Efficient Memory Safety for TinyOS:
http://www.cs.utah.edu/~coop/me/pubs/sensys07.htm
Safe TinyOS
原文:http://tinyos.stanford.edu/tinyos-wiki/index.php/Safe_TinyOS
What is Safe TinyOS?
Safe TinyOS is a modified compilation toolchain, released as part of TinyOS 2.1, that uses the Deputy compiler (now part of the Ivy project) to enforce type and memory safety at run time on motes. Together, type and memory safety give Java-like guarantees to TinyOS programs: pointer and array errors are trapped before they occur.
Untrapped memory errors are particularly harmful in sensor network applications because:
- Lacking memory protection hardware and a user/kernel boundary, it is easy to accidentally corrupt OS code; these errors are never trapped as segmentation violations.
- Since memory objects are tightly packed, walking even one element past the end of an array is likely to corrupt RAM that your application relies on.
- Since deployed motes are difficult to debug, and since other failure modes exist (concurrency errors, battery outage, connectivity failure, etc.) it can be very hard to tell when memory corruption is the root cause of a serious deployment problem.
- They might permit someone to own your motes
Safe TinyOS is described in detail in a paper from SenSys 2007: Efficient Memory Safety for TinyOS.
Supported Platforms
Safe compilation is supported by default by TinyOS 2.1 for these platforms:
- Mica2
- MicaZ
- TelosB
Building a Safe Application
Getting started is easy:
cd $TOSROOT/apps/Blink make micaz safe
Since Blink contains no known safety errors, the safe version of Blink should act just like its unsafe counterpart.
As a sanity check, you may want to make sure that the safe version of Blink has a larger code size than the unsafe version:
$ make micaz mkdir -p build/micaz compiling BlinkAppC to a micaz binary ncc -o build/micaz/main.exe .... compiled BlinkAppC to build/micaz/main.exe 2052 bytes in ROM 51 bytes in RAM avr-objcopy --output-target=srec build/micaz/main.exe build/micaz/main.srec avr-objcopy --output-target=ihex build/micaz/main.exe build/micaz/main.ihex writing TOS image $ make micaz safe mkdir -p build/micaz compiling BlinkAppC to a micaz binary ncc -o build/micaz/main.exe ... compiled BlinkAppC to build/micaz/main.exe 2562 bytes in ROM 51 bytes in RAM avr-objcopy --output-target=srec build/micaz/main.exe build/micaz/main.srec avr-objcopy --output-target=ihex build/micaz/main.exe build/micaz/main.ihex writing TOS image
The ROM overhead is due to bounds-checking code inserted by the Deputy compiler. Safe TinyOS does not have any RAM overhead.
Interpreting Safety Violations
TinyOS 2.1 comes with an application that is designed to fail; it is here:
$TOSROOT/apps/tutorials/BlinkFail
Looking at BlinkFailC.nc, it is easy to see that the 11th time the Timer1.fired() is signaled, it will access out of bounds storage.
First, run the unsafe version of this application:
make micaz install
In general, it is hard to predict the consequences of a memory safety violation in a TinyOS program. In many cases, the observable symptom is node crash. That is what should happen after running the unsafe BlinkFail for a few seconds.
Next run the safe version:
make micaz safe install
Prior to accessing a[10], the Safe TinyOS failure handler is invoked, resulting in lots of LED activity. The purpose of this activity is to give you a "FLID" (fault location identifier) indicating the source-code location of the faulting access.
A FLID is a sequence of 8 base-4 digits. To read a FLID, wait for the mote's LEDs to "roll," or rapidly blink several times in a 1-2-3 sequence. Next, some number of LEDs will be lit: write down this number. Usually the first digit or two of a FLID will be zero. Following each digit of the FLID, all three LEDs will be very briefly illuminated to serve as a separator. After all 8 digits have been presented, the LEDs will again roll and then the FLID is repeated. Once you have the sequence of 8 digits, it can be decoded as follows:
tos-decode-flid flid-file flid
The flid-file argument specifies a file that helps the script map from the flid to an error message. Building a safe application causes this file to be placed here:
build/[platform]/flids.txt
For example, for the BlinkFail application we get:
[regehr@babel BlinkFail]$ tos-decode-flid ./build/micaz/flids.txt 00001020 Deputy error message for flid 0x0048: BlinkFailC__a <= BlinkFailC__a + BlinkFailC__i++ + 1 (with no overflow): BlinkFailC.nc:70: Assertion failed in CPtrArithAccess: BlinkFailC__a + BlinkFailC__i++ + 1 <= BlinkFailC__a + 10 (with no overflow)
This confirms what we see by inspecting BlinkFailC.nc: line 70 is a safety problem.
Safety Violations in a Simulator
A simulator offers a pleasant alternative to debugging safety violations on real hardware. To try this out, first download and install the CVS version of Avrora.
