Re: [klee-dev] Working with fixed memory locations.
Hi Daniel, Thanks a lot for your suggestion. I've tried it and it solved the issue. However, it moved it down to the solvers. Now it is STP and Z3 who get their stacks blown up. See the backtrace here: https://pastebin.com/raw/WE2M0bJf I also set `ulimit -s unlimited` and now they all seem to be running. This is what I am observing: For the Z3 solver, KLEE has been running for ~25hs now and kept a steady memory usage of ~35GiB. I don't know if it's making progress or if it is stuck. Launch params: *--optimize-array=all --simplify-sym-indices --solver-backend=z3* For the STP solver, KLEE has been running for ~22hs (launch params: *--optimize-array=all --solver-backend=stp*), and I now realize that the behavior I described in the previous email was due to the *-use-forked-server *default option. What I am seeing is that new processes get spawned and consume progressively more memory. Right now the child process runs for 50 minutes and consumes 155GiB of memory. Here are some questions: (1) Does it make sense to disable the forked server that wouldn't change anything (in terms of speed and resource usage). (2) Is there a way to see how these solvers are making progress? (3) I ran the same experiment, but with *klee_assume(input[0] == 'x'), input[i] = solution[i]* (i.e: setting an assumption on one byte of the input and making the rest concrete). I expected this to finish sooner with an answer, but it seems like it doesn't. Does this indicate that there's something else symbolic besides my *klee_make_symbolic* call? Best Regards, Marco On Sun, Jun 19, 2022 at 11:04 AM Daniel Schemmel wrote: > Hi Marco, > > I had a look at your stack trace, and the crash might have been caused by > KLEE when building the solver query before passing it to the actual solver. > Could you check out https://github.com/klee/klee/pull/1523 and see if > that changes anything for you? > > Best, > Daniel > On 2022-06-18 01:12, Marco Vanotti wrote: > > Hi Cristian, > > Thanks for your answer. I have tried *--optimize-array=all*, but that > didn't fix the problem :(. > > It would be a better user experience to get an error message instead of a > segfault. In any case, if this is stopping because it's running out of > memory, is there a way to remove that restriction? My server still had a > few GiB to spare . > > > On Fri, Jun 17, 2022 at 2:53 AM Cristian Cadar > wrote: > >> Hi Marco, you seem to be reaching an issue with the solver, which is >> having trouble reasoning about the huge symbolic array (requiring >> excessive time and memory). You should try to shrink that array if >> possible. You can also try --optimize-array=all, but it might not help >> in your case. >> >> Best, >> Cristian >> >> On 17/06/2022 05:02, Marco Vanotti wrote: >> > After letting it run for a few hours I've observed that klee spawns a >> > subprocess that keeps growing on memory until it reaches ~100GiB and >> > then it stops and restarts again. >> > Nothing is being printed indicating an error, but I'm not sure if the >> > behavior is normal. This is with KLEE from the docker container. >> > >> > I've tried building KLEE from source, both with STP and Z3 support, and >> > running my program makes it crash with a segfault :( >> > >> > Here is the backtrace for the crash with the STP solver: >> > https://pastebin.com/raw/xpf9D9VD <https://pastebin.com/raw/xpf9D9VD> >> > >> > Best Regards, >> > Marco >> > >> > On Thu, Jun 16, 2022 at 3:48 PM Marco Vanotti > > <mailto:mvano...@dc.uba.ar>> wrote: >> > >> > Hi Martin, Manuel, >> > >> > Thanks for your answer :) ! >> > >> > On Thu, Jun 16, 2022 at 1:19 PM Nowack, Martin >> > mailto:m.now...@imperial.ac.uk>> wrote: >> > >> > Hi Marco, >> > >> > Maybe the following helps you: >> > >> https://github.com/klee/klee/blob/292600cf54d5fd73278f67a4f98c2f955cbdaa10/test/Feature/DefineFixedObject.c >> > < >> https://github.com/klee/klee/blob/292600cf54d5fd73278f67a4f98c2f955cbdaa10/test/Feature/DefineFixedObject.c >> > >> > >> > >> > This seems to be what I am looking for, thanks!. I tried using it >> > for small variables and it works. However, for a big object >> > (0x256000 bytes) it shows the following warning: >> > >> > *KLEE: WARNING ONCE*: flushing 2449408 bytes on read, may be slow >> > and/or crash: MO195[2449408] allocated at main(): call void >> > @klee_define_fixed_object(i8* inttoptr (i64 8404992 to i8*), i64 >
Re: [klee-dev] Working with fixed memory locations.
