svn commit: r302854 - head/sys/kern

[Bruce Evans]

On Fri, 15 Jul 2016, Alexey Dokuchaev wrote:

> On Fri, Jul 15, 2016 at 07:19:22AM +1000, Bruce Evans wrote:
>> On Fri, 15 Jul 2016, Bruce Evans wrote:
>>>> Log:
>>>>  Let DDB's buf printer handle NULL pointers in the buf page array.
>>>
>>> I noticed some other bugs in this code:
>>
>> Oops, that was supposed to be a private reply.
>
> I'm glad it leaked: %p abuse is unfortunately not that rare, and getting
> developers' attention is a good thing.  E.g., every time I do kldstat(8)
> on my PowerPC box I sigh:
>
> % kldstat
> Id Refs Address    Size     Name
> 1    7 0x100000 e0c378   kernel
> 2    1 0xd20dd000 1d000    tmpfs.ko
> 3    2 0xd2114000 18000    geom_sched.ko
> 4    1 0xd2131000 13000    gsched_rr.ko

I think some mail program further broke the formatting -- it omitted leading
spaces so nothing lines up.  After fixing this, only the narrow Address for
Id 1 messes up the alignment.

Output on amd64 shows another problem.  All kernel pointers are above
0xf000000000000000, so %p prints them all with the same width, but this
width is is too wide -- the 'f's  provide no useful info.  Also, kldstat
hard-codes the i386 width of 2+8 in the header, so the Size and Name
columns are consistently misaligned on amd64.

%p is more usable in the kernel.  One of its undocumented features
there is that field widths work with it.  This is sometimes used to
avoid the above bug.  Kernel ddb used to use %8p for all pointers.
The 8 in this is hard-coded for 32-bit arches, so was wrong in another
way.  ddb now uses %p and hard-codes the width of this in the header
as 2+8 or 2+16 depending on __LP64__ and is more broken than before
for small pointer values.  The only important table where this is
used is in ps.  Spaces in this table is wasted to print just 1
pointer -- the relatively useless wchan.

I sometimes compress kernel pointers in output and trace buffers by
omitting leading 1 or 0 bits and trailing 0 bits as well as leading
0x.

User ps is documented to trim even non-redundant leading parts of the
address for *wchan (0x80324000 -> 324000 in the example), but the code
hasn't matched the man page since 2001.  In 4.4BSD-Lite2, this as
implemented using &~ KERNBASE.  That wasn't portable.  The first fix
replaced the 32-bit %08lx format.  Now the format is %0lx.  This moves
the bugs much as in ddb ps, except it doesn't use %p and doesn't print
the 0x prefix.

Bruce