eal: don't use default library path for static binaries
diff mbox series

Message ID 20200626145957.722089-1-bruce.richardson@intel.com
State Accepted, archived
Delegated to: Thomas Monjalon
Headers show
Series
  • eal: don't use default library path for static binaries
Related show

Checks

Context Check Description
ci/iol-testing success Testing PASS
ci/travis-robot success Travis build: passed
ci/Intel-compilation success Compilation OK
ci/iol-mellanox-Performance success Performance Testing PASS
ci/iol-nxp-Performance success Performance Testing PASS
ci/iol-intel-Performance success Performance Testing PASS
ci/iol-broadcom-Performance success Performance Testing PASS
ci/checkpatch success coding style OK

Commit Message

Bruce Richardson June 26, 2020, 2:59 p.m. UTC
When using statically linked DPDK binaries, the EAL checks the default PMD
path and tries to load any drivers there, despite the fact that all drivers
are normally linked into the binary.  This behaviour can cause issues if
the PMD path and lib dir is configured to a non-standard location which is
not in the ld.so.conf paths, e.g. a build with prefix set to a home
directory location. In a case such as this, EAL will try and
(unnecessarily) load the .so driver files but that load will fail as their
dependent libraries, such as ethdev, for example, will not be found.

Because of this, it is better if statically linked DPDK apps do not load
drivers from the standard paths automatically. The user can always have
this behaviour by explicitly specifying the path using -d flag, if so
desired.

Not loading the libraries automatically can also prevent potential issues
with a user building and running a statically-linked DPDK binary based off
a private copy of DPDK, while there exists on the same machine a
system-wide installation of DPDK in the default locations. Without this
change, the system-installed drivers will be loaded to the binary alongside
the statically-linked drivers, which is not what the user would have
intended.

To detect whether we are in a statically or dynamically linked binary, we
can have EAL try to get a dlopen handle to its own shared library, by
calling dlopen with the RTLD_NOLOAD flag. This will return NULL if there is
no such shared lib loaded i.e. the code is executing from a static library,
or a handle to the lib if it is loaded.

Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
---
 lib/librte_eal/common/eal_common_options.c | 11 +++++++++--
 1 file changed, 9 insertions(+), 2 deletions(-)

Comments

Stephen Hemminger June 26, 2020, 5:05 p.m. UTC | #1
On Fri, 26 Jun 2020 15:59:57 +0100
Bruce Richardson <bruce.richardson@intel.com> wrote:

> When using statically linked DPDK binaries, the EAL checks the default PMD
> path and tries to load any drivers there, despite the fact that all drivers
> are normally linked into the binary.  This behaviour can cause issues if
> the PMD path and lib dir is configured to a non-standard location which is
> not in the ld.so.conf paths, e.g. a build with prefix set to a home
> directory location. In a case such as this, EAL will try and
> (unnecessarily) load the .so driver files but that load will fail as their
> dependent libraries, such as ethdev, for example, will not be found.
> 
> Because of this, it is better if statically linked DPDK apps do not load
> drivers from the standard paths automatically. The user can always have
> this behaviour by explicitly specifying the path using -d flag, if so
> desired.
> 
> Not loading the libraries automatically can also prevent potential issues
> with a user building and running a statically-linked DPDK binary based off
> a private copy of DPDK, while there exists on the same machine a
> system-wide installation of DPDK in the default locations. Without this
> change, the system-installed drivers will be loaded to the binary alongside
> the statically-linked drivers, which is not what the user would have
> intended.
> 
> To detect whether we are in a statically or dynamically linked binary, we
> can have EAL try to get a dlopen handle to its own shared library, by
> calling dlopen with the RTLD_NOLOAD flag. This will return NULL if there is
> no such shared lib loaded i.e. the code is executing from a static library,
> or a handle to the lib if it is loaded.
> 
> Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>

