| Message ID | b925e11d639726afbaaeea5aeaa58572b3aacf8e.1618753862.git.leonro@nvidia.com (mailing list archive) |
|---|---|
| State | Changes Requested |
| Delegated to: | Jason Gunthorpe |
| Headers | show |
| Series | CMA fixes | expand |
On Sun, Apr 18, 2021 at 04:55:53PM +0300, Leon Romanovsky wrote: > From: Shay Drory <shayd@nvidia.com> > > rdma_listen() checks if device already attached to rdma_id_priv, > based on the response the its decide to what to listen, however > this is different when the listeners are canceled. > > This leads to a mismatch between rdma_listen() and cma_cancel_operation(), > and causes to bellow wild-memory-access. Fix it by aligning rdma_listen() > according to the cma_cancel_operation(). So this is happening because the error unwind in rdma_bind_addr() is taking the exit path and calling cma_release_dev()? This allows rdma_listen() to be called with a bogus device pointer which precipitates this UAF during destroy. However, I think rdma_bind_addr() should not allow the bogus device pointer to leak out at all, since the ULP could see it. It really is invalid to have it present no matter what. This would make cma_release_dev() and _cma_attach_to_dev() symmetrical - what do you think? diff --git a/drivers/infiniband/core/cma.c b/drivers/infiniband/core/cma.c index 2dc302a83014ae..91f6d968b46f65 100644 --- a/drivers/infiniband/core/cma.c +++ b/drivers/infiniband/core/cma.c @@ -474,6 +474,7 @@ static void cma_release_dev(struct rdma_id_private *id_priv) list_del(&id_priv->list); cma_dev_put(id_priv->cma_dev); id_priv->cma_dev = NULL; + id_priv->id.device = NULL; if (id_priv->id.route.addr.dev_addr.sgid_attr) { rdma_put_gid_attr(id_priv->id.route.addr.dev_addr.sgid_attr); id_priv->id.route.addr.dev_addr.sgid_attr = NULL;
On 4/22/2021 14:28, Jason Gunthorpe wrote: > On Sun, Apr 18, 2021 at 04:55:53PM +0300, Leon Romanovsky wrote: >> From: Shay Drory <shayd@nvidia.com> >> >> rdma_listen() checks if device already attached to rdma_id_priv, >> based on the response the its decide to what to listen, however >> this is different when the listeners are canceled. >> >> This leads to a mismatch between rdma_listen() and cma_cancel_operation(), >> and causes to bellow wild-memory-access. Fix it by aligning rdma_listen() >> according to the cma_cancel_operation(). > So this is happening because the error unwind in rdma_bind_addr() is > taking the exit path and calling cma_release_dev()? > > This allows rdma_listen() to be called with a bogus device pointer > which precipitates this UAF during destroy. > > However, I think rdma_bind_addr() should not allow the bogus device > pointer to leak out at all, since the ULP could see it. It really is > invalid to have it present no matter what. > > This would make cma_release_dev() and _cma_attach_to_dev() > symmetrical - what do you think? > > diff --git a/drivers/infiniband/core/cma.c b/drivers/infiniband/core/cma.c > index 2dc302a83014ae..91f6d968b46f65 100644 > --- a/drivers/infiniband/core/cma.c > +++ b/drivers/infiniband/core/cma.c > @@ -474,6 +474,7 @@ static void cma_release_dev(struct rdma_id_private *id_priv) > list_del(&id_priv->list); > cma_dev_put(id_priv->cma_dev); > id_priv->cma_dev = NULL; > + id_priv->id.device = NULL; > if (id_priv->id.route.addr.dev_addr.sgid_attr) { > rdma_put_gid_attr(id_priv->id.route.addr.dev_addr.sgid_attr); > id_priv->id.route.addr.dev_addr.sgid_attr = NULL; I try that. this will break restrack_del() since restrack_del() is using id_priv->id.device and is being called before restrack_del(): _destroy_id <https://elixir.bootlin.com/linux/v5.11/C/ident/_destroy_id>() { if (id_priv <https://elixir.bootlin.com/linux/v5.11/C/ident/id_priv>->cma_dev <https://elixir.bootlin.com/linux/v5.11/C/ident/cma_dev>) cma_release_dev <https://elixir.bootlin.com/linux/v5.11/C/ident/cma_release_dev>(id_priv <https://elixir.bootlin.com/linux/v5.11/C/ident/id_priv>); ... rdma_restrack_del <https://elixir.bootlin.com/linux/v5.11/C/ident/rdma_restrack_del>(&id_priv <https://elixir.bootlin.com/linux/v5.11/C/ident/id_priv>->res); }
