PROBLEMS - third_party/openssl - Git at Google

 * System libcrypto.dylib and libssl.dylib are used by system ld on MacOS X.


     NOTE: The problem described here only applies when OpenSSL isn't built
     with shared library support (i.e. without the "shared" configuration
     option).  If you build with shared library support, you will have no
     problems as long as you set up DYLD_LIBRARY_PATH properly at all times.


 This is really a misfeature in ld, which seems to look for .dylib libraries
 along the whole library path before it bothers looking for .a libraries.  This
 means that -L switches won't matter unless OpenSSL is built with shared
 library support.

 The workaround may be to change the following lines in apps/Makefile and
 test/Makefile:

   LIBCRYPTO=-L.. -lcrypto
   LIBSSL=-L.. -lssl

 to:

   LIBCRYPTO=../libcrypto.a
   LIBSSL=../libssl.a

 It's possible that something similar is needed for shared library support
 as well.  That hasn't been well tested yet.


 Another solution that many seem to recommend is to move the libraries
 /usr/lib/libcrypto.0.9.dylib, /usr/lib/libssl.0.9.dylib to a different
 directory, build and install OpenSSL and anything that depends on your
 build, then move libcrypto.0.9.dylib and libssl.0.9.dylib back to their
 original places.  Note that the version numbers on those two libraries
 may differ on your machine.


 As long as Apple doesn't fix the problem with ld, this problem building
 OpenSSL will remain as is. Well, the problem was addressed in 0.9.8f by
 passing -Wl,-search_paths_first, but it's unknown if the flag was
 supported from the initial MacOS X release.


 * Parallell make leads to errors

 While running tests, running a parallell make is a bad idea.  Many test
 scripts use the same name for output and input files, which means different
 will interfere with each other and lead to test failure.

 The solution is simple for now: don't run parallell make when testing.


 * Bugs in gcc triggered

 - According to a problem report, there are bugs in gcc 3.0 that are
   triggered by some of the code in OpenSSL, more specifically in
   PEM_get_EVP_CIPHER_INFO().  The triggering code is the following:

 	header+=11;
 	if (*header != '4') return(0); header++;
 	if (*header != ',') return(0); header++;

   What happens is that gcc might optimize a little too agressively, and
   you end up with an extra incrementation when *header != '4'.

   We recommend that you upgrade gcc to as high a 3.x version as you can.

 - According to multiple problem reports, some of our message digest
   implementations trigger bug[s] in code optimizer in gcc 3.3 for sparc64
   and gcc 2.96 for ppc. Former fails to complete RIPEMD160 test, while
   latter - SHA one.

   The recomendation is to upgrade your compiler. This naturally applies to
   other similar cases.

 - There is a subtle Solaris x86-specific gcc run-time environment bug, which
   "falls between" OpenSSL [0.9.8 and later], Solaris ld and GCC. The bug
   manifests itself as Segmentation Fault upon early application start-up.
   The problem can be worked around by patching the environment according to
   http://www.openssl.org/~appro/values.c.

 * solaris64-sparcv9-cc SHA-1 performance with WorkShop 6 compiler.

 As subject suggests SHA-1 might perform poorly (4 times slower)
 if compiled with WorkShop 6 compiler and -xarch=v9. The cause for
 this seems to be the fact that compiler emits multiplication to
 perform shift operations:-( To work the problem around configure
 with './Configure solaris64-sparcv9-cc -DMD32_REG_T=int'.

 * Problems with hp-parisc2-cc target when used with "no-asm" flag

 When using the hp-parisc2-cc target, wrong bignum code is generated.
 This is due to the SIXTY_FOUR_BIT build being compiled with the +O3
 aggressive optimization.
 The problem manifests itself by the BN_kronecker test hanging in an
 endless loop. Reason: the BN_kronecker test calls BN_generate_prime()
 which itself hangs. The reason could be tracked down to the bn_mul_comba8()
 function in bn_asm.c. At some occasions the higher 32bit value of r[7]
 is off by 1 (meaning: calculated=shouldbe+1). Further analysis failed,
 as no debugger support possible at +O3 and additional fprintf()'s
 introduced fixed the bug, therefore it is most likely a bug in the
 optimizer.
 The bug was found in the BN_kronecker test but may also lead to
 failures in other parts of the code.
 (See Ticket #426.)

 Workaround: modify the target to +O2 when building with no-asm.