But what if the majority of the DPDK is statically linked but the
application wants also load a dynamically linked driver?
Bruce Richardson June 29, 2020, 9:11 a.m. UTC | #2
On Fri, Jun 26, 2020 at 10:05:42AM -0700, Stephen Hemminger wrote:
> On Fri, 26 Jun 2020 15:59:57 +0100
> Bruce Richardson <bruce.richardson@intel.com> wrote:
> 
> > When using statically linked DPDK binaries, the EAL checks the default PMD
> > path and tries to load any drivers there, despite the fact that all drivers
> > are normally linked into the binary.  This behaviour can cause issues if
> > the PMD path and lib dir is configured to a non-standard location which is
> > not in the ld.so.conf paths, e.g. a build with prefix set to a home
> > directory location. In a case such as this, EAL will try and
> > (unnecessarily) load the .so driver files but that load will fail as their
> > dependent libraries, such as ethdev, for example, will not be found.
> > 
> > Because of this, it is better if statically linked DPDK apps do not load
> > drivers from the standard paths automatically. The user can always have
> > this behaviour by explicitly specifying the path using -d flag, if so
> > desired.
> > 
> > Not loading the libraries automatically can also prevent potential issues
> > with a user building and running a statically-linked DPDK binary based off
> > a private copy of DPDK, while there exists on the same machine a
> > system-wide installation of DPDK in the default locations. Without this
> > change, the system-installed drivers will be loaded to the binary alongside
> > the statically-linked drivers, which is not what the user would have
> > intended.
> > 
> > To detect whether we are in a statically or dynamically linked binary, we
> > can have EAL try to get a dlopen handle to its own shared library, by
> > calling dlopen with the RTLD_NOLOAD flag. This will return NULL if there is
> > no such shared lib loaded i.e. the code is executing from a static library,
> > or a handle to the lib if it is loaded.
> > 
> > Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
> 
> But what if the majority of the DPDK is statically linked but the
> application wants also load a dynamically linked driver?
> 
They use the -d flag as now. The only change here is that we don't
*automatically* (and silently) attempt to load all drivers from a system
location when you have a static binary.
Bruce Richardson June 29, 2020, 9:19 a.m. UTC | #3
On Mon, Jun 29, 2020 at 10:11:58AM +0100, Bruce Richardson wrote:
> On Fri, Jun 26, 2020 at 10:05:42AM -0700, Stephen Hemminger wrote:
> > On Fri, 26 Jun 2020 15:59:57 +0100
> > Bruce Richardson <bruce.richardson@intel.com> wrote:
> > 
> > > When using statically linked DPDK binaries, the EAL checks the default PMD
> > > path and tries to load any drivers there, despite the fact that all drivers
> > > are normally linked into the binary.  This behaviour can cause issues if
> > > the PMD path and lib dir is configured to a non-standard location which is
> > > not in the ld.so.conf paths, e.g. a build with prefix set to a home
> > > directory location. In a case such as this, EAL will try and
> > > (unnecessarily) load the .so driver files but that load will fail as their
> > > dependent libraries, such as ethdev, for example, will not be found.
> > > 
> > > Because of this, it is better if statically linked DPDK apps do not load
> > > drivers from the standard paths automatically. The user can always have
> > > this behaviour by explicitly specifying the path using -d flag, if so
> > > desired.
> > > 
> > > Not loading the libraries automatically can also prevent potential issues
> > > with a user building and running a statically-linked DPDK binary based off
> > > a private copy of DPDK, while there exists on the same machine a
> > > system-wide installation of DPDK in the default locations. Without this
> > > change, the system-installed drivers will be loaded to the binary alongside
> > > the statically-linked drivers, which is not what the user would have
> > > intended.
> > > 
> > > To detect whether we are in a statically or dynamically linked binary, we
> > > can have EAL try to get a dlopen handle to its own shared library, by
> > > calling dlopen with the RTLD_NOLOAD flag. This will return NULL if there is
> > > no such shared lib loaded i.e. the code is executing from a static library,
> > > or a handle to the lib if it is loaded.
> > > 
> > > Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
> > 
> > But what if the majority of the DPDK is statically linked but the
> > application wants also load a dynamically linked driver?
> > 
> They use the -d flag as now. The only change here is that we don't
> *automatically* (and silently) attempt to load all drivers from a system
> location when you have a static binary.

I'd also make a couple of additional points:
1. If you have a static app and you have extra drivers in your EAL_PMD_PATH
directory you have no way of preventing the loading of them.

2. Since all DPDK apps try to load all files in that directory, all one has
to do is put a non-loadable file into the DPDK PMD directory and suddenly
all DPDK apps on the system will fail to run.  [Patchset
http://patches.dpdk.org/project/dpdk/list/?series=10553 will also help
here]

3. Since this is a behaviour change, perhaps it needs to be deferred to
20.11? Ideally I think we should fix this now, because I think the current
behaviour doesn't make sense and causes more problems than it solves, but
if it needs to be deferred, so be it.