On Thu, Apr 22, 2021 at 03:44:55PM +0300, Shay Drory wrote: > On 4/22/2021 14:28, Jason Gunthorpe wrote: > > > On Sun, Apr 18, 2021 at 04:55:53PM +0300, Leon Romanovsky wrote: > > > From: Shay Drory <shayd@nvidia.com> > > > > > > rdma_listen() checks if device already attached to rdma_id_priv, > > > based on the response the its decide to what to listen, however > > > this is different when the listeners are canceled. > > > > > > This leads to a mismatch between rdma_listen() and cma_cancel_operation(), > > > and causes to bellow wild-memory-access. Fix it by aligning rdma_listen() > > > according to the cma_cancel_operation(). > > So this is happening because the error unwind in rdma_bind_addr() is > > taking the exit path and calling cma_release_dev()? > > > > This allows rdma_listen() to be called with a bogus device pointer > > which precipitates this UAF during destroy. > > > > However, I think rdma_bind_addr() should not allow the bogus device > > pointer to leak out at all, since the ULP could see it. It really is > > invalid to have it present no matter what. > > > > This would make cma_release_dev() and _cma_attach_to_dev() > > symmetrical - what do you think? > > > > diff --git a/drivers/infiniband/core/cma.c b/drivers/infiniband/core/cma.c > > index 2dc302a83014ae..91f6d968b46f65 100644 > > +++ b/drivers/infiniband/core/cma.c > > @@ -474,6 +474,7 @@ static void cma_release_dev(struct rdma_id_private *id_priv) > > list_del(&id_priv->list); > > cma_dev_put(id_priv->cma_dev); > > id_priv->cma_dev = NULL; > > + id_priv->id.device = NULL; > > if (id_priv->id.route.addr.dev_addr.sgid_attr) { > > rdma_put_gid_attr(id_priv->id.route.addr.dev_addr.sgid_attr); > > id_priv->id.route.addr.dev_addr.sgid_attr = NULL; > > I try that. this will break restrack_del() since restrack_del() is > using id_priv->id.device and is being called before restrack_del(): Oh that is another bug, once cma_release_dev() is called there is no refcount protecting the id.device and any access to it is invalid. The order of rdma_restrack_del should be moved to be ahead of the cma_release_dev, and we also can't have a restrack without a cma_dev in the first place Jason
On Thu, Apr 22, 2021 at 09:51:35AM -0300, Jason Gunthorpe wrote: > On Thu, Apr 22, 2021 at 03:44:55PM +0300, Shay Drory wrote: > > On 4/22/2021 14:28, Jason Gunthorpe wrote: > > > > > On Sun, Apr 18, 2021 at 04:55:53PM +0300, Leon Romanovsky wrote: > > > > From: Shay Drory <shayd@nvidia.com> > > > > > > > > rdma_listen() checks if device already attached to rdma_id_priv, > > > > based on the response the its decide to what to listen, however > > > > this is different when the listeners are canceled. > > > > > > > > This leads to a mismatch between rdma_listen() and cma_cancel_operation(), > > > > and causes to bellow wild-memory-access. Fix it by aligning rdma_listen() > > > > according to the cma_cancel_operation(). > > > So this is happening because the error unwind in rdma_bind_addr() is > > > taking the exit path and calling cma_release_dev()? > > > > > > This allows rdma_listen() to be called with a bogus device pointer > > > which precipitates this UAF during destroy. > > > > > > However, I think rdma_bind_addr() should not allow the bogus device > > > pointer to leak out at all, since