Re: [PATCH v4 2/9] iommu: Replace sva_iommu with iommu_attach_handle
On Sun, Apr 28, 2024 at 06:22:28PM +0800, Baolu Lu wrote: > /* A bond already exists, just take a reference`. */ > handle = iommu_attach_handle_get(group, iommu_mm->pasid); > if (handle) { > if (handle->domain->iopf_handler != iommu_sva_iopf_handler) > { > ret = -EBUSY; > goto out_unlock; > } > > refcount_inc(>users); > mutex_unlock(_sva_lock); > return handle; > } > > But it appears that this code is not lock safe. If the domain on the > PASID is not a SVA domain, the check of "handle->domain->iopf_handler != > iommu_sva_iopf_handler" could result in a use-after-free issue as the > other thread might detach the domain in between the fetch and check > lines. For the above you just need to pass in the iommu_sva_iopf_handler as an argument to attach_handle_get() and have it check it under the xa_lock. The whole thing is already protected under the ugly sva_lock. Ideally it would be protected by the group mutex.. Jason
Re: [PATCH v4 2/9] iommu: Replace sva_iommu with iommu_attach_handle
On 4/29/24 10:39 AM, Tian, Kevin wrote: From: Baolu Lu Sent: Sunday, April 28, 2024 6:22 PM On 2024/4/10 7:48, Jason Gunthorpe wrote: On Tue, Apr 09, 2024 at 10:11:28AM +0800, Baolu Lu wrote: On 4/8/24 10:19 PM, Jason Gunthorpe wrote: On Sat, Apr 06, 2024 at 02:09:34PM +0800, Baolu Lu wrote: On 4/3/24 7:59 PM, Jason Gunthorpe wrote: On Wed, Apr 03, 2024 at 09:15:12AM +0800, Lu Baolu wrote: + /* A bond already exists, just take a reference`. */ + handle = iommu_attach_handle_get(group, iommu_mm- pasid); + if (handle) { + mutex_unlock(_sva_lock); + return handle; } At least in this context this is not enough we need to ensure that the domain on the PASID is actually an SVA domain and it was installed by this mechanism, not an iommufd domain for instance. ie you probably need a type field in the iommu_attach_handle to tell what the priv is. Otherwise this seems like a great idea! Yes, you are right. For the SVA case, I will add the following changes. The IOMMUFD path will also need such enhancement. I will update it in the next version. The only use for this is the PRI callbacks right? Maybe instead of adding a handle type let's just check domain->iopf_handler ? Ie SVA will pass _sva_iopf_handler as its "type" Sorry that I don't fully understand the proposal here. I was talking specifically about the type field you suggested adding to the handle struct. Instead of adding a type field check the domain->iopf_handler to determine the domain and thus handle type. The problem is that the context code (SVA, IOMMUFD, etc.) needs to make sure that the attach handle is really what it has installed during domain attachment. The context code needs some mechanism to include some kind of "owner cookie" in the attach handle, so that it could check against it later for valid use. Right, you have a derived struct for each user and you need a way to check if casting from the general handle struct to the derived struct is OK. I'm suggesting using domain->iopf_handle as the type key. After removing the refcount from the attach handle, I am trying to make the code look like this, /* A bond already exists, just take a reference`. */ handle = iommu_attach_handle_get(group, iommu_mm->pasid); if (handle) { if (handle->domain->iopf_handler != iommu_sva_iopf_handler) { ret = -EBUSY; goto out_unlock; } refcount_inc(>users); mutex_unlock(_sva_lock); return handle; } But it appears that this code is not lock safe. If the domain on the PASID is not a SVA domain, the check of "handle->domain->iopf_handler != iommu_sva_iopf_handler" could result in a use-after-free issue as the other thread might detach the domain in between the fetch and check lines. Probably we still need to keep the refcount in the attach handle? What about Jason's another comment in his original replies? " Though I'm not convinced the refcount should be elevated into the core structure. The prior patch I showed you where the caller can provide the memory for the handle and we don't have a priv would make it easy to put the refcount in a SVA dervied handle struct without more allocation. Then we don't need this weirdness. " That sounds like we'll need a iommu_sva like structure to hold its own refcnt. Then we don't need this type check and refcnt in the core. The problem I'm facing isn't about who allocates the handle memory. Instead, there's no mechanism to synchronize access between two threads. One thread might remove the handle while another fetches and reads a member of its structure. A similar issue exists with iommu_get_domain_for_dev_pasid(). It fetches and returns a domain, but there's no guarantee that the domain will *not* be freed while the caller is still using it. One reason I introduced the reference count for attach handles is to potentially replace iommu_get_domain_for_dev_pasid(), allowing the domain to be accessible without any potential UAF issue. Best regards, baolu
RE: [PATCH v4 2/9] iommu: Replace sva_iommu with iommu_attach_handle
