This revised patch set is rebased onto linux-mmotm.
I have a new mail provider now, but they seem just as bad as the last, so I'm
going to try to split the recpients in half and pray that this goes through.
I'm sorry to those of you who have gotten partial patch sets and I hope to get
this fixed soon. (If anybody knows of a descent email service for developers
who send patches, please let me know.)
Summary
=======
This patch set improves on Andrea Arcangeli's original Red-Black Tree
implementation by adding generic search and insert functions with
complete support for:
o leftmost - keeps a pointer to the leftmost (lowest value) node cached
in your container struct
o rightmost - ditto for rightmost (greatest value)
o count - optionally update an count variable when you perform inserts
or deletes
o unique or non-unique keys
o find and insert "near" functions - when you already have a node that
is likely near another one you want to search for
o type-safe wrapper interface available via pre-processor macro
Outstanding Issues
==================
General
-------
o Need something in Documents to explain generic rbtrees.
o Due to a bug in gcc's optimizer, extra instructions are generated in various
places. Pavel Pisa has provided me a possible work-around that should be
examined more closely to see if it can be working in (Discussed in
Performance section).
o Doc-comments are missing or out of date in some places for the new
ins_compare field of struct rb_relationship (including at least one code
example).
Selftests
---------
o In-kernel test module not completed.
o Userspace selftest's Makefile should run modules_prepare in KERNELDIR.
o Validation in self-tests doesn't yet cover tests for
- insert_near
- find_{first,last,next,prev}
o Selftest scripts need better portability (maybe solved? we'll see)
o It would be nice to have some fault-injection in test code to verify that
CONFIG_DEBUG_GRBTREE and CONFIG_DEBUG_GRBTREE_VALIDATE (and it's
RB_VERIFY_INTEGRITY counterpart flag) catch the errors they are supposed to.
Undecided (Opinions Requested!)
-------------------------------
o With the exception of the rb_node & rb_root structs, "Layer 2" of the code
(see below) completely abstracts away the underlying red-black tree
mechanism. The structs rb_node and rb_root can also be abstracted away via
a typeset or some other mechanism. Thus, should the "Layer 2" code be
separated from "Layer 1" and renamed "Generic Tree (gtree)" or some such,
paving the way for an alternate tree implementation in the future?
o Do we need RB_INSERT_DUPE_RIGHT? (see the last patch)
Theory of Operation
===================
Historically, genericity in C meant function pointers, the overhead of a
function call and the inability of the compiler to optimize code across
the function call boundary. GCC has been getting better and better at
optimization and determining when a value is a compile-time constant and
compiling it out. As of gcc 4.6, it has finally reached a point where
it's possible to have generic search & insert cores that optimize
exactly as well as if they were hand-coded. (see also gcc man page:
-findirect-inlining)
This implementation actually consists of two layers written on top of the
existing rbtree implementation.
Layer 1: Type-Specific (But Not Type-Safe)
------------------------------------------
The first layer consists of enum rb_flags, struct rb_relationship and
some generic inline functions(see patch for doc comments).