the ULP could see it. It really is > > > invalid to have it present no matter what. > > > > > > This would make cma_release_dev() and _cma_attach_to_dev() > > > symmetrical - what do you think? > > > > > > diff --git a/drivers/infiniband/core/cma.c b/drivers/infiniband/core/cma.c > > > index 2dc302a83014ae..91f6d968b46f65 100644 > > > +++ b/drivers/infiniband/core/cma.c > > > @@ -474,6 +474,7 @@ static void cma_release_dev(struct rdma_id_private *id_priv) > > > list_del(&id_priv->list); > > > cma_dev_put(id_priv->cma_dev); > > > id_priv->cma_dev = NULL; > > > + id_priv->id.device = NULL; > > > if (id_priv->id.route.addr.dev_addr.sgid_attr) { > > > rdma_put_gid_attr(id_priv->id.route.addr.dev_addr.sgid_attr); > > > id_priv->id.route.addr.dev_addr.sgid_attr = NULL; > > > > I try that. this will break restrack_del() since restrack_del() is > > using id_priv->id.device and is being called before restrack_del(): > > Oh that is another bug, once cma_release_dev() is called there is no > refcount protecting the id.device and any access to it is invalid. > > The order of rdma_restrack_del should be moved to be ahead of the > cma_release_dev, and we also can't have a restrack without a cma_dev > in the first place We have restrack per-cmd_id and not per-cma_dev. > > Jason
On Thu, Apr 22, 2021 at 04:01:20PM +0300, Leon Romanovsky wrote: > On Thu, Apr 22, 2021 at 09:51:35AM -0300, Jason Gunthorpe wrote: > > On Thu, Apr 22, 2021 at 03:44:55PM +0300, Shay Drory wrote: > > > On 4/22/2021 14:28, Jason Gunthorpe wrote: > > > > > > > On Sun, Apr 18, 2021 at 04:55:53PM +0300, Leon Romanovsky wrote: > > > > > From: Shay Drory <shayd@nvidia.com> > > > > > > > > > > rdma_listen() checks if device already attached to rdma_id_priv, > > > > > based on the response the its decide to what to listen, however > > > > > this is different when the listeners are canceled. > > > > > > > > > > This leads to a mismatch between rdma_listen() and cma_cancel_operation(), > > > > > and causes to bellow wild-memory-access. Fix it by aligning rdma_listen() > > > > > according to the cma_cancel_operation(). > > > > So this is happening because the error unwind in rdma_bind_addr() is > > > > taking the exit path and calling cma_release_dev()? > > > > > > > > This allows rdma_listen() to be called with a bogus device pointer > > > > which precipitates this UAF during destroy. > > > > > > > > However, I think rdma_bind_addr() should not allow the bogus device > > > > pointer to leak out at all, since the ULP could see it. It really is > > > > invalid to have it present no matter what. > > > > > > > > This would make cma_release_dev() and _cma_attach_to_dev() > > > > symmetrical - what do you think? > > > > > > > > diff --git a/drivers/infiniband/core/cma.c b/drivers/infiniband/core/cma.c > > > > index 2dc302a83014ae..91f6d968b46f65 100644 > > > > +++ b/drivers/infiniband/core/cma.c > > > > @@ -474,6 +474,7 @@ static void cma_release_dev(struct rdma_id_private *id_priv) > > > > list_del(&id_priv->list); > > > > cma_dev_put(id_priv->cma_dev); > > > > id_priv->cma_dev = NULL; > > > > + id_priv->id.device = NULL; > > > > if (id_priv->id.route.addr.dev_addr.sgid_attr) { > > > > rdma_put_gid_attr(id_priv->id.route.addr.dev_addr.sgid_attr); > > > > id_priv->id.route.addr.dev_addr.sgid_attr = NULL; > > > > > > I try that. this will break restrack_del() since restrack_del() is > > > using id_priv->id.device and is being called before restrack_del(): > > > > Oh that is another bug, once cma_release_dev() is called there is no > > refcount protecting the id.device and any access to it is invalid. > > > > The order of rdma_restrack_del should be moved to be ahead of the > > cma_release_dev, and we also can't have a restrack without a cma_dev > > in the first place > > We have restrack per-cmd_id and not per-cma_dev. No, restrack has this: dev = res_to_dev(res); if (WARN_ON(!dev)) And here dev will be NULL if cma_dev isn't set Jasno
