On 7/14/22 8:09 AM, Jose E. Marchesi wrote:
Hi Yonghong.
On 7/7/22 1:24 PM, Jose E. Marchesi wrote:
Hi Yonghong.
On 6/21/22 9:12 AM, Jose E. Marchesi wrote:
On 6/17/22 10:18 AM, Jose E. Marchesi wrote:
Hi Yonghong.
On 6/15/22 1:57 PM, David Faust wrote:
On 6/14/22 22:53, Yonghong Song wrote:
On 6/7/22 2:43 PM, David Faust wrote:
Hello,
This patch series adds support for:
- Two new C-language-level attributes that allow to associate (to "annotate" or
to "tag") particular declarations and types with arbitrary strings. As
explained below, this is intended to be used to, for example,
characterize
certain pointer types.
- The conveyance of that information in the DWARF output in the form of a new
DIE: DW_TAG_GNU_annotation.
- The conveyance of that information in the BTF output in the form of two new
kinds of BTF objects: BTF_KIND_DECL_TAG and BTF_KIND_TYPE_TAG.
All of these facilities are being added to the eBPF ecosystem, and support for
them exists in some form in LLVM.
Purpose
=======
1) Addition of C-family language constructs (attributes) to specify free-text
tags on certain language elements, such as struct fields.
The purpose of these annotations is to provide additional information
about
types, variables, and function parameters of interest to the kernel. A
driving use case is to tag pointer types within the linux kernel and
eBPF
programs with additional semantic information, such as '__user' or
'__rcu'.
For example, consider the linux kernel function do_execve with the
following declaration:
static int do_execve(struct filename *filename,
const char __user *const __user *__argv,
const char __user *const __user *__envp);
Here, __user could be defined with these annotations to record
semantic
information about the pointer parameters (e.g., they are
user-provided) in
DWARF and BTF information. Other kernel facilites such as the eBPF
verifier
can read the tags and make use of the information.
2) Conveying the tags in the generated DWARF debug info.
The main motivation for emitting the tags in DWARF is that the Linux
kernel
generates its BTF information via pahole, using DWARF as a source:
+--------+ BTF BTF +----------+
| pahole |-------> vmlinux.btf ------->| verifier |
+--------+ +----------+
^ ^
| |
DWARF | BTF |
| |
vmlinux +-------------+
module1.ko | BPF program |
module2.ko +-------------+
...
This is because:
a) Unlike GCC, LLVM will only generate BTF for BPF programs.
b) GCC can generate BTF for whatever target with -gbtf, but there is
no
support for linking/deduplicating BTF in the linker.
In the scenario above, the verifier needs access to the pointer tags
of
both the kernel types/declarations (conveyed in the DWARF and
translated
to BTF by pahole) and those of the BPF program (available directly in
BTF).
Another motivation for having the tag information in DWARF, unrelated
to
BPF and BTF, is that the drgn project (another DWARF consumer) also
wants
to benefit from these tags in order to differentiate between different
kinds of pointers in the kernel.
3) Conveying the tags in the generated BTF debug info.
This is easy: the main purpose of having this info in BTF is for the
compiled eBPF programs. The kernel verifier can then access the tags
of pointers used by the eBPF programs.
For more information about these tags and the motivation behind them, please
refer to the following linux kernel discussions:
https://lore.kernel.org/bpf/20210914223004.244411-1-...@fb.com/
https://lore.kernel.org/bpf/20211012164838.3345699-1-...@fb.com/
https://lore.kernel.org/bpf/20211112012604.1504583-1-...@fb.com/
Implementation Overview
=======================
To enable these annotations, two new C language attributes are added:
__attribute__((debug_annotate_decl("foo"))) and
__attribute__((debug_annotate_type("bar"))). Both attributes accept a single
arbitrary string constant argument, which will be recorded in the generated
DWARF and/or BTF debug information. They have no effect on code generation.
Note that we are not using the same attribute names as LLVM (btf_decl_tag and
btf_type_tag, respectively). While these attributes are functionally very
similar, they have grown beyond purely BTF-specific uses, so inclusion of "btf"
in the attribute name seems misleading.
DWARF support is enabled via a new DW_TAG_GNU_annotation. When generating DWARF,
declarations and types will be checked for the corresponding attributes. If
present, a DW_TAG_GNU_annotation DIE will be created as a child of the DIE for
the annotated type or declaration, one for each tag. These DIEs link the
arbitrary tag value to the item they annotate.
