llvm dev - Mar 2018 - [pre-RFC] Data races in concurrent ThinLTO processes

Hello,

I am sending the following proposal to discuss issues and solutions regarding
data races in concurrent ThinLTO processes.

This caught my attention when we encountered a race condition in ThinLTO with
caching.
While looking into how ThinLTO deals with the problem of cache files
reads/writes/deletes I spotted a lot of problems: some of them are related to
data races, others - to file/resource sharing between different processes. I
wanted to point out these problems and start the discussion about potential
solutions. I would like to get your feedback.


Problem #1: In certain common situations, 'rename' can fail, causing a
data race later on.
Look at the following code in the function 'write' in
ThinLTOCodeGenerator.cpp

std::error_code EC   sys::fs::createTemporaryFile("Thin",
"tmp.o", TempFD, TempFilename);
if (EC) {
  errs() << "Error: " << EC.message() <<
"\n";
  report_fatal_error("ThinLTO: Can't get a temporary file");
}
{
  raw_fd_ostream OS(TempFD, /* ShouldClose */ true);
  OS << OutputBuffer.getBuffer();
}
// Rename to final destination (hopefully race condition won't matter here)
EC = sys::fs::rename(TempFilename, EntryPath);

Compared to a race-prone direct write of a buffer to a file, this code avoids a
data race by first writing a buffer to a temp file and then renaming that temp
file to become the final file in the cache. After r315079, when 'rename'
became more POSIX-compliant, this scheme guarantees atomicity of writing a file
to the cache,
(except on Wine, where (in some cases) non-atomic ::MoveFileExW is used).

However, 'rename' might fail and return an error code in several
different situations. If this happens, we will fall back to a direct write to
the file, which is neither atomic, nor avoids race conditions (causing problem
#3).
An example where 'rename' can fail on both Windows and Unix-like
systems, causing us to fall back to using non-atomic write, is problem #2.

Solution:
All solutions for problem #3 (see below) will take care of problem #1.


Problem #2 (Use of POSIX-compliant 'rename' function is not always
suitable for ThinLTO)
We just encountered this issue in Sony. For ThinLTO, we use the function
'rename', which after r315079 doesn't support renaming across the
different logical volumes on Windows.
Let's say a user has a %TEMP% directory on volume C:\, but he/she passed the
caching directory name located on volume D:\, then 'rename' fails. Since
we don't support renaming across different volumes after r315079, we then
fall back to writing to the cache file directly (which is not atomic) and hit a
data race.
We anticipate that this will be a common situation for our users, many of whom
are Windows "power users" and have multiple disks for different
purposes.

I think this problem doesn't only affect Windows. The same issue might
happen on Linux/MacOS, if the user's $TEMP/$TMP directory is located in a
different file system than the user-specified caching directory.

Solution #1 (not so good):
The patch below solves this problem on Windows, however, we will lose the
advantage that 'rename' gained in r315079 (i.e., its atomicity), which
is not good.

This patch assumes that the function 'rename' in Windows/Path.inc is
expected to work when the source and destination are located of different
logical volumes or different file systems. Note that function 'rename'
is not ThinLTO-specific and might have some other consumers who want this
behavior (which was the behavior before r315079). However, it seems that this
assumption has changed later to make 'rename' more POSIX-compliant. In
this case, we should add a comment to the function 'rename' so that its
users knew that renaming across different volumes is not supported by design. Or
we could have two different functions, one POSIX-compliant, not allowing renames
across different volumes, another non-POSIX-compliant.

Before r315079, the function 'rename', (in the case when the destination
file isn't opened), boiled down to calling the function
'MoveFileExW' with the MOVEFILE_COPY_ALLOWED flag set, allowing renaming
files across the volumes.
The current implementation of the function 'rename' (in
'rename_internal'), essentially calls
'SetFileInformationByHandle'. This function is fast and atomic, so in a
sense, it's an improvement over the previously used 'MoveFileExW'.
However, 'SetFileInformationByHandle' doesn't work when the source
and destination are located on the different volumes.

My fix just falls back to calling 'MoveFileExW' when
'SetFileInformationByHandle' returns an error
'ERROR_NOT_SAME_DEVICE'.


+    // Wine doesn't support SetFileInformationByHandle in rename_internal.
+    // Rename_internal doesn't work accross different disks. In both of
these
+    // cases fall back to using MoveFileEx.
     if (EC =-        std::error_code(ERROR_CALL_NOT_IMPLEMENTED,
std::system_category())) {
-      // Wine doesn't support SetFileInformationByHandle in
rename_internal.
-      // Fall back to MoveFileEx.
+        std::error_code(ERROR_CALL_NOT_IMPLEMENTED, std::system_category()) ||
+        EC == std::error_code(ERROR_NOT_SAME_DEVICE, std::system_category())) {
       SmallVector<wchar_t, MAX_PATH> WideFrom;
       if (std::error_code EC2 = realPathFromHandle(FromHandle, WideFrom))
         return EC2;
       if (::MoveFileExW(WideFrom.begin(), WideTo.begin(),
-                        MOVEFILE_REPLACE_EXISTING))
+                        MOVEFILE_COPY_ALLOWED | MOVEFILE_REPLACE_EXISTING))
         return std::error_code();
       return mapWindowsError(GetLastError());
     }

Note, that when the source and destination are located on the different logical
volumes, we will effectively perform a copy, followed by a delete, which are not
atomic operations. Since copying to a different volume might be quite time
consuming, we also increase the probability that some other process starts to
rename to the same destination file.

So, this fix for 'rename' is not good for ThinLTO (we should do
something else in this case). 'rename' is a generic function and has
many different users, but unless renaming across the volumes is needed by
someone else, other than ThinLTO, this patch shouldn't be accepted.


Solution #2 (better)
Here I only try to fix a ThinLTO issue, not a generic 'rename' issue as
in solution #1:

For ThinLTO+caching we don't have to create temp files in %TEMP% or %TMP%
directories (or $TMP/$TEMP). We can create them in the same location where the
cached files are stored or in a 'temp' subfolder within this folder.
With this approach:

-  If the patch described in Solution #1 gets accepted (for the sake of other
consumers), then from ThinLTO's perspective 'rename' will stay
atomic.

-  If the patch doesn't get accepted, then rename won't fail, since we
only rename the files within the same directory.

Solution #3
All the solutions for problem #3 (see below) will take care of problem #2.


Problem #3 (data race)
Look at the following code in ThinLTOCodeGenerator.cpp. This code gets executed
if the 'rename' function failed (e.g., because of the problem #1 or
problem #2 described above).

EC = sys::fs::rename(TempFilename, EntryPath);
if (EC) {
  sys::fs::remove(TempFilename);
  raw_fd_ostream OS(EntryPath, EC, sys::fs::F_None);
  if (EC)
    report_fatal_error(Twine("Failed to open ") + EntryPath +
                       " to save cached entry\n");
  OS << OutputBuffer.getBuffer();            // Two ThinLTO processes can
write to the same file here
                                             // causing data race.

Here, we have a data race problem. We actually stumbled across it while testing
one of the games with ThinLTO+caching.

We need to find a generic solution to solve data race problem, while ensuring
'atomicity' of writing to cache files. I'm not a ThinLTO expert and
I might not be aware of some ThinLTO constraints. Let me know which problems you
see with the two solutions below. Hopefully, it will trigger a further
discussion.


Solution #0 (it won't require much modifications):

(a)   If 'rename' fails, do not write into the cache directly (we
don't want to have non-atomic writes!).

(b)   If 'rename' fails, try to read from the cache.

auto ReloadedBufferOrErr = CacheEntry.tryLoadingBuffer();

(c)   If reading from the cache fails simply use the file that you just compiled
directly (bypassing the cache entirely).



This solution might cause cache misses (causing recompilations), but at least it
should prevent data races when two ThinLTO processes write to the same file
(which can produce invalid cache!). Correctness is more important than
optimization. It's worth noticing another shortage of this solution: unlike
generic solution #1, #2 and #3 described below this particular solution
won't solve the problem #4.



BTW, we should do something with WINE support in 'rename' (which is
non-atomic). I didn't want to list it as a separate problem here, but it is
a problem. I could file a separate bug, if necessary.



Solution #1 (naïve generic solution):
(a) If the process wants to write to the cache file, it opens it with
'exclusive read/write' access.
(b) If a process wants to write to the cache file that is 'exclusively'
opened by some other process, we could assume that the cache file will be
successfully created by the first process and simply return from 'write'
function. Different processes writing to the cache file of the same name, are
writing the same content, right?
(c) If the process needs to read from the cache file, it might wait a bit while
the file is open for 'exclusive write'. If within a certain period the
cache file doesn't get released by a writer or gets removed by a pruner -
oh, well, we have a hit miss. After all, using cache is just an acceleration.
Correctness is more important.
(d) If we are concerned about data integrity of a cache file (e.g., a writer
started to write to a file and something bad happened in the middle), we could
do additional checks (I suspect that this might be already implemented in
ThinLTO). A writer could add a unique hash at the end of the cache file or a
CRC32 for each section of the file, which could be an indication that the write
to this file was successful. A reader checks that this unique hash or a CRC32
checksum matches its expectations.

I'm sure that there are good reasons why this "naive" solution
wasn't implemented in the first place. If this solution is picked by LLVM
community, I could write a working prototype for it that we could apply to
ThinLTO.