 * Problems building shared libraries on SCO OpenServer Release 5.0.6
   with gcc 2.95.3

 The symptoms appear when running the test suite, more specifically
 test/ectest, with the following result:

 OSSL_LIBPATH="`cd ..; pwd`"; LD_LIBRARY_PATH="$OSSL_LIBPATH:$LD_LIBRARY_PATH"; DYLD_LIBRARY_PATH="$OSSL_LIBPATH:$DYLD_LIBRARY_PATH"; SHLIB_PATH="$OSSL_LIBPATH:$SHLIB_PATH"; LIBPATH="$OSSL_LIBPATH:$LIBPATH"; if [ "debug-sco5-gcc" = "Cygwin" ]; then PATH="${LIBPATH}:$PATH"; fi; export LD_LIBRARY_PATH DYLD_LIBRARY_PATH SHLIB_PATH LIBPATH PATH; ./ectest
 ectest.c:186: ABORT

 The cause of the problem seems to be that isxdigit(), called from
 BN_hex2bn(), returns 0 on a perfectly legitimate hex digit.  Further
 investigation shows that any of the isxxx() macros return 0 on any
 input.  A direct look in the information array that the isxxx() use,
 called __ctype, shows that it contains all zeroes...

 Taking a look at the newly created libcrypto.so with nm, one can see
 that the variable __ctype is defined in libcrypto's .bss (which
 explains why it is filled with zeroes):

 $ nm -Pg libcrypto.so | grep __ctype
 __ctype B 0011659c
 __ctype2 U

 Curiously, __ctype2 is undefined, in spite of being declared in
 /usr/include/ctype.h in exactly the same way as __ctype.

 Any information helping to solve this issue would be deeply
 appreciated.

 NOTE: building non-shared doesn't come with this problem.

 * ULTRIX build fails with shell errors, such as "bad substitution"
   and "test: argument expected"

 The problem is caused by ULTRIX /bin/sh supporting only original
 Bourne shell syntax/semantics, and the trouble is that the vast
 majority is so accustomed to more modern syntax, that very few
 people [if any] would recognize the ancient syntax even as valid.
 This inevitably results in non-trivial scripts breaking on ULTRIX,
 and OpenSSL isn't an exclusion. Fortunately there is workaround,
 hire /bin/ksh to do the job /bin/sh fails to do.

 1. Trick make(1) to use /bin/ksh by setting up following environ-
    ment variables *prior* you execute ./Configure and make:

 	PROG_ENV=POSIX
 	MAKESHELL=/bin/ksh
 	export PROG_ENV MAKESHELL

    or if your shell is csh-compatible:

 	setenv PROG_ENV POSIX
 	setenv MAKESHELL /bin/ksh

 2. Trick /bin/sh to use alternative expression evaluator. Create
    following 'test' script for example in /tmp:

 	#!/bin/ksh
 	${0##*/} "$@"

    Then 'chmod a+x /tmp/test; ln /tmp/test /tmp/[' and *prepend*
    your $PATH with chosen location, e.g. PATH=/tmp:$PATH. Alter-
    natively just replace system /bin/test and /bin/[ with the
    above script.

 * hpux64-ia64-cc fails blowfish test.

 Compiler bug, presumably at particular patch level. It should be noted
 that same compiler generates correct 32-bit code, a.k.a. hpux-ia64-cc
 target. Drop optimization level to +O2 when compiling 64-bit bf_skey.o.

 * no-engines generates errors.

 Unfortunately, the 'no-engines' configuration option currently doesn't
 work properly.  Use 'no-hw' and you'll will at least get no hardware
 support.  We'll see how we fix that on OpenSSL versions past 0.9.8.

 * 'make test' fails in BN_sqr [commonly with "error 139" denoting SIGSEGV]
   if elder GNU binutils were deployed to link shared libcrypto.so.

 As subject suggests the failure is caused by a bug in elder binutils,
 either as or ld, and was observed on FreeBSD and Linux. There are two
 options. First is naturally to upgrade binutils, the second one - to
 reconfigure with additional no-sse2 [or 386] option passed to ./config.

 * If configured with ./config no-dso, toolkit still gets linked with -ldl,
   which most notably poses a problem when linking with dietlibc.

 We don't have framework to associate -ldl with no-dso, therefore the only
 way is to edit Makefile right after ./config no-dso and remove -ldl from
 EX_LIBS line.

 * hpux-parisc2-cc no-asm build fails with SEGV in ECDSA/DH.

 Compiler bug, presumably at particular patch level. Remaining
 hpux*-parisc*-cc configurations can be affected too. Drop optimization
 level to +O2 when compiling bn_nist.o.

 * solaris64-sparcv9-cc link failure

 Solaris 8 ar can fail to maintain symbol table in .a, which results in
 link failures. Apply 109147-09 or later or modify Makefile generated
 by ./Configure solaris64-sparcv9-cc and replace RANLIB assignment with

 	RANLIB= /usr/ccs/bin/ar rs
	* System libcrypto.dylib and libssl.dylib are used by system ld on MacOS X.


