Re: [PATCH RESEND] mm/hugetlb: Don't call region_abort if region_chg fails

From: Andrew Morton
Date: Wed Mar 29 2017 - 17:18:57 EST


On Wed, 29 Mar 2017 14:08:02 -0700 Mike Kravetz <mike.kravetz@xxxxxxxxxx> wrote:

> Resending because of typo in Andrew's e-mail when first sent
>
> Changes to hugetlbfs reservation maps is a two step process. The first
> step is a call to region_chg to determine what needs to be changed, and
> prepare that change. This should be followed by a call to call to
> region_add to commit the change, or region_abort to abort the change.
>
> The error path in hugetlb_reserve_pages called region_abort after a
> failed call to region_chg. As a result, the adds_in_progress counter
> in the reservation map is off by 1. This is caught by a VM_BUG_ON
> in resv_map_release when the reservation map is freed.
>
> syzkaller fuzzer found this bug, that resulted in the following:

I'll change the above to

: syzkaller fuzzer (when using an injected kmalloc failure) found this bug,
: that resulted in the following:

it's important, because this bug won't be triggered (at all easily, at
least) in real-world workloads.

> kernel BUG at mm/hugetlb.c:742!
> Call Trace:
> hugetlbfs_evict_inode+0x7b/0xa0 fs/hugetlbfs/inode.c:493
> evict+0x481/0x920 fs/inode.c:553
> iput_final fs/inode.c:1515 [inline]
> iput+0x62b/0xa20 fs/inode.c:1542
> hugetlb_file_setup+0x593/0x9f0 fs/hugetlbfs/inode.c:1306
> newseg+0x422/0xd30 ipc/shm.c:575
> ipcget_new ipc/util.c:285 [inline]
> ipcget+0x21e/0x580 ipc/util.c:639
> SYSC_shmget ipc/shm.c:673 [inline]
> SyS_shmget+0x158/0x230 ipc/shm.c:657
> entry_SYSCALL_64_fastpath+0x1f/0xc2
> RIP: resv_map_release+0x265/0x330 mm/hugetlb.c:742