[PATCH v2] Cygwin: spawn: Fix segfalt when too many command line args are specified.

[PATCH v2] Cygwin: spawn: Fix segfalt when too many command line args are specified.

[Takashi Yano]

Previously, the number of command line args was not checked for
cygwin process. Due to this, segmentation fault was caused if too
many command line args are specified.
https://cygwin.com/pipermail/cygwin/2023-August/254333.html

Since char *argv[argc + 1] is placed on the stack in dll_crt0_1(),
STATUS_STACK_OVERFLOW occurs if the stack does not have enough
space.

With this patch, the total length of the arguments and the size of
argv[] is restricted to 1/4 of total stack size for the process, and
spawnve() returns E2BIG if the size exceeds the limit.

Reported-by: Ed Morton <mortoneccc@comcast.net>
Signed-off-by: Takashi Yano <takashi.yano@nifty.ne.jp>
---
 winsup/cygwin/release/3.4.9 |  3 +++
 winsup/cygwin/spawn.cc      | 43 ++++++++++++++++++++++++++++++++++++-
 winsup/cygwin/sysconf.cc    |  9 +++++++-
 3 files changed, 53 insertions(+), 2 deletions(-)

diff --git a/winsup/cygwin/release/3.4.9 b/winsup/cygwin/release/3.4.9
index 2f2da9e13..53c4e5fc8 100644
--- a/winsup/cygwin/release/3.4.9
+++ b/winsup/cygwin/release/3.4.9
@@ -8,3 +8,6 @@ Bug Fixes
 - For the time being, disable creating special files using mknod/mkfifo
   on NFS.
   Addresses: https://cygwin.com/pipermail/cygwin/2023-August/254266.html
+
+- Fix segfault when too many command line args are specified.
+  Addresses: https://cygwin.com/pipermail/cygwin/2023-August/254333.html
diff --git a/winsup/cygwin/spawn.cc b/winsup/cygwin/spawn.cc
index 32ba5b377..06b62cd42 100644
--- a/winsup/cygwin/spawn.cc
+++ b/winsup/cygwin/spawn.cc
@@ -273,6 +273,29 @@ extern "C" void __posix_spawn_sem_release (void *sem, int error);
 
 extern DWORD mutex_timeout; /* defined in fhandler_termios.cc */
 
+static size_t
+get_stack_size (const WCHAR *filename)
+{
+  HANDLE h;
+  h = CreateFileW (filename, GENERIC_READ, FILE_SHARE_READ,
+		   NULL, OPEN_EXISTING, 0, NULL);
+  char buf[1024];
+  DWORD n;
+  ReadFile (h, buf, sizeof (buf), &n, 0);
+  CloseHandle (h);
+  IMAGE_NT_HEADERS32 *p = (IMAGE_NT_HEADERS32 *) memmem (buf, n, "PE\0\0", 4);
+  if (!p)
+    return 0;
+  if ((char *) &p->OptionalHeader.SizeOfStackCommit > buf + n)
+    return 0; /* buf[] is not enough */
+  if (p->OptionalHeader.Magic == IMAGE_NT_OPTIONAL_HDR32_MAGIC)
+    return p->OptionalHeader.SizeOfStackReserve;
+  IMAGE_NT_HEADERS64 *p64 = (IMAGE_NT_HEADERS64 *) p;
+  if ((char *) &p64->OptionalHeader.SizeOfStackCommit > buf + n)
+    return 0; /* buf[] is not enough */
+  return p64->OptionalHeader.SizeOfStackReserve;
+}
+
 int
 child_info_spawn::worker (const char *prog_arg, const char *const *argv,
 			  const char *const envp[], int mode,
@@ -340,8 +363,9 @@ child_info_spawn::worker (const char *prog_arg, const char *const *argv,
 	 We need to quote any argument that has whitespace or embedded "'s.  */
 
       int ac;
+      size_t arg_len = 0;
       for (ac = 0; argv[ac]; ac++)
-	/* nothing */;
+	arg_len += strlen (argv[ac]) + 1;
 
       int err;
       const char *ext;
@@ -610,6 +634,23 @@ child_info_spawn::worker (const char *prog_arg, const char *const *argv,
 	    }
 	}
 
+      if (iscygwin ())
+	{
+	  size_t child_stack_size = get_stack_size (runpath);
+	  /* char *argv[] will be placed on the stack in dll_crt0_1(), so
+	     restrict total argument length and the size of argv[] to 1/4
+	     of total stack size. */
+	  arg_len = max (arg_len, sizeof (char *) * ac);
+	  bool too_many_args = child_stack_size ?
+	    arg_len > child_stack_size / 4 : ac >= MAXWINCMDLEN;
+	  if (too_many_args)
+	    {
+	      set_errno (E2BIG);
+	      res = -1;
+	      __leave;
+	    }
+	}
+
       bool no_pcon = mode != _P_OVERLAY && mode != _P_WAIT;
       term_spawn_worker.setup (iscygwin (), handle (fileno_stdin, false),
 			       runpath, no_pcon, reset_sendsig, envblock);
diff --git a/winsup/cygwin/sysconf.cc b/winsup/cygwin/sysconf.cc
index 2db92e4de..6cb2aecd0 100644
--- a/winsup/cygwin/sysconf.cc
+++ b/winsup/cygwin/sysconf.cc
@@ -21,6 +21,13 @@ details. */
 #include "cpuid.h"
 #include "clock.h"
 