	NOTE: The problem described here only applies when OpenSSL isn't built
	with shared library support (i.e. without the "shared" configuration
	option). If you build with shared library support, you will have no
	problems as long as you set up DYLD_LIBRARY_PATH properly at all times.


	This is really a misfeature in ld, which seems to look for .dylib libraries
	along the whole library path before it bothers looking for .a libraries. This
	means that -L switches won't matter unless OpenSSL is built with shared
	library support.

	The workaround may be to change the following lines in apps/Makefile and
	test/Makefile:

	LIBCRYPTO=-L.. -lcrypto
	LIBSSL=-L.. -lssl

	to:

	LIBCRYPTO=../libcrypto.a
	LIBSSL=../libssl.a

	It's possible that something similar is needed for shared library support
	as well. That hasn't been well tested yet.


	Another solution that many seem to recommend is to move the libraries
	/usr/lib/libcrypto.0.9.dylib, /usr/lib/libssl.0.9.dylib to a different
	directory, build and install OpenSSL and anything that depends on your
	build, then move libcrypto.0.9.dylib and libssl.0.9.dylib back to their
	original places. Note that the version numbers on those two libraries
	may differ on your machine.


	As long as Apple doesn't fix the problem with ld, this problem building
	OpenSSL will remain as is. Well, the problem was addressed in 0.9.8f by
	passing -Wl,-search_paths_first, but it's unknown if the flag was
	supported from the initial MacOS X release.


	* Parallell make leads to errors

	While running tests, running a parallell make is a bad idea. Many test
	scripts use the same name for output and input files, which means different
	will interfere with each other and lead to test failure.

	The solution is simple for now: don't run parallell make when testing.


	* Bugs in gcc triggered

	- According to a problem report, there are bugs in gcc 3.0 that are
	triggered by some of the code in OpenSSL, more specifically in
	PEM_get_EVP_CIPHER_INFO(). The triggering code is the following:

	header+=11;
	if (*header != '4') return(0); header++;
	if (*header != ',') return(0); header++;

	What happens is that gcc might optimize a little too agressively, and
	you end up with an extra incrementation when *header != '4'.

	We recommend that you upgrade gcc to as high a 3.x version as you can.

	- According to multiple problem reports, some of our message digest
	implementations trigger bug[s] in code optimizer in gcc 3.3 for sparc64
	and gcc 2.96 for ppc. Former fails to complete RIPEMD160 test, while
	latter - SHA one.

	The recomendation is to upgrade your compiler. This naturally applies to
	other similar cases.

	- There is a subtle Solaris x86-specific gcc run-time environment bug, which
	"falls between" OpenSSL [0.9.8 and later], Solaris ld and GCC. The bug
	manifests itself as Segmentation Fault upon early application start-up.
	The problem can be worked around by patching the environment according to
	http://www.openssl.org/~appro/values.c.

	* solaris64-sparcv9-cc SHA-1 performance with WorkShop 6 compiler.

	As subject suggests SHA-1 might perform poorly (4 times slower)
	if compiled with WorkShop 6 compiler and -xarch=v9. The cause for
	this seems to be the fact that compiler emits multiplication to
	perform shift operations:-( To work the problem around configure
	with './Configure solaris64-sparcv9-cc -DMD32_REG_T=int'.

	* Problems with hp-parisc2-cc target when used with "no-asm" flag

	When using the hp-parisc2-cc target, wrong bignum code is generated.
	This is due to the SIXTY_FOUR_BIT build being compiled with the +O3
	aggressive optimization.
	The problem manifests itself by the BN_kronecker test hanging in an
	endless loop. Reason: the BN_kronecker test calls BN_generate_prime()
	which itself hangs. The reason could be tracked down to the bn_mul_comba8()
	function in bn_asm.c. At some occasions the higher 32bit value of r[7]
	is off by 1 (meaning: calculated=shouldbe+1). Further analysis failed,
	as no debugger support possible at +O3 and additional fprintf()'s
	introduced fixed the bug, therefore it is most likely a bug in the
	optimizer.
	The bug was found in the BN_kronecker test but may also lead to
	failures in other parts of the code.
	(See Ticket #426.)

	Workaround: modify the target to +O2 when building with no-asm.