Solution #2 (better generic solution):
It looks like cache file sharing between several ThinLTO processes is a classic
readers-writers problem.
https://en.wikipedia.org/wiki/Readers%E2%80%93writers_problem

We could use read-write locks for global inter-process file sharing.
https://en.wikipedia.org/wiki/Readers%E2%80%93writer_lock

ThinLTO's writer (creator/deleter) must acquire a write lock before writing
to a file and release a write lock when it's done writing.
ThinLTO's user (reader) must acquire a read lock before reading a file and
release a read lock when it's done reading. On a Unix-like system,
read-write locks are part of POSIX-threads (pthreads). There is an
implementation of Slim read-write locks on Windows, but unfortunately for
in-process only (i.e., the locks can't be shared among different processes),
so it won't work for us.
https://msdn.microsoft.com/en-us/library/windows/desktop/aa904937(v=vs.85).aspx

However, on Windows, we could implement read-write locks ourselves, using a
combination of a named mutex and a named semaphore (any unique global name could
be used for creating a named mutex/semaphore, the simplest one will be the cache
file name itself).

Here is an example of an implementation:
http://www.zachburlingame.com/2011/06/simple-reader-writer-lock-in-c-on-win32/

If this solution is picked, I could write a working prototype for it that we
could apply to ThinLTO.

Solution #3 (hybrid generic solution)
Basically, use solution #2 as a base and integrate some optimizations and
completeness checking features from solution #1 (1.b and 1.d respectively).

Problem #4 (cache misses could have been avoided via synchronization).

With our current approach we might often race for the cache file resources and
as a result have cache misses.
Of course, it's not as dangerous as the data race problem #3, but it
definitely lowers the efficiency of caching.

(a) The reader checks that a file exists, but then the pruner deletes it before
the reader had a chance to read it. Cache miss.
(b) The reader starts reading a file (block by block) and at this time the
pruner removes the file. The read might fail. This is OS-specific, but likely a
cache miss.
(c) If the writer will rename a temp file into a file that is currently being
read by a reader, I don't expect anything good out of it (the behavior is
OS-specific). In the best case scenario, the read will fail, in the worst one
the reader will read the wrong content. So, it's a cache miss or a
correctness issue.
(d) This is not a very likely event, but what if two processes are trying to
rename to the same file at the same time?

Solution:
Solutions #1, #2 and #3 for problem #3 will take care of problem #4.


(Potential) problem #5:
The implementation of the function 'rename' on Windows used by ThinLTO
heavily depends on the function CreateFileW or, to be more exact, on its flag
FILE_SHARE_DELETE being supported and working correctly on *all* the file
systems that can be mounted on Windows. Is the FILE_SHARE_DELETE flag supported
on all non-native Windows file systems that we care about? Is it supported on
WINE? On FAT32?

If the flag FILE_SHARE_DELETE is not supported, the 'rename' fails, and
we have problem #3.

Solution:
All the solutions for problem #3 will take care of problem #5.


Please let me know what do you think about the problems/solutions discussed
here. Hopefully this pre-RFC shapes into a RFC soon.

Thank you!
Katya.
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20180322/f0450ec4/attachment.html>