+#define DEFAULT_STACKGUARD (wincap.def_guard_page_size() + wincap.page_size ())
+static long
+get_arg_max (int in)
+{
+  return (long) (get_rlimit_stack () + DEFAULT_STACKGUARD) / 4;
+}
+
 static long
 get_page_size (int in)
 {
@@ -485,7 +492,7 @@ static struct
     };
 } sca[] =
 {
-  {cons, {c:ARG_MAX}},			/*   0, _SC_ARG_MAX */
+  {func, {f:get_arg_max}},		/*   0, _SC_ARG_MAX */
   {cons, {c:CHILD_MAX}},		/*   1, _SC_CHILD_MAX */
   {cons, {c:CLOCKS_PER_SEC}},		/*   2, _SC_CLK_TCK */
   {cons, {c:NGROUPS_MAX}},		/*   3, _SC_NGROUPS_MAX */
-- 
2.39.0

Re: [PATCH v2] Cygwin: spawn: Fix segfalt when too many command line args are specified.

[Corinna Vinschen]

On Aug 28 18:46, Takashi Yano wrote:
> Previously, the number of command line args was not checked for
> cygwin process. Due to this, segmentation fault was caused if too
> many command line args are specified.
> https://cygwin.com/pipermail/cygwin/2023-August/254333.html
> 
> Since char *argv[argc + 1] is placed on the stack in dll_crt0_1(),
> STATUS_STACK_OVERFLOW occurs if the stack does not have enough
> space.
> 
> With this patch, the total length of the arguments and the size of
> argv[] is restricted to 1/4 of total stack size for the process, and
> spawnve() returns E2BIG if the size exceeds the limit.
> [...]
> +static size_t
> +get_stack_size (const WCHAR *filename)
> +{
> +  HANDLE h;
> +  h = CreateFileW (filename, GENERIC_READ, FILE_SHARE_READ,
> +		   NULL, OPEN_EXISTING, 0, NULL);
> +  char buf[1024];
> +  DWORD n;
> +  ReadFile (h, buf, sizeof (buf), &n, 0);
> +  CloseHandle (h);
> +  IMAGE_NT_HEADERS32 *p = (IMAGE_NT_HEADERS32 *) memmem (buf, n, "PE\0\0", 4);
> +  if (!p)
> +    return 0;
> +  if ((char *) &p->OptionalHeader.SizeOfStackCommit > buf + n)
> +    return 0; /* buf[] is not enough */
> +  if (p->OptionalHeader.Magic == IMAGE_NT_OPTIONAL_HDR32_MAGIC)
> +    return p->OptionalHeader.SizeOfStackReserve;
> +  IMAGE_NT_HEADERS64 *p64 = (IMAGE_NT_HEADERS64 *) p;
> +  if ((char *) &p64->OptionalHeader.SizeOfStackCommit > buf + n)
> +    return 0; /* buf[] is not enough */
> +  return p64->OptionalHeader.SizeOfStackReserve;
> +}

Sorry, but this proposal is too complicated, IMHO.

Checking the stacksize in the file header for each single execve
is quite a bit time consuming, isn't it?

The question is rather, why storing argv on the stack at all?  I guess
the original idea was that argv is always a rather overseeable number.
But it doesn't have to stay on the stack.

I tried this simple patch:

diff --git a/winsup/cygwin/dcrt0.cc b/winsup/cygwin/dcrt0.cc
index 49b7a44aeb15..961dea4ab993 100644
--- a/winsup/cygwin/dcrt0.cc
+++ b/winsup/cygwin/dcrt0.cc
@@ -978,11 +978,8 @@ dll_crt0_1 (void *)
 	 a change to an element of argv[] it does not affect Cygwin's argv.
 	 Changing the the contents of what argv[n] points to will still
 	 affect Cygwin.  This is similar (but not exactly like) Linux. */
-      char *newargv[__argc + 1];
-      char **nav = newargv;
-      char **oav = __argv;
-      while ((*nav++ = *oav++) != NULL)
-	continue;
+      char **newargv = (char **) malloc ((__argc + 1) * sizeof (char **));
+      memcpy (newargv, __argv, (__argc + 1) * sizeof (char **));
       /* Handle any signals which may have arrived */
       sig_dispatch_pending (false);
       _my_tls.call_signal_handler ();

and the testcase `LC_ALL=C sed 's/x/y/' $(seq 1000000)' simply ran
as desired.  Combined with a bit of error checking...

> diff --git a/winsup/cygwin/sysconf.cc b/winsup/cygwin/sysconf.cc
> index 2db92e4de..6cb2aecd0 100644
> --- a/winsup/cygwin/sysconf.cc
> +++ b/winsup/cygwin/sysconf.cc
> @@ -21,6 +21,13 @@ details. */
>  #include "cpuid.h"
>  #include "clock.h"
>  
> +#define DEFAULT_STACKGUARD (wincap.def_guard_page_size() + wincap.page_size ())
> +static long
> +get_arg_max (int in)
> +{
> +  return (long) (get_rlimit_stack () + DEFAULT_STACKGUARD) / 4;
> +}
> +
>  static long
>  get_page_size (int in)
>  {
> @@ -485,7 +492,7 @@ static struct
>      };
>  } sca[] =
>  {
> -  {cons, {c:ARG_MAX}},			/*   0, _SC_ARG_MAX */
> +  {func, {f:get_arg_max}},		/*   0, _SC_ARG_MAX */
>    {cons, {c:CHILD_MAX}},		/*   1, _SC_CHILD_MAX */
>    {cons, {c:CLOCKS_PER_SEC}},		/*   2, _SC_CLK_TCK */
>    {cons, {c:NGROUPS_MAX}},		/*   3, _SC_NGROUPS_MAX */
> -- 
> 2.39.0

Along these lines, there's no reason to couple SC_ARG_MAX to the
size of the stack.  I'd propose to return the value 2097152.  It's
the default value returned by getconf ARG_MAX on LInx as well.