On Thu, Apr 22, 2021 at 10:02:18AM -0300, Jason Gunthorpe wrote: > On Thu, Apr 22, 2021 at 04:01:20PM +0300, Leon Romanovsky wrote: > > On Thu, Apr 22, 2021 at 09:51:35AM -0300, Jason Gunthorpe wrote: > > > On Thu, Apr 22, 2021 at 03:44:55PM +0300, Shay Drory wrote: > > > > On 4/22/2021 14:28, Jason Gunthorpe wrote: > > > > > > > > > On Sun, Apr 18, 2021 at 04:55:53PM +0300, Leon Romanovsky wrote: > > > > > > From: Shay Drory <shayd@nvidia.com> > > > > > > > > > > > > rdma_listen() checks if device already attached to rdma_id_priv, > > > > > > based on the response the its decide to what to listen, however > > > > > > this is different when the listeners are canceled. > > > > > > > > > > > > This leads to a mismatch between rdma_listen() and cma_cancel_operation(), > > > > > > and causes to bellow wild-memory-access. Fix it by aligning rdma_listen() > > > > > > according to the cma_cancel_operation(). > > > > > So this is happening because the error unwind in rdma_bind_addr() is > > > > > taking the exit path and calling cma_release_dev()? > > > > > > > > > > This allows rdma_listen() to be called with a bogus device pointer > > > > > which precipitates this UAF during destroy. > > > > > > > > > > However, I think rdma_bind_addr() should not allow the bogus device > > > > > pointer to leak out at all, since the ULP could see it. It really is > > > > > invalid to have it present no matter what. > > > > > > > > > > This would make cma_release_dev() and _cma_attach_to_dev() > > > > > symmetrical - what do you think? > > > > > > > > > > diff --git a/drivers/infiniband/core/cma.c b/drivers/infiniband/core/cma.c > > > > > index 2dc302a83014ae..91f6d968b46f65 100644 > > > > > +++ b/drivers/infiniband/core/cma.c > > > > > @@ -474,6 +474,7 @@ static void cma_release_dev(struct rdma_id_private *id_priv) > > > > > list_del(&id_priv->list); > > > > > cma_dev_put(id_priv->cma_dev); > > > > > id_priv->cma_dev = NULL; > > > > > + id_priv->id.device = NULL; > > > > > if (id_priv->id.route.addr.dev_addr.sgid_attr) { > > > > > rdma_put_gid_attr(id_priv->id.route.addr.dev_addr.sgid_attr); > > > > > id_priv->id.route.addr.dev_addr.sgid_attr = NULL; > > > > > > > > I try that. this will break restrack_del() since restrack_del() is > > > > using id_priv->id.device and is being called before restrack_del(): > > > > > > Oh that is another bug, once cma_release_dev() is called there is no > > > refcount protecting the id.device and any access to it is invalid. > > > > > > The order of rdma_restrack_del should be moved to be ahead of the > > > cma_release_dev, and we also can't have a restrack without a cma_dev > > > in the first place > > > > We have restrack per-cmd_id and not per-cma_dev. > > No, restrack has this: > > dev = res_to_dev(res); > if (WARN_ON(!dev)) > > And here dev will be NULL if cma_dev isn't set 127 static struct ib_device *res_to_dev(struct rdma_restrack_entry *res) 128 { <...> 136 case RDMA_RESTRACK_CM_ID: 137 return container_of(res, struct rdma_id_private, 138 res)->id.device; ^^^^^ it is not cma_dev Thanks
