Re: [PATCH 1/3] CodingStyle updates

From: Randy Dunlap
Date: Fri Sep 28 2007 - 17:48:23 EST

On Fri, 28 Sep 2007 17:32:00 -0400 Erez Zadok wrote:

> 1. Updates chapter 13 (printing kernel messages) to expand on the use of
> pr_debug()/pr_info(), what to avoid, and how to hook your debug code with
> kernel.h.
> 2. New chapter 19, branch prediction optimizations, discusses the whole
> un/likely issue.
> Cc: "Kok, Auke" <auke-jan.h.kok@xxxxxxxxx>
> Cc: Kyle Moffett <mrmacman_g4@xxxxxxx>
> Cc: Jan Engelhardt <jengelh@xxxxxxxxxxxxxxx>
> Cc: Adrian Bunk <bunk@xxxxxxxxxx>
> Cc: roel <12o3l@xxxxxxxxxx>
> Signed-off-by: Erez Zadok <ezk@xxxxxxxxxxxxx>

A few comments below...

Acked-by: Randy Dunlap <randy.dunlap@xxxxxxxxxx>

> ---
> Documentation/CodingStyle | 88 +++++++++++++++++++++++++++++++++++++++++++-
> 1 files changed, 86 insertions(+), 2 deletions(-)
> diff --git a/Documentation/CodingStyle b/Documentation/CodingStyle
> index 7f1730f..00b29e4 100644
> --- a/Documentation/CodingStyle
> +++ b/Documentation/CodingStyle
> @@ -779,6 +797,69 @@ includes markers for indentation and mode configuration. People may use their
> own custom mode, or may have some other magic method for making indentation
> work correctly.
> + Chapter 19: branch prediction optimizations
> +
> +The kernel includes macros called likely() and unlikely(), which can be used
> +as hints to the compiler to optimize branch prediction. They operate by
> +asking gcc to shuffle the code around so that the more favorable outcome
> +executes linearly, avoiding a JMP instruction; this can improve cache
> +pipeline efficiency. For technical details how these macros work, see the
> +References section at the end of this document.
> +
> +An example use of this as as follows:
> +
> + ptr = kmalloc(size, GFP_KERNEL);
> + if (unlikely(!ptr))
> + ...
> +
> +or
> + err = some_function(...);
> + if (likely(!err))
> + ...
> +
> +The main two problems with using un/likely() are that (a) programmers can
> +easily be wrong about their code's likelihood to take one branch
> +vs. another, and (b) on average, gcc will do a much better job optimizing
> +branches that the programmer can. The benefit on some systems for
> +predicting correctly can be in saving a few instructions. But the penalty
> +for wrong use of un/likely() can be very significant (possibly dozens of
> +instructions), as you may be doing a JMP instruction, hurting your pipeline
> +cache, EACH time you get to the branch in question!
> +
> +Therefore, use un/likely() sparingly, consider it primarily for hot paths,
> +use it only when you are certain that the condition in question rarely
> +happens, be sure that it happens with roughly the same probability under
> +most/all user conditions. One rule of thumb suggested is that the
> +probability of the branch un/taken should exceed 99% (although some would
> +consider 95% as well). Of course, beware of silly mistakes such as
> +intending to use likely() and using unlikely() instead.
> +
> +A good example of this is the above kmalloc(GFP_KERNEL) call. The chances
> +of kmalloc() returning NULL are rather small, because even if it doesn't
> +have memory to return to you at the moment, with GFP_KERNEL/__GFP_WAIT
> +passed, kmalloc() will wait and suspend your thread, while it goes off to
> +find some free memory (purging caches, flushing buffers, etc.). In other
> +words, kmalloc() tries very hard to give you the memory you asked for by the
> +time it return.


> +
> +Consider the next, bad example. Suppose you're developing a file system

^comma seems odd here

> +which performs logically different actions on different types of entities:
> +files, directories, symlinks, devices, etc. and you use this code:
> +
> + if (unlikely(S_ISBLK(mode))
> + ...
> +
> +On first glance, the above use of unlikely() seems right. After all, most
> +file system objects are files and directories, and very few of them tend to
> +be block devices. So this optimization should work well, no? Although it's
> +true that it'll work well for most users, what about some user who happens
> +to have a file system with lots of block devices? Or what if the user has
> +only one block device object on the file system, but the user has an
> +application which causes the above conditional to be traversed very
> +frequently (e.g., a shell script that deals with devices)? Such users will
> +be penalized heavily for going [sic] down the wrong path... Therefore, you
> +should consider also whether a seemingly-rare condition is indeed rare ALL
> +the time.
> Appendix I: References
> @@ -804,6 +885,9 @@ language C, URL:
> Kernel CodingStyle, by greg@xxxxxxxxx at OLS 2002:
> +FAQ/LikelyUnlikely:
> +
> +
> --

Phaedrus says that Quality is about caring.
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at
Please read the FAQ at