Corinna

Re: [PATCH v2] Cygwin: spawn: Fix segfalt when too many command line args are specified.

[Corinna Vinschen]

On Aug 28 12:57, Corinna Vinschen wrote:
> On Aug 28 18:46, Takashi Yano wrote:
> > Previously, the number of command line args was not checked for
> > cygwin process. Due to this, segmentation fault was caused if too
> > many command line args are specified.
> > https://cygwin.com/pipermail/cygwin/2023-August/254333.html
> > 
> > Since char *argv[argc + 1] is placed on the stack in dll_crt0_1(),
> > STATUS_STACK_OVERFLOW occurs if the stack does not have enough
> > space.
> > 
> > With this patch, the total length of the arguments and the size of
> > argv[] is restricted to 1/4 of total stack size for the process, and
> > spawnve() returns E2BIG if the size exceeds the limit.
> > [...]
> I tried this simple patch:
> 
> diff --git a/winsup/cygwin/dcrt0.cc b/winsup/cygwin/dcrt0.cc
> index 49b7a44aeb15..961dea4ab993 100644
> --- a/winsup/cygwin/dcrt0.cc
> +++ b/winsup/cygwin/dcrt0.cc
> @@ -978,11 +978,8 @@ dll_crt0_1 (void *)
>  	 a change to an element of argv[] it does not affect Cygwin's argv.
>  	 Changing the the contents of what argv[n] points to will still
>  	 affect Cygwin.  This is similar (but not exactly like) Linux. */
> -      char *newargv[__argc + 1];
> -      char **nav = newargv;
> -      char **oav = __argv;
> -      while ((*nav++ = *oav++) != NULL)
> -	continue;
> +      char **newargv = (char **) malloc ((__argc + 1) * sizeof (char **));
> +      memcpy (newargv, __argv, (__argc + 1) * sizeof (char **));
>        /* Handle any signals which may have arrived */
>        sig_dispatch_pending (false);
>        _my_tls.call_signal_handler ();
> 
> and the testcase `LC_ALL=C sed 's/x/y/' $(seq 1000000)' simply ran
> as desired.  Combined with a bit of error checking...

We may also get away with storing it in the Windows heap, but I didn't
test this .

> Along these lines, there's no reason to couple SC_ARG_MAX to the
> size of the stack.  I'd propose to return the value 2097152.  It's
> the default value returned by getconf ARG_MAX on LInx as well.

Oh, and we can carefully check in child_info_spawn::worker that
the args are not taking more space than the value returned by
SC_ARG_MAX and return E2BIG if so.  We do this when checking the
args for non-Cygwin apps anyway.


Thanks,
Corinna

Re: [PATCH v2] Cygwin: spawn: Fix segfalt when too many command line args are specified.

[Corinna Vinschen]

On Aug 28 13:09, Corinna Vinschen wrote:
> On Aug 28 12:57, Corinna Vinschen wrote:
> > On Aug 28 18:46, Takashi Yano wrote:
> > > Previously, the number of command line args was not checked for
> > > cygwin process. Due to this, segmentation fault was caused if too
> > > many command line args are specified.
> > > https://cygwin.com/pipermail/cygwin/2023-August/254333.html
> > > 
> > > Since char *argv[argc + 1] is placed on the stack in dll_crt0_1(),
> > > STATUS_STACK_OVERFLOW occurs if the stack does not have enough
> > > space.
> > > 
> > > With this patch, the total length of the arguments and the size of
> > > argv[] is restricted to 1/4 of total stack size for the process, and
> > > spawnve() returns E2BIG if the size exceeds the limit.
> > > [...]
> > I tried this simple patch:
> > 
> > diff --git a/winsup/cygwin/dcrt0.cc b/winsup/cygwin/dcrt0.cc
> > index 49b7a44aeb15..961dea4ab993 100644
> > --- a/winsup/cygwin/dcrt0.cc
> > +++ b/winsup/cygwin/dcrt0.cc
> > @@ -978,11 +978,8 @@ dll_crt0_1 (void *)
> >  	 a change to an element of argv[] it does not affect Cygwin's argv.
> >  	 Changing the the contents of what argv[n] points to will still
> >  	 affect Cygwin.  This is similar (but not exactly like) Linux. */
> > -      char *newargv[__argc + 1];
> > -      char **nav = newargv;
> > -      char **oav = __argv;
> > -      while ((*nav++ = *oav++) != NULL)
> > -	continue;
> > +      char **newargv = (char **) malloc ((__argc + 1) * sizeof (char **));
> > +      memcpy (newargv, __argv, (__argc + 1) * sizeof (char **));
> >        /* Handle any signals which may have arrived */
> >        sig_dispatch_pending (false);
> >        _my_tls.call_signal_handler ();
> > 
> > and the testcase `LC_ALL=C sed 's/x/y/' $(seq 1000000)' simply ran
> > as desired.  Combined with a bit of error checking...
> 
> We may also get away with storing it in the Windows heap, but I didn't
> test this .

No, we don't. The Windows heap would not be inherited by a forked
child and argv would be lost.  Sounds not great.


Corinna

Re: [PATCH v2] Cygwin: spawn: Fix segfalt when too many command line args are specified.