Next: build a safe version of BlinkFail, make a copy with the .elf extension, and run it in Avrora:
cd $TOSROOT/apps/tutorials/BlinkFail make micaz safe cp build/micaz/main.exe main.elf java avrora.Main -simulation=sensor-network -platform=micaz -monitors=leds,break,sleep,interrupts main.elf
The Avrora output should contain text like this:
0 54435382 break instruction @ 0x0B9E, r30:r31 = 0x0048 0 54435382 @ deputy_fail_noreturn_fast 0 54435382 @ deputy_fail_mayreturn 0 54435382 @ VirtualizeTimerC$0$fireTimers 0 54435382 @ AlarmToTimerC$0$fired$runTask
The callstack leading up to the safety violation is shown. The value loaded into the register pair r30:r31 (the AVR's Z register) is the FLID.
Understanding Safe Code
Safe code contains annotations that express invariants in such a way that the Deputy compiler can perform appropriate static and dynamic checking. Most annotations appear in nesC interface files and in global variable declarations, but sometimes annotations are needed inside functions.
Annotations that you will find in TinyOS code are:
-
ONE
- A pointer that always refers to a single object, similar to a C++ reference.
-
ONE_NOK
- Same as ONE but may be NULL.
-
COUNT(n)
- A pointer that always refers to a block of at least n objects.
-
COUNT_NOK(n)
- Same as COUNT but may be NULL.
-
BND(n,m)
- A pointer p such that n≤p<m, and that is aligned with respect to n and m.
-
BND_NOK(n,m)
- Same as BND but may be NULL.
-
TCAST(type,expr)
- A trusted cast, which tells Deputy to just trust the programmer. This is needed to perform casts that are safe, but are beyond the reach of Deputy's type system. Pointer casts in the network stack (e.g. getHeader() and getFooter() code) tend to fall into this category.
-
TRUSTEDBLOCK
- Code that is completely trusted (i.e., ignored by Deputy). This is used in very few places, and should be avoided when possible.
The quick reference and user manual for Deputy are the best starting points for learning to use its type system, but note that the Safe TinyOS annotations are slightly different from the original Deputy ones.
Writing Safe Code
Writing safe code is generally very easy: in the common case annotations are only needed in declarations for pointers. The best way to get started is to compile regular nesC code in safe mode and see what warnings/errors are emitted by the Deputy compiler. Then, start annotating the code until it compiles cleanly and runs properly. In a small minority of cases, it is necessary to do some refactoring.
Important: To get safe compilation, modules must be marked with nesC's @safe() attribute (see #Mixing Safe and Unsafe Code below). Modules lacking this annotation are trusted code: they are compiled without safety checks.
The non-null annotations should be used whenever possible, for two reasons. First, they serve as good documentation. Second, they generally result in more efficient code since dereferences of these pointers need not be preceded by a NULL check.
The quick reference and user manual for Deputy are the best starting points for learning to use its type system, but note that the Safe TinyOS annotations are slightly different from the original Deputy ones.
Safe Code Examples
To declare msg as a variable of type "pointer to message_t" that always points to a valid message_t:
message_t* ONE msg = ...;
This pointer is never NULL and pointer arithmetic cannot be performed on it.
To declare uartQueue as an array containing UART_QUEUE_LEN elements, where the type of each element is "pointer to message_t" and each element of the array must at all times either be NULL or refer to a valid message_t:
message_t* ONE_NOK uartQueue[UART_QUEUE_LEN];
This pointer can be used in array expressions as follows:
p = uartQueue[-1]; // illegal p = uartQueue[0]; // legal p = uartQueue[1]; // legal p = uartQueue[UART_QUEUE_LEN-1]; // legal p = uartQueue[UART_QUEUE_LEN]; // illegal
To declare reqBuf as a pointer that is at all times either NULL or pointing to a valid block of at least reqBytes uint8_ts:
uint8_t * COUNT_NOK(reqBytes) reqBuf;
To declare pos as a pointer to uint16_t that is always either NULL or refers to a valid uint16_t inside the block of memory delimited by buffer and buffer+count:
norace uint16_t * BND_NOK(buffer, buffer+count) pos;
Multiply indirect pointers (rarely seen in TinyOS) require annotations at each level of indirection. For example, to declare pp as a pointer to pointer to integer such that both pp and anything it points to are valid at all times:
int * ONE * ONE pp = ...;
The TCAST macro evaluates its second argument and turns it into a pointer of the type of its first argument without any safety checks:
cc2420_header_t* ONE x = TCAST(cc2420_header_t* ONE, (uint8_t *)msg + offsetof(message_t, data) - sizeof( cc2420_header_t ));
Understanding Interface Annotations
The TinyOS Core Working Group decided that it was important to avoid hurting the readability of the primary TinyOS interfaces by cluttering them up with safety annotations. Therefore, the Safe TinyOS annotations in the interfaces use a different syntax than the one generally used for code in modules. To see how this works, consider this example from $TOSDIR/interfaces/Packet.nc:
/** * Return a pointer to a protocol's payload region in a packet. * If the caller intends to write to the payload region then * the <tt>len</tt> parameter must reflect the maximum required * length. If the caller (only) wants to read from the payload * region, then <tt>len</tt> may be set to the value of * payloadLength(). If the payload region is smaller than * <tt>len</tt> this command returns NULL. The offset where * the payload region starts within a packet is fixed, i.e. for * a given <tt>msg</tt> this command will always return the same * pointer or NULL. * * @param 'message_t* ONE msg' the packet * @param len the length of payload required * @return 'void* COUNT_NOK(len)' a pointer to the packet's data payload for this layer * or NULL if <tt>len</tt> is too big */ command void* getPayload(message_t* msg, uint8_t len);
Instead of inlining the Safe TinyOS annotations into the getPayload() prototype, the annotations are placed in the nesdoc comment. The syntax above is equivalent to
command void* COUNT_NOK(len) getPayload(message_t* ONE msg, uint8_t len);
It would be prudent to avoid mixing the two kinds of annotation syntax within a single function.
