在 2023/6/22 5:54, Dave Chinner 写道:
> On Wed, Jun 21, 2023 at 05:25:27PM +0800, Wu Guanghao wrote:
>> We encountered a segfault while testing the mkfs.xfs + iscsi.
>>
>> (gdb) bt
>> #0 libxfs_log_sb (tp=0xaaaafaea0630) at xfs_sb.c:810
>> #1 0x0000aaaaca991468 in __xfs_trans_commit (tp=<optimized out>, tp@entry=0xaaaafaea0630, regrant=regrant@entry=true) at trans.c:995
>> #2 0x0000aaaaca991790 in libxfs_trans_roll (tpp=tpp@entry=0xfffffe1f3018) at trans.c:103
>> #3 0x0000aaaaca9bcde8 in xfs_dialloc_roll (agibp=0xaaaafaea2fa0, tpp=0xfffffe1f31c8) at xfs_ialloc.c:1561
>> #4 xfs_dialloc_try_ag (ok_alloc=true, new_ino=<synthetic pointer>, parent=0, pag=0xaaaafaea0210, tpp=0xfffffe1f31c8) at xfs_ialloc.c:1698
>> #5 xfs_dialloc (tpp=tpp@entry=0xfffffe1f31c8, parent=0, mode=mode@entry=16877, new_ino=new_ino@entry=0xfffffe1f3128) at xfs_ialloc.c:1776
>> #6 0x0000aaaaca9925b0 in libxfs_dir_ialloc (tpp=tpp@entry=0xfffffe1f31c8, dp=dp@entry=0x0, mode=mode@entry=16877, nlink=nlink@entry=1, rdev=rdev@entry=0, cr=cr@entry=0xfffffe1f31d0,
>> fsx=fsx@entry=0xfffffe1f36a4, ipp=ipp@entry=0xfffffe1f31c0) at util.c:525
>> #7 0x0000aaaaca988fac in parseproto (mp=0xfffffe1f36c8, pip=0x0, fsxp=0xfffffe1f36a4, pp=0xfffffe1f3370, name=0x0) at proto.c:552
>> #8 0x0000aaaaca9867a4 in main (argc=<optimized out>, argv=<optimized out>) at xfs_mkfs.c:4217
>>
>> (gdb) p bp
>> $1 = 0x0
>>
>> ```
>> void
>> xfs_log_sb(
>> struct xfs_trans *tp)
>> {
>> // iscsi offline
>> ...
>> // failed to read sb, bp = NULL
>> struct xfs_buf *bp = xfs_trans_getsb(tp);
>> ...
>> }
>> ```
>>
>> When writing data to sb, if the device is abnormal at this time,
>> the bp may be empty. Using it without checking will result in
>> a segfault.
>
> xfs_trans_getsb() is not supposed to fail. In the kernel code (which
> this is a copy of) it can't fail because the superblock buffer is
> always pinned in memory at mount time and so is *never read from the
> storage* after mount.
>
Thank you for your suggestion. Later, I will send a patch for the V2 patch.
Thanks,
Guanghao
> Hence something similar needs to be in userspace with libxfs_getsb()
> so that the superblock is only read when setting up the initial
> mount state in libxfs....
>
>> diff --git a/libxfs/xfs_attr_leaf.c b/libxfs/xfs_attr_leaf.c
>> index 6cac2531..73079df1 100644
>> --- a/libxfs/xfs_attr_leaf.c
>> +++ b/libxfs/xfs_attr_leaf.c
>> @@ -668,7 +668,7 @@ xfs_sbversion_add_attr2(
>> spin_lock(&mp->m_sb_lock);
>> xfs_add_attr2(mp);
>> spin_unlock(&mp->m_sb_lock);
>> - xfs_log_sb(tp);
>> + ASSERT(!xfs_log_sb(tp));
>
> FWIW, that's never a valid conversion nor a valid way to handle
> something that can fail. That turns the code into code that is only
> executed on debug builds, and it will panic the debug build rather
> than handle the error.....
>
> -Dave.
>