Hi Katya

Thanks for investigating this. Here is my thought inline.
> On Mar 22, 2018, at 1:32 AM, katya.romanova at sony.com wrote:
> 
>  
> Hello,
>  
> I am sending the following proposal to discuss issues and solutions
regarding data races in concurrent ThinLTO processes.
>  
> This caught my attention when we encountered a race condition in ThinLTO
with caching.
> While looking into how ThinLTO deals with the problem of cache files
reads/writes/deletes I spotted a lot of problems: some of them are related to
data races, others - to file/resource sharing between different processes. I
wanted to point out these problems and start the discussion about potential
solutions. I would like to get your feedback.
>  
>  
> Problem #1: In certain common situations, ‘rename’ can fail, causing a data
race later on.
> Look at the following code in the function ‘write’ in
ThinLTOCodeGenerator.cpp
>  
> std::error_code EC >   sys::fs::createTemporaryFile("Thin",
"tmp.o", TempFD, TempFilename);
> if (EC) {
>   errs() << "Error: " << EC.message() <<
"\n";
>   report_fatal_error("ThinLTO: Can't get a temporary file");
> }
> {
>   raw_fd_ostream OS(TempFD, /* ShouldClose */ true);
>   OS << OutputBuffer.getBuffer();
> }
> // Rename to final destination (hopefully race condition won't matter
here)
> EC = sys::fs::rename(TempFilename, EntryPath);    
>  
> Compared to a race-prone direct write of a buffer to a file, this code
avoids a data race by first writing a buffer to a temp file and then renaming
that temp file to become the final file in the cache. After r315079, when
‘rename’ became more POSIX-compliant, this scheme guarantees atomicity of
writing a file to the cache,
> (except on Wine, where (in some cases) non-atomic ::MoveFileExW is used).
>  
> However, ‘rename’ might fail and return an error code in several different
situations. If this happens, we will fall back to a direct write to the file,
which is neither atomic, nor avoids race conditions (causing problem #3).
> An example where 'rename' can fail on both Windows and Unix-like
systems, causing us to fall back to using non-atomic write, is problem #2.
>  
> Solution:
> All solutions for problem #3 (see below) will take care of problem #1.
I don't think there is a race here but it can probably be optimized. I
don't see LTOCodeGenerator fall back to non-atomic write in the code. If the
rename failed, it will just fatal_error and exit. Maybe a better solution is
just leave the temp file and use it as output. The file will be recompiled if
needed again (turns into a cache miss).
>  
>  
> Problem #2 (Use of POSIX-compliant ‘rename’ function is not always suitable
for ThinLTO)
> We just encountered this issue in Sony. For ThinLTO, we use the function
‘rename’, which after r315079 doesn’t support renaming across the different
logical volumes on Windows.
> Let’s say a user has a %TEMP% directory on volume C:\, but he/she passed
the caching directory name located on volume D:\, then ‘rename’ fails. Since we
don’t support renaming across different volumes after r315079, we then fall back
to writing to the cache file directly (which is not atomic) and hit a data race.
> We anticipate that this will be a common situation for our users, many of
whom are Windows “power users” and have multiple disks for different purposes.
>  
> I think this problem doesn’t only affect Windows. The same issue might
happen on Linux/MacOS, if the user's $TEMP/$TMP directory is located in a
different file system than the user-specified caching directory.
>  
> Solution #1 (not so good):
> The patch below solves this problem on Windows, however, we will lose the
advantage that ‘rename’ gained in r315079 (i.e., its atomicity), which is not
good.
>  
> This patch assumes that the function 'rename' in Windows/Path.inc
is expected to work when the source and destination are located of different
logical volumes or different file systems. Note that function ‘rename’ is not
ThinLTO-specific and might have some other consumers who want this behavior
(which was the behavior before r315079). However, it seems that this assumption
has changed later to make ‘rename’ more POSIX-compliant. In this case, we should
add a comment to the function 'rename' so that its users knew that
renaming across different volumes is not supported by design. Or we could have
two different functions, one POSIX-compliant, not allowing renames across
different volumes, another non-POSIX-compliant.
>  
> Before r315079, the function 'rename', (in the case when the
destination file isn't opened), boiled down to calling the function
'MoveFileExW' with the MOVEFILE_COPY_ALLOWED flag set, allowing renaming
files across the volumes.
> The current implementation of the function ‘rename’ (in
'rename_internal'), essentially calls
'SetFileInformationByHandle'. This function is fast and atomic, so in a
sense, it's an improvement over the previously used 'MoveFileExW'.
However, 'SetFileInformationByHandle' doesn't work when the source
and destination are located on the different volumes.
>  
> My fix just falls back to calling 'MoveFileExW' when
'SetFileInformationByHandle' returns an error
'ERROR_NOT_SAME_DEVICE'.
>  
>  
> +    // Wine doesn't support SetFileInformationByHandle in
rename_internal.
> +    // Rename_internal doesn't work accross different disks. In both
of these
> +    // cases fall back to using MoveFileEx.
>      if (EC => -        std::error_code(ERROR_CALL_NOT_IMPLEMENTED,
std::system_category())) {
> -      // Wine doesn't support SetFileInformationByHandle in
rename_internal.
> -      // Fall back to MoveFileEx.
> +        std::error_code(ERROR_CALL_NOT_IMPLEMENTED,
std::system_category()) ||
> +        EC == std::error_code(ERROR_NOT_SAME_DEVICE,
std::system_category())) {
>        SmallVector<wchar_t, MAX_PATH> WideFrom;
>        if (std::error_code EC2 = realPathFromHandle(FromHandle, WideFrom))
>          return EC2;
>        if (::MoveFileExW(WideFrom.begin(), WideTo.begin(),
> -                        MOVEFILE_REPLACE_EXISTING))
> +                        MOVEFILE_COPY_ALLOWED |
MOVEFILE_REPLACE_EXISTING))
>          return std::error_code();
>        return mapWindowsError(GetLastError());
>      }
>  
> Note, that when the source and destination are located on the different
logical volumes, we will effectively perform a copy, followed by a delete, which
are not atomic operations. Since copying to a different volume might be quite
time consuming, we also increase the probability that some other process starts
to rename to the same destination file.
>  
> So, this fix for ‘rename’ is not good for ThinLTO (we should do something
else in this case). ‘rename’ is a generic function and has many different users,
but unless renaming across the volumes is needed by someone else, other than
ThinLTO, this patch shouldn’t be accepted.
>  
>  
> Solution #2 (better)
> Here I only try to fix a ThinLTO issue, not a generic ‘rename’ issue as in
solution #1:
>  
> For ThinLTO+caching we don’t have to create temp files in %TEMP% or %TMP%
directories (or $TMP/$TEMP). We can create them in the same location where the
cached files are stored or in a ‘temp’ subfolder within this folder. With this
approach:
> -  If the patch described in Solution #1 gets accepted (for the sake of
other consumers), then from ThinLTO’s perspective ‘rename’ will stay atomic.
> -  If the patch doesn’t get accepted, then rename won’t fail, since we only
rename the files within the same directory.
>  
> Solution #3
> All the solutions for problem #3 (see below) will take care of problem #2.
I am surprised that we are not doing #2 already. Rename across the volumes is
not atomic thus it can cause issues.
>  
>  
> Problem #3 (data race)
> Look at the following code in ThinLTOCodeGenerator.cpp. This code gets
executed if the ‘rename’ function failed (e.g., because of the problem #1 or
problem #2 described above).
>  
> EC = sys::fs::rename(TempFilename, EntryPath);
> if (EC) {
>   sys::fs::remove(TempFilename);
>   raw_fd_ostream OS(EntryPath, EC, sys::fs::F_None);
>   if (EC)
>     report_fatal_error(Twine("Failed to open ") + EntryPath +
>                        " to save cached entry\n");
>   OS << OutputBuffer.getBuffer();            // Two ThinLTO processes
can write to the same file here
>                                              // causing data race.
>  
> Here, we have a data race problem. We actually stumbled across it while
testing one of the games with ThinLTO+caching.
>  
> We need to find a generic solution to solve data race problem, while
ensuring ‘atomicity’ of writing to cache files. I’m not a ThinLTO expert and I
might not be aware of some ThinLTO constraints. Let me know which problems you
see with the two solutions below. Hopefully, it will trigger a further
discussion.
>  
>  
> Solution #0 (it won’t require much modifications):
> (a)   If ‘rename’ fails, do not write into the cache directly (we don’t
want to have non-atomic writes!).
> (b)   If ‘rename’ fails, try to read from the cache.            
> auto ReloadedBufferOrErr = CacheEntry.tryLoadingBuffer();
> (c)   If reading from the cache fails simply use the file that you just
compiled directly (bypassing the cache entirely).
>  
> This solution might cause cache misses (causing recompilations), but at
least it should prevent data races when two ThinLTO processes write to the same
file (which can produce invalid cache!). Correctness is more important than
optimization. It’s worth noticing another shortage of this solution: unlike
generic solution #1, #2 and #3 described below this particular solution won’t
solve the problem #4.
>  
> BTW, we should do something with WINE support in ‘rename’ (which is
non-atomic). I didn’t want to list it as a separate problem here, but it is a
problem. I could file a separate bug, if necessary.
>  
>  
> Solution #1 (naïve generic solution):
> (a) If the process wants to write to the cache file, it opens it with
'exclusive read/write' access.
> (b) If a process wants to write to the cache file that is ‘exclusively’
opened by some other process, we could assume that the cache file will be
successfully created by the first process and simply return from ‘write’
function. Different processes writing to the cache file of the same name, are
writing the same content, right?
> (c) If the process needs to read from the cache file, it might wait a bit
while the file is open for 'exclusive write'. If within a certain period
the cache file doesn’t get released by a writer or gets removed by a pruner –
oh, well, we have a hit miss. After all, using cache is just an acceleration.
Correctness is more important.
> (d) If we are concerned about data integrity of a cache file (e.g., a
writer started to write to a file and something bad happened in the middle), we
could do additional checks (I suspect that this might be already implemented in
ThinLTO). A writer could add a unique hash at the end of the cache file or a
CRC32 for each section of the file, which could be an indication that the write
to this file was successful. A reader checks that this unique hash or a CRC32
checksum matches its expectations.
>  
> I'm sure that there are good reasons why this "naive"
solution wasn't implemented in the first place. If this solution is picked
by LLVM community, I could write a working prototype for it that we could apply
to ThinLTO.
>  
>  
> Solution #2 (better generic solution):
> It looks like cache file sharing between several ThinLTO processes is a
classic readers-writers problem.
> https://en.wikipedia.org/wiki/Readers%E2%80%93writers_problem
<https://en.wikipedia.org/wiki/Readers%E2%80%93writers_problem>
>  
> We could use read-write locks for global inter-process file sharing.
> https://en.wikipedia.org/wiki/Readers%E2%80%93writer_lock
<https://en.wikipedia.org/wiki/Readers%E2%80%93writer_lock>
>  
> ThinLTO's writer (creator/deleter) must acquire a write lock before
writing to a file and release a write lock when it's done writing.
> ThinLTO’s user (reader) must acquire a read lock before reading a file and
release a read lock when it's done reading. On a Unix-like system,
read-write locks are part of POSIX-threads (pthreads). There is an
implementation of Slim read-write locks on Windows, but unfortunately for
in-process only (i.e., the locks can’t be shared among different processes), so
it won’t work for us.
>
https://msdn.microsoft.com/en-us/library/windows/desktop/aa904937(v=vs.85).aspx
<https://msdn.microsoft.com/en-us/library/windows/desktop/aa904937(v=vs.85).aspx>
>  
> However, on Windows, we could implement read-write locks ourselves, using a
combination of a named mutex and a named semaphore (any unique global name could
be used for creating a named mutex/semaphore, the simplest one will be the cache
file name itself).
>  
> Here is an example of an implementation:
>
http://www.zachburlingame.com/2011/06/simple-reader-writer-lock-in-c-on-win32/
<http://www.zachburlingame.com/2011/06/simple-reader-writer-lock-in-c-on-win32/>
>  
> If this solution is picked, I could write a working prototype for it that
we could apply to ThinLTO.
>  
> Solution #3 (hybrid generic solution)
> Basically, use solution #2 as a base and integrate some optimizations and
completeness checking features from solution #1 (1.b and 1.d respectively).
>  
> Problem #4 (cache misses could have been avoided via synchronization).
>  
> With our current approach we might often race for the cache file resources
and as a result have cache misses.
> Of course, it's not as dangerous as the data race problem #3, but it
definitely lowers the efficiency of caching.
>  
> (a) The reader checks that a file exists, but then the pruner deletes it
before the reader had a chance to read it. Cache miss.
> (b) The reader starts reading a file (block by block) and at this time the
pruner removes the file. The read might fail. This is OS-specific, but likely a
cache miss.
> (c) If the writer will rename a temp file into a file that is currently
being read by a reader, I don't expect anything good out of it (the behavior
is OS-specific). In the best case scenario, the read will fail, in the worst one
the reader will read the wrong content. So, it's a cache miss or a
correctness issue.
> (d) This is not a very likely event, but what if two processes are trying
to rename to the same file at the same time?
>  
> Solution:
> Solutions #1, #2 and #3 for problem #3 will take care of problem #4.
>  
>  
> (Potential) problem #5:
> The implementation of the function ‘rename’ on Windows used by ThinLTO
heavily depends on the function CreateFileW or, to be more exact, on its flag
FILE_SHARE_DELETE being supported and working correctly on *all* the file
systems that can be mounted on Windows. Is the FILE_SHARE_DELETE flag supported
on all non-native Windows file systems that we care about? Is it supported on
WINE? On FAT32?
>  
> If the flag FILE_SHARE_DELETE is not supported, the ‘rename’ fails, and we
have problem #3.
>  
> Solution:
> All the solutions for problem #3 will take care of problem #5.
>  
>  
> Please let me know what do you think about the problems/solutions discussed
here. Hopefully this pre-RFC shapes into a RFC soon.
I am not really concerned with all the problems, except Problem #2. From my
understanding, LTO cache is for incremental build only. For every file in the
cache, there should be one single writer, otherwise, it is a hash collision and
will result in miscompile (or link error most likely). The cached file will only
be used when you rebuild the binary. Lots of data races are in situations that
are not supported.

For its current design, you cannot shared the LTO cache between two linking. For
example, if you are building llc and libLTO from LLVM and they shared many same
static archives with common LTO object files, you still need to use two
different LTO caches.

Steven

>  
> Thank you!
> Katya.
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20180322/a2cafa6b/attachment-0001.html>

Hi Katya,

Please take a look at how caching is implemented in the new LTO API:
http://llvm-cs.pcc.me.uk/lib/LTO/Caching.cpp\

I believe that these problems are solved there.

I would be happy to consider a patch porting this implementation to the
legacy API.

