Introduction

sparse is a useful tool for static analysis of the Linux kernel.

However, it tends to generate a lot of output. A run on a reasonably small powerpc defconfig gives over 3000 lines of output. This - and the fact that running with make -jNN will give you output in unpredictable order - makes running a diff over the results quite tricky. Furthermore, a code change that moves code around will lead to sparse warnings changing line numbers, meaning diff will double-count them.

This python script aims to make that a bit easier.

Designed specifically with sparse in mind, this script:

• eliminates exact matches
• eliminates warnings that are the same except for a change in line number
• comfortably deals with warnings that are split across multiple lines

  For example:

  drivers/scsi/lpfc/lpfc_scsi.c:5606:30: warning: incorrect type in assignment (different base types)
  drivers/scsi/lpfc/lpfc_scsi.c:5606:30:    expected int [signed] memory_flags
  drivers/scsi/lpfc/lpfc_scsi.c:5606:30:    got restricted gfp_t
      

• prints the output in the unified diff format we all know and love

Usage

./smart-sparse-diff.py <old log> <new log>

Currently no flags are accepted. However, if you want more verbose output, you can edit the python file and set verbose to True.

Example

~/d/smart-sparse-diff> ./smart-sparse-diff.py  pseries_le_defconfig-234.log pseries_le_defconfig-238.log 
-kernel/cgroup.c:3145:9: warning: context imbalance in 'cgroup_lock_and_drain_offline' - wrong count at exit
+kernel/cgroup/cgroup.c:2598:9: warning: context imbalance in 'cgroup_lock_and_drain_offline' - wrong count at exit
-kernel/config_data.h:XX:10: warning: trying to concatenate 20558-character string (8191 bytes max)
+kernel/config_data.h:XX:10: warning: trying to concatenate 20575-character string (8191 bytes max)
+kernel/rcu/tree.c:XX:6: warning: symbol 'rcu_dynticks_curr_cpu_in_eqs' was not declared. Should it be static?
+kernel/rcu/tree_trace.c:335:9: warning: incorrect type in argument 1 (different modifiers) expected struct raw_spinlock [usertype] *lock got struct raw_spinlock [noderef] *<noident>

XX marks where line-insensitive matching has been attempted.

Limitations & Bugs

Limitations based on input limitations

Often you’ll find compiler/linker warnings interspersed with sparse output. Hopefully they’re present in both the old and new files are will be removed in the exact line match phase, but if not you may either get warnings about them or you may find somewhat odd output.

Limitations based on sparse limitations

Sparse also has some trouble with warnings that are interactions between C and header files. For example:

./arch/powerpc/include/asm/spinlock.h:168:9: warning: context imbalance in 'i8259_irq' - unexpected unlock

There’s no i8259_irq in spinlock.h: instead there’s a file that includes that header file and uses something in that file that causes a warning. Sparse isn’t smart enough to blame the correct file here.

Other bugs

There are probably other bugs.

We should pick up cases where the number of occurrences of a warning changes, but I haven’t tested it properly. This is particularly relevant to header file-related warnings.

Requirements, etc

This uses Python 3, and supports mypy static typing.