[Takashi Yano]

On Mon, 28 Aug 2023 13:09:18 +0200
Corinna Vinschen wrote:
> On Aug 28 12:57, Corinna Vinschen wrote:
> > On Aug 28 18:46, Takashi Yano wrote:
> > > Previously, the number of command line args was not checked for
> > > cygwin process. Due to this, segmentation fault was caused if too
> > > many command line args are specified.
> > > https://cygwin.com/pipermail/cygwin/2023-August/254333.html
> > > 
> > > Since char *argv[argc + 1] is placed on the stack in dll_crt0_1(),
> > > STATUS_STACK_OVERFLOW occurs if the stack does not have enough
> > > space.
> > > 
> > > With this patch, the total length of the arguments and the size of
> > > argv[] is restricted to 1/4 of total stack size for the process, and
> > > spawnve() returns E2BIG if the size exceeds the limit.
> > > [...]
> > I tried this simple patch:
> > 
> > diff --git a/winsup/cygwin/dcrt0.cc b/winsup/cygwin/dcrt0.cc
> > index 49b7a44aeb15..961dea4ab993 100644
> > --- a/winsup/cygwin/dcrt0.cc
> > +++ b/winsup/cygwin/dcrt0.cc
> > @@ -978,11 +978,8 @@ dll_crt0_1 (void *)
> >  	 a change to an element of argv[] it does not affect Cygwin's argv.
> >  	 Changing the the contents of what argv[n] points to will still
> >  	 affect Cygwin.  This is similar (but not exactly like) Linux. */
> > -      char *newargv[__argc + 1];
> > -      char **nav = newargv;
> > -      char **oav = __argv;
> > -      while ((*nav++ = *oav++) != NULL)
> > -	continue;
> > +      char **newargv = (char **) malloc ((__argc + 1) * sizeof (char **));
> > +      memcpy (newargv, __argv, (__argc + 1) * sizeof (char **));
> >        /* Handle any signals which may have arrived */
> >        sig_dispatch_pending (false);
> >        _my_tls.call_signal_handler ();
> > 
> > and the testcase `LC_ALL=C sed 's/x/y/' $(seq 1000000)' simply ran
> > as desired.  Combined with a bit of error checking...
> 
> We may also get away with storing it in the Windows heap, but I didn't
> test this .
> 
> > Along these lines, there's no reason to couple SC_ARG_MAX to the
> > size of the stack.  I'd propose to return the value 2097152.  It's
> > the default value returned by getconf ARG_MAX on LInx as well.
> 
> Oh, and we can carefully check in child_info_spawn::worker that
> the args are not taking more space than the value returned by
> SC_ARG_MAX and return E2BIG if so.  We do this when checking the
> args for non-Cygwin apps anyway.

As for non-cygwin apps, the check seems to be done in linebuf::fromargv()
in winf.cc. Further, dcr0.cc code does not affect to non-cygwin app.
So my patch check only for cygwin apps case.

-- 
Takashi Yano <takashi.yano@nifty.ne.jp>

Re: [PATCH v2] Cygwin: spawn: Fix segfalt when too many command line args are specified.

[Takashi Yano]

On Mon, 28 Aug 2023 12:57:19 +0200
Corinna Vinschen wrote:
> On Aug 28 18:46, Takashi Yano wrote:
> > Previously, the number of command line args was not checked for
> > cygwin process. Due to this, segmentation fault was caused if too
> > many command line args are specified.
> > https://cygwin.com/pipermail/cygwin/2023-August/254333.html
> > 
> > Since char *argv[argc + 1] is placed on the stack in dll_crt0_1(),
> > STATUS_STACK_OVERFLOW occurs if the stack does not have enough
> > space.
> > 
> > With this patch, the total length of the arguments and the size of
> > argv[] is restricted to 1/4 of total stack size for the process, and
> > spawnve() returns E2BIG if the size exceeds the limit.
> > [...]
> > +static size_t
> > +get_stack_size (const WCHAR *filename)
> > +{
> > +  HANDLE h;
> > +  h = CreateFileW (filename, GENERIC_READ, FILE_SHARE_READ,
> > +		   NULL, OPEN_EXISTING, 0, NULL);
> > +  char buf[1024];
> > +  DWORD n;
> > +  ReadFile (h, buf, sizeof (buf), &n, 0);
> > +  CloseHandle (h);
> > +  IMAGE_NT_HEADERS32 *p = (IMAGE_NT_HEADERS32 *) memmem (buf, n, "PE\0\0", 4);
> > +  if (!p)
> > +    return 0;
> > +  if ((char *) &p->OptionalHeader.SizeOfStackCommit > buf + n)
> > +    return 0; /* buf[] is not enough */
> > +  if (p->OptionalHeader.Magic == IMAGE_NT_OPTIONAL_HDR32_MAGIC)
> > +    return p->OptionalHeader.SizeOfStackReserve;
> > +  IMAGE_NT_HEADERS64 *p64 = (IMAGE_NT_HEADERS64 *) p;
> > +  if ((char *) &p64->OptionalHeader.SizeOfStackCommit > buf + n)
> > +    return 0; /* buf[] is not enough */
> > +  return p64->OptionalHeader.SizeOfStackReserve;
> > +}
> 
> Sorry, but this proposal is too complicated, IMHO.
> 
> Checking the stacksize in the file header for each single execve
> is quite a bit time consuming, isn't it?

I did this because the stack size of an executable can vary by
peflags -x command or -Wl,--stack=xxxxx option.

