| Message ID | 20210727205855.411487-5-keescook@chromium.org |
|---|---|
| State | New |
| Headers | show |
| Series | Introduce strict memcpy() bounds checking | expand |
On Tue, Jul 27, 2021 at 01:57:55PM -0700, Kees Cook wrote: > Kernel code has a regular need to describe groups of members within a > structure usually when they need to be copied or initialized separately > from the rest of the surrounding structure. The generally accepted design > pattern in C is to use a named sub-struct: > > struct foo { > int one; > struct { > int two; > int three; > } thing; > int four; > }; > > This would allow for traditional references and sizing: > > memcpy(&dst.thing, &src.thing, sizeof(dst.thing)); > > However, doing this would mean that referencing struct members enclosed > by such named structs would always require including the sub-struct name > in identifiers: > > do_something(dst.thing.three); > > This has tended to be quite inflexible, especially when such groupings > need to be added to established code which causes huge naming churn. > Three workarounds exist in the kernel for this problem, and each have > other negative properties. > > To avoid the naming churn, there is a design pattern of adding macro > aliases for the named struct: > > #define f_three thing.three > > This ends up polluting the global namespace, and makes it difficult to > search for identifiers. > > Another common work-around in kernel code avoids the pollution by avoiding > the named struct entirely, instead identifying the group's boundaries using > either a pair of empty anonymous structs of a pair of zero-element arrays: > > struct foo { > int one; > struct { } start; > int two; > int three; > struct { } finish; > int four; > }; > > struct foo { > int one; > int start[0]; > int two; > int three; > int finish[0]; > int four; > }; > > This allows code to avoid needing to use a sub-struct name for member > references within the surrounding structure, but loses the benefits of > being able to actually use such a struct, making it rather fragile. Using > these requires open-coded calculation of sizes and offsets. The efforts > made to avoid common mistakes include lots of comments, or adding various > BUILD_BUG_ON()s. Such code is left with no way for the compiler to reason > about the boundaries (e.g. the "start" object looks like it's 0 bytes > in length and is not structurally associated with "finish"), making bounds > checking depend on open-coded calculations: > > if (length > offsetof(struct foo, finish) - > offsetof(struct foo, start)) > return -EINVAL; > memcpy(&dst.start, &src.start, length); > > However, the vast majority of places in the kernel that operate on > groups of members do so without any identification of the grouping, > relying either on comments or implicit knowledge of the struct contents, > which is even harder for the compiler to reason about, and results in > even more fragile manual sizing, usually depending on member locations > outside of the region (e.g. to copy "two" and "three", use the start of > "four" to find the size): > > BUILD_BUG_ON((offsetof(struct foo, four) < > offsetof(struct foo, two)) || > (offsetof(struct foo, four) < > offsetof(struct foo, three)); > if (length > offsetof(struct foo, four) - > offsetof(struct foo, two)) > return -EINVAL; > memcpy(&dst.two, &src.two, length); > > And both of the prior two idioms additionally appear to write beyond the > end of the referenced struct member, forcing the compiler to ignore any > attempt to perform bounds checking. > > In order to have a regular programmatic way to describe a struct > region that can be used for references and sizing, can be examined for > bounds checking, avoids forcing the use of intermediate identifiers, > and avoids polluting the global namespace, introduce the struct_group() > macro. This macro wraps the member declarations to create an anonymous > union of an anonymous struct (no intermediate name) and a named struct > (for references and sizing): > > struct foo { > int one; > struct_group(thing, > int two, > int three, > ); > int four; > }; > > if (length > sizeof(src.thing)) > return -EINVAL; > memcpy(&dst.thing, &src.thing, length); > do_something(dst.three); > > There are some rare cases where the resulting struct_group() needs > attributes added, so struct_group_attr() is also introduced to allow > for specifying struct attributes (e.g. __align(x) or __packed). > > Co-developed-by: Keith Packard <keithpac@amazon.com> > Signed-off-by: Keith Packard <keithpac@amazon.com> > Signed-off-by: Kees Cook <keescook@chromium.org> Acked-by: Gustavo A. R. Silva <gustavoars@kernel.org> Love it! :) Thanks -- Gustavo > --- > include/linux/stddef.h | 34 ++++++++++++++++++++++++++++++++++ > 1 file changed, 34 insertions(+) > > diff --git a/include/linux/stddef.h b/include/linux/stddef.h > index 998a4ba28eba..cf7f866944f9 100644 > --- a/include/linux/stddef.h > +++ b/include/linux/stddef.h > @@ -36,4 +36,38 @@ enum { > #define offsetofend(TYPE, MEMBER) \ > (offsetof(TYPE, MEMBER) + sizeof_field(TYPE, MEMBER)) > > +/** > + * struct_group_attr(NAME, ATTRS, MEMBERS) > + * > + * Used to create an anonymous union of two