For example, the following variable declaration:
#define __typetag1 __attribute__((debug_annotate_type ("typetag1")))
#define __decltag1 __attribute__((debug_annotate_decl ("decltag1")))
#define __decltag2 __attribute__((debug_annotate_decl ("decltag2")))
int * __typetag1 x __decltag1 __decltag2;
Based on the above example
static int do_execve(struct filename *filename,
const char __user *const __user *__argv,
const char __user *const __user *__envp);
Should the above example should be the below?
int __typetag1 * x __decltag1 __decltag2
This example is not related to the one above. It is just meant to
show the behavior of both attributes. My apologies for not making
that clear.
Okay, it should be fine if the dwarf debug_info is shown.
Produces the following DWARF information:
<1><1e>: Abbrev Number: 3 (DW_TAG_variable)
<1f> DW_AT_name : x
<21> DW_AT_decl_file : 1
<22> DW_AT_decl_line : 7
<23> DW_AT_decl_column : 18
<24> DW_AT_type : <0x49>
<28> DW_AT_external : 1
<28> DW_AT_location : 9 byte block: 3 0 0 0 0 0 0 0 0
(DW_OP_addr: 0)
<32> DW_AT_sibling : <0x49>
<2><36>: Abbrev Number: 1 (User TAG value: 0x6000)
<37> DW_AT_name : (indirect string, offset: 0xd6):
debug_annotate_decl
<3b> DW_AT_const_value : (indirect string, offset: 0xcd): decltag2
<2><3f>: Abbrev Number: 1 (User TAG value: 0x6000)
<40> DW_AT_name : (indirect string, offset: 0xd6):
debug_annotate_decl
<44> DW_AT_const_value : (indirect string, offset: 0x0): decltag1
<2><48>: Abbrev Number: 0
<1><49>: Abbrev Number: 4 (DW_TAG_pointer_type)
<4a> DW_AT_byte_size : 8
<4b> DW_AT_type : <0x5d>
<4f> DW_AT_sibling : <0x5d>
<2><53>: Abbrev Number: 1 (User TAG value: 0x6000)
<54> DW_AT_name : (indirect string, offset: 0x9):
debug_annotate_type
<58> DW_AT_const_value : (indirect string, offset: 0x1d): typetag1
<2><5c>: Abbrev Number: 0
<1><5d>: Abbrev Number: 5 (DW_TAG_base_type)
<5e> DW_AT_byte_size : 4
<5f> DW_AT_encoding : 5 (signed)
<60> DW_AT_name : int
<1><64>: Abbrev Number: 0
This shows the info in .debug_abbrev. What I mean is to
show the related info in .debug_info section which seems more useful to
understand the relationships between different tags. Maybe this is due
to that I am not fully understanding what <1>/<2> means in <1><49> and
<2><53> etc.
I think that dump actually shows .debug_info, with the abbrevs
expanded...
Anyway, it seems to us that the root of this problem is the fact the
kernel sparse annotations, such as address_space(__user), are:
1) To be processed by an external kernel-specific tool (
https://sparse.docs.kernel.org/en/latest/annotations.html) and not a
C compiler, and therefore,
2) Not quite the same than compiler attributes (despite the way they
look.) In particular, they seem to assume an ordering different than
of GNU attributes: in some cases given the same written order, they
refer to different things!. Which is quite unfortunate :(
Yes, currently __user/__kernel macros (implemented with address_space
attribute) are processed by macros.
Now, if I understood properly, you plan to change the definition of
__user and __kernel in the kernel sources in order to generate the tag
compiler attributes, correct?
Right. The original __user definition likes:
# define __user __attribute__((noderef, address_space(__user)))
The new attribute looks like
# define BTF_TYPE_TAG(value) __attribute__((btf_type_tag(#value)))
# define __user BTF_TYPE_TAG(user)
Ok I see. So the kernel will stop using sparse attributes to
implement
__user and __kernel and start using compiler attributes for tags
instead.
Is that the reason why LLVM implements what we assume to be the
sparse
ordering, and not the correct GNU attributes ordering, for the tag
attributes?
Note that __user attributes apply to pointee's and not pointers.
Just like
const int *p;
the 'const' is not applied to pointer 'p', but the pointee of 'p'.
What current llvm dwarf generation with
pointer
<--- btf_type_tag
is just ONE implementation. As I said earlier, I am okay to
have dwarf implementation like
p->btf_type_tag->const->int.
If you can propose an implementation like this in dwarf. I can propose
to change implementation in llvm.
I think we are miscommunicating.
Looks like there is a divergence on what attributes apply to what
language entities between the sparse compiler and GCC/LLVM. How to
represent that in DWARF is a different matter.
For this example:
int __typetag1 * __typetag2 __typetag3 * g;
a) GCC associates __typetag1 with the pointer-to-pointer-to-int.
b) LLVM associates __typetag1 to pointer-to-int.
