_prepare_main()
do_notify_resume()
get_signal()
task_work_run()
create_worker_cb()
create_io_worker()
copy_process()
dup_task_struct()
arch_dup_task_struct()
flush_all_to_thread()
save_all()
if (tsk->thread.regs->msr & MSR_FP)
save_fpu()
# f0 is clobbered and potentially live in userspace
Note the above discussion applies equally to save_altivec().
Fixes: 8792468da5e1 ("powerpc: Add the ability to save FPU without giving it up")
Cc: stable@vger.kernel.org # v4.6+
Closes: https://lore.kernel.org/all/480932026.45576726.1699374859845.JavaMail.zimbra@raptorengineeringinc.com/
Closes: https://lore.kernel.org/linuxppc-dev/480221078.47953493.1700206777956.JavaMail.zimbra@raptorengineeringinc.com/
Tested-by: Timothy Pearson <tpearson@raptorengineering.com>
Tested-by: Jens Axboe <axboe@kernel.dk>
Signed-off-by: Timothy Pearson <tpearson@raptorengineering.com>
[mpe: Reword change log to describe exact path of corruption & other minor tweaks]
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://msgid.link/1921539696.48534988.1700407082933.JavaMail.zimbra@raptorengineeringinc.com
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
[ Upstream commit daa9ada2093ed23d52b4c1fe6e13cf78f55cc85f ]
Erhard reported that his G5 was crashing with v6.6-rc kernels:
mpic: Setting up HT PICs workarounds for U3/U4
BUG: Unable to handle kernel data access at 0xfeffbb62ffec65fe
Faulting instruction address: 0xc00000000005dc40
Oops: Kernel access of bad area, sig: 11 [#1]
BE PAGE_SIZE=4K MMU=Hash SMP NR_CPUS=2 PowerMac
Modules linked in:
CPU: 0 PID: 0 Comm: swapper/0 Tainted: G T 6.6.0-rc3-PMacGS #1
Hardware name: PowerMac11,2 PPC970MP 0x440101 PowerMac
NIP: c00000000005dc40 LR: c000000000066660 CTR: c000000000007730
REGS: c0000000022bf510 TRAP: 0380 Tainted: G T (6.6.0-rc3-PMacGS)
MSR: 9000000000001032 <SF,HV,ME,IR,DR,RI> CR: 44004242 XER: 00000000
IRQMASK: 3
GPR00: 0000000000000000 c0000000022bf7b0 c0000000010c0b00 00000000000001ac
GPR04: 0000000003c80000 0000000000000300 c0000000f20001ae 0000000000000300
GPR08: 0000000000000006 feffbb62ffec65ff 0000000000000001 0000000000000000
GPR12: 9000000000001032 c000000002362000 c000000000f76b80 000000000349ecd8
GPR16: 0000000002367ba8 0000000002367f08 0000000000000006 0000000000000000
GPR20: 00000000000001ac c000000000f6f920 c0000000022cd985 000000000000000c
GPR24: 0000000000000300 00000003b0a3691d c0003e008030000e 0000000000000000
GPR28: c00000000000000c c0000000f20001ee feffbb62ffec65fe 00000000000001ac
NIP hash_page_do_lazy_icache+0x50/0x100
LR __hash_page_4K+0x420/0x590
Call Trace:
hash_page_mm+0x364/0x6f0
do_hash_fault+0x114/0x2b0
data_access_common_virt+0x198/0x1f0
--- interrupt: 300 at mpic_init+0x4bc/0x10c4
NIP: c000000002020a5c LR: c000000002020a04 CTR: 0000000000000000
REGS: c0000000022bf9f0 TRAP: 0300 Tainted: G T (6.6.0-rc3-PMacGS)
MSR: 9000000000001032 <SF,HV,ME,IR,DR,RI> CR: 24004248 XER: 00000000
DAR: c0003e008030000e DSISR: 40000000 IRQMASK: 1
...
NIP mpic_init+0x4bc/0x10c4
LR mpic_init+0x464/0x10c4
--- interrupt: 300
pmac_setup_one_mpic+0x258/0x2dc
pmac_pic_init+0x28c/0x3d8
init_IRQ+0x90/0x140
start_kernel+0x57c/0x78c
start_here_common+0x1c/0x20
A bisect pointed to the breakage beginning with commit 9fee28baa601 ("powerpc:
implement the new page table range API").
Analysis of the oops pointed to a struct page with a corrupted
compound_head being loaded via page_folio() -> _compound_head() in
hash_page_do_lazy_icache().
The access by the mpic code is to an MMIO address, so the expectation
is that the struct page for that address would be initialised by
init_unavailable_range(), as pointed out by Aneesh.
Instrumentation showed that was not the case, which eventually lead to
the realisation that pfn_valid() was returning false for that address,
causing the struct page to not be initialised.
Because the system is using FLATMEM, the version of pfn_valid() in
memory_model.h is used:
static inline int pfn_valid(unsigned long pfn)
{
...
return pfn >= pfn_offset && (pfn - pfn_offset) < max_mapnr;
}
Which relies on max_mapnr being initialised. Early in boot max_mapnr is
zero meaning no PFNs are valid.
max_mapnr is initialised in mem_init() called via:
start_kernel()
mm_core_init() # init/main.c:928
mem_init()
But that is too late for the usage in init_unavailable_range() called via:
start_kernel()
setup_arch() # init/main.c:893
paging_init()
free_area_init()
init_unavailable_range()
Although max_mapnr is currently set in mem_init(), the value is actually
already available much earlier, as soon as mem_topology_setup() has
completed, which is also before paging_init() is called. So move the
initialisation there, which causes paging_init() to correctly initialise
the struct page and fixes the bug.
This bug seems to have been lurking for years, but went unnoticed
because the pre-folio code was inspecting the uninitialised page->flags
but not dereferencing it.
Thanks to Erhard and Aneesh for help debugging.
Reported-by: Erhard Furtner <erhard_f@mailbox.org>
Closes: https://lore.kernel.org/all/20230929132750.3cd98452@yea/
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://msgid.link/20231023112500.1550208-1-mpe@ellerman.id.au
Signed-off-by: Sasha Levin <sashal@kernel.org>
Timothy Pearson