Hi Manuel, On Fri, Jun 17, 2022 at 3:29 AM Carrasco, Manuel G < m.carra...@imperial.ac.uk> wrote: > Hi Marco, > > I wasn't aware of the klee_define_fixed_object intrinsic. > > This is a question on my side regarding KLEE. If an array is always > read/written using constant accesses, are the "getArrayForUpdate" calls > strictly necessary? However, I'm not sure if this description fits your > blob. Is it a constant buffer after its initialization? Is it symbolically > indexed? How is it initialized? > No, the blob is not constant. It represents a virtual machine which holds the instructions, memory and registers. My program is just an emulator. I point it to the entry point inside the memory region and it starts executing. I've heard that KLEE, and symbolic execution in general, are not that great in this scenario. However, the machine that I am interpreting is extremely simple. Best, Marco > Best, > Manuel. > -- > *From:* klee-dev-boun...@imperial.ac.uk > on behalf of Cristian Cadar > *Sent:* 17 June 2022 10:53 > *To:* mvano...@dc.uba.ar ; Nowack, Martin < > m.now...@imperial.ac.uk> > *Cc:* klee-dev > *Subject:* Re: [klee-dev] Working with fixed memory locations. > > Hi Marco, you seem to be reaching an issue with the solver, which is > having trouble reasoning about the huge symbolic array (requiring > excessive time and memory). You should try to shrink that array if > possible. You can also try --optimize-array=all, but it might not help > in your case. > > Best, > Cristian > > On 17/06/2022 05:02, Marco Vanotti wrote: > > After letting it run for a few hours I've observed that klee spawns a > > subprocess that keeps growing on memory until it reaches ~100GiB and > > then it stops and restarts again. > > Nothing is being printed indicating an error, but I'm not sure if the > > behavior is normal. This is with KLEE from the docker container. > > > > I've tried building KLEE from source, both with STP and Z3 support, and > > running my program makes it crash with a segfault :( > > > > Here is the backtrace for the crash with the STP solver: > > https://pastebin.com/raw/xpf9D9VD <https://pastebin.com/raw/xpf9D9VD> > > > > Best Regards, > > Marco > > > > On Thu, Jun 16, 2022 at 3:48 PM Marco Vanotti > <mailto:mvano...@dc.uba.ar >> wrote: > > > > Hi Martin, Manuel, > > > > Thanks for your answer :) ! > > > > On Thu, Jun 16, 2022 at 1:19 PM Nowack, Martin > > mailto:m.now...@imperial.ac.uk > >> wrote: > > > > Hi Marco, > > > > Maybe the following helps you: > > > https://github.com/klee/klee/blob/292600cf54d5fd73278f67a4f98c2f955cbdaa10/test/Feature/DefineFixedObject.c > > < > https://github.com/klee/klee/blob/292600cf54d5fd73278f67a4f98c2f955cbdaa10/test/Feature/DefineFixedObject.c > > > > > > > > This seems to be what I am looking for, thanks!. I tried using it > > for small variables and it works. However, for a big object > > (0x256000 bytes) it shows the following warning: > > > > *KLEE: WARNING ONCE*: flushing 2449408 bytes on read, may be slow > > and/or crash: MO195[2449408] allocated at main(): call void > > @klee_define_fixed_object(i8* inttoptr (i64 8404992 to i8*), i64 > > 2449408), !dbg !171 > > KLEE is still running, so maybe it just means it is slow. > > > > I went with the approach of having my blob as a global variable, and > > then `memcpy` it into the address after calling define_fixed_object. > > > > Best, > > Martin > > > >> On 16. Jun 2022, at 20:43, Carrasco, Manuel G > >> mailto:m.carra...@imperial.ac.uk > >> > >> wrote: > >> > >> Hi Marco! > >> > >> I have a program that when compiled, adds a program header > >> that loads a data blob into a fixed memory location. > >> > >> I'm sorry to ask, but could you explain a bit more how this > >> works? At first glance, I'd say that if any of this happens on > >> a stage later than LLVM-IR, it may be hard to mimic in KLEE. > > > > I have a bunch of files that I add as .incbin into a section, and > > then my linker scripts put them in a fixed address when it links the > > program altogether. I think there is no way this would work with > > LLVM IR. > > > >> > >> As far as I understand, when KLEE
Re: [klee-dev] Working with fixed memory locations.