Peter

On Thu, Mar 22, 2018 at 1:32 AM, <katya.romanova at sony.com> wrote:
>
>
> Hello,
>
>
>
> I am sending the following proposal to discuss issues and solutions
> regarding data races in concurrent ThinLTO processes.
>
>
>
> This caught my attention when we encountered a race condition in ThinLTO
> with caching.
>
> While looking into how ThinLTO deals with the problem of cache files
> reads/writes/deletes I spotted a lot of problems: some of them are related
> to data races, others - to file/resource sharing between different
> processes. I wanted to point out these problems and start the discussion
> about potential solutions. I would like to get your feedback.
>
>
>
>
>
> *Problem #1: In certain common situations, ‘rename’ can fail, causing a
> data race later on.*
>
> Look at the following code in the function ‘write’ in
> ThinLTOCodeGenerator.cpp
>
>
>
> *std::error_code EC =*
>
> *  sys::fs::createTemporaryFile("Thin", "tmp.o",
TempFD, TempFilename);*
>
> *if (EC) {*
>
> *  errs() << "Error: " << EC.message() <<
"\n";*
>
> *  report_fatal_error("ThinLTO: Can't get a temporary
file");*
>
> *}*
>
> *{*
>
> *  raw_fd_ostream OS(TempFD, /* ShouldClose */ true);*
>
> *  OS << OutputBuffer.getBuffer();*
>
> *}*
>
> *// Rename to final destination (hopefully race condition won't matter
> here)*
>
> *EC = sys::fs::rename(TempFilename, EntryPath);     *
>
>
>
> Compared to a race-prone direct write of a buffer to a file, this code
> avoids a data race by first writing a buffer to a temp file and then
> renaming that temp file to become the final file in the cache. After
> r315079, when ‘rename’ became more POSIX-compliant, this scheme guarantees
> atomicity of writing a file to the cache,
>
> (except on Wine, where (in some cases) non-atomic ::MoveFileExW is used).
>
>
>
> However, ‘rename’ might fail and return an error code in several different
> situations. If this happens, we will fall back to a direct write to the
> file, which is neither atomic, nor avoids race conditions (causing problem
> #3).
>
> An example where 'rename' can fail on both Windows and Unix-like
systems,
> causing us to fall back to using non-atomic write, is problem #2.
>
>
>
> *Solution:*
>
> All solutions for problem #3 (see below) will take care of problem #1.
>
>
>
>
>
> *Problem #2 (Use of POSIX-compliant ‘rename’ function is not always
> suitable for ThinLTO)*
>
> We just encountered this issue in Sony. For ThinLTO, we use the function
> ‘rename’, which after r315079 doesn’t support renaming across the different
> logical volumes on Windows.
>
> Let’s say a user has a %TEMP% directory on volume C:\, but he/she passed
> the caching directory name located on volume D:\, then ‘rename’ fails.
> Since we don’t support renaming across different volumes after r315079, we
> then fall back to writing to the cache file directly (which is not atomic)
> and hit a data race.
>
> We anticipate that this will be a common situation for our users, many of
> whom are Windows “power users” and have multiple disks for different
> purposes.
>
>
>
> I think this problem doesn’t only affect Windows. The same issue might
> happen on Linux/MacOS, if the user's $TEMP/$TMP directory is located in
a
> different file system than the user-specified caching directory.
>
>
>
> *Solution #1 (not so good):*
>
> The patch below solves this problem on Windows, however, we will lose the
> advantage that ‘rename’ gained in r315079 (i.e., its atomicity), which is
> not good.
>
>
>
> This patch assumes that the function 'rename' in Windows/Path.inc
is
> expected to work when the source and destination are located of different
> logical volumes or different file systems. Note that function ‘rename’ is
> not ThinLTO-specific and might have some other consumers who want this
> behavior (which was the behavior before r315079). However, it seems that
> this assumption has changed later to make ‘rename’ more POSIX-compliant. In
> this case, we should add a comment to the function 'rename' so that
its
> users knew that renaming across different volumes is not supported by
> design. Or we could have two different functions, one POSIX-compliant, not
> allowing renames across different volumes, another non-POSIX-compliant.
>
>
>
> Before r315079, the function 'rename', (in the case when the
destination
> file isn't opened), boiled down to calling the function
'MoveFileExW' with
> the MOVEFILE_COPY_ALLOWED flag set, allowing renaming files across the
> volumes.
>
> The current implementation of the function ‘rename’ (in
> 'rename_internal'), essentially calls
'SetFileInformationByHandle'. This
> function is fast and atomic, so in a sense, it's an improvement over
the
> previously used 'MoveFileExW'. However,
'SetFileInformationByHandle'
> doesn't work when the source and destination are located on the
different
> volumes.
>
>
>
> My fix just falls back to calling 'MoveFileExW' when
> 'SetFileInformationByHandle' returns an error
'ERROR_NOT_SAME_DEVICE'.
>
>
>
>
>
> *+    // Wine doesn't support SetFileInformationByHandle in
> rename_internal.*
>
> *+    // Rename_internal doesn't work accross different disks. In both
of
> these*
>
> *+    // cases fall back to using MoveFileEx. *
>
> *     if (EC ==*
>
> *-        std::error_code(ERROR_CALL_NOT_IMPLEMENTED,
> std::system_category())) {*
>
> *-      // Wine doesn't support SetFileInformationByHandle in
> rename_internal.*
>
> *-      // Fall back to MoveFileEx.*
>
> *+        std::error_code(ERROR_CALL_NOT_IMPLEMENTED,
> std::system_category()) ||*
>
> *+        EC == std::error_code(ERROR_NOT_SAME_DEVICE,
> std::system_category())) {*
>
> *       SmallVector<wchar_t, MAX_PATH> WideFrom;*
>
> *       if (std::error_code EC2 = realPathFromHandle(FromHandle,
> WideFrom))*
>
> *         return EC2;*
>
> *       if (::MoveFileExW(WideFrom.begin(), WideTo.begin(),*
>
> *-                        MOVEFILE_REPLACE_EXISTING))*
>
> *+                        MOVEFILE_COPY_ALLOWED |
> MOVEFILE_REPLACE_EXISTING))*
>
> *         return std::error_code();*
>
> *       return mapWindowsError(GetLastError());*
>
> *     }*
>
>
>
> Note, that when the source and destination are located on the different
> logical volumes, we will effectively perform a copy, followed by a delete,
> which are not atomic operations. Since copying to a different volume might
> be quite time consuming, we also increase the probability that some other
> process starts to rename to the same destination file.
>
>
>
> So, this fix for ‘rename’ is not good for ThinLTO (we should do something
> else in this case). ‘rename’ is a generic function and has many different
> users, but unless renaming across the volumes is needed by someone else,
> other than ThinLTO, this patch shouldn’t be accepted.
>
>
>
>
>
> *Solution #2 (better)*
>
> Here I only try to fix a ThinLTO issue, not a generic ‘rename’ issue as in
> solution #1:
>
>
>
> For ThinLTO+caching we don’t have to create temp files in %TEMP% or %TMP%
> directories (or $TMP/$TEMP). We can create them in the same location where
> the cached files are stored or in a ‘temp’ subfolder within this folder.
> With this approach:
>
> -  If the patch described in Solution #1 gets accepted (for the sake of
> other consumers), then from ThinLTO’s perspective ‘rename’ will stay
> atomic.
>
> -  If the patch doesn’t get accepted, then rename won’t fail, since we
> only rename the files within the same directory.
>
>
>
> *Solution #3*
>
> All the solutions for problem #3 (see below) will take care of problem #2.
>
>
>
>
>
> *Problem #3 (data race)*
>
> Look at the following code in ThinLTOCodeGenerator.cpp. This code gets
> executed if the ‘rename’ function failed (e.g., because of the problem #1
> or problem #2 described above).
>
>
>
> *EC = sys::fs::rename(TempFilename, EntryPath);*
>
> *if (EC) {*
>
> *  sys::fs::remove(TempFilename);*
>
> *  raw_fd_ostream OS(EntryPath, EC, sys::fs::F_None);*
>
> *  if (EC)*
>
> *    report_fatal_error(Twine("Failed to open ") + EntryPath +*
>
> *                       " to save cached entry\n");*
>
> *  OS << OutputBuffer.getBuffer();            **// Two ThinLTO
processes
> can write to the same file here *
>
> *                                             // causing data race.*
>
>
>
> Here, we have a data race problem. We actually stumbled across it while
> testing one of the games with ThinLTO+caching.
>
>
>
> We need to find a generic solution to solve data race problem, while
> ensuring ‘atomicity’ of writing to cache files. I’m not a ThinLTO expert
> and I might not be aware of some ThinLTO constraints. Let me know which
> problems you see with the two solutions below. Hopefully, it will trigger a
> further discussion.
>
>
>
>
>
> *Solution #0 (it won’t require much modifications): *
>
> (a)   If ‘rename’ fails, do not write into the cache directly (we don’t
> want to have non-atomic writes!).
>
> (b)   If ‘rename’ fails, try to read from the cache.
>
> *auto ReloadedBufferOrErr = CacheEntry.tryLoadingBuffer();*
>
> *(c)*   If reading from the cache fails simply use the file that you just
> compiled directly (bypassing the cache entirely).
>
>
>
> This solution might cause cache misses (causing recompilations), but at
> least it should prevent data races when two ThinLTO processes write to the
> same file (which can produce invalid cache!). Correctness is more important
> than optimization. It’s worth noticing another shortage of this solution:
> unlike generic solution #1, #2 and #3 described below this particular
> solution won’t solve the problem #4.
>
>
>
> BTW, we should do something with WINE support in ‘rename’ (which is
> non-atomic). I didn’t want to list it as a separate problem here, but it is
> a problem. I could file a separate bug, if necessary.
>
>
>
>
>
> *Solution #1 (naïve generic solution): *
>
> (a) If the process wants to write to the cache file, it opens it with
> 'exclusive read/write' access.
>
> (b) If a process wants to write to the cache file that is ‘exclusively’
> opened by some other process, we could assume that the cache file will be
> successfully created by the first process and simply return from ‘write’
> function. Different processes writing to the cache file of the same name,
> are writing the same content, right?
>
> (c) If the process needs to read from the cache file, it might wait a bit
> while the file is open for 'exclusive write'. If within a certain
period
> the cache file doesn’t get released by a writer or gets removed by a pruner
> – oh, well, we have a hit miss. After all, using cache is just an
> acceleration. Correctness is more important.
>
> (d) If we are concerned about data integrity of a cache file (e.g., a
> writer started to write to a file and something bad happened in the
> middle), we could do additional checks (I suspect that this might be
> already implemented in ThinLTO). A writer could add a unique hash at the
> end of the cache file or a CRC32 for each section of the file, which could
> be an indication that the write to this file was successful. A reader
> checks that this unique hash or a CRC32 checksum matches its expectations.
>
>
>
> I'm sure that there are good reasons why this "naive"
solution wasn't
> implemented in the first place. If this solution is picked by LLVM
> community, I could write a working prototype for it that we could apply to
> ThinLTO.
>
>
>
>
>
> *Solution #2 (better generic solution):*
>
> It looks like cache file sharing between several ThinLTO processes is a
> classic readers-writers problem.
>
> https://en.wikipedia.org/wiki/Readers%E2%80%93writers_problem
>
>
>
> We could use read-write locks for global inter-process file sharing.
>
> https://en.wikipedia.org/wiki/Readers%E2%80%93writer_lock
>
>
>
> ThinLTO's writer (creator/deleter) must acquire a write lock before
> writing to a file and release a write lock when it's done writing.
>
> ThinLTO’s user (reader) must acquire a read lock before reading a file and
> release a read lock when it's done reading. On a Unix-like system,
> read-write locks are part of POSIX-threads (pthreads). There is an
> implementation of Slim read-write locks on Windows, but unfortunately for
> in-process only (i.e., the locks can’t be shared among different
> processes), so it won’t work for us.
>
> https://msdn.microsoft.com/en-us/library/windows/desktop/
> aa904937(v=vs.85).aspx
>
>
>
> However, on Windows, we could implement read-write locks ourselves, using
> a combination of a named mutex and a named semaphore (any unique global
> name could be used for creating a named mutex/semaphore, the simplest one
> will be the cache file name itself).
>
>
>
> Here is an example of an implementation:
>
> http://www.zachburlingame.com/2011/06/simple-reader-writer-
> lock-in-c-on-win32/
>
>
>
> If this solution is picked, I could write a working prototype for it that
> we could apply to ThinLTO.
>
>
>
> *Solution #3 (hybrid generic solution)*
>
> Basically, use solution #2 as a base and integrate some optimizations and
> completeness checking features from solution #1 (1.b and 1.d respectively).
>
>
>
> *Problem #4 (cache misses could have been avoided via synchronization).*
>
>
>
> With our current approach we might often race for the cache file resources
> and as a result have cache misses.
>
> Of course, it's not as dangerous as the data race problem #3, but it
> definitely lowers the efficiency of caching.
>
>
>
> (a) The reader checks that a file exists, but then the pruner deletes it
> before the reader had a chance to read it. Cache miss.
>
> (b) The reader starts reading a file (block by block) and at this time the
> pruner removes the file. The read might fail. This is OS-specific, but
> likely a cache miss.
>
> (c) If the writer will rename a temp file into a file that is currently
> being read by a reader, I don't expect anything good out of it (the
> behavior is OS-specific). In the best case scenario, the read will fail, in
> the worst one the reader will read the wrong content. So, it's a cache
miss
> or a correctness issue.
>
> (d) This is not a very likely event, but what if two processes are trying
> to rename to the same file at the same time?
>
>
>
> *Solution:*
>
> Solutions #1, #2 and #3 for problem #3 will take care of problem #4.
>
>
>
>
>
> *(Potential) problem #5:*
>
> The implementation of the function ‘rename’ on Windows used by ThinLTO
> heavily depends on the function CreateFileW or, to be more exact, on its
> flag FILE_SHARE_DELETE being supported and working correctly on *all* the
> file systems that can be mounted on Windows. Is the FILE_SHARE_DELETE flag
> supported on all non-native Windows file systems that we care about? Is it
> supported on WINE? On FAT32?
>
>
>
> If the flag FILE_SHARE_DELETE is not supported, the ‘rename’ fails, and we
> have problem #3.
>
>
>
> *Solution:*
>
> All the solutions for problem #3 will take care of problem #5.
>
>
>
>
>
> Please let me know what do you think about the problems/solutions
> discussed here. Hopefully this pre-RFC shapes into a RFC soon.
>
>
>
> Thank you!
>
> Katya.
>