enum rb_flags {
RB_HAS_LEFTMOST = 0x00000001,
RB_HAS_RIGHTMOST = 0x00000002,
RB_HAS_COUNT = 0x00000004,
RB_UNIQUE_KEYS = 0x00000008,
RB_INSERT_REPLACES = 0x00000010,
RB_IS_AUGMENTED = 0x00000040,
RB_VERIFY_USAGE = 0x00000080,
RB_VERIFY_INTEGRITY = 0x00000100
};
struct rb_relationship {
ssize_t root_offset;
ssize_t left_offset;
ssize_t right_offset;
ssize_t count_offset;
ssize_t node_offset;
ssize_t key_offset;
int flags;
const rb_compare_f compare; /* comparitor for lookups */
const rb_compare_f ins_compare; /* comparitor for inserts */
unsigned key_size;
};
/* these function for use on all trees */
struct rb_node *rb_find(
struct rb_root *root,
const void *key,
const struct rb_relationship *rel);
struct rb_node *rb_find_near(
struct rb_node *from,
const void *key,
const struct rb_relationship *rel);
struct rb_node *rb_insert(
struct rb_root *root,
struct rb_node *node,
const struct rb_relationship *rel);
struct rb_node *rb_insert_near(
struct rb_root *root,
struct rb_node *start,
struct rb_node *node,
const struct rb_relationship *rel);
void rb_remove( struct rb_root *root,
struct rb_node *node,
const struct rb_relationship *rel);
/* these function for use on trees with non-unique keys */
struct rb_node *rb_find_first(
struct rb_root *root,
const void *key,
const struct rb_relationship *rel);
struct rb_node *rb_find_last(
struct rb_root *root,
const void *key,
const struct rb_relationship *rel);
struct rb_node *rb_find_next(
const struct rb_node *node,
const struct rb_relationship *rel)
struct rb_node *rb_find_prev(
const struct rb_node *node,
const struct rb_relationship *rel)
Using this layer involves initializing a const struct rb_relationship
variable with compile-time constant values and feeding its "address" to
the generic inline functions. The trick being, that (when gcc behaves
properly) it never creates a struct rb_relationship variable, stores an
initializer in the data section of the object file or passes a struct
rb_relationship pointer. Instead, gcc "optimizes out" out the struct,
and uses the compile-time constant values to dictate how the inline
functions will expand.
Thus, this structure can be thought of both as a database's DDL (data
definition language), defining the relationship between two entities and the
template parameters to a C++ templatized function that controls how the
template function is instantiated. This creates type-specific functions,
although type-safety is still not achieved (e.g., you can pass a pointer to
any rb_node you like).
To simplify usage, you can initialize your struct rb_relationship variable
with the RB_RELATIONSHIP macro, feeding it your types, member names and flags
and it will calculate the offsets for you. See doc comments in patch for
examples of using this layer (either with or without the RB_RELATIONSHIP
macro).
Layer 2: Type-Safety
--------------------
In order to achieve type-safety of a generic interface in C, we must delve
deep into the darkened Swamps of The Preprocessor and confront the Prince of
Darkness himself: Big Ugly Macro. To be fair, there is an alternative
solution (discussed in History & Design Goals), the so-called "x-macro" or
"supermacro" where you #define some pre-processor values and include an
unguarded header file. With 17 parameters, I choose this solution for its
ease of use and brevity, but it's an area worth debate (some of which you can
find here if you wish: http://lwn.net/Articles/501876).
So this second layer allows you to use a single macro to define your
relationship as well as type-safe wrapper functions all in one go.
RB_DEFINE_INTERFACE(
prefix,
cont_type, root, left, right, count,
obj_type, node, key,
flags, compare, ins_compare,
find_mod, insert_mod, find_near_mod, insert_near_mod)
To avoid needing multiple versions of the macro, we use a paradigm where
optional values can be left empty. (See RB_DEFINE_INTERFACE doc comments for
details.) Thus, if your container doesn't need to know leftmost, you leave
the parameter empty. Here's a quick example:
struct container {
struct rb_root root;
struct rb_node *leftmost;
unsigned long count;
};
struct object {
struct rb_node node;
long key;
};
static inline long compare_long(const long *a, const long *b)
{
return *a - *b;
}
RB_DEFINE_INTERFACE(
my_objects,
struct container, root, leftmost, /* no rightmost */, count,
struct object, node, key,
RB_UNIQUE_KEYS | RB_INSERT_REPLACES, compare_long, compare_long,
,,,)
This will do some type-checking, create the struct rb_relationship and
the following static __always_inline wrapper functions. (Note that
"my_objects" is the prefix used in the example above. It will be
whatever you pass as the first parameter to the RB_DEFINE_INTERFACE
macro.)