Hi Cristian, Thanks for your answer. I have tried *--optimize-array=all*, but that didn't fix the problem :(. It would be a better user experience to get an error message instead of a segfault. In any case, if this is stopping because it's running out of memory, is there a way to remove that restriction? My server still had a few GiB to spare . On Fri, Jun 17, 2022 at 2:53 AM Cristian Cadar wrote: > Hi Marco, you seem to be reaching an issue with the solver, which is > having trouble reasoning about the huge symbolic array (requiring > excessive time and memory). You should try to shrink that array if > possible. You can also try --optimize-array=all, but it might not help > in your case. > > Best, > Cristian > > On 17/06/2022 05:02, Marco Vanotti wrote: > > After letting it run for a few hours I've observed that klee spawns a > > subprocess that keeps growing on memory until it reaches ~100GiB and > > then it stops and restarts again. > > Nothing is being printed indicating an error, but I'm not sure if the > > behavior is normal. This is with KLEE from the docker container. > > > > I've tried building KLEE from source, both with STP and Z3 support, and > > running my program makes it crash with a segfault :( > > > > Here is the backtrace for the crash with the STP solver: > > https://pastebin.com/raw/xpf9D9VD <https://pastebin.com/raw/xpf9D9VD> > > > > Best Regards, > > Marco > > > > On Thu, Jun 16, 2022 at 3:48 PM Marco Vanotti > <mailto:mvano...@dc.uba.ar>> wrote: > > > > Hi Martin, Manuel, > > > > Thanks for your answer :) ! > > > > On Thu, Jun 16, 2022 at 1:19 PM Nowack, Martin > > mailto:m.now...@imperial.ac.uk>> wrote: > > > > Hi Marco, > > > > Maybe the following helps you: > > > https://github.com/klee/klee/blob/292600cf54d5fd73278f67a4f98c2f955cbdaa10/test/Feature/DefineFixedObject.c > > < > https://github.com/klee/klee/blob/292600cf54d5fd73278f67a4f98c2f955cbdaa10/test/Feature/DefineFixedObject.c > > > > > > > > This seems to be what I am looking for, thanks!. I tried using it > > for small variables and it works. However, for a big object > > (0x256000 bytes) it shows the following warning: > > > > *KLEE: WARNING ONCE*: flushing 2449408 bytes on read, may be slow > > and/or crash: MO195[2449408] allocated at main(): call void > > @klee_define_fixed_object(i8* inttoptr (i64 8404992 to i8*), i64 > > 2449408), !dbg !171 > > KLEE is still running, so maybe it just means it is slow. > > > > I went with the approach of having my blob as a global variable, and > > then `memcpy` it into the address after calling define_fixed_object. > > > > Best, > > Martin > > > >> On 16. Jun 2022, at 20:43, Carrasco, Manuel G > >> mailto:m.carra...@imperial.ac.uk>> > >> wrote: > >> > >> Hi Marco! > >> > >> I have a program that when compiled, adds a program header > >> that loads a data blob into a fixed memory location. > >> > >> I'm sorry to ask, but could you explain a bit more how this > >> works? At first glance, I'd say that if any of this happens on > >> a stage later than LLVM-IR, it may be hard to mimic in KLEE. > > > > I have a bunch of files that I add as .incbin into a section, and > > then my linker scripts put them in a fixed address when it links the > > program altogether. I think there is no way this would work with > > LLVM IR. > > > >> > >> As far as I understand, when KLEEexecutes a LLVM-IR load > >> instruction > >> < > https://github.com/klee/klee/blob/master/lib/Core/Executor.cpp#L2722>, > >> it will try tofind > >> < > https://github.com/klee/klee/blob/master/lib/Core/Executor.cpp#L4191>the > >> MemoryObjects (more than one if it is a symbolic pointer) that > >> contain the address. Conceptually, you want KLEE to somehow > >> have a MemoryObject at the hardcoded address. > >> > >> One way to go could be modelling this as a LLVM-IR > >> GlobalVariable at your fixed address with the content of your > >> blob. If this makes sense, you may want to check thisfunction > >> < > https://github.com/klee/klee/blob/mas
