Messages in this thread Patch in this message | | | Date | Wed, 2 Feb 2022 13:07:27 +0800 | From | Feng Tang <> | Subject | Re: Possible bug for ZSTD kernel decompressing |
| |
Hi Nick,
On Mon, Jan 31, 2022 at 08:31:10PM +0000, Nick Terrell wrote: > > > > On Jan 27, 2022, at 8:53 PM, Feng Tang <feng.tang@intel.com> wrote: > > > > Hi All, > > > > Recently 0Day reported a 32bit i386 kernel decompression failure for my > > patch [1], which essentially increase the kernel data section's size > > from 19MB to 53MB, with error message: > > > > early console in setup code > > early console in extract_kernel > > input_data: 0x05077079 > > input_len: 0x00f8a633 > > output: 0x01000000 > > output_len: 0x045c4328 > > kernel_total_size: 0x05040000 > > needed_size: 0x05040000 > > > > Decompressing Linux... > > > > ZSTD-compressed data is corrupt > > > > -- System haltedBUG: kernel hang in boot stage > > > > From debug, it is likely a problem of ZSTD decompression code, as when I > > reverted my patch and hacked to increase the size of kernel data > > section by 32MB, the same error will happen. > > > > Some other hints are: > > * same i386 config with lz4 and xz algo can boot > > * X86_64 + zstd also boots fine > > > > This could be reproduced by qemu cmd: > > > > qemu-system-i386 -machine pc -cpu host -enable-kvm -kernel bzImage -m 2048m -smp 4 -serial stdio --append "earlyprintk=ttyS0,115200 console=ttyS0,115200" > > > > i386 kernel config is attached, and the debug patch as below: > > --- > > diff --git a/init/main.c b/init/main.c > > index 767ee2672176..873f40ddf96e 100644 > > --- a/init/main.c > > +++ b/init/main.c > > @@ -162,6 +162,10 @@ static size_t initargs_offs; > > static char *execute_command; > > static char *ramdisk_execute_command = "/init"; > > > > +#define DT_SIZE 8192000 > > +static unsigned long tbuf[DT_SIZE] = { 1, 2, 3, 4, }; > > + > > /* > > * Used to generate warnings if static_key manipulation functions are used > > * before jump_label_init is called. > > @@ -690,6 +694,11 @@ noinline void __ref rest_init(void) > > struct task_struct *tsk; > > int pid; > > > > + unsigned long i, j; > > + for (i = 0; i < DT_SIZE; i++) > > + j += tbuf[i]; > > + printk("j = 0x%x\n", j); > > + > > rcu_scheduler_starting(); > > /* > > * We need to spawn init first so that it obtains pid 1, however > > > > Please let me know if you need more info. > > > > [1.] https://lore.kernel.org/lkml/1627456900-42743-1-git-send-email-feng.tang@intel.com/ > > I've been unable to reproduce this issue using the provided patch + config based on > Linux v5.17-rc2. > > What version of Linux are you testing on? Zstd was updated in v5.16, so if you're not > testing on v5.16 or later, can you please re-test on v5.17-rc2? The original report I got is against commit 8cd7c588decf "mm/vmscan: throttle reclaim until some writeback completes if congested" which is post 5.15.
I just retested and the issue can _not_ be reproduced against 5.17-rc2. Thanks for the check and fix, and sorry for not trying latest kernel.
- Feng
| |