-- 
-- 
Peter
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20180322/a425547e/attachment.html>

Hi Steven,
Look at my replies inline (below your comments).
Katya.

From: stevenwu at apple.com <stevenwu at apple.com>
Sent: Thursday, March 22, 2018 4:46 PM
To: Romanova, Katya <katya.romanova at sony.com>
Cc: Teresa Johnson <tejohnson at google.com>; Mehdi AMINI <joker.eph at
gmail.com>; Rafael Avila de Espindola <rafael.espindola at gmail.com>;
Peter Collingbourne <peter at pcc.me.uk>; Hans Wennborg <hans at
chromium.org>; Reid Kleckner <rnk at google.com>; bd1976llvm at
gmail.com; llvm-dev at lists.llvm.org
Subject: Re: [pre-RFC] Data races in concurrent ThinLTO processes

Hi Katya

Thanks for investigating this. Here is my thought inline.


On Mar 22, 2018, at 1:32 AM, katya.romanova at sony.com<mailto:katya.romanova
at sony.com> wrote:


Hello,

I am sending the following proposal to discuss issues and solutions regarding
data races in concurrent ThinLTO processes.

This caught my attention when we encountered a race condition in ThinLTO with
caching.
While looking into how ThinLTO deals with the problem of cache files
reads/writes/deletes I spotted a lot of problems: some of them are related to
data races, others - to file/resource sharing between different processes. I
wanted to point out these problems and start the discussion about potential
solutions. I would like to get your feedback.


Problem #1: In certain common situations, ‘rename’ can fail, causing a data race
later on.
Look at the following code in the function ‘write’ in ThinLTOCodeGenerator.cpp

std::error_code EC   sys::fs::createTemporaryFile("Thin",
"tmp.o", TempFD, TempFilename);
if (EC) {
  errs() << "Error: " << EC.message() <<
"\n";
  report_fatal_error("ThinLTO: Can't get a temporary file");
}
{
  raw_fd_ostream OS(TempFD, /* ShouldClose */ true);
  OS << OutputBuffer.getBuffer();
}
// Rename to final destination (hopefully race condition won't matter here)
EC = sys::fs::rename(TempFilename, EntryPath);

Compared to a race-prone direct write of a buffer to a file, this code avoids a
data race by first writing a buffer to a temp file and then renaming that temp
file to become the final file in the cache. After r315079, when ‘rename’ became
more POSIX-compliant, this scheme guarantees atomicity of writing a file to the
cache,
(except on Wine, where (in some cases) non-atomic ::MoveFileExW is used).

However, ‘rename’ might fail and return an error code in several different
situations. If this happens, we will fall back to a direct write to the file,
which is neither atomic, nor avoids race conditions (causing problem #3).
An example where 'rename' can fail on both Windows and Unix-like
systems, causing us to fall back to using non-atomic write, is problem #2.

Solution:
All solutions for problem #3 (see below) will take care of problem #1.

I don't think there is a race here but it can probably be optimized. I
don't see LTOCodeGenerator fall back to non-atomic write in the code. If the
rename failed, it will just fatal_error and exit. Maybe a better solution is
just leave the temp file and use it as output. The file will be recompiled if
needed again (turns into a cache miss).

I think there is a non-atomic write (and race) here. See below. If rename
failed, we remove the temp file and write to the cache file directly (OS
<< OutputBuffer.getBuffer()).

    // Rename to final destination (hopefully race condition won't matter
here)
    EC = sys::fs::rename(TempFilename, EntryPath);
    if (EC) {
      sys::fs::remove(TempFilename);
      raw_fd_ostream OS(EntryPath, EC, sys::fs::F_None);
      if (EC)
        report_fatal_error(Twine("Failed to open ") + EntryPath +
                           " to save cached entry\n");
      OS << OutputBuffer.getBuffer();
    }
  }

What we can do the following way:
Option 1: If the rename fails, we could simply use OutputBuffer (which has the
same content as our temp file), bypassing the cache altogether.
Option 2: We could try reading from the existing cache first, and then, if this
read fails, we could use OutputBuffer (in this case, we will be closer to the
existing code). I don’t understand why do we have to read from the existing
cache first (even if we just created this cache). I suspect this was done for a
reason. There is a comment below with the explanation, but it doesn’t make it
more clear to me. If this reason/explanation makes sense, then we should do
Option 2. This is exactly what I previously proposed in Solution #0 for Problem
#3.