> The question is rather, why storing argv on the stack at all?  I guess
> the original idea was that argv is always a rather overseeable number.
> But it doesn't have to stay on the stack.
> 
> I tried this simple patch:
> 
> diff --git a/winsup/cygwin/dcrt0.cc b/winsup/cygwin/dcrt0.cc
> index 49b7a44aeb15..961dea4ab993 100644
> --- a/winsup/cygwin/dcrt0.cc
> +++ b/winsup/cygwin/dcrt0.cc
> @@ -978,11 +978,8 @@ dll_crt0_1 (void *)
>  	 a change to an element of argv[] it does not affect Cygwin's argv.
>  	 Changing the the contents of what argv[n] points to will still
>  	 affect Cygwin.  This is similar (but not exactly like) Linux. */
> -      char *newargv[__argc + 1];
> -      char **nav = newargv;
> -      char **oav = __argv;
> -      while ((*nav++ = *oav++) != NULL)
> -	continue;
> +      char **newargv = (char **) malloc ((__argc + 1) * sizeof (char **));
> +      memcpy (newargv, __argv, (__argc + 1) * sizeof (char **));
>        /* Handle any signals which may have arrived */
>        sig_dispatch_pending (false);
>        _my_tls.call_signal_handler ();
> 
> and the testcase `LC_ALL=C sed 's/x/y/' $(seq 1000000)' simply ran
> as desired.  Combined with a bit of error checking...
> 
> > diff --git a/winsup/cygwin/sysconf.cc b/winsup/cygwin/sysconf.cc
> > index 2db92e4de..6cb2aecd0 100644
> > --- a/winsup/cygwin/sysconf.cc
> > +++ b/winsup/cygwin/sysconf.cc
> > @@ -21,6 +21,13 @@ details. */
> >  #include "cpuid.h"
> >  #include "clock.h"
> >  
> > +#define DEFAULT_STACKGUARD (wincap.def_guard_page_size() + wincap.page_size ())
> > +static long
> > +get_arg_max (int in)
> > +{
> > +  return (long) (get_rlimit_stack () + DEFAULT_STACKGUARD) / 4;
> > +}
> > +
> >  static long
> >  get_page_size (int in)
> >  {
> > @@ -485,7 +492,7 @@ static struct
> >      };
> >  } sca[] =
> >  {
> > -  {cons, {c:ARG_MAX}},			/*   0, _SC_ARG_MAX */
> > +  {func, {f:get_arg_max}},		/*   0, _SC_ARG_MAX */
> >    {cons, {c:CHILD_MAX}},		/*   1, _SC_CHILD_MAX */
> >    {cons, {c:CLOCKS_PER_SEC}},		/*   2, _SC_CLK_TCK */
> >    {cons, {c:NGROUPS_MAX}},		/*   3, _SC_NGROUPS_MAX */
> > -- 
> > 2.39.0
> 
> Along these lines, there's no reason to couple SC_ARG_MAX to the
> size of the stack.  I'd propose to return the value 2097152.  It's
> the default value returned by getconf ARG_MAX on LInx as well.

The default value of 2097152 in Linux comes from 1/4 of the stack
size. It varies by ulimit -s command. So I tried to imitate it.

-- 
Takashi Yano <takashi.yano@nifty.ne.jp>

Re: [PATCH v2] Cygwin: spawn: Fix segfalt when too many command line args are specified.

[Takashi Yano]