On Thu, Apr 22, 2021 at 05:58:18PM +0300, Leon Romanovsky wrote: > On Thu, Apr 22, 2021 at 10:02:18AM -0300, Jason Gunthorpe wrote: > > On Thu, Apr 22, 2021 at 04:01:20PM +0300, Leon Romanovsky wrote: > > > On Thu, Apr 22, 2021 at 09:51:35AM -0300, Jason Gunthorpe wrote: > > > > On Thu, Apr 22, 2021 at 03:44:55PM +0300, Shay Drory wrote: > > > > > On 4/22/2021 14:28, Jason Gunthorpe wrote: > > > > > > > > > > > On Sun, Apr 18, 2021 at 04:55:53PM +0300, Leon Romanovsky wrote: > > > > > > > From: Shay Drory <shayd@nvidia.com> > > > > > > > > > > > > > > rdma_listen() checks if device already attached to rdma_id_priv, > > > > > > > based on the response the its decide to what to listen, however > > > > > > > this is different when the listeners are canceled. > > > > > > > > > > > > > > This leads to a mismatch between rdma_listen() and cma_cancel_operation(), > > > > > > > and causes to bellow wild-memory-access. Fix it by aligning rdma_listen() > > > > > > > according to the cma_cancel_operation(). > > > > > > So this is happening because the error unwind in rdma_bind_addr() is > > > > > > taking the exit path and calling cma_release_dev()? > > > > > > > > > > > > This allows rdma_listen() to be called with a bogus device pointer > > > > > > which precipitates this UAF during destroy. > > > > > > > > > > > > However, I think rdma_bind_addr() should not allow the bogus device > > > > > > pointer to leak out at all, since the ULP could see it. It really is > > > > > > invalid to have it present no matter what. > > > > > > > > > > > > This would make cma_release_dev() and _cma_attach_to_dev() > > > > > > symmetrical - what do you think? > > > > > > > > > > > > diff --git a/drivers/infiniband/core/cma.c b/drivers/infiniband/core/cma.c > > > > > > index 2dc302a83014ae..91f6d968b46f65 100644 > > > > > > +++ b/drivers/infiniband/core/cma.c > > > > > > @@ -474,6 +474,7 @@ static void cma_release_dev(struct rdma_id_private *id_priv) > > > > > > list_del(&id_priv->list); > > > > > > cma_dev_put(id_priv->cma_dev); > > > > > > id_priv->cma_dev = NULL; > > > > > > + id_priv->id.device = NULL; > > > > > > if (id_priv->id.route.addr.dev_addr.sgid_attr) { > > > > > > rdma_put_gid_attr(id_priv->id.route.addr.dev_addr.sgid_attr); > > > > > > id_priv->id.route.addr.dev_addr.sgid_attr = NULL; > > > > > > > > > > I try that. this will break restrack_del() since restrack_del() is > > > > > using id_priv->id.device and is being called before restrack_del(): > > > > > > > > Oh that is another bug, once cma_release_dev() is called there is no > > > > refcount protecting the id.device and any access to it is invalid. > > > > > > > > The order of rdma_restrack_del should be moved to be ahead of the > > > > cma_release_dev, and we also can't have a restrack without a cma_dev > > > > in the first place > > > > > > We have restrack per-cmd_id and not per-cma_dev. > > > > No, restrack has this: > > > > dev = res_to_dev(res); > > if (WARN_ON(!dev)) > > > > And here dev will be NULL if cma_dev isn't set > > 127 static struct ib_device *res_to_dev(struct rdma_restrack_entry *res) > 128 { > > <...> > > 136 case RDMA_RESTRACK_CM_ID: > 137 return container_of(res, struct rdma_id_private, > 138 res)->id.device; > ^^^^^ it is not cma_dev The invariant is that priv.id.device == priv.cma_dev->device (this de-normalization of data exists only to allow priv to be in a private header) If cma_dev == NULL then id.device == NULL as cma_Dev was the thing preventing the pointer from being free'd. Jason