Where:
a) Is the expected behavior of a compiler attributes, as documented
in
the GCC manual.
b) Is presumably what the sparse compiler expects, but _not_ the
ordering expected for a compiler GNU attribute.
So, if the kernel source __user and __kernel annotations (which
currently expand to sparse attributes) follow the sparse ordering, and
you want to implement __user and __kernel in terms of compiler
attributes instead (the annotation attributes) then you will have to:
1) Fix LLVM to implement the usual ordering for these attributes and
2) fix the kernel sources to use that ordering
[Incidentally, the same applies to another "ex-sparse" attribute you
have in the kernel and also implemented in LLVM with a weird ordering:
the address_space attribute.]
For 2), it may be possible to write a coccinnelle script to generate
the
patch...
I don't think (2) (to change kernel source for different attr ordering)
will work. So the only thing we can do is in compiler/pahole except
macro replacement in kernel.
I looked at sparse and its parser. Wanted to be sure the ordering
it
uses to interpret sparse annotations (such as address_space, alignment,
etc) is definitely _not_ the same ordering used by __attribute__ in C
compilers.
It is very different indeed and the same can be said about how
sparse
interprets other modifiers like `const': in sparse both `int const *foo'
and `int *const foo' parse to a constant pointer to int, for example.
I am not to judge how sparse handles its annotations. It may be
very
well and pertinent for its particular purpose.
But I am not sure if it is reasonable to expect C compilers to
implement
certain type __attributes__ to parse differently, just because it
happens these attributes are reused from sparse annotations in a
particular program (in this case the kernel.) The debug_annotate_decl
and debug_annotate_type attributes are not even intended to be
kernel-specific.
So, if changing the kernel sources is not an option (why btw, other
than
being a PITA?) at this point I really don't know what else to suggest :/
Any suggestion from the front-end people?
Just want to understand the overall picture. So gcc can still emit
BTF properly with btf_type_tag right? The issue we are talking about
here is about the dwarf, right?
If by "properly" you mean how sparse handles its annotations, then not
really.
The issue we are talking about is rather a language-level one: to what
entity/type the compiler attribute applies.
So, for:
int __attribute__((debug_annotate_decl("user"))) *foo;
GCC will apply the attribute to the int type, following the rules for
type attributes (sparse would apply the annotation to the *int type
instead). The emitted debug info (be it DWARF or BTF) will reflect
that, no more no less :/
I don't know what does this 'apply the attribute to the int' mean.
In current clang implementation it means the following dwarf chains
from right to left
variable 'foo'
type: ptr
base type: attr_type: attr
underlying type: int
So the type chain is foo -> ptr -> attr -> int
If this is the case, we might have
a partial solution here.
- gcc emits BTF for vmlinux
Note that for emitting BTF for vmlinux we would need support in the
linker to merge and deduplicate BTF, which at the moment we don't have.
This should be okay. pahole will merge and deduplicate btf. In pahole
'-j' mode, each thread will convert each .o file dwarf to btf, and
then pahole will merge and deduplicate btf.
- gcc emits dwarf for vmlinux ignoring btf_type_tag
- in pahole, vmlinux BTF is amended with some additional misc things.
Although there are some use cases to have btf_type_tag in dwarf, but
that can be workarouned with BTF + dwarf both of which are generated
by the compiler. Not elegent, but probably works.
Does this make sense?
If that is so, we have quite a problem here: I don't think we can
change
the way GCC handles GNU-like attributes just because the kernel sources
want to hook on these __user/__kernel sparse annotations to generate the
compiler tags, even if we could mayhaps get GCC to handle
debug_annotate_type and debug_annotate_decl differently. Some would say
doing so would perpetuate the mistake instead of fixing it...
Is my understanding correct?
Let us just say that the btf_type_tag attribute applies to pointees.
Does this help?
Maybe you can also show what dwarf debug_info looks like
I am not sure what you mean. This is the .debug_info section as output
by readelf -w. I did trim some information not relevant to the discussion
such as the DW_TAG_compile_unit DIE, for brevity.
In the case of BTF, the annotations are recorded in two type kinds recently
added to the BTF specification: BTF_KIND_DECL_TAG and BTF_KIND_TYPE_TAG.
The above example declaration prodcues the following BTF information:
[1] INT 'int' size=4 bits_offset=0 nr_bits=32 encoding=SIGNED
[2] PTR '(anon)' type_id=3
[3] TYPE_TAG 'typetag1' type_id=1
[4] DECL_TAG 'decltag1' type_id=6 component_idx=-1
[5] DECL_TAG 'decltag2' type_id=6 component_idx=-1
[6] VAR 'x' type_id=2, linkage=global
[7] DATASEC '.bss' size=0 vlen=1
type_id=6 offset=0 size=8 (VAR 'x')
[...]