struct object *my_objects_find(
struct container *cont,
const typeof(((struct object *)0)->key) *_key);
struct object *my_objects_insert(
struct container *cont,
struct object *obj);
struct object *my_objects_find_near(
struct object *near,
const typeof(((struct object *)0)->key) *_key);
struct object *my_objects_insert_near(
struct container *cont,
struct object *near,
struct object *obj);
void my_objects_remove(struct container *cont, struct object *obj);
struct object *my_objects_find_first(
struct container *cont,
const typeof(((struct object *)0)->key) *_key);
struct object *my_objects_find_last(
struct container *cont,
const typeof(((struct object *)0)->key) *_key);
struct object *my_objects_find_next(const struct object *obj);
struct object *my_objects_find_last(const struct object *obj);
struct object *my_objects_next(const struct object *obj);
struct object *my_objects_prev(const struct object *obj);
struct object *my_objects_first(struct container *cont);
struct object *my_objects_last(struct container *cont);
Each of these are each declared static __always_inline. However, you can
change the modifiers for the first four (find, insert, find_near and
insert_near) by populating any of the last 4 parameters with the function
modifiers of the respective function (when empty, they default to static
__always_inline).
Not only does this layer give you type-safety, it removes almost all of
the implementation details of the rbtree from the code using it, thus
making it easier to replace the underlying algorithm at some later
date.
Compare Functions
-----------------
Because equality is unimportant when doing inserts into a tree with duplicate
keys, struct rb_relationship's ins_compare field can be set to a greater-than
function for better performance. Using the example in the section above as a
model, this is what it would look like:
static inline long compare_long(const long *a, const long *b)
...
static inline long greater_long(const long *a, const long *b)
{
return *a > *b;
}
RB_DEFINE_INTERFACE(
my_objects,
struct container, root, leftmost, /* no rightmost */, count,
struct object, node, key,
0, compare_long, greater_long,
,,,)
History & Design Goals
======================
I've been through many iterations of various techniques searching for the
perfect "clean" implementation and finally settled on having a huge macro
expand to wrapper functions after exhausting all other alternatives. The trick
is that what one person considers a "clean" implementation is a bit of a value
judgment. So by "clean", I mean balancing these requirements:
1.) minimal dependence on pre-processor
2.) avoiding pre-processor expanded code that will break debug
information (backtraces)
3.) optimal encapsulation of the details of your rbtree in minimal
source code (this is where you define the relationship between your
container and contained objects, their types, keys, rather or not
non-unique objects are allowed, etc.) -- preferably eliminating
duplication of these details entirely.
4.) offering a complete feature-set in a single implementation (not
multiple functions when various features are used)
5.) perfect optimization -- the generic function must be exactly as
efficient as the hand-coded version
By those standards, the "cleanest" implementation I had come up with
actually used a different mechanism: defining an anonymous interface
struct something like this:
/* generic non-type-safe function */
static __always_inline void *__generic_func(void *obj);
struct { \
out_type *(*const func)(in_type *obj); \
} name = { \
.func = (out_type *(*const)(in_type *obj))__generic_func;\
}
/* usage looks like this: */
DEFINE_INTERFACE(solution_a, struct something, struct something_else);
struct something *s;
struct something_else *se;
se = solution_a.func(s);
Sadly, while solution_a.func(s) optimizes perfectly in 4.6, it completely
bombed in 4.5 and prior -- the call by struct-member-function-pointer is never
inlined and nothing passed to it is every considered a compile-time constant
(again, see gcc's docs on -findirect-inline). Because of the implementation
of the generic functions, this bloated the code unacceptably (3x larger).
Thus, I finally settled on the current RB_DEFINE_INTERFACE macro, which is
massive, but optimizes perfectly in 4.6+ and close enough in 4.5 and prior
(prior to 4.6, the compare function is never inlined).
The other alternative I briefly considered was to have a header file
that is only included after #defining all of these parameters, relying
primarily on cpp rather than cc & compile-time constants to fill in the
relationship details (the "x-macro" approach). While this mechanism
would perform better on older compilers and never break backtraces, in
the end, I just couldn't stomach it. Aside from that, it would make
using the interface almost as verbose as hand-coding it yourself.