/Bruce
Burakov, Anatoly June 29, 2020, 2:53 p.m. UTC | #4
On 26-Jun-20 3:59 PM, Bruce Richardson wrote:
> When using statically linked DPDK binaries, the EAL checks the default PMD
> path and tries to load any drivers there, despite the fact that all drivers
> are normally linked into the binary.  This behaviour can cause issues if
> the PMD path and lib dir is configured to a non-standard location which is
> not in the ld.so.conf paths, e.g. a build with prefix set to a home
> directory location. In a case such as this, EAL will try and
> (unnecessarily) load the .so driver files but that load will fail as their
> dependent libraries, such as ethdev, for example, will not be found.
> 
> Because of this, it is better if statically linked DPDK apps do not load
> drivers from the standard paths automatically. The user can always have
> this behaviour by explicitly specifying the path using -d flag, if so
> desired.
> 
> Not loading the libraries automatically can also prevent potential issues
> with a user building and running a statically-linked DPDK binary based off
> a private copy of DPDK, while there exists on the same machine a
> system-wide installation of DPDK in the default locations. Without this
> change, the system-installed drivers will be loaded to the binary alongside
> the statically-linked drivers, which is not what the user would have
> intended.
> 
> To detect whether we are in a statically or dynamically linked binary, we
> can have EAL try to get a dlopen handle to its own shared library, by
> calling dlopen with the RTLD_NOLOAD flag. This will return NULL if there is
> no such shared lib loaded i.e. the code is executing from a static library,
> or a handle to the lib if it is loaded.
> 
> Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
> ---

+1, been bitten by this more than... ummm... well, a lot.

Acked-by: Anatoly Burakov <anatoly.burakov@intel.com>
Stephen Hemminger June 29, 2020, 3:41 p.m. UTC | #5
On Mon, 29 Jun 2020 10:19:10 +0100
Bruce Richardson <bruce.richardson@intel.com> wrote:

> On Mon, Jun 29, 2020 at 10:11:58AM +0100, Bruce Richardson wrote:
> > On Fri, Jun 26, 2020 at 10:05:42AM -0700, Stephen Hemminger wrote:  
> > > On Fri, 26 Jun 2020 15:59:57 +0100
> > > Bruce Richardson <bruce.richardson@intel.com> wrote:
> > >   
> > > > When using statically linked DPDK binaries, the EAL checks the default PMD
> > > > path and tries to load any drivers there, despite the fact that all drivers
> > > > are normally linked into the binary.  This behaviour can cause issues if
> > > > the PMD path and lib dir is configured to a non-standard location which is
> > > > not in the ld.so.conf paths, e.g. a build with prefix set to a home
> > > > directory location. In a case such as this, EAL will try and
> > > > (unnecessarily) load the .so driver files but that load will fail as their
> > > > dependent libraries, such as ethdev, for example, will not be found.
> > > > 
> > > > Because of this, it is better if statically linked DPDK apps do not load
> > > > drivers from the standard paths automatically. The user can always have
> > > > this behaviour by explicitly specifying the path using -d flag, if so
> > > > desired.
> > > > 
> > > > Not loading the libraries automatically can also prevent potential issues
> > > > with a user building and running a statically-linked DPDK binary based off
> > > > a private copy of DPDK, while there exists on the same machine a
> > > > system-wide installation of DPDK in the default locations. Without this
> > > > change, the system-installed drivers will be loaded to the binary alongside
> > > > the statically-linked drivers, which is not what the user would have
> > > > intended.
> > > > 
> > > > To detect whether we are in a statically or dynamically linked binary, we
> > > > can have EAL try to get a dlopen handle to its own shared library, by
> > > > calling dlopen with the RTLD_NOLOAD flag. This will return NULL if there is
> > > > no such shared lib loaded i.e. the code is executing from a static library,
> > > > or a handle to the lib if it is loaded.
> > > > 
> > > > Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>  
> > > 
> > > But what if the majority of the DPDK is statically linked but the
> > > application wants also load a dynamically linked driver?
> > >   
> > They use the -d flag as now. The only change here is that we don't
> > *automatically* (and silently) attempt to load all drivers from a system
> > location when you have a static binary.  
> 
> I'd also make a couple of additional points:
> 1. If you have a static app and you have extra drivers in your EAL_PMD_PATH
> directory you have no way of preventing the loading of them.
> 
> 2. Since all DPDK apps try to load all files in that directory, all one has
> to do is put a non-loadable file into the DPDK PMD directory and suddenly
> all DPDK apps on the system will fail to run.  [Patchset
> http://patches.dpdk.org/project/dpdk/list/?series=10553 will also help
> here]
> 
> 3. Since this is a behaviour change, perhaps it needs to be deferred to
> 20.11? Ideally I think we should fix this now, because I think the current
> behaviour doesn't make sense and causes more problems than it solves, but
> if it needs to be deferred, so be it.


Thanks, just trying to poke the corners of this change.
Please make sure the documentation and web site match the behavior.