       // Commit to the cache (if enabled)
        CacheEntry.write(*OutputBuffer);

        if (SavedObjectsDirectoryPath.empty()) {
          // We need to generated a memory buffer for the linker.
          if (!CacheEntryPath.empty()) {
            // Cache is enabled, reload from the cache
            // We do this to lower memory pressuree: the buffer is on the heap
            // and releasing it frees memory that can be used for the next input
            // file. The final binary link will read from the VFS cache
           // (hopefully!) or from disk if the memory pressure wasn't too
high.                   <- I don’t understand why we are reading from the
cache file (via tryLoadingBuffer), though we already have the same content in
OutputBuffer?
            auto ReloadedBufferOrErr = CacheEntry.tryLoadingBuffer();           
<-The explanation says this is done to reduce the memory pressure. But it’s
not clear to me why this helps to reduce memory pressure, since we still have to
keep
            if (auto EC = ReloadedBufferOrErr.getError()) {                     
<- the content of the cache file in a buffer, pointed to by
ReloadedBufferOrErr).
              // On error, keeping the preexisting buffer and printing a
              // diagnostic is more friendly than just crashing.
              errs() << "error: can't reload cached file
'" << CacheEntryPath
                     << "': " << EC.message() <<
"\n";
            } else {
              OutputBuffer = std::move(*ReloadedBufferOrErr);
            }
          }




Problem #2 (Use of POSIX-compliant ‘rename’ function is not always suitable for
ThinLTO)
We just encountered this issue in Sony. For ThinLTO, we use the function
‘rename’, which after r315079 doesn’t support renaming across the different
logical volumes on Windows.
Let’s say a user has a %TEMP% directory on volume C:\, but he/she passed the
caching directory name located on volume D:\, then ‘rename’ fails. Since we
don’t support renaming across different volumes after r315079, we then fall back
to writing to the cache file directly (which is not atomic) and hit a data race.
We anticipate that this will be a common situation for our users, many of whom
are Windows “power users” and have multiple disks for different purposes.

I think this problem doesn’t only affect Windows. The same issue might happen on
Linux/MacOS, if the user's $TEMP/$TMP directory is located in a different
file system than the user-specified caching directory.

Solution #1 (not so good):
The patch below solves this problem on Windows, however, we will lose the
advantage that ‘rename’ gained in r315079 (i.e., its atomicity), which is not
good.

This patch assumes that the function 'rename' in Windows/Path.inc is
expected to work when the source and destination are located of different
logical volumes or different file systems. Note that function ‘rename’ is not
ThinLTO-specific and might have some other consumers who want this behavior
(which was the behavior before r315079). However, it seems that this assumption
has changed later to make ‘rename’ more POSIX-compliant. In this case, we should
add a comment to the function 'rename' so that its users knew that
renaming across different volumes is not supported by design. Or we could have
two different functions, one POSIX-compliant, not allowing renames across
different volumes, another non-POSIX-compliant.

Before r315079, the function 'rename', (in the case when the destination
file isn't opened), boiled down to calling the function
'MoveFileExW' with the MOVEFILE_COPY_ALLOWED flag set, allowing renaming
files across the volumes.
The current implementation of the function ‘rename’ (in
'rename_internal'), essentially calls
'SetFileInformationByHandle'. This function is fast and atomic, so in a
sense, it's an improvement over the previously used 'MoveFileExW'.
However, 'SetFileInformationByHandle' doesn't work when the source
and destination are located on the different volumes.

My fix just falls back to calling 'MoveFileExW' when
'SetFileInformationByHandle' returns an error
'ERROR_NOT_SAME_DEVICE'.


+    // Wine doesn't support SetFileInformationByHandle in rename_internal.
+    // Rename_internal doesn't work accross different disks. In both of
these
+    // cases fall back to using MoveFileEx.
     if (EC =-        std::error_code(ERROR_CALL_NOT_IMPLEMENTED,
std::system_category())) {
-      // Wine doesn't support SetFileInformationByHandle in
rename_internal.
-      // Fall back to MoveFileEx.
+        std::error_code(ERROR_CALL_NOT_IMPLEMENTED, std::system_category()) ||
+        EC == std::error_code(ERROR_NOT_SAME_DEVICE, std::system_category())) {
       SmallVector<wchar_t, MAX_PATH> WideFrom;
       if (std::error_code EC2 = realPathFromHandle(FromHandle, WideFrom))
         return EC2;
       if (::MoveFileExW(WideFrom.begin(), WideTo.begin(),
-                        MOVEFILE_REPLACE_EXISTING))
+                        MOVEFILE_COPY_ALLOWED | MOVEFILE_REPLACE_EXISTING))
         return std::error_code();
       return mapWindowsError(GetLastError());
     }

Note, that when the source and destination are located on the different logical
volumes, we will effectively perform a copy, followed by a delete, which are not
atomic operations. Since copying to a different volume might be quite time
consuming, we also increase the probability that some other process starts to
rename to the same destination file.

So, this fix for ‘rename’ is not good for ThinLTO (we should do something else
in this case). ‘rename’ is a generic function and has many different users, but
unless renaming across the volumes is needed by someone else, other than
ThinLTO, this patch shouldn’t be accepted.


Solution #2 (better)
Here I only try to fix a ThinLTO issue, not a generic ‘rename’ issue as in
solution #1:

For ThinLTO+caching we don’t have to create temp files in %TEMP% or %TMP%
directories (or $TMP/$TEMP). We can create them in the same location where the
cached files are stored or in a ‘temp’ subfolder within this folder. With this
approach:
-  If the patch described in Solution #1 gets accepted (for the sake of other
consumers), then from ThinLTO’s perspective ‘rename’ will stay atomic.
-  If the patch doesn’t get accepted, then rename won’t fail, since we only
rename the files within the same directory.

Solution #3
All the solutions for problem #3 (see below) will take care of problem #2.

I am surprised that we are not doing #2 already. Rename across the volumes is
not atomic thus it can cause issues.

Yes, we are calling createTemporaryFile, which creates a file in the location
pointed by $TEMP, $TMP, etc.
        sys::fs::createTemporaryFile("Thin", "tmp.o",
TempFD, TempFilename);
So, it looks like this problem is applicable to Unix-like systems too (assuming
that your $TEMP and cache directory are in the different file systems).



Problem #3 (data race)
Look at the following code in ThinLTOCodeGenerator.cpp. This code gets executed
if the ‘rename’ function failed (e.g., because of the problem #1 or problem #2
described above).

EC = sys::fs::rename(TempFilename, EntryPath);
if (EC) {
  sys::fs::remove(TempFilename);
  raw_fd_ostream OS(EntryPath, EC, sys::fs::F_None);
  if (EC)
    report_fatal_error(Twine("Failed to open ") + EntryPath +
                       " to save cached entry\n");
  OS << OutputBuffer.getBuffer();            // Two ThinLTO processes can
write to the same file here
                                             // causing data race.

Here, we have a data race problem. We actually stumbled across it while testing
one of the games with ThinLTO+caching.

We need to find a generic solution to solve data race problem, while ensuring
‘atomicity’ of writing to cache files. I’m not a ThinLTO expert and I might not
be aware of some ThinLTO constraints. Let me know which problems you see with
the two solutions below. Hopefully, it will trigger a further discussion.


Solution #0 (it won’t require much modifications):
(a)   If ‘rename’ fails, do not write into the cache directly (we don’t want to
have non-atomic writes!).
(b)   If ‘rename’ fails, try to read from the cache.
auto ReloadedBufferOrErr = CacheEntry.tryLoadingBuffer();
(c)   If reading from the cache fails simply use the file that you just compiled
directly (bypassing the cache entirely).

This solution might cause cache misses (causing recompilations), but at least it
should prevent data races when two ThinLTO processes write to the same file
(which can produce invalid cache!). Correctness is more important than
optimization. It’s worth noticing another shortage of this solution: unlike
generic solution #1, #2 and #3 described below this particular solution won’t
solve the problem #4.

BTW, we should do something with WINE support in ‘rename’ (which is non-atomic).
I didn’t want to list it as a separate problem here, but it is a problem. I
could file a separate bug, if necessary.


Solution #1 (naïve generic solution):
(a) If the process wants to write to the cache file, it opens it with
'exclusive read/write' access.
(b) If a process wants to write to the cache file that is ‘exclusively’ opened
by some other process, we could assume that the cache file will be successfully
created by the first process and simply return from ‘write’ function. Different
processes writing to the cache file of the same name, are writing the same
content, right?
(c) If the process needs to read from the cache file, it might wait a bit while
the file is open for 'exclusive write'. If within a certain period the
cache file doesn’t get released by a writer or gets removed by a pruner – oh,
well, we have a hit miss. After all, using cache is just an acceleration.
Correctness is more important.
(d) If we are concerned about data integrity of a cache file (e.g., a writer
started to write to a file and something bad happened in the middle), we could
do additional checks (I suspect that this might be already implemented in
ThinLTO). A writer could add a unique hash at the end of the cache file or a
CRC32 for each section of the file, which could be an indication that the write
to this file was successful. A reader checks that this unique hash or a CRC32
checksum matches its expectations.

I'm sure that there are good reasons why this "naive" solution
wasn't implemented in the first place. If this solution is picked by LLVM
community, I could write a working prototype for it that we could apply to
ThinLTO.


Solution #2 (better generic solution):
It looks like cache file sharing between several ThinLTO processes is a classic
readers-writers problem.
https://en.wikipedia.org/wiki/Readers%E2%80%93writers_problem

We could use read-write locks for global inter-process file sharing.
https://en.wikipedia.org/wiki/Readers%E2%80%93writer_lock

ThinLTO's writer (creator/deleter) must acquire a write lock before writing
to a file and release a write lock when it's done writing.
ThinLTO’s user (reader) must acquire a read lock before reading a file and
release a read lock when it's done reading. On a Unix-like system,
read-write locks are part of POSIX-threads (pthreads). There is an
implementation of Slim read-write locks on Windows, but unfortunately for
in-process only (i.e., the locks can’t be shared among different processes), so
it won’t work for us.
https://msdn.microsoft.com/en-us/library/windows/desktop/aa904937(v=vs.85).aspx

However, on Windows, we could implement read-write locks ourselves, using a
combination of a named mutex and a named semaphore (any unique global name could
be used for creating a named mutex/semaphore, the simplest one will be the cache
file name itself).