On Mon, 28 Aug 2023 21:09:53 +0900
Takashi Yano wrote:
> On Mon, 28 Aug 2023 12:57:19 +0200
> Corinna Vinschen wrote:
> > On Aug 28 18:46, Takashi Yano wrote:
> > > Previously, the number of command line args was not checked for
> > > cygwin process. Due to this, segmentation fault was caused if too
> > > many command line args are specified.
> > > https://cygwin.com/pipermail/cygwin/2023-August/254333.html
> > > 
> > > Since char *argv[argc + 1] is placed on the stack in dll_crt0_1(),
> > > STATUS_STACK_OVERFLOW occurs if the stack does not have enough
> > > space.
> > > 
> > > With this patch, the total length of the arguments and the size of
> > > argv[] is restricted to 1/4 of total stack size for the process, and
> > > spawnve() returns E2BIG if the size exceeds the limit.
> > > [...]
> > > +static size_t
> > > +get_stack_size (const WCHAR *filename)
> > > +{
> > > +  HANDLE h;
> > > +  h = CreateFileW (filename, GENERIC_READ, FILE_SHARE_READ,
> > > +		   NULL, OPEN_EXISTING, 0, NULL);
> > > +  char buf[1024];
> > > +  DWORD n;
> > > +  ReadFile (h, buf, sizeof (buf), &n, 0);
> > > +  CloseHandle (h);
> > > +  IMAGE_NT_HEADERS32 *p = (IMAGE_NT_HEADERS32 *) memmem (buf, n, "PE\0\0", 4);
> > > +  if (!p)
> > > +    return 0;
> > > +  if ((char *) &p->OptionalHeader.SizeOfStackCommit > buf + n)
> > > +    return 0; /* buf[] is not enough */
> > > +  if (p->OptionalHeader.Magic == IMAGE_NT_OPTIONAL_HDR32_MAGIC)
> > > +    return p->OptionalHeader.SizeOfStackReserve;
> > > +  IMAGE_NT_HEADERS64 *p64 = (IMAGE_NT_HEADERS64 *) p;
> > > +  if ((char *) &p64->OptionalHeader.SizeOfStackCommit > buf + n)
> > > +    return 0; /* buf[] is not enough */
> > > +  return p64->OptionalHeader.SizeOfStackReserve;
> > > +}
> > 
> > Sorry, but this proposal is too complicated, IMHO.
> > 
> > Checking the stacksize in the file header for each single execve
> > is quite a bit time consuming, isn't it?
> 
> I did this because the stack size of an executable can vary by
> peflags -x command or -Wl,--stack=xxxxx option.
> 
> > The question is rather, why storing argv on the stack at all?  I guess
> > the original idea was that argv is always a rather overseeable number.
> > But it doesn't have to stay on the stack.
> > 
> > I tried this simple patch:
> > 
> > diff --git a/winsup/cygwin/dcrt0.cc b/winsup/cygwin/dcrt0.cc
> > index 49b7a44aeb15..961dea4ab993 100644
> > --- a/winsup/cygwin/dcrt0.cc
> > +++ b/winsup/cygwin/dcrt0.cc
> > @@ -978,11 +978,8 @@ dll_crt0_1 (void *)
> >  	 a change to an element of argv[] it does not affect Cygwin's argv.
> >  	 Changing the the contents of what argv[n] points to will still
> >  	 affect Cygwin.  This is similar (but not exactly like) Linux. */
> > -      char *newargv[__argc + 1];
> > -      char **nav = newargv;
> > -      char **oav = __argv;
> > -      while ((*nav++ = *oav++) != NULL)
> > -	continue;
> > +      char **newargv = (char **) malloc ((__argc + 1) * sizeof (char **));
> > +      memcpy (newargv, __argv, (__argc + 1) * sizeof (char **));
> >        /* Handle any signals which may have arrived */
> >        sig_dispatch_pending (false);
> >        _my_tls.call_signal_handler ();
> > 
> > and the testcase `LC_ALL=C sed 's/x/y/' $(seq 1000000)' simply ran
> > as desired.  Combined with a bit of error checking...
> > 
> > > diff --git a/winsup/cygwin/sysconf.cc b/winsup/cygwin/sysconf.cc
> > > index 2db92e4de..6cb2aecd0 100644
> > > --- a/winsup/cygwin/sysconf.cc
> > > +++ b/winsup/cygwin/sysconf.cc
> > > @@ -21,6 +21,13 @@ details. */
> > >  #include "cpuid.h"
> > >  #include "clock.h"
> > >  
> > > +#define DEFAULT_STACKGUARD (wincap.def_guard_page_size() + wincap.page_size ())
> > > +static long
> > > +get_arg_max (int in)
> > > +{
> > > +  return (long) (get_rlimit_stack () + DEFAULT_STACKGUARD) / 4;
> > > +}
> > > +
> > >  static long
> > >  get_page_size (int in)
> > >  {
> > > @@ -485,7 +492,7 @@ static struct
> > >      };
> > >  } sca[] =
> > >  {
> > > -  {cons, {c:ARG_MAX}},			/*   0, _SC_ARG_MAX */
> > > +  {func, {f:get_arg_max}},		/*   0, _SC_ARG_MAX */
> > >    {cons, {c:CHILD_MAX}},		/*   1, _SC_CHILD_MAX */
> > >    {cons, {c:CLOCKS_PER_SEC}},		/*   2, _SC_CLK_TCK */
> > >    {cons, {c:NGROUPS_MAX}},		/*   3, _SC_NGROUPS_MAX */
> > > -- 
> > > 2.39.0
> > 
> > Along these lines, there's no reason to couple SC_ARG_MAX to the
> > size of the stack.  I'd propose to return the value 2097152.  It's
> > the default value returned by getconf ARG_MAX on LInx as well.
> 
> The default value of 2097152 in Linux comes from 1/4 of the stack
> size. It varies by ulimit -s command. So I tried to imitate it.

The value 2097152 is too large for default stack size of cygwin.
It is whole of the stack size.

-- 
Takashi Yano <takashi.yano@nifty.ne.jp>

Re: [PATCH v2] Cygwin: spawn: Fix segfalt when too many command line args are specified.

[Takashi Yano]