Acked-by: Stephen Hemminger <stephen@networkplumber.org>
Bruce Richardson June 29, 2020, 4:15 p.m. UTC | #6
On Mon, Jun 29, 2020 at 08:41:46AM -0700, Stephen Hemminger wrote:
> On Mon, 29 Jun 2020 10:19:10 +0100
> Bruce Richardson <bruce.richardson@intel.com> wrote:
> 
> > On Mon, Jun 29, 2020 at 10:11:58AM +0100, Bruce Richardson wrote:
> > > On Fri, Jun 26, 2020 at 10:05:42AM -0700, Stephen Hemminger wrote:  
> > > > On Fri, 26 Jun 2020 15:59:57 +0100
> > > > Bruce Richardson <bruce.richardson@intel.com> wrote:
> > > >   
> > > > > When using statically linked DPDK binaries, the EAL checks the default PMD
> > > > > path and tries to load any drivers there, despite the fact that all drivers
> > > > > are normally linked into the binary.  This behaviour can cause issues if
> > > > > the PMD path and lib dir is configured to a non-standard location which is
> > > > > not in the ld.so.conf paths, e.g. a build with prefix set to a home
> > > > > directory location. In a case such as this, EAL will try and
> > > > > (unnecessarily) load the .so driver files but that load will fail as their
> > > > > dependent libraries, such as ethdev, for example, will not be found.
> > > > > 
> > > > > Because of this, it is better if statically linked DPDK apps do not load
> > > > > drivers from the standard paths automatically. The user can always have
> > > > > this behaviour by explicitly specifying the path using -d flag, if so
> > > > > desired.
> > > > > 
> > > > > Not loading the libraries automatically can also prevent potential issues
> > > > > with a user building and running a statically-linked DPDK binary based off
> > > > > a private copy of DPDK, while there exists on the same machine a
> > > > > system-wide installation of DPDK in the default locations. Without this
> > > > > change, the system-installed drivers will be loaded to the binary alongside
> > > > > the statically-linked drivers, which is not what the user would have
> > > > > intended.
> > > > > 
> > > > > To detect whether we are in a statically or dynamically linked binary, we
> > > > > can have EAL try to get a dlopen handle to its own shared library, by
> > > > > calling dlopen with the RTLD_NOLOAD flag. This will return NULL if there is
> > > > > no such shared lib loaded i.e. the code is executing from a static library,
> > > > > or a handle to the lib if it is loaded.
> > > > > 
> > > > > Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>  
> > > > 
> > > > But what if the majority of the DPDK is statically linked but the
> > > > application wants also load a dynamically linked driver?
> > > >   
> > > They use the -d flag as now. The only change here is that we don't
> > > *automatically* (and silently) attempt to load all drivers from a system
> > > location when you have a static binary.  
> > 
> > I'd also make a couple of additional points:
> > 1. If you have a static app and you have extra drivers in your EAL_PMD_PATH
> > directory you have no way of preventing the loading of them.
> > 
> > 2. Since all DPDK apps try to load all files in that directory, all one has
> > to do is put a non-loadable file into the DPDK PMD directory and suddenly
> > all DPDK apps on the system will fail to run.  [Patchset
> > http://patches.dpdk.org/project/dpdk/list/?series=10553 will also help
> > here]
> > 
> > 3. Since this is a behaviour change, perhaps it needs to be deferred to
> > 20.11? Ideally I think we should fix this now, because I think the current
> > behaviour doesn't make sense and causes more problems than it solves, but
> > if it needs to be deferred, so be it.
> 
> 
> Thanks, just trying to poke the corners of this change.
> Please make sure the documentation and web site match the behavior.
> 
> Acked-by: Stephen Hemminger <stephen@networkplumber.org>

This appears to be a gap, since I looked through the docs and couldn't find
any reference to this behaviour at all. However, I may well have missed
something so please flag if so.