Re: [klee-dev] Working with fixed memory locations.
After letting it run for a few hours I've observed that klee spawns a subprocess that keeps growing on memory until it reaches ~100GiB and then it stops and restarts again. Nothing is being printed indicating an error, but I'm not sure if the behavior is normal. This is with KLEE from the docker container. I've tried building KLEE from source, both with STP and Z3 support, and running my program makes it crash with a segfault :( Here is the backtrace for the crash with the STP solver: https://pastebin.com/raw/xpf9D9VD Best Regards, Marco On Thu, Jun 16, 2022 at 3:48 PM Marco Vanotti wrote: > Hi Martin, Manuel, > > Thanks for your answer :) ! > > On Thu, Jun 16, 2022 at 1:19 PM Nowack, Martin > wrote: > >> Hi Marco, >> >> Maybe the following helps you: >> >> https://github.com/klee/klee/blob/292600cf54d5fd73278f67a4f98c2f955cbdaa10/test/Feature/DefineFixedObject.c >> > > This seems to be what I am looking for, thanks!. I tried using it for > small variables and it works. However, for a big object (0x256000 bytes) it > shows the following warning: > > *KLEE: WARNING ONCE*: flushing 2449408 bytes on read, may be slow and/or > crash: MO195[2449408] allocated at main(): call void > @klee_define_fixed_object(i8* inttoptr (i64 8404992 to i8*), i64 2449408), > !dbg !171 > > KLEE is still running, so maybe it just means it is slow. > > I went with the approach of having my blob as a global variable, and then > `memcpy` it into the address after calling define_fixed_object. > > Best, >> Martin >> >> On 16. Jun 2022, at 20:43, Carrasco, Manuel G >> wrote: >> >> Hi Marco! >> >> I have a program that when compiled, adds a program header that loads a >> data blob into a fixed memory location. >> >> I'm sorry to ask, but could you explain a bit more how this works? At >> first glance, I'd say that if any of this happens on a stage later than >> LLVM-IR, it may be hard to mimic in KLEE. >> >> I have a bunch of files that I add as .incbin into a section, and then my > linker scripts put them in a fixed address when it links the program > altogether. I think there is no way this would work with LLVM IR. > >> >> As far as I understand, when KLEE executes a LLVM-IR load instruction >> <https://github.com/klee/klee/blob/master/lib/Core/Executor.cpp#L2722>, >> it will try to find >> <https://github.com/klee/klee/blob/master/lib/Core/Executor.cpp#L4191> the >> MemoryObjects (more than one if it is a symbolic pointer) that contain the >> address. Conceptually, you want KLEE to somehow have a MemoryObject at the >> hardcoded address. >> >> One way to go could be modelling this as a LLVM-IR GlobalVariable at your >> fixed address with the content of your blob. If this makes sense, you may >> want to check this function >> <https://github.com/klee/klee/blob/master/lib/Core/Executor.cpp#L648> and >> addExternalObject perhaps as well. >> >> Thanks! This looks interesting, but I am a bit puzzled about how to go > with this. Should I recompile KLEE to add support for my use case? I > checked on the MemoryManager class and it seems like it just allocates > stuff at whatever place is available. > >> >> I apologise if you already know this! >> >> > I did not know any of that :) This is the second time I am using KLEE, and > the first one was a big failure :P > > Thanks! > Marco > > > >> >> Best regards, >> Manuel. >> >> -- >> *From:* klee-dev-boun...