Monday, 29 August 2011

Finding unwanted global variables and functions with dwarves

The dwarves package contains a set of useful tools that use the DWARF information placed in the ELF binaries by the compiler. Utilities in the dwarves package include:

pahole: This will find alignment holes in structs and classes in languages such as C/C++.  With careful repacking one can achieve better cache hits.  I could have done with this when optimising some code a few years back...

codiff:  This is a diff like tool use to compare the effect a change in source code can create on the compiled code.

pfunct:  This displays information about functions, inlines, goto labels, function size, number of variables and much more.

pdwtags: A DWARF information pretty-printer.

pglobal:  Dump out global symbols.

prefcnt:  A DWARF tags usage count.

dtagnames: Will lists tag names.

So, using pglobal, I was able to quickly check which variables I had made global (or accidentally not made them static!) on some code that I was developing as follows:

pglobal -v progname

and the same for functions:

pglobal -f progname

Easy!  Obviously these tools only work if the DWARF information is not stripped out.

All in all, these are really useful tools to know about and will help me in producing better code in the future.

Wednesday, 24 August 2011

Firmware Test Suite full documentation.

I've now competed the documentation of the Firmware Test Suite and this include documenting each of the 50+ tests (which was a bit of a typing marathon).  Each test has a brief description of what the test covers, example output from the test, how to run the test (possibly with different options) and explanations of test failure messages.

For example of the per-test documentation, check out the the suspend/resume test page and the ACPI tables test page.

I hope this is useful!

Monday, 22 August 2011

Assume Nothing.

When I was a very junior software engineer working on Fortran 77 signal processing modules on MicroVaxes, PDP-11s, Masscomps and old 286 PCs I was given some very wise words by the owner of the company:   "Assume nothing".   This has stuck with me for nearly quarter of a century.  It is pithy, easy to remember and is so true for software engineering.

1. "Assume nothing" makes me look up details when I'm not 100% sure.
2. "Assume nothing" means that I double check my facts when I think I'm 100% sure.
3. "Assume nothing" makes me question even the so called 'obvious'.  "Of course it will work.." turns into "are we sure it will work for every possible case?"
4. "Assume nothing" makes me dot the i's and cross the t's.
5. "Assume nothing" keeps me sceptical, which is useful as there is a lot of stupidity masquerading as knowledge on the internet.

I could ramble on. However enough said. Just assume nothing, it will keep you out of a lot of trouble.

How x86 computers boot up

Gustavo Duarte has written a concise and very readable article describing how computers boot up.  Well worth reading.

Friday, 19 August 2011

Which version of the GCC compiled my program?

Here's a quick one-liner to find out which version of GCC was used to compile some code:

readelf -p .comment a.out

String dump of section '.comment':
  [     0]  GCC: (Ubuntu/Linaro 4.6.1-7ubuntu1) 4.6.1
  [    2a]  GCC: (Ubuntu/Linaro 4.6.1-6ubuntu6) 4.6.1

Thursday, 18 August 2011

Debugging S3 suspend/resume using SystemTap and minimodem.

Some problems are a little challenging to debug and require sometimes a bit of lateral thinking to solve.   One particular issue is when suspend/resume locks up and one has no idea where or why because the console has is suspended and any debug messages just don't appear.

In the past I've had to use techniques like flashing keyboard LEDs, making the PC speaker beep or even forcible rebooting the machine at known points to be able to get some idea of roughly where a hang has occurred.   This is fine, but it is tedious since we can only emit a few bits of state per iteration.   Saving state is difficult since when a machine locks up one has to reboot it and one looses debug state.   One technique is to squirrel away debug state in the real time clock (RTC) which allows one to store twenty or so bits of state, which is still quite tough going.

One project I've been working on is to use the power of system tap to instrument the entire suspend/resume code paths - every time a function is entered a hash of the name is generated and stored in the RTC.  If the machine hangs, one can then grab this hash out of the RTC can compare this to the known function names in /proc/kallsyms, and hopefully this will give some idea of where we got to before the machine hung.

However, what would be really useful is the ability to print out more debug state during suspend/resume in real time.   Normally I approach this by using a USB/serial cable and capturing console messages via this mechanism.  However, once USB is suspended, this provides no more information.

One solution I'm now using is with Kamal Mostafa's minimodem.  This wonderful tool is an implementation of a software modem and can send and receive data by emulating a Bell-type or RTTY FSK modem.  It allows me to transmit characters at 110 to 300 baud over a standard PC speaker and reliably receive them on a host machine.  If the wind is in the right direction, one can transmit at higher speeds with an audio cable plugged in the headphone jack of the transmitter and into the microphone socket on the receiver if hardware allows.

The 8254 Programmable Interval-timer on a PC can be used to generate a square wave at a predefined frequency and can be connected to the PC speaker to emit a beep.  Sending data using the speaker to minimodem is a case of sending a 500ms leader tone, then emitting characters.  Each character has a 1 baud space tone, followed by 8 bits (least significant bit first) with a zero being a 1 baud space tone and a 1 being represented by a 1 baud mark tone, and the a trailing bunch of stop bits.