Here is an example of an implementation:
http://www.zachburlingame.com/2011/06/simple-reader-writer-lock-in-c-on-win32/

If this solution is picked, I could write a working prototype for it that we
could apply to ThinLTO.

Solution #3 (hybrid generic solution)
Basically, use solution #2 as a base and integrate some optimizations and
completeness checking features from solution #1 (1.b and 1.d respectively).

Problem #4 (cache misses could have been avoided via synchronization).

With our current approach we might often race for the cache file resources and
as a result have cache misses.
Of course, it's not as dangerous as the data race problem #3, but it
definitely lowers the efficiency of caching.

(a) The reader checks that a file exists, but then the pruner deletes it before
the reader had a chance to read it. Cache miss.
(b) The reader starts reading a file (block by block) and at this time the
pruner removes the file. The read might fail. This is OS-specific, but likely a
cache miss.
(c) If the writer will rename a temp file into a file that is currently being
read by a reader, I don't expect anything good out of it (the behavior is
OS-specific). In the best case scenario, the read will fail, in the worst one
the reader will read the wrong content. So, it's a cache miss or a
correctness issue.
(d) This is not a very likely event, but what if two processes are trying to
rename to the same file at the same time?

Solution:
Solutions #1, #2 and #3 for problem #3 will take care of problem #4.


(Potential) problem #5:
The implementation of the function ‘rename’ on Windows used by ThinLTO heavily
depends on the function CreateFileW or, to be more exact, on its flag
FILE_SHARE_DELETE being supported and working correctly on *all* the file
systems that can be mounted on Windows. Is the FILE_SHARE_DELETE flag supported
on all non-native Windows file systems that we care about? Is it supported on
WINE? On FAT32?

If the flag FILE_SHARE_DELETE is not supported, the ‘rename’ fails, and we have
problem #3.

Solution:
All the solutions for problem #3 will take care of problem #5.


Please let me know what do you think about the problems/solutions discussed
here. Hopefully this pre-RFC shapes into a RFC soon.

I am not really concerned with all the problems, except Problem #2. From my
understanding, LTO cache is for incremental build only.
Yes.

For every file in the cache, there should be one single writer, otherwise, it is
a hash collision and will result in miscompile (or link error most likely). The
cached file will only be used when you rebuild the binary. Lots of data races
are in situations that are not supported.
Are you aware of any other data races when we share LTO cache between two links,
except the one that we identified above? Please let me know. I’d rather know all
the potential problems/limitation in advance.

For its current design, you cannot shared the LTO cache between two linking. For
example, if you are building llc and libLTO from LLVM and they shared many same
static archives with common LTO object files, you still need to use two
different LTO caches.
But what if we fix this particular problem? We could either do it the way Peter
implemented it for the new LTO API or the way we discussed above. I assume that
after fixing this, will we be able to share the same cache between 2 links.

The question is, how exactly we should fix it for C LTO API? I’m aware of only
two major C LTO API stakeholders (Apple and Sony), so most likely nobody else
cares much about legacy C API. Peter generously proposed to port the non-racy
algorithm from new LTO API to the legacy C API. I looked into new LTO API today
and it looks that the data race is fixed there, though the code is more heavy
C++ and will be harder to read. Alternatively, I could write a simpler patch for
legacy C LTO API, doing what we discussed above (i.e. creating temp files in
cache directory and if renaming of the temp into a cache file failed, bypass
reading from the cache file).

What do you think?


Steven



Thank you!
Katya.

-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20180327/af9e39af/attachment-0001.html>