@imperial.ac.uk >> on behalf of Marco Vanotti >> *Sent:* 16 June 2022 18:55 >> *To:* klee-dev >> *Subject:* [klee-dev] Working with fixed memory locations. >> >> Hi klee-dev! >> >> I am new to KLEE, and have a question about using it with one of my >> programs. >> >> I have a program that when compiled, adds a program header that loads a >> data blob into a fixed memory location. >> >> This means that my program has this fixed memory location hardcoded all >> around the place (also this blob has references to itself). >> >> I would like to load my program in KLEE to get a better understanding of >> how it works. The problem I am facing is that I have no idea how to make >> KLEE understand that I need this blob mapped in that address. >> >> This are the things I've tried: >> >> * Using wllvm/gclang to get the full program linked together, following >> my link script, then extracting the bc and running that with KLEE. This >> didn't work. KLEE complains that the pointers are invalid. >> >&g
Re: [klee-dev] Working with fixed memory locations.
Hi Martin, Manuel, Thanks for your answer :) ! On Thu, Jun 16, 2022 at 1:19 PM Nowack, Martin wrote: > Hi Marco, > > Maybe the following helps you: > > https://github.com/klee/klee/blob/292600cf54d5fd73278f67a4f98c2f955cbdaa10/test/Feature/DefineFixedObject.c > This seems to be what I am looking for, thanks!. I tried using it for small variables and it works. However, for a big object (0x256000 bytes) it shows the following warning: *KLEE: WARNING ONCE*: flushing 2449408 bytes on read, may be slow and/or crash: MO195[2449408] allocated at main(): call void @klee_define_fixed_object(i8* inttoptr (i64 8404992 to i8*), i64 2449408), !dbg !171 KLEE is still running, so maybe it just means it is slow. I went with the approach of having my blob as a global variable, and then `memcpy` it into the address after calling define_fixed_object. Best, > Martin > > On 16. Jun 2022, at 20:43, Carrasco, Manuel G > wrote: > > Hi Marco! > > I have a program that when compiled, adds a program header that loads a > data blob into a fixed memory location. > > I'm sorry to ask, but could you explain a bit more how this works? At > first glance, I'd say that if any of this happens on a stage later than > LLVM-IR, it may be hard to mimic in KLEE. > > I have a bunch of files that I add as .incbin into a section, and then my linker scripts put them in a fixed address when it links the program altogether. I think there is no way this would work with LLVM IR. > > As far as I understand, when KLEE executes a LLVM-IR load instruction > <https://github.com/klee/klee/blob/master/lib/Core/Executor.cpp#L2722>, > it will try to find > <https://github.com/klee/klee/blob/master/lib/Core/Executor.cpp#L4191> the > MemoryObjects (more than one if it is a symbolic pointer) that contain the > address. Conceptually, you want KLEE to somehow have a MemoryObject at the > hardcoded address. > > One way to go could be modelling this as a LLVM-IR GlobalVariable at your > fixed address with the content of your blob. If this makes sense, you may > want to check this function > <https://github.com/klee/klee/blob/master/lib/Core/Executor.cpp#L648> and > addExternalObject perhaps as well. > > Thanks! This looks interesting, but I am a bit puzzled about how to go with this. Should I recompile KLEE to add support for my use case? I checked on the MemoryManager class and it seems like it just allocates stuff at whatever place is available. > > I apologise if you already know this! > > I did not know any of that :) This is the second time I am using KLEE, and the first one was a big failure :P Thanks! Marco > > Best regards, > Manuel. > > -- > *From:* klee-dev-boun...