Regards,
/Bruce
Thomas Monjalon July 5, 2020, 6:01 p.m. UTC | #7
29/06/2020 18:15, Bruce Richardson:
> On Mon, Jun 29, 2020 at 08:41:46AM -0700, Stephen Hemminger wrote:
> > Bruce Richardson <bruce.richardson@intel.com> wrote:
> > > On Mon, Jun 29, 2020 at 10:11:58AM +0100, Bruce Richardson wrote:
> > > > On Fri, Jun 26, 2020 at 10:05:42AM -0700, Stephen Hemminger wrote:  
> > > > > Bruce Richardson <bruce.richardson@intel.com> wrote:
> > > > >   
> > > > > > When using statically linked DPDK binaries, the EAL checks the default PMD
> > > > > > path and tries to load any drivers there, despite the fact that all drivers
> > > > > > are normally linked into the binary.  This behaviour can cause issues if
> > > > > > the PMD path and lib dir is configured to a non-standard location which is
> > > > > > not in the ld.so.conf paths, e.g. a build with prefix set to a home
> > > > > > directory location. In a case such as this, EAL will try and
> > > > > > (unnecessarily) load the .so driver files but that load will fail as their
> > > > > > dependent libraries, such as ethdev, for example, will not be found.
> > > > > > 
> > > > > > Because of this, it is better if statically linked DPDK apps do not load
> > > > > > drivers from the standard paths automatically. The user can always have
> > > > > > this behaviour by explicitly specifying the path using -d flag, if so
> > > > > > desired.
> > > > > > 
> > > > > > Not loading the libraries automatically can also prevent potential issues
> > > > > > with a user building and running a statically-linked DPDK binary based off
> > > > > > a private copy of DPDK, while there exists on the same machine a
> > > > > > system-wide installation of DPDK in the default locations. Without this
> > > > > > change, the system-installed drivers will be loaded to the binary alongside
> > > > > > the statically-linked drivers, which is not what the user would have
> > > > > > intended.
> > > > > > 
> > > > > > To detect whether we are in a statically or dynamically linked binary, we
> > > > > > can have EAL try to get a dlopen handle to its own shared library, by
> > > > > > calling dlopen with the RTLD_NOLOAD flag. This will return NULL if there is
> > > > > > no such shared lib loaded i.e. the code is executing from a static library,
> > > > > > or a handle to the lib if it is loaded.
> > > > > > 
> > > > > > Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>  
> > > > > 
> > > > > But what if the majority of the DPDK is statically linked but the
> > > > > application wants also load a dynamically linked driver?
> > > > >   
> > > > They use the -d flag as now. The only change here is that we don't
> > > > *automatically* (and silently) attempt to load all drivers from a system
> > > > location when you have a static binary.  
> > > 
> > > I'd also make a couple of additional points:
> > > 1. If you have a static app and you have extra drivers in your EAL_PMD_PATH
> > > directory you have no way of preventing the loading of them.
> > > 
> > > 2. Since all DPDK apps try to load all files in that directory, all one has
> > > to do is put a non-loadable file into the DPDK PMD directory and suddenly
> > > all DPDK apps on the system will fail to run.  [Patchset
> > > http://patches.dpdk.org/project/dpdk/list/?series=10553 will also help
> > > here]
> > > 
> > > 3. Since this is a behaviour change, perhaps it needs to be deferred to
> > > 20.11? Ideally I think we should fix this now, because I think the current
> > > behaviour doesn't make sense and causes more problems than it solves, but
> > > if it needs to be deferred, so be it.
> > 
> > 
> > Thanks, just trying to poke the corners of this change.
> > Please make sure the documentation and web site match the behavior.
> > 
> > Acked-by: Stephen Hemminger <stephen@networkplumber.org>
> 
> This appears to be a gap, since I looked through the docs and couldn't find
> any reference to this behaviour at all. However, I may well have missed
> something so please flag if so.

Yes we are lacking some doc around the basics of EAL (and also ethdev).
This is a behaviour change, but I prefer applying it in 20.08,
so we can get some user feedback before the LTS release.

Applied, thanks

Patch
diff mbox series

diff --git a/lib/librte_eal/common/eal_common_options.c b/lib/librte_eal/common/eal_common_options.c
index 0901b493c..1d958d11e 100644
--- a/lib/librte_eal/common/eal_common_options.c
+++ b/lib/librte_eal/common/eal_common_options.c
@@ -447,8 +447,15 @@  eal_plugins_init(void)
 	struct shared_driver *solib = NULL;
 	struct stat sb;
 
-	if (*default_solib_dir != '\0' && stat(default_solib_dir, &sb) == 0 &&
-				S_ISDIR(sb.st_mode))
+	/* if we are not statically linked, add default driver loading
+	 * path if it exists as a directory.
+	 * (Using dlopen with NOLOAD flag on eal, will return NULL if the eal
+	 * shared library is not already loaded i.e. it's statically linked.)
+	 */
+	if (dlopen("librte_eal.so", RTLD_LAZY | RTLD_NOLOAD) != NULL &&
+			*default_solib_dir != '\0' &&
+			stat(default_solib_dir, &sb) == 0 &&
+			S_ISDIR(sb.st_mode))
 		eal_plugin_add(default_solib_dir);
 
 	TAILQ_FOREACH(solib, &solib_list, next) {