	* Problems building shared libraries on SCO OpenServer Release 5.0.6
	with gcc 2.95.3

	The symptoms appear when running the test suite, more specifically
	test/ectest, with the following result:

	OSSL_LIBPATH="`cd ..; pwd`"; LD_LIBRARY_PATH="$OSSL_LIBPATH:$LD_LIBRARY_PATH"; DYLD_LIBRARY_PATH="$OSSL_LIBPATH:$DYLD_LIBRARY_PATH"; SHLIB_PATH="$OSSL_LIBPATH:$SHLIB_PATH"; LIBPATH="$OSSL_LIBPATH:$LIBPATH"; if [ "debug-sco5-gcc" = "Cygwin" ]; then PATH="${LIBPATH}:$PATH"; fi; export LD_LIBRARY_PATH DYLD_LIBRARY_PATH SHLIB_PATH LIBPATH PATH; ./ectest
	ectest.c:186: ABORT

	The cause of the problem seems to be that isxdigit(), called from
	BN_hex2bn(), returns 0 on a perfectly legitimate hex digit. Further
	investigation shows that any of the isxxx() macros return 0 on any
	input. A direct look in the information array that the isxxx() use,
	called __ctype, shows that it contains all zeroes...

	Taking a look at the newly created libcrypto.so with nm, one can see
	that the variable __ctype is defined in libcrypto's .bss (which
	explains why it is filled with zeroes):

	$ nm -Pg libcrypto.so \| grep __ctype
	__ctype B 0011659c
	__ctype2 U

	Curiously, __ctype2 is undefined, in spite of being declared in
	/usr/include/ctype.h in exactly the same way as __ctype.

	Any information helping to solve this issue would be deeply
	appreciated.

	NOTE: building non-shared doesn't come with this problem.

	* ULTRIX build fails with shell errors, such as "bad substitution"
	and "test: argument expected"

	The problem is caused by ULTRIX /bin/sh supporting only original
	Bourne shell syntax/semantics, and the trouble is that the vast
	majority is so accustomed to more modern syntax, that very few
	people [if any] would recognize the ancient syntax even as valid.
	This inevitably results in non-trivial scripts breaking on ULTRIX,
	and OpenSSL isn't an exclusion. Fortunately there is workaround,
	hire /bin/ksh to do the job /bin/sh fails to do.

	1. Trick make(1) to use /bin/ksh by setting up following environ-
	ment variables prior you execute ./Configure and make:

	PROG_ENV=POSIX
	MAKESHELL=/bin/ksh
	export PROG_ENV MAKESHELL

	or if your shell is csh-compatible:

	setenv PROG_ENV POSIX
	setenv MAKESHELL /bin/ksh

	2. Trick /bin/sh to use alternative expression evaluator. Create
	following 'test' script for example in /tmp:

	#!/bin/ksh
	${0##*/} "$@"

	Then 'chmod a+x /tmp/test; ln /tmp/test /tmp/[' and prepend
	your $PATH with chosen location, e.g. PATH=/tmp:$PATH. Alter-
	natively just replace system /bin/test and /bin/[ with the
	above script.

	* hpux64-ia64-cc fails blowfish test.

	Compiler bug, presumably at particular patch level. It should be noted
	that same compiler generates correct 32-bit code, a.k.a. hpux-ia64-cc
	target. Drop optimization level to +O2 when compiling 64-bit bf_skey.o.

	* no-engines generates errors.

	Unfortunately, the 'no-engines' configuration option currently doesn't
	work properly. Use 'no-hw' and you'll will at least get no hardware
	support. We'll see how we fix that on OpenSSL versions past 0.9.8.

	* 'make test' fails in BN_sqr [commonly with "error 139" denoting SIGSEGV]
	if elder GNU binutils were deployed to link shared libcrypto.so.

	As subject suggests the failure is caused by a bug in elder binutils,
	either as or ld, and was observed on FreeBSD and Linux. There are two
	options. First is naturally to upgrade binutils, the second one - to
	reconfigure with additional no-sse2 [or 386] option passed to ./config.

	* If configured with ./config no-dso, toolkit still gets linked with -ldl,
	which most notably poses a problem when linking with dietlibc.

	We don't have framework to associate -ldl with no-dso, therefore the only
	way is to edit Makefile right after ./config no-dso and remove -ldl from
	EX_LIBS line.

	* hpux-parisc2-cc no-asm build fails with SEGV in ECDSA/DH.

	Compiler bug, presumably at particular patch level. Remaining
	hpux-parisc-cc configurations can be affected too. Drop optimization
	level to +O2 when compiling bn_nist.o.

	* solaris64-sparcv9-cc link failure

	Solaris 8 ar can fail to maintain symbol table in .a, which results in
	link failures. Apply 109147-09 or later or modify Makefile generated
	by ./Configure solaris64-sparcv9-cc and replace RANLIB assignment with

	RANLIB= /usr/ccs/bin/ar rs