structs with identical > + * layout and size: one anonymous and one named. The former can be > + * used normally without sub-struct naming, and the latter can be > + * used to reason about the start, end, and size of the group of > + * struct members. Includes structure attributes argument. > + * > + * @NAME: The name of the mirrored sub-struct > + * @ATTRS: Any struct attributes (normally empty) > + * @MEMBERS: The member declarations for the mirrored structs > + */ > +#define struct_group_attr(NAME, ATTRS, MEMBERS) \ > + union { \ > + struct { MEMBERS } ATTRS; \ > + struct { MEMBERS } ATTRS NAME; \ > + } > + > +/** > + * struct_group(NAME, MEMBERS) > + * > + * Used to create an anonymous union of two structs with identical > + * layout and size: one anonymous and one named. The former can be > + * used normally without sub-struct naming, and the latter can be > + * used to reason about the start, end, and size of the group of > + * struct members. > + * > + * @NAME: The name of the mirrored sub-struct > + * @MEMBERS: The member declarations for the mirrored structs > + */ > +#define struct_group(NAME, MEMBERS) \ > + struct_group_attr(NAME, /* no attrs */, MEMBERS) > + > #endif > -- > 2.30.2 >
On 27/07/2021 22.57, Kees Cook wrote: > In order to have a regular programmatic way to describe a struct > region that can be used for references and sizing, can be examined for > bounds checking, avoids forcing the use of intermediate identifiers, > and avoids polluting the global namespace, introduce the struct_group() > macro. This macro wraps the member declarations to create an anonymous > union of an anonymous struct (no intermediate name) and a named struct > (for references and sizing): > > struct foo { > int one; > struct_group(thing, > int two, > int three, > ); > int four; > }; That example won't compile, the commas after two and three should be semicolons. And your implementation relies on MEMBERS not containing any comma tokens, but as int a, b, c, d; is a valid way to declare multiple members, consider making MEMBERS variadic #define struct_group(NAME, MEMBERS...) to have it slurp up every subsequent argument and make that work. > > Co-developed-by: Keith Packard <keithpac@amazon.com> > Signed-off-by: Keith Packard <keithpac@amazon.com> > Signed-off-by: Kees Cook <keescook@chromium.org> > --- > include/linux/stddef.h | 34 ++++++++++++++++++++++++++++++++++ Bikeshedding a bit, but do we need to add 34 lines that need to be preprocessed to virtually each and every translation unit [as opposed to adding a struct_group.h header]? Oh well, you need it for struct skbuff.h, so it would be pulled in by a lot regardless :( Rasmus
On Wed, Jul 28, 2021 at 12:54:18PM +0200, Rasmus Villemoes wrote: > On 27/07/2021 22.57, Kees Cook wrote: > > > In order to have a regular programmatic way to describe a struct > > region that can be used for references and sizing, can be examined for > > bounds checking, avoids forcing the use of intermediate identifiers, > > and avoids polluting the global namespace, introduce the struct_group() > > macro. This macro wraps the member declarations to create an anonymous > > union of an anonymous struct (no intermediate name) and a named struct > > (for references and sizing): > > > > struct foo { > > int one; > > struct_group(thing, > > int two, > > int three, > > ); > > int four; > > }; > > That example won't compile, the commas after two and three should be > semicolons. Oops, yes, thanks. This is why I shouldn't write code that doesn't first go through a compiler. ;) > And your implementation relies on MEMBERS not containing any comma > tokens, but as > > int a, b, c, d; > > is a valid way to declare multiple members, consider making MEMBERS > variadic > > #define struct_group(NAME, MEMBERS...) > > to have it slurp up every subsequent argument and make that work. Ah! Perfect, thank you. I totally forgot I could do it that way. > > > > > Co-developed-by: Keith Packard <keithpac@amazon.com> > > Signed-off-by: Keith Packard <keithpac@amazon.com> > > Signed-off-by: Kees Cook <keescook@chromium.org> > > --- > > include/linux/stddef.h | 34 ++++++++++++++++++++++++++++++++++ > > Bikeshedding a bit, but do we need to add 34 lines that need to be > preprocessed to virtually each and every translation unit [as opposed to > adding a struct_group.h header]? Oh well, you need it for struct > skbuff.h, so it would be pulled in by a lot regardless :( My instinct is to make these kinds of helpers "always available" (like sizeof_field(), etc), but I have no strong opinion on where it should live. If the consensus is to move it, I certainly can! :) -Kees