Performance
===========
Here are the results of performance tests run on v5 of this patch set (against
v3.5 kernel) on an AMD Phenom 9850. This is a reformatted version of what
tools/testing/selftests/grbtree/user/gen_report.sh outputs. Test results vary
quite a bit dependent upon the selected features.
For all of these tests, I used the following parameters:
key range 0-4095
key type u32
object_count 2048
repititions 131,072
node_size 24 bytes
object_size 32 bytes
total data size 65,536 bytes
num insertions 268,435,456
Below is a summary of the performance drop using generic rbtrees on various
ranges of compilers. (negative values are performance improvements)
GCC versions Best Worst
3.4 - 4.0 35% 80%
4.1 - 4.5 18% 23%
4.6 - 4.7 -7% 5%
The tables below list the time in seconds it took to execute the tests on each
compiler and the difference between the generic and specific (i.e.,
hand-coded) test results (from v5 of patches against the 3.5 kernel).
Duplicate keys (no leftmost, rightmost or count)
Compiler Generic Specific Performance Loss
gcc-3.4.6 33.41 18.78 77.94%
gcc-4.0.4 32.36 17.94 80.37%
gcc-4.1.2 23.11 17.76 30.14%
gcc-4.2.4 22.97 17.83 28.84%
gcc-4.3.6 23.07 17.78 29.79%
gcc-4.4.7 21.88 17.64 24.03%
gcc-4.5.4 21.75 17.54 23.99%
gcc-4.6.3 16.84 16.82 0.10%
gcc-4.7.1 16.79 16.68 0.66%
Duplicate keys, use leftmost (no rightmost or count)
Compiler Generic Specific Performance Loss
gcc-3.4.6 33.54 22.57 48.63%
gcc-4.0.4 32.82 22.16 48.07%
gcc-4.1.2 27.30 22.77 19.93%
gcc-4.2.4 27.41 22.86 19.95%
gcc-4.3.6 28.65 23.03 24.38%
gcc-4.4.7 27.03 21.41 26.24%
gcc-4.5.4 26.69 22.48 18.71%
gcc-4.6.3 21.58 21.53 0.24%
gcc-4.7.1 22.40 22.23 0.77%
Duplicate keys, use leftmost, rightmost and count
Compiler Generic Specific Performance Loss
gcc-3.4.6 33.49 22.70 47.52%
gcc-4.0.4 33.19 23.71 39.94%
gcc-4.1.2 29.03 23.76 22.18%
gcc-4.2.4 28.59 23.82 20.04%
gcc-4.3.6 29.69 23.94 24.01%
gcc-4.4.7 28.62 23.89 19.79%
gcc-4.5.4 28.73 23.54 22.04%
gcc-4.6.3 23.82 23.70 0.51%
gcc-4.7.1 23.84 23.94 -0.40%
Unique keys (no leftmost, rightmost or count)
Compiler Generic Specific Performance Loss
gcc-3.4.6 29.38 19.94 47.33%
gcc-4.0.4 28.85 21.14 36.48%
gcc-4.1.2 25.16 20.30 23.95%
gcc-4.2.4 25.26 20.50 23.23%
gcc-4.3.6 25.41 20.82 22.02%
gcc-4.4.7 26.12 20.68 26.33%
gcc-4.5.4 25.29 20.31 24.54%
gcc-4.6.3 21.57 20.35 6.01%
gcc-4.7.1 20.98 20.20 3.88%
Unique keys, use leftmost (no rightmost or count)
Compiler Generic Specific Performance Loss
gcc-3.4.6 29.50 20.96 40.76%
gcc-4.0.4 28.93 20.90 38.41%
gcc-4.1.2 26.26 22.29 17.80%
gcc-4.2.4 25.49 22.05 15.61%
gcc-4.3.6 26.55 22.25 19.34%
gcc-4.4.7 28.90 22.24 29.92%
gcc-4.5.4 26.85 21.86 22.80%
gcc-4.6.3 22.95 22.06 4.03%
gcc-4.7.1 22.56 21.48 5.01%
Unique keys, use leftmost, rightmost and count
Compiler Generic Specific Performance Loss
gcc-3.4.6 29.48 20.91 40.97%
gcc-4.0.4 29.37 21.72 35.20%
gcc-4.1.2 25.25 23.10 9.29%
gcc-4.2.4 26.17 22.35 17.13%
gcc-4.3.6 26.34 22.30 18.10%
gcc-4.4.7 25.24 22.43 12.51%
gcc-4.5.4 25.58 23.07 10.89%
gcc-4.6.3 21.79 23.50 -7.29%
gcc-4.7.1 23.27 25.08 -7.22%
I've done an analysis of the gcc 4.7.1-generated code and discovered the
following flaws in the generic insert function.