@imperial.ac.uk > on behalf of Marco Vanotti > *Sent:* 16 June 2022 18:55 > *To:* klee-dev > *Subject:* [klee-dev] Working with fixed memory locations. > > Hi klee-dev! > > I am new to KLEE, and have a question about using it with one of my > programs. > > I have a program that when compiled, adds a program header that loads a > data blob into a fixed memory location. > > This means that my program has this fixed memory location hardcoded all > around the place (also this blob has references to itself). > > I would like to load my program in KLEE to get a better understanding of > how it works. The problem I am facing is that I have no idea how to make > KLEE understand that I need this blob mapped in that address. > > This are the things I've tried: > > * Using wllvm/gclang to get the full program linked together, following my > link script, then extracting the bc and running that with KLEE. This didn't > work. KLEE complains that the pointers are invalid. > > * Manually embedding the blob into my program as an array, then calling > `mmap` with `MAP_FIXED` to map the area that I want and copying over the > blob. > > The issue here is that MAP_FIXED returns EPERM because probably the > address range I am trying to map is already mapped. > > > * Setting the KLEE deterministic allocations to encompass the range that I > care about, then doing a big `malloc` and making sure that my range is > inside that malloc chunk. > > For this last one, I am using flags like: > --allocate-determ --allocate-determ-start-address=8404992 > --allocate-determ-size=3145728 > > One of the things that I see is that KLEE fails to mmap big chunks (in the > order of 100MiB). But even if I decrease the size, I still get failures > when I try to assert things like: > > uintptr_t malloc_addr = (uintptr_t) malloc(malloc_size); > klee_assert(BASE_ADDR >= malloc_addr); > klee_assert(BASE_ADDR < malloc_addr + malloc_size); > > -- > > Something that might be
[klee-dev] Working with fixed memory locations.
Hi klee-dev! I am new to KLEE, and have a question about using it with one of my programs. I have a program that when compiled, adds a program header that loads a data blob into a fixed memory location. This means that my program has this fixed memory location hardcoded all around the place (also this blob has references to itself). I would like to load my program in KLEE to get a better understanding of how it works. The problem I am facing is that I have no idea how to make KLEE understand that I need this blob mapped in that address. This are the things I've tried: * Using wllvm/gclang to get the full program linked together, following my link script, then extracting the bc and running that with KLEE. This didn't work. KLEE complains that the pointers are invalid. * Manually embedding the blob into my program as an array, then calling `mmap` with `MAP_FIXED` to map the area that I want and copying over the blob. The issue here is that MAP_FIXED returns EPERM because probably the address range I am trying to map is already mapped. * Setting the KLEE deterministic allocations to encompass the range that I care about, then doing a big `malloc` and making sure that my range is inside that malloc chunk. For this last one, I am using flags like: --allocate-determ --allocate-determ-start-address=8404992 --allocate-determ-size=3145728 One of the things that I see is that KLEE fails to mmap big chunks (in the order of 100MiB). But even if I decrease the size, I still get failures when I try to assert things like: uintptr_t malloc_addr = (uintptr_t) malloc(malloc_size); klee_assert(BASE_ADDR >= malloc_addr); klee_assert(BASE_ADDR < malloc_addr + malloc_size); -- Something that might be relevant is that in reality I need two of these blobs loaded into different regions of memory, but so far I can't even get to load one. And they are not too far apart from each other, so if, for example, the malloc approach works, I could just increase the size and make the two allocations. One thing that might complicate things, is that these addresses might collide with where KLEE tries to load the program. I don't know how to deal with that either. Any advice on how to tune KLEE for this use case? Best Regards, Marco ___ klee-dev mailing list klee-dev@imperial.ac.uk https://mailman.ic.ac.uk/mailman/listinfo/klee-dev