diff --git a/include/linux/stddef.h b/include/linux/stddef.h index 998a4ba28eba..cf7f866944f9 100644 --- a/include/linux/stddef.h +++ b/include/linux/stddef.h @@ -36,4 +36,38 @@ enum { #define offsetofend(TYPE, MEMBER) \ (offsetof(TYPE, MEMBER) + sizeof_field(TYPE, MEMBER)) +/** + * struct_group_attr(NAME, ATTRS, MEMBERS) + * + * Used to create an anonymous union of two structs with identical + * layout and size: one anonymous and one named. The former can be + * used normally without sub-struct naming, and the latter can be + * used to reason about the start, end, and size of the group of + * struct members. Includes structure attributes argument. + * + * @NAME: The name of the mirrored sub-struct + * @ATTRS: Any struct attributes (normally empty) + * @MEMBERS: The member declarations for the mirrored structs + */ +#define struct_group_attr(NAME, ATTRS, MEMBERS) \ + union { \ + struct { MEMBERS } ATTRS; \ + struct { MEMBERS } ATTRS NAME; \ + } + +/** + * struct_group(NAME, MEMBERS) + * + * Used to create an anonymous union of two structs with identical + * layout and size: one anonymous and one named. The former can be + * used normally without sub-struct naming, and the latter can be + * used to reason about the start, end, and size of the group of + * struct members. + * + * @NAME: The name of the mirrored sub-struct + * @MEMBERS: The member declarations for the mirrored structs + */ +#define struct_group(NAME, MEMBERS) \ + struct_group_attr(NAME, /* no attrs */, MEMBERS) + #endif
Kernel code has a regular need to describe groups of members within a structure usually when they need to be copied or initialized separately from the rest of the surrounding structure. The generally accepted design pattern in C is to use a named sub-struct: struct foo { int one; struct { int two; int three; } thing; int four; }; This would allow for traditional references and sizing: memcpy(&dst.thing, &src.thing, sizeof(dst.thing)); However, doing this would mean that referencing struct members enclosed by such named structs would always require including the sub-struct name in identifiers: do_something(dst.thing.three); This has tended to be quite inflexible, especially when such groupings need to be added to established code which causes huge naming churn. Three workarounds exist in the kernel for this problem, and each have other negative properties. To avoid the naming churn, there is a design pattern of adding macro aliases for the named struct: #define f_three thing.three This ends up polluting the global namespace, and makes it difficult to search for identifiers. Another common work-around in kernel code avoids the pollution by avoiding the named struct entirely, instead identifying the group's boundaries using either a pair of empty anonymous structs of a pair of zero-element arrays: struct foo { int one; struct { } start; int two; int three; struct { } finish; int four; }; struct foo { int one; int start[0]; int two; int three; int finish[0]; int four; }; This allows code to avoid needing to use a sub-struct name for member references within the surrounding structure, but loses the benefits of being able to actually use such a struct, making it rather fragile. Using these requires open-coded calculation of sizes and offsets. The efforts made to avoid common mistakes include lots of comments, or adding various BUILD_BUG_ON()s. Such code is left with no way for the compiler to reason about the boundaries (e.g. the "start" object looks like it's 0 bytes in length and is not structurally associated with "finish"), making bounds checking depend on open-coded calculations: if (length > offsetof(struct foo, finish) - offsetof(struct foo, start)) return -EINVAL; memcpy(&dst.start, &src.start, length); However, the vast majority of places in the kernel that operate on groups of members do so without any identification of the grouping, relying either on comments or implicit knowledge of the struct contents, which is even harder for the compiler to reason about, and results in even more fragile manual sizing, usually depending on member locations outside of the region (e.g. to copy "two" and "three", use the start of "four" to find the size): BUILD_BUG_ON((offsetof(struct foo, four) < offsetof(struct foo, two)) || (offsetof(struct foo, four) < offsetof(struct foo, three)); if (length > offsetof(struct foo, four) - offsetof(struct foo, two)) return -EINVAL; memcpy(&dst.two, &src.two, length); And both of the prior two idioms additionally appear to write beyond the end of the referenced struct member, forcing the compiler to ignore any attempt to perform bounds checking. In order to have a regular programmatic way to describe a struct region that can be used for references and sizing, can be examined for bounds checking, avoids forcing the use of intermediate identifiers, and avoids polluting the global namespace, introduce the struct_group() macro. This macro wraps the member declarations to create an anonymous union of an anonymous struct (no intermediate name) and a named struct (for references and sizing): struct foo { int one; struct_group(thing, int two, int three, ); int four; }; if (length > sizeof(src.thing)) return -EINVAL; memcpy(&dst.thing, &src.thing, length); do_something(dst.three); There are some rare cases where the resulting struct_group() needs attributes added, so struct_group_attr() is also introduced to allow for specifying struct attributes (e.g. __align(x) or __packed). Co-developed-by: Keith Packard <keithpac@amazon.com> Signed-off-by: Keith Packard <keithpac@amazon.com> Signed-off-by: Kees Cook <keescook@chromium.org> --- include/linux/stddef.h | 34 ++++++++++++++++++++++++++++++++++ 1 file changed, 34 insertions(+)