So using a prototype driver written by Kamal, I tweaked the code and put it into my suspend/resume SystemTap script and now I can dump out messages over the PC speaker and decode them using minimodem.  300 baud may not be speedy, but I am able to now instrument and trace through the entire suspend/resume path.

The SystemTap scripts are "work-in-progress" (i.e. if it breaks you keep the pieces), but can be found in my pmdebug git repo git://  The README file gives a quick run down of how to use this script and I have written up a full set of instructions.

The caveat to this is that one requires a PC where one can beep the PC speaker using the PIT.  Lots of modern machines seem to either have this disabled, or the volume somehow under the control of the Intel HDA audio driver.  Anyhow, kudos to Kamal for providing minimodem and giving me the prototype kernel driver to allow me to plug this into a SystemTap scrip.

Wednesday, 17 August 2011

Fragmentation on ext2/ext3/ext4 filesystems

If you want know how fragmented a file is on an ext2/ext3/ext4 filesystem there are a couple of methods of finding out.  One method is to use hdparm, but one needs CAP_SYS_RAWIO capability to do so, hence run with sudo:

sudo hdparm --fibmap /boot/initrd.img-2.6.38-9-generic

 filesystem blocksize 4096, begins at LBA 24000512; assuming 512 byte sectors.
 byte_offset  begin_LBA    end_LBA    sectors
           0   35747840   35764223      16384
     8388608   39942144   39958527      16384
    16777216   40954664   40957423       2760

Alternatively, one can use the filefrag utility (part of the e2fsprogs package) for reporting the number of extents:

filefrag /boot/initrd.img-2.6.38-9-generic
/boot/initrd.img-2.6.38-9-generic: 3 extents found

..or more verbosely:

filefrag -v /boot/initrd.img-2.6.38-9-generic
Filesystem type is: ef53
File size of /boot/initrd.img-2.6.38-9-generic is 18189027 (4441 blocks, blocksize 4096)
 ext logical physical expected length flags
   0       0  1468416            2048
   1    2048  1992704  1470463   2048
   2    4096  2119269  1994751    345 eof
/boot/initrd.img-2.6.38-9-generic: 3 extents found

Well, that's useful. So, going one step further, how many free extents are available on the filesystem? Well, e2freefrag is the tool for this:

sudo e2freefrag /dev/sda1
Device: /dev/sda1
Blocksize: 4096 bytes
Total blocks: 2999808
Free blocks: 1669815 (55.7%)

Min. free extent: 4 KB
Max. free extent: 1370924 KB
Avg. free extent: 6160 KB
Num. free extent: 1084

Extent Size Range :  Free extents   Free Blocks  Percent
    4K...    8K-  :           450           450    0.03%
    8K...   16K-  :            97           227    0.01%
   16K...   32K-  :            99           527    0.03%
   32K...   64K-  :           134          1475    0.09%
   64K...  128K-  :            96          2193    0.13%
  128K...  256K-  :            50          2235    0.13%
  256K...  512K-  :            30          2643    0.16%
  512K... 1024K-  :            36          6523    0.39%
    1M...    2M-  :            36         13125    0.79%
    2M...    4M-  :            14          9197    0.55%
    4M...    8M-  :            15         18841    1.13%
    8M...   16M-  :             8         17515    1.05%
   16M...   32M-  :             7         38276    2.29%
   32M...   64M-  :             3         39409    2.36%
   64M...  128M-  :             3         67865    4.06%
  512M... 1024M-  :             4        802922   48.08%
    1G...    2G-  :             2        646392   38.71%

..thus reminding me to do some housekeeping and remove some junk from my file system... :-)

Saturday, 13 August 2011

What's new in the Firmware Test Suite for Oneiric 11.10?

The Firmware Test Suite (fwts) is still a relatively new tool and hence this cycle I've still been adding some features and fixing bugs.  I've been running fwts against large data sets to soak test the tool to catch a lot of stupid corner cases (e.g. broken ACPI tables). Also, I am focused on getting some better documentation written (this is still "work in progress").

New tests for the Oneiric 11.10 release are as follows:

    Sanity check tables against the MultiProcessor Specification (MPS). For more information about MPS, see the wikipedia MPS page.

    Dump annotated MPS tables.

    Sanity check Model Specific Registers across all CPUs. Does some form of MSR default field sanity checking.

    Very simple suspend (S3) power checks.  This puts the machine into suspend and attempts to measure the power consumed while suspended. Generally this test gives more accurate results the longer one suspends the machine.  Your mileage may vary on this test.

     Hex dump of the Extended BIOS Data Area.

In addition to the above, the fwts "method" test is now expanded to evaluate and exercise over 90 ACPI objects and methods.

One can also join the fwts mailing list by going to the project page and subscribing.