Date: Tue, 4 Jan 2022 16:05:53 +0100 From: Peter <pmc@citylink.dinoex.sub.org> To: freebsd-stable@freebsd.org, markj@freebsd.org, jtl@freebsd.org Subject: Re: dtrace bitfields failure (was: 12.3-RC1 fails ...) Message-ID: <YdRiUXviQK8hCBhC@gate.intra.daemon.contact>
next in thread | raw e-mail | index | archive | help
Hija,
sadly, I was too early in agreeing that the two patches
22082f15f9
68396709e7
together do solve the issue. They only do on a certain assumption,
which does not hold true in all cases.
Let's look at https://reviews.freebsd.org/D27213
This is the code in question that will trigger the action:
if (dst_type == CTF_ERR && name[0] != '\0' &&
(hep = ctf_hash_lookup(&src_fp->ctf_names, src_fp, name,
strlen(name))) != NULL &&
src_type != (ctf_id_t)hep->h_type) {
What happens here: in the case of a bitfield type we need to also
copy the corresponding intrinsic type. This condition here checks for
the case and also should deliver that respective intrinsic type
into the "hep" variable.
But this depends on the assumption that the intrinsic type appears
first in the "src_fp" container, so that the hash will point to it.
And that is not necessarily true; it depends on what options you have
in your kernel config.
For instance, with my custom kernel, things look like this:
$ ctfdump -t kernel.full
- Types ----------------------------------------------------------------------
[1] STRUCT (anon) (8 bytes)
sle_next type=262 off=0
[2] STRUCT (anon) (8 bytes)
stqe_next type=262 off=0
[3] UNION (anon) (8 bytes)
m_next type=262 off=0
m_slist type=1 off=0
m_stailq type=2 off=0
[4] UNION (anon) (8 bytes)
m_nextpkt type=262 off=0
m_slistpkt type=1 off=0
m_stailqpkt type=2 off=0
<5> INTEGER char encoding=SIGNED CHAR offset=0 bits=8
<6> POINTER (anon) refers to 5
<7> TYPEDEF caddr_t refers to 6
<8> INTEGER int encoding=SIGNED offset=0 bits=32
<9> TYPEDEF __int32_t refers to 8
<10> TYPEDEF int32_t refers to 9
[11] INTEGER unsigned int encoding=0x0 offset=0 bits=8
[12] INTEGER unsigned int encoding=0x0 offset=0 bits=24
[13] STRUCT (anon) (8 bytes)
cstqe_next type=229 off=0
<14> POINTER (anon) refers to 229
[15] STRUCT (anon) (16 bytes)
le_next type=229 off=0
le_prev type=14 off=64
<16> INTEGER long encoding=SIGNED offset=0 bits=64
<17> ARRAY (anon) content: 5 index: 16 nelems: 16
<18> INTEGER unsigned int encoding=0x0 offset=0 bits=32
<19> TYPEDEF u_int refers to 18
[etc.etc.]
As we can see, this one has the bitfield types as #11 and #12, and
the intrinsic type as #18. And consequentially things do fail.
I currently do not know what is the culprit. Has the linking stage of
the kernel a flaw? Or is the patch D27213 based on a wrong assumption?
I hope You guys can answer that. For now I changed the patch D27213
to cover the case, so that things do work.
Further details on request.
cheerio,
PMc
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?YdRiUXviQK8hCBhC>