On Mon, 28 Aug 2023 12:57:19 +0200
Corinna Vinschen wrote:
> On Aug 28 18:46, Takashi Yano wrote:
> > Previously, the number of command line args was not checked for
> > cygwin process. Due to this, segmentation fault was caused if too
> > many command line args are specified.
> > https://cygwin.com/pipermail/cygwin/2023-August/254333.html
> > 
> > Since char *argv[argc + 1] is placed on the stack in dll_crt0_1(),
> > STATUS_STACK_OVERFLOW occurs if the stack does not have enough
> > space.
> > 
> > With this patch, the total length of the arguments and the size of
> > argv[] is restricted to 1/4 of total stack size for the process, and
> > spawnve() returns E2BIG if the size exceeds the limit.
> > [...]
> > +static size_t
> > +get_stack_size (const WCHAR *filename)
> > +{
> > +  HANDLE h;
> > +  h = CreateFileW (filename, GENERIC_READ, FILE_SHARE_READ,
> > +		   NULL, OPEN_EXISTING, 0, NULL);
> > +  char buf[1024];
> > +  DWORD n;
> > +  ReadFile (h, buf, sizeof (buf), &n, 0);
> > +  CloseHandle (h);
> > +  IMAGE_NT_HEADERS32 *p = (IMAGE_NT_HEADERS32 *) memmem (buf, n, "PE\0\0", 4);
> > +  if (!p)
> > +    return 0;
> > +  if ((char *) &p->OptionalHeader.SizeOfStackCommit > buf + n)
> > +    return 0; /* buf[] is not enough */
> > +  if (p->OptionalHeader.Magic == IMAGE_NT_OPTIONAL_HDR32_MAGIC)
> > +    return p->OptionalHeader.SizeOfStackReserve;
> > +  IMAGE_NT_HEADERS64 *p64 = (IMAGE_NT_HEADERS64 *) p;
> > +  if ((char *) &p64->OptionalHeader.SizeOfStackCommit > buf + n)
> > +    return 0; /* buf[] is not enough */
> > +  return p64->OptionalHeader.SizeOfStackReserve;
> > +}
> 
> Sorry, but this proposal is too complicated, IMHO.
> 
> Checking the stacksize in the file header for each single execve
> is quite a bit time consuming, isn't it?
> 
> The question is rather, why storing argv on the stack at all?  I guess
> the original idea was that argv is always a rather overseeable number.
> But it doesn't have to stay on the stack.
> 
> I tried this simple patch:
> 
> diff --git a/winsup/cygwin/dcrt0.cc b/winsup/cygwin/dcrt0.cc
> index 49b7a44aeb15..961dea4ab993 100644
> --- a/winsup/cygwin/dcrt0.cc
> +++ b/winsup/cygwin/dcrt0.cc
> @@ -978,11 +978,8 @@ dll_crt0_1 (void *)
>  	 a change to an element of argv[] it does not affect Cygwin's argv.
>  	 Changing the the contents of what argv[n] points to will still
>  	 affect Cygwin.  This is similar (but not exactly like) Linux. */
> -      char *newargv[__argc + 1];
> -      char **nav = newargv;
> -      char **oav = __argv;
> -      while ((*nav++ = *oav++) != NULL)
> -	continue;
> +      char **newargv = (char **) malloc ((__argc + 1) * sizeof (char **));
> +      memcpy (newargv, __argv, (__argc + 1) * sizeof (char **));
>        /* Handle any signals which may have arrived */
>        sig_dispatch_pending (false);
>        _my_tls.call_signal_handler ();
> 
> and the testcase `LC_ALL=C sed 's/x/y/' $(seq 1000000)' simply ran
> as desired.  Combined with a bit of error checking...
> 
> > diff --git a/winsup/cygwin/sysconf.cc b/winsup/cygwin/sysconf.cc
> > index 2db92e4de..6cb2aecd0 100644
> > --- a/winsup/cygwin/sysconf.cc
> > +++ b/winsup/cygwin/sysconf.cc
> > @@ -21,6 +21,13 @@ details. */
> >  #include "cpuid.h"
> >  #include "clock.h"
> >  
> > +#define DEFAULT_STACKGUARD (wincap.def_guard_page_size() + wincap.page_size ())
> > +static long
> > +get_arg_max (int in)
> > +{
> > +  return (long) (get_rlimit_stack () + DEFAULT_STACKGUARD) / 4;
> > +}
> > +
> >  static long
> >  get_page_size (int in)
> >  {
> > @@ -485,7 +492,7 @@ static struct
> >      };
> >  } sca[] =
> >  {
> > -  {cons, {c:ARG_MAX}},			/*   0, _SC_ARG_MAX */
> > +  {func, {f:get_arg_max}},		/*   0, _SC_ARG_MAX */
> >    {cons, {c:CHILD_MAX}},		/*   1, _SC_CHILD_MAX */
> >    {cons, {c:CLOCKS_PER_SEC}},		/*   2, _SC_CLK_TCK */
> >    {cons, {c:NGROUPS_MAX}},		/*   3, _SC_NGROUPS_MAX */
> > -- 
> > 2.39.0
> 
> Along these lines, there's no reason to couple SC_ARG_MAX to the
> size of the stack.  I'd propose to return the value 2097152.  It's
> the default value returned by getconf ARG_MAX on LInx as well.

What happens if user allocate stack whose size is not enough for
2097152?

-- 
Takashi Yano <takashi.yano@nifty.ne.jp>

Re: [PATCH v2] Cygwin: spawn: Fix segfalt when too many command line args are specified.

[Takashi Yano]