On Thu, Apr 22, 2021 at 01:18:09PM -0300, Jason Gunthorpe wrote: > On Thu, Apr 22, 2021 at 05:58:18PM +0300, Leon Romanovsky wrote: > > On Thu, Apr 22, 2021 at 10:02:18AM -0300, Jason Gunthorpe wrote: > > > On Thu, Apr 22, 2021 at 04:01:20PM +0300, Leon Romanovsky wrote: > > > > On Thu, Apr 22, 2021 at 09:51:35AM -0300, Jason Gunthorpe wrote: > > > > > On Thu, Apr 22, 2021 at 03:44:55PM +0300, Shay Drory wrote: > > > > > > On 4/22/2021 14:28, Jason Gunthorpe wrote: > > > > > > > > > > > > > On Sun, Apr 18, 2021 at 04:55:53PM +0300, Leon Romanovsky wrote: > > > > > > > > From: Shay Drory <shayd@nvidia.com> > > > > > > > > > > > > > > > > rdma_listen() checks if device already attached to rdma_id_priv, > > > > > > > > based on the response the its decide to what to listen, however > > > > > > > > this is different when the listeners are canceled. > > > > > > > > > > > > > > > > This leads to a mismatch between rdma_listen() and cma_cancel_operation(), > > > > > > > > and causes to bellow wild-memory-access. Fix it by aligning rdma_listen() > > > > > > > > according to the cma_cancel_operation(). > > > > > > > So this is happening because the error unwind in rdma_bind_addr() is > > > > > > > taking the exit path and calling cma_release_dev()? > > > > > > > > > > > > > > This allows rdma_listen() to be called with a bogus device pointer > > > > > > > which precipitates this UAF during destroy. > > > > > > > > > > > > > > However, I think rdma_bind_addr() should not allow the bogus device > > > > > > > pointer to leak out at all, since the ULP could see it. It really is > > > > > > > invalid to have it present no matter what. > > > > > > > > > > > > > > This would make cma_release_dev() and _cma_attach_to_dev() > > > > > > > symmetrical - what do you think? > > > > > > > > > > > > > > diff --git a/drivers/infiniband/core/cma.c b/drivers/infiniband/core/cma.c > > > > > > > index 2dc302a83014ae..91f6d968b46f65 100644 > > > > > > > +++ b/drivers/infiniband/core/cma.c > > > > > > > @@ -474,6 +474,7 @@ static void cma_release_dev(struct rdma_id_private *id_priv) > > > > > > > list_del(&id_priv->list); > > > > > > > cma_dev_put(id_priv->cma_dev); > > > > > > > id_priv->cma_dev = NULL; > > > > > > > + id_priv->id.device = NULL; > > > > > > > if (id_priv->id.route.addr.dev_addr.sgid_attr) { > > > > > > > rdma_put_gid_attr(id_priv->id.route.addr.dev_addr.sgid_attr); > > > > > > > id_priv->id.route.addr.dev_addr.sgid_attr = NULL; > > > > > > > > > > > > I try that. this will break restrack_del() since restrack_del() is > > > > > > using id_priv->id.device and is being called before restrack_del(): > > > > > > > > > > Oh that is another bug, once cma_release_dev() is called there is no > > > > > refcount protecting the id.device and any access to it is invalid. > > > > > > > > > > The order of rdma_restrack_del should be moved to be ahead of the > > > > > cma_release_dev, and we also can't have a restrack without a cma_dev > > > > > in the first place > > > > > > > > We have restrack per-cmd_id and not per-cma_dev. > > > > > > No, restrack has this: > > > > > > dev = res_to_dev(res); > > > if (WARN_ON(!dev)) > > > > > > And here dev will be NULL if cma_dev isn't set > > > > 127 static struct ib_device *res_to_dev(struct rdma_restrack_entry *res) > > 128 { > > > > <...> > > > > 136 case RDMA_RESTRACK_CM_ID: > > 137 return container_of(res, struct rdma_id_private, > > 138 res)->id.device; > > ^^^^^ it is not cma_dev > > The invariant is that > > priv.id.device == priv.cma_dev->device > > (this de-normalization of data exists only to allow priv to be in a > private header) > > If cma_dev == NULL then id.device == NULL as cma_Dev was the thing > preventing the pointer from being free'd. Let's see what I can do here. Thanks > > Jason
diff --git a/drivers/infiniband/core/cma.c b/drivers/infiniband/core/cma.c index 2dc302a83014..cc990adaf2b5 100644 --- a/drivers/infiniband/core/cma.c +++ b/drivers/infiniband/core/cma.c @@ -3768,7 +3768,7 @@ int rdma_listen(struct rdma_cm_id *id, int backlog) } id_priv->backlog = backlog; - if (id->device) { + if (id_priv->cma_dev) { if (rdma_cap_ib_cm(id->device, 1)) { ret = cma_ib_listen(id_priv); if (ret)