1. Key of inserted object being read repeatedly. Instead of reading the value
of the inserted key once, at the start of the function, the key is read
prior to each comparision. I'm guessing that this is because optimizer
makes the faulty assumption that the value could change throughout the
course of execution. This costs us one extra instruction each iteration of
the loop as we search the tree (32-bit key).
mov 0x18(%rax),%edx
A work-around is in place to eliminate this problem on gcc 4.6.0 and later
if your key size is 16, 32 or 64 bits, which manages to get gcc to store
the key of the supplied object in a regsiter at the start of the function
preventing us a performance loss of roughly 4%.
2. Due to gcc bug 3507 (http://gcc.gnu.org/bugzilla/show_bug.cgi?id=3507),
this code:
long diff = a - b;
if (diff > 0)
do_gt();
else if (diff < 0)
do_lt();
else
do_eq();
Optimizes more poorly than this code:
if (a > b)
do_gt();
else if (b < a)
do_lt();
else
do_eq();
So instead of the key compare happening like this (64-bit key):
cmp 0x18(%rax),%rsi
We get this:
mov %rsi,%rdx
sub 0x18(%rax),%rdx
cmp $0x0,%rdx
The results can be slightly worse when the key type isn't the same as long.
With a signed 32-bit key (s32) on x86_64, gcc thinks it needs to convert
the difference to a 64-bit long.
mov %esi,%edx
sub 0x18(%rax),%edx
movslq %edx,%rdx
cmp $0x0,%rdx
Not only is this 2-3 extra instruction, it also uses one extra register,
which in turn forces gcc to use an r8-15 register in other places, which
requires larger opcodes. Also, this only occurs when using the normal
compare function (doesn't occur when using 'greater'). So this affects
inserts on trees with unique keys and all lookups.
Q&A
===
Q: Why did you add BUILD_BUG_ON_NON_CONST() and
BUILD_BUG_ON_NON_CONST42()?
A: There were initially enough BUILD_BUG_ON(!__builtin_constant_p(arg))
calls to warrant it having a macro for it. However, I've since
discovered that using __builtin_constant_p on a struct member did not
behave very consistently, so after writing some test programs &
scripts, and refining 200k+ test results, I graphed out basically
where __builtin_constant_p() worked and didn't. As it turns out,
using it on struct members is fragile until gcc 4.2, so
BUILD_BUG_ON_NON_CONST42() is intended for use with struct members.
Q: Why empty parameters?
What is IFF_EMPTY() for?
Why don't you just pass zero instead of an empty parameter?
A: Support for caching the left- & right-most nodes in the tree as well
as maintaining a count variable are all optional. Passing the offset
value directly not only means more characters of code to use the
RB_RELATIONSHIP and RB_DEFINE_INTERFACE macros (because now you'll
have to invoke the offsetof macro, supplying your struct types
again), but the offset may actually be zero, so passing zero as "I'm
not using this feature" wont work. (This is the reason why the flags
RB_HAS_LEFTMOST, et. al. exist.) Thus, you would also need to
manually pass the appropriate rb_flag value to specify that you're
using the feature. All of this means more copy, paste & edit code
that is error-prone and a maintenance nightmare. This implementation
allows the caller to pass the name of the struct member or leave the
parameter empty to mean "I'm not using this feature", thus
eliminating all of these other complications.