Hi Corinna,

On Mon, 28 Aug 2023 21:27:09 +0900
Takashi Yano wrote:
> On Mon, 28 Aug 2023 12:57:19 +0200
> Corinna Vinschen wrote:
> > On Aug 28 18:46, Takashi Yano wrote:
> > > Previously, the number of command line args was not checked for
> > > cygwin process. Due to this, segmentation fault was caused if too
> > > many command line args are specified.
> > > https://cygwin.com/pipermail/cygwin/2023-August/254333.html
> > > 
> > > Since char *argv[argc + 1] is placed on the stack in dll_crt0_1(),
> > > STATUS_STACK_OVERFLOW occurs if the stack does not have enough
> > > space.
> > > 
> > > With this patch, the total length of the arguments and the size of
> > > argv[] is restricted to 1/4 of total stack size for the process, and
> > > spawnve() returns E2BIG if the size exceeds the limit.
> > > [...]
> > > +static size_t
> > > +get_stack_size (const WCHAR *filename)
> > > +{
> > > +  HANDLE h;
> > > +  h = CreateFileW (filename, GENERIC_READ, FILE_SHARE_READ,
> > > +		   NULL, OPEN_EXISTING, 0, NULL);
> > > +  char buf[1024];
> > > +  DWORD n;
> > > +  ReadFile (h, buf, sizeof (buf), &n, 0);
> > > +  CloseHandle (h);
> > > +  IMAGE_NT_HEADERS32 *p = (IMAGE_NT_HEADERS32 *) memmem (buf, n, "PE\0\0", 4);
> > > +  if (!p)
> > > +    return 0;
> > > +  if ((char *) &p->OptionalHeader.SizeOfStackCommit > buf + n)
> > > +    return 0; /* buf[] is not enough */
> > > +  if (p->OptionalHeader.Magic == IMAGE_NT_OPTIONAL_HDR32_MAGIC)
> > > +    return p->OptionalHeader.SizeOfStackReserve;
> > > +  IMAGE_NT_HEADERS64 *p64 = (IMAGE_NT_HEADERS64 *) p;
> > > +  if ((char *) &p64->OptionalHeader.SizeOfStackCommit > buf + n)
> > > +    return 0; /* buf[] is not enough */
> > > +  return p64->OptionalHeader.SizeOfStackReserve;
> > > +}
> > 
> > Sorry, but this proposal is too complicated, IMHO.
> > 
> > Checking the stacksize in the file header for each single execve
> > is quite a bit time consuming, isn't it?
> > 
> > The question is rather, why storing argv on the stack at all?  I guess
> > the original idea was that argv is always a rather overseeable number.
> > But it doesn't have to stay on the stack.
> > 
> > I tried this simple patch:
> > 
> > diff --git a/winsup/cygwin/dcrt0.cc b/winsup/cygwin/dcrt0.cc
> > index 49b7a44aeb15..961dea4ab993 100644
> > --- a/winsup/cygwin/dcrt0.cc
> > +++ b/winsup/cygwin/dcrt0.cc
> > @@ -978,11 +978,8 @@ dll_crt0_1 (void *)
> >  	 a change to an element of argv[] it does not affect Cygwin's argv.
> >  	 Changing the the contents of what argv[n] points to will still
> >  	 affect Cygwin.  This is similar (but not exactly like) Linux. */
> > -      char *newargv[__argc + 1];
> > -      char **nav = newargv;
> > -      char **oav = __argv;
> > -      while ((*nav++ = *oav++) != NULL)
> > -	continue;
> > +      char **newargv = (char **) malloc ((__argc + 1) * sizeof (char **));
> > +      memcpy (newargv, __argv, (__argc + 1) * sizeof (char **));
> >        /* Handle any signals which may have arrived */
> >        sig_dispatch_pending (false);
> >        _my_tls.call_signal_handler ();
> > 
> > and the testcase `LC_ALL=C sed 's/x/y/' $(seq 1000000)' simply ran
> > as desired.  Combined with a bit of error checking...
> > 
> > > diff --git a/winsup/cygwin/sysconf.cc b/winsup/cygwin/sysconf.cc
> > > index 2db92e4de..6cb2aecd0 100644
> > > --- a/winsup/cygwin/sysconf.cc
> > > +++ b/winsup/cygwin/sysconf.cc
> > > @@ -21,6 +21,13 @@ details. */
> > >  #include "cpuid.h"
> > >  #include "clock.h"
> > >  
> > > +#define DEFAULT_STACKGUARD (wincap.def_guard_page_size() + wincap.page_size ())
> > > +static long
> > > +get_arg_max (int in)
> > > +{
> > > +  return (long) (get_rlimit_stack () + DEFAULT_STACKGUARD) / 4;
> > > +}
> > > +
> > >  static long
> > >  get_page_size (int in)
> > >  {
> > > @@ -485,7 +492,7 @@ static struct
> > >      };
> > >  } sca[] =
> > >  {
> > > -  {cons, {c:ARG_MAX}},			/*   0, _SC_ARG_MAX */
> > > +  {func, {f:get_arg_max}},		/*   0, _SC_ARG_MAX */
> > >    {cons, {c:CHILD_MAX}},		/*   1, _SC_CHILD_MAX */
> > >    {cons, {c:CLOCKS_PER_SEC}},		/*   2, _SC_CLK_TCK */
> > >    {cons, {c:NGROUPS_MAX}},		/*   3, _SC_NGROUPS_MAX */
> > > -- 
> > > 2.39.0
> > 
> > Along these lines, there's no reason to couple SC_ARG_MAX to the
> > size of the stack.  I'd propose to return the value 2097152.  It's
> > the default value returned by getconf ARG_MAX on LInx as well.
> 
> What happens if user allocate stack whose size is not enough for
> 2097152?

I think I got your point. If char *newargv[__argc+1] is placed
in heap, we don't care of stack size for the argument length.
Right?

I'll revise the patch accordingly. Please look v3 patch.

-- 
Takashi Yano <takashi.yano@nifty.ne.jp>

Re: [PATCH v2] Cygwin: spawn: Fix segfalt when too many command line args are specified.

[Takashi Yano]

On Mon, 28 Aug 2023 12:57:19 +0200
Corinna Vinschen wrote:
> -      char *newargv[__argc + 1];
> -      char **nav = newargv;
> -      char **oav = __argv;
> -      while ((*nav++ = *oav++) != NULL)
> -	continue;
> +      char **newargv = (char **) malloc ((__argc + 1) * sizeof (char **));
> +      memcpy (newargv, __argv, (__argc + 1) * sizeof (char **));
>        /* Handle any signals which may have arrived */
>        sig_dispatch_pending (false);
>        _my_tls.call_signal_handler ();

Shouldn't this be sizeof (char *), not sizeof (char **)?

-- 
Takashi Yano <takashi.yano@nifty.ne.jp>