Le jeu. 22 mars 2018 à 16:46, Steven Wu <stevenwu at apple.com> a écrit :
> Hi Katya
>
> Thanks for investigating this. Here is my thought inline.
>
> On Mar 22, 2018, at 1:32 AM, katya.romanova at sony.com wrote:
>
>
> Hello,
>
> I am sending the following proposal to discuss issues and solutions
> regarding data races in concurrent ThinLTO processes.
>
> This caught my attention when we encountered a race condition in ThinLTO
> with caching.
> While looking into how ThinLTO deals with the problem of cache files
> reads/writes/deletes I spotted a lot of problems: some of them are related
> to data races, others - to file/resource sharing between different
> processes. I wanted to point out these problems and start the discussion
> about potential solutions. I would like to get your feedback.
>
>
> *Problem #1: In certain common situations, ‘rename’ can fail, causing a
> data race later on.*
> Look at the following code in the function ‘write’ in
> ThinLTOCodeGenerator.cpp
>
> *std::error_code EC =*
> *  sys::fs::createTemporaryFile("Thin", "tmp.o",
TempFD, TempFilename);*
> *if (EC) {*
> *  errs() << "Error: " << EC.message() <<
"\n";*
> *  report_fatal_error("ThinLTO: Can't get a temporary
file");*
> *}*
> *{*
> *  raw_fd_ostream OS(TempFD, /* ShouldClose */ true);*
> *  OS << OutputBuffer.getBuffer();*
> *}*
> *// Rename to final destination (hopefully race condition won't matter
> here)*
> *EC = sys::fs::rename(TempFilename, EntryPath);    *
>
> Compared to a race-prone direct write of a buffer to a file, this code
> avoids a data race by first writing a buffer to a temp file and then
> renaming that temp file to become the final file in the cache. After
> r315079, when ‘rename’ became more POSIX-compliant, this scheme guarantees
> atomicity of writing a file to the cache,
> (except on Wine, where (in some cases) non-atomic ::MoveFileExW is used).
>
> However, ‘rename’ might fail and return an error code in several different
> situations. If this happens, we will fall back to a direct write to the
> file, which is neither atomic, nor avoids race conditions (causing problem
> #3).
> An example where 'rename' can fail on both Windows and Unix-like
systems,
> causing us to fall back to using non-atomic write, is problem #2.
>
> *Solution:*
> All solutions for problem #3 (see below) will take care of problem #1.
>
>
> I don't think there is a race here but it can probably be optimized. I
> don't see LTOCodeGenerator fall back to non-atomic write in the code.
If
> the rename failed, it will just fatal_error and exit. Maybe a better
> solution is just leave the temp file and use it as output. The file will be
> recompiled if needed again (turns into a cache miss).
>
>
>
> *Problem #2 (Use of POSIX-compliant ‘rename’ function is not always
> suitable for ThinLTO)*
> We just encountered this issue in Sony. For ThinLTO, we use the function
> ‘rename’, which after r315079 doesn’t support renaming across the different
> logical volumes on Windows.
> Let’s say a user has a %TEMP% directory on volume C:\, but he/she passed
> the caching directory name located on volume D:\, then ‘rename’ fails.
> Since we don’t support renaming across different volumes after r315079, we
> then fall back to writing to the cache file directly (which is not atomic)
> and hit a data race.
> We anticipate that this will be a common situation for our users, many of
> whom are Windows “power users” and have multiple disks for different
> purposes.
>
> I think this problem doesn’t only affect Windows. The same issue might
> happen on Linux/MacOS, if the user's $TEMP/$TMP directory is located in
a
> different file system than the user-specified caching directory.
>
> *Solution #1 (not so good):*
> The patch below solves this problem on Windows, however, we will lose the
> advantage that ‘rename’ gained in r315079 (i.e., its atomicity), which is
> not good.
>
> This patch assumes that the function 'rename' in Windows/Path.inc
is
> expected to work when the source and destination are located of different
> logical volumes or different file systems. Note that function ‘rename’ is
> not ThinLTO-specific and might have some other consumers who want this
> behavior (which was the behavior before r315079). However, it seems that
> this assumption has changed later to make ‘rename’ more POSIX-compliant. In
> this case, we should add a comment to the function 'rename' so that
its
> users knew that renaming across different volumes is not supported by
> design. Or we could have two different functions, one POSIX-compliant, not
> allowing renames across different volumes, another non-POSIX-compliant.
>
> Before r315079, the function 'rename', (in the case when the
destination
> file isn't opened), boiled down to calling the function
'MoveFileExW' with
> the MOVEFILE_COPY_ALLOWED flag set, allowing renaming files across the
> volumes.
> The current implementation of the function ‘rename’ (in
> 'rename_internal'), essentially calls
'SetFileInformationByHandle'. This
> function is fast and atomic, so in a sense, it's an improvement over
the
> previously used 'MoveFileExW'. However,
'SetFileInformationByHandle'
> doesn't work when the source and destination are located on the
different
> volumes.
>
> My fix just falls back to calling 'MoveFileExW' when
> 'SetFileInformationByHandle' returns an error
'ERROR_NOT_SAME_DEVICE'.
>
>
> *+    // Wine doesn't support SetFileInformationByHandle in
> rename_internal.*
> *+    // Rename_internal doesn't work accross different disks. In both
of
> these*
> *+    // cases fall back to using MoveFileEx.*
> *     if (EC ==*
> *-        std::error_code(ERROR_CALL_NOT_IMPLEMENTED,
> std::system_category())) {*
> *-      // Wine doesn't support SetFileInformationByHandle in
> rename_internal.*
> *-      // Fall back to MoveFileEx.*
> *+        std::error_code(ERROR_CALL_NOT_IMPLEMENTED,
> std::system_category()) ||*
> *+        EC == std::error_code(ERROR_NOT_SAME_DEVICE,
> std::system_category())) {*
> *       SmallVector<wchar_t, MAX_PATH> WideFrom;*
> *       if (std::error_code EC2 = realPathFromHandle(FromHandle,
> WideFrom))*
> *         return EC2;*
> *       if (::MoveFileExW(WideFrom.begin(), WideTo.begin(),*
> *-                        MOVEFILE_REPLACE_EXISTING))*
> *+                        MOVEFILE_COPY_ALLOWED |
> MOVEFILE_REPLACE_EXISTING))*
> *         return std::error_code();*
> *       return mapWindowsError(GetLastError());*
> *     }*
>
> Note, that when the source and destination are located on the different
> logical volumes, we will effectively perform a copy, followed by a delete,
> which are not atomic operations. Since copying to a different volume might
> be quite time consuming, we also increase the probability that some other
> process starts to rename to the same destination file.
>
> So, this fix for ‘rename’ is not good for ThinLTO (we should do something
> else in this case). ‘rename’ is a generic function and has many different
> users, but unless renaming across the volumes is needed by someone else,
> other than ThinLTO, this patch shouldn’t be accepted.
>
>
> *Solution #2 (better)*
> Here I only try to fix a ThinLTO issue, not a generic ‘rename’ issue as in
> solution #1:
>
> For ThinLTO+caching we don’t have to create temp files in %TEMP% or %TMP%
> directories (or $TMP/$TEMP). We can create them in the same location where
> the cached files are stored or in a ‘temp’ subfolder within this folder.
> With this approach:
> -  If the patch described in Solution #1 gets accepted (for the sake of
> other consumers), then from ThinLTO’s perspective ‘rename’ will stay
atomic.
> -  If the patch doesn’t get accepted, then rename won’t fail, since we
> only rename the files within the same directory.
>
> *Solution #3*
> All the solutions for problem #3 (see below) will take care of problem #2.
>
>
> I am surprised that we are not doing #2 already. Rename across the volumes
> is not atomic thus it can cause issues.
>
>
>
> *Problem #3 (data race)*
> Look at the following code in ThinLTOCodeGenerator.cpp. This code gets
> executed if the ‘rename’ function failed (e.g., because of the problem #1
> or problem #2 described above).
>
> *EC = sys::fs::rename(TempFilename, EntryPath);*
> *if (EC) {*
> *  sys::fs::remove(TempFilename);*
> *  raw_fd_ostream OS(EntryPath, EC, sys::fs::F_None);*
> *  if (EC)*
> *    report_fatal_error(Twine("Failed to open ") + EntryPath +*
> *                       " to save cached entry\n");*
> *  OS << OutputBuffer.getBuffer();            **// Two ThinLTO
processes
> can write to the same file here*
> *                                             // causing data race.*
>
> Here, we have a data race problem. We actually stumbled across it while
> testing one of the games with ThinLTO+caching.
>
> We need to find a generic solution to solve data race problem, while
> ensuring ‘atomicity’ of writing to cache files. I’m not a ThinLTO expert
> and I might not be aware of some ThinLTO constraints. Let me know which
> problems you see with the two solutions below. Hopefully, it will trigger a
> further discussion.
>
>
> *Solution #0 (it won’t require much modifications):*
> (a)   If ‘rename’ fails, do not write into the cache directly (we don’t
> want to have non-atomic writes!).
> (b)   If ‘rename’ fails, try to read from the cache.
> *auto ReloadedBufferOrErr = CacheEntry.tryLoadingBuffer();*
> *(c)*   If reading from the cache fails simply use the file that you just
> compiled directly (bypassing the cache entirely).
>
> This solution might cause cache misses (causing recompilations), but at
> least it should prevent data races when two ThinLTO processes write to the
> same file (which can produce invalid cache!). Correctness is more important
> than optimization. It’s worth noticing another shortage of this solution:
> unlike generic solution #1, #2 and #3 described below this particular
> solution won’t solve the problem #4.
>
> BTW, we should do something with WINE support in ‘rename’ (which is
> non-atomic). I didn’t want to list it as a separate problem here, but it is
> a problem. I could file a separate bug, if necessary.
>
>
> *Solution #1 (naïve generic solution):*
> (a) If the process wants to write to the cache file, it opens it with
> 'exclusive read/write' access.
> (b) If a process wants to write to the cache file that is ‘exclusively’
> opened by some other process, we could assume that the cache file will be
> successfully created by the first process and simply return from ‘write’
> function. Different processes writing to the cache file of the same name,
> are writing the same content, right?
> (c) If the process needs to read from the cache file, it might wait a bit
> while the file is open for 'exclusive write'. If within a certain
period
> the cache file doesn’t get released by a writer or gets removed by a pruner
> – oh, well, we have a hit miss. After all, using cache is just an
> acceleration. Correctness is more important.
> (d) If we are concerned about data integrity of a cache file (e.g., a
> writer started to write to a file and something bad happened in the
> middle), we could do additional checks (I suspect that this might be
> already implemented in ThinLTO). A writer could add a unique hash at the
> end of the cache file or a CRC32 for each section of the file, which could
> be an indication that the write to this file was successful. A reader
> checks that this unique hash or a CRC32 checksum matches its expectations.
>
> I'm sure that there are good reasons why this "naive"
solution wasn't
> implemented in the first place. If this solution is picked by LLVM
> community, I could write a working prototype for it that we could apply to
> ThinLTO.
>
>
> *Solution #2 (better generic solution):*
> It looks like cache file sharing between several ThinLTO processes is a
> classic readers-writers problem.
> https://en.wikipedia.org/wiki/Readers%E2%80%93writers_problem
>
> We could use read-write locks for global inter-process file sharing.
> https://en.wikipedia.org/wiki/Readers%E2%80%93writer_lock
>
> ThinLTO's writer (creator/deleter) must acquire a write lock before
> writing to a file and release a write lock when it's done writing.
> ThinLTO’s user (reader) must acquire a read lock before reading a file and
> release a read lock when it's done reading. On a Unix-like system,
> read-write locks are part of POSIX-threads (pthreads). There is an
> implementation of Slim read-write locks on Windows, but unfortunately for
> in-process only (i.e., the locks can’t be shared among different
> processes), so it won’t work for us.
>
>
https://msdn.microsoft.com/en-us/library/windows/desktop/aa904937(v=vs.85).aspx
>
> However, on Windows, we could implement read-write locks ourselves, using
> a combination of a named mutex and a named semaphore (any unique global
> name could be used for creating a named mutex/semaphore, the simplest one
> will be the cache file name itself).
>
> Here is an example of an implementation:
>
>
http://www.zachburlingame.com/2011/06/simple-reader-writer-lock-in-c-on-win32/
>
> If this solution is picked, I could write a working prototype for it that
> we could apply to ThinLTO.
>
> *Solution #3 (hybrid generic solution)*
> Basically, use solution #2 as a base and integrate some optimizations and
> completeness checking features from solution #1 (1.b and 1.d respectively).
>
> *Problem #4 (cache misses could have been avoided via synchronization).*
>
> With our current approach we might often race for the cache file resources
> and as a result have cache misses.
> Of course, it's not as dangerous as the data race problem #3, but it
> definitely lowers the efficiency of caching.
>
> (a) The reader checks that a file exists, but then the pruner deletes it
> before the reader had a chance to read it. Cache miss.
> (b) The reader starts reading a file (block by block) and at this time the
> pruner removes the file. The read might fail. This is OS-specific, but
> likely a cache miss.
> (c) If the writer will rename a temp file into a file that is currently
> being read by a reader, I don't expect anything good out of it (the
> behavior is OS-specific). In the best case scenario, the read will fail, in
> the worst one the reader will read the wrong content. So, it's a cache
miss
> or a correctness issue.
> (d) This is not a very likely event, but what if two processes are trying
> to rename to the same file at the same time?
>
> *Solution:*
> Solutions #1, #2 and #3 for problem #3 will take care of problem #4.
>
>
> *(Potential) problem #5:*
> The implementation of the function ‘rename’ on Windows used by ThinLTO
> heavily depends on the function CreateFileW or, to be more exact, on its
> flag FILE_SHARE_DELETE being supported and working correctly on *all* the
> file systems that can be mounted on Windows. Is the FILE_SHARE_DELETE flag
> supported on all non-native Windows file systems that we care about? Is it
> supported on WINE? On FAT32?
>
> If the flag FILE_SHARE_DELETE is not supported, the ‘rename’ fails, and we
> have problem #3.
>
> *Solution:*
> All the solutions for problem #3 will take care of problem #5.
>
>
> Please let me know what do you think about the problems/solutions
> discussed here. Hopefully this pre-RFC shapes into a RFC soon.
>
>
> I am not really concerned with all the problems, except Problem #2. From
> my understanding, LTO cache is for incremental build only. For every file
> in the cache, there should be one single writer, otherwise, it is a hash
> collision and will result in miscompile (or link error most likely). The
> cached file will only be used when you rebuild the binary. Lots of data
> races are in situations that are not supported.
>
> For its current design, you cannot shared the LTO cache between two
> linking. For example, if you are building llc and libLTO from LLVM and they
> shared many same static archives with common LTO object files, you still
> need to use two different LTO caches.
>
Why do we need to use two different caches?
In my memory a ThinLTO build of LLVM does use a single cache and share and
reuse objects across binaries if possible.

Best,

-- 
Mehdi
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20180327/07f4a97e/attachment.html>