> From: Baolu Lu > Sent: Sunday, April 28, 2024 6:22 PM > > On 2024/4/10 7:48, Jason Gunthorpe wrote: > > On Tue, Apr 09, 2024 at 10:11:28AM +0800, Baolu Lu wrote: > >> On 4/8/24 10:19 PM, Jason Gunthorpe wrote: > >>> On Sat, Apr 06, 2024 at 02:09:34PM +0800, Baolu Lu wrote: > On 4/3/24 7:59 PM, Jason Gunthorpe wrote: > > On Wed, Apr 03, 2024 at 09:15:12AM +0800, Lu Baolu wrote: > >> + /* A bond already exists, just take a reference`. */ > >> + handle = iommu_attach_handle_get(group, iommu_mm- > >pasid); > >> + if (handle) { > >> + mutex_unlock(_sva_lock); > >> + return handle; > >>} > > At least in this context this is not enough we need to ensure that the > > domain on the PASID is actually an SVA domain and it was installed by > > this mechanism, not an iommufd domain for instance. > > > > ie you probably need a type field in the iommu_attach_handle to tell > > what the priv is. > > > > Otherwise this seems like a great idea! > Yes, you are right. For the SVA case, I will add the following changes. > The IOMMUFD path will also need such enhancement. I will update it in > the next version. > >>> The only use for this is the PRI callbacks right? Maybe instead of > >>> adding a handle type let's just check domain->iopf_handler ? > >>> > >>> Ie SVA will pass _sva_iopf_handler as its "type" > >> Sorry that I don't fully understand the proposal here. > > I was talking specifically about the type field you suggested adding > > to the handle struct. > > > > Instead of adding a type field check the domain->iopf_handler to > > determine the domain and thus handle type. > > > >> The problem is that the context code (SVA, IOMMUFD, etc.) needs to > make > >> sure that the attach handle is really what it has installed during > >> domain attachment. The context code needs some mechanism to include > some > >> kind of "owner cookie" in the attach handle, so that it could check > >> against it later for valid use. > > Right, you have a derived struct for each user and you need a way to > > check if casting from the general handle struct to the derived struct > > is OK. > > > > I'm suggesting using domain->iopf_handle as the type key. > > After removing the refcount from the attach handle, I am trying to make > the code look like this, > > /* A bond already exists, just take a reference`. */ > handle = iommu_attach_handle_get(group, iommu_mm->pasid); > if (handle) { > if (handle->domain->iopf_handler != > iommu_sva_iopf_handler) { > ret = -EBUSY; > goto out_unlock; > } > > refcount_inc(>users); > mutex_unlock(_sva_lock); > return handle; > } > > But it appears that this code is not lock safe. If the domain on the > PASID is not a SVA domain, the check of "handle->domain->iopf_handler != > iommu_sva_iopf_handler" could result in a use-after-free issue as the > other thread might detach the domain in between the fetch and check > lines. > > Probably we still need to keep the refcount in the attach handle? > What about Jason's another comment in his original replies? " Though I'm not convinced the refcount should be elevated into the core structure. The prior patch I showed you where the caller can provide the memory for the handle and we don't have a priv would make it easy to put the refcount in a SVA dervied handle struct without more allocation. Then we don't need this weirdness. " That sounds like we'll need a iommu_sva like structure to hold its own refcnt. Then we don't need this type check and refcnt in the core.
Re: [PATCH v4 2/9] iommu: Replace sva_iommu with iommu_attach_handle
On 2024/4/10 7:48, Jason Gunthorpe wrote: On Tue, Apr 09, 2024 at 10:11:28AM +0800, Baolu Lu wrote: On 4/8/24 10:19 PM, Jason Gunthorpe wrote: On Sat, Apr 06, 2024 at 02:09:34PM +0800, Baolu Lu wrote: On 4/3/24 7:59 PM, Jason Gunthorpe wrote: On Wed, Apr 03, 2024 at 09:15:12AM +0800, Lu Baolu wrote: + /* A bond already exists, just take a reference`. */ + handle = iommu_attach_handle_get(group, iommu_mm->pasid); + if (handle) { + mutex_unlock(_sva_lock); + return handle; } At least in this context this is not enough we need to ensure that the domain on the PASID is actually an SVA domain and it was installed by this mechanism, not an iommufd domain for instance. ie you probably need a type field in the iommu_attach_handle to tell what the priv is. Otherwise this seems like a great idea! Yes, you are right. For the SVA case, I will add the following changes. The IOMMUFD path will also need such enhancement. I will update it in the next version. The only use for this is the PRI callbacks right? Maybe instead of adding a handle type let's just check domain->iopf_handler ? Ie SVA will pass _sva_iopf_handler as its "type" Sorry that I don't fully understand the proposal here. I was talking specifically about the type field you suggested adding to the handle struct. Instead of adding a type field check the domain->iopf_handler to determine the domain and thus handle type. The problem is that the context code (SVA, IOMMUFD, etc.) needs to make sure that the attach handle is really what it has installed during domain attachment. The context code needs some mechanism to include some kind of "owner cookie" in the attach handle, so that it could check against it later for valid use. Right, you have a derived struct for each user and you need a way to check if casting from the general handle struct to the derived struct is OK. I'm suggesting using domain->iopf_handle as the type key. After removing the refcount from the attach handle, I am trying to make the code look like this, /* A bond already exists, just take a reference`. */ handle = iommu_attach_handle_get(group, iommu_mm->pasid); if (handle) { if (handle->domain->iopf_handler != iommu_sva_iopf_handler) { ret = -EBUSY; goto out_unlock; } refcount_inc(>users); mutex_unlock(_sva_lock); return handle; } But it appears that this code is not lock safe. If the domain on the PASID is not a SVA domain, the check of "handle->domain->iopf_handler != iommu_sva_iopf_handler" could result in a use-after-free issue as the other thread might detach the domain in between the fetch and check lines. Probably we still need to keep the refcount in the attach handle? Best regards, baolu
Re: [PATCH v4 2/9] iommu: Replace sva_iommu with iommu_attach_handle
On 4/10/24 7:48 AM, Jason Gunthorpe wrote: On Tue, Apr 09, 2024 at 10:11:28AM +0800, Baolu Lu wrote: On 4/8/24 10:19 PM, Jason Gunthorpe wrote: On Sat, Apr 06, 2024 at 02:