The current convention (which is not automatically enforced) is to use the nesdoc annotation syntax in tos/interfaces and the inline syntax elsewhere.
Mixing Safe and Unsafe Code
By default, a TinyOS module is compiled as unsafe code. In other words, it is trusted code and Deputy will not insert any checks into it. The TinyOS Core Working Group decided on unsafe as the default compilation mode in order to make it easier to transition applications to Safe TinyOS. This default can be overridden using the nesC compiler command line option -fnesc-default-safe.
A module with the @safe() attribute is compiled in safe mode: Deputy checks are inserted. For example in tos/system/SimpleArbiterP.nc starts with:
generic module SimpleArbiterP() @safe() {
A module with the @unsafe() attribute is explicitly exempt from Deputy's checking.
In general, the following conventions should be respected:
- Use @safe() for any module that has been "Deputized," meaning that it contains sufficient annotations to compile and execute properly in safe mode
- Use @unsafe() for any module that fundamentally cannot be compiled in safe mode (at the time this is written, no such modules exist)
- Leave out annotations for modules that have not yet been Deputized
When compiling without Safe TinyOS (e.g. make micaz), safety annotations are ignored.
An application that contains no unsafe modules is globally safe: any safety error that can be caught by Deputy, will be caught. The easiest way to check if an application that you are using or writing is globally safe is to inspect its app.c forDEPUTY_TRUSTEDBLOCK annotations, which are inserted by the nesC compiler to tell Deputy to avoid compiling functions from unsafe modules as safe code. For example:
[regehr@babel Blink]$ cd $TOSROOT/apps/Blink [regehr@babel Blink]$ make micaz safe mkdir -p build/micaz compiling BlinkAppC to a micaz binary ncc -o build/micaz/main.exe -DSAFE_TINYOS ... compiled BlinkAppC to build/micaz/main.exe 2562 bytes in ROM 51 bytes in RAM avr-objcopy --output-target=srec build/micaz/main.exe build/micaz/main.srec avr-objcopy --output-target=ihex build/micaz/main.exe build/micaz/main.ihex writing TOS image [regehr@babel Blink]$ grep -c DEPUTY_TRUSTEDBLOCK build/micaz/app.c 14
This shows that the app.c for Blink on MicaZ contains 14 DEPUTY_TRUSTEDBLOCK annotations. However, a detailed inspection shows that almost all of this trusted code is in the failure handlers themselves. It makes no sense to compile these as safe code.
Customizing the Failure Handler
The default safety violation handler is here:
$TOSDIR/lib/safe/SafeFailureHandlerP.nc
This code could be changed to notify a base station that a safety violation occurred, log a FLID to flash, reboot the mote, or whatever you like. Ideally, failure handling code should be atomic, should be independent of the state of the mote's RAM and hardware devices at the time a violation happens, and should not call into any safe code. Thus, it is highly unlikely that you could successfully send a packet reporting a fault that occurred in the TinyOS network stack without having written a small, customized network driver whose specific purpose is to report safety errors in the main TinyOS network stack.
Feedback
Please send problem reports and other feedback to tinyos-help or tinyos-devel, as appropriate.
To Do
For this document:
- Factor some of this information out into a TEP
- Explain how to find and understand the Deputy output
- Split out stuff for 2.1 vs. CVS HEAD (where there's a difference)
- Show examples of how to refactor code to be Deputy-friendly
For Safe TinyOS in general:
- Support intelmote2, shimmer, and iris
- Commit the Utah patch for safe TOSSIM
- Although both the Deputy compiler and gcc both attempt to remove unneeded safety checks, there is still plenty of room for improvement. We'd like advanced compiler techniques (e.g. calling out to a modern decision procedure) for showing that safety checks definitely do or do not fire
- Write an Avrora monitor that goes ahead and automatically decodes the FLID
- Write an mspsim hack that does the same
- Add a nesC compiler option that prints a list of unsafe modules linked into an app
- Integrate Safe TinyOS with a stack depth analysis tool to avoid unsafety through stack overflow
- Solve the problem of unsafe accesses through pointers to dead stack frames
- Eliminate as much trusted code (TCAST and TRUSTEDBLOCK) as possible
Let us know if you'd like to help.