Q: Using huge macro like RB_DEFINE_INTERFACE prone to usage errors that
create crappy error messages and have zero type-safety. (not really a
question)
A: True. However, much of this is mitigated by creating an
__rb_sanity_check_##name function that is never called, but will
generate meaningful error messages for most mistakes (incorrect
struct member types, etc.)
Q: The traditional boolean comparitor passed to for sorted sets is a less_than
function, why are you using 'greater than'?
A: This decision is purely for optimization purposes, as compare and
greather_than are interchangable when we don't care about equality.
However, this may become a moot point if we can't get gcc to properly
optimize code using the compare function, and switch to a pair of
equals/less functions.
Revision History
===============
New in v6:
o Rebased onto linux-next.
o Removed augmented suport which is now redundant. This should give us a
little speed back on those older compilers.
o Renamed CONFIG_RBTREE* to CONFIG_GRBTREE* to avoid conflicts with Michel
Lespinasse's new code.
o Added support to selftests for a payload on the test objects to see how
performance changes as the size of the objects grow.
o Various other enhancements to test code & scripts.
o Dumped my email provider for silently refusing to send to more than 20
recipients (but sending emails 20 random recipients anyway).
New in v5:
o Added a ability to specify a different compare function for inserts. This
is more efficient on trees with duplicate keys, since you can use a boolean
"greater than" function.
o Added an optimization to generate better code where key size is 16, 32 or 64
bits.
o Add test & validation framework (CONFIG_DEBUG_RBTREE and
CONFIG_DEBUG_RBTREE_VALIDATE)
o Fixed bugs in kernel-doc so that API documentation generates correctly.
o Add userspace test program & scripts.
o Fixed a lot of typos
o Cleaned up and completed kernel-doc comments
New in v4:
o Added type-safe wrapper functions for rb_{next,prev,first,last}
to RB_DEFINE_INTERFACE. Naming is the same as other type-safe
functions (e.g., prefix##_first wraps rb_first). (thanks Pavel Pisa
for the suggestion)
o Added rb_find_{first,next,last,prev} (for non-unique trees) to find
the first or last occurrence of a key and iterate through them.
Type-safe wrapper functions also added to RB_DEFINE_INTERFACE. (thanks
again Pavel Pisa)
o Added support for an unsigned long count member of the container
struct that will be updated upon insertions & deletions.
o Improve sanity checks performed by RB_DEFINE_INTERFACE -- error
messages are now more specific and clearer. Type safety for compare
function is now enforced.
o Completed implementation of insert_near (still untested).
o Completed testing for find_near. Performance is something like
O(log distance * 2 + 1), so if your start node is a bit closer than
half way across the tree, find_near will be about the same speed as
find. If it is further, it will be slower. Either way, it is larger
than a normal find (which should be taken into account), so should
only be used when you are fairly certain your target objects is near
the start.
o Added support for specifying modifiers for functions generated by
RB_DEFINE_INTERFACE. This adds 4 more parameters, but is probably
better than forcing the user to write their own wrapper functions to
macro-generated wrapper functions, just to change their function
attributes.
o Added run-time versions of all of the __rb_xxx_to_xxx inline
functions, for use in those conditions where someone may actually need
to access these using a run-time struct rb_relatinoship value.
o Performed compile tests on gcc 3.4.6 - 4.7.0 and tweaked BUILD_BUG_ON*
macros to not fail on any of these compilers.
New in v3:
o Moved compare & augment functions back into struct rb_relationship
after discovering that calling them will be inlined in gcc 4.6+ if the
function is flattened.
o Improved doc comments.
o Solved problem of compare function not being checked for
type-correctness by adding a __sanity_check_##name() function to
__RB_DEFINE_INTERFACE that generates usable errors when there's a type
or member name problem in the macro parameters. This is helpful since
the errors produced when the RB_RELATIONSHIP macro expands were quite
terrible.
New in v2:
o Added RB_RELATIONSHIP macro (thanks Peter Zijlstra for the
suggestions).
o Added RB_DEFINE_INTERFACE macro.
--
To unsubscribe from this list: send the line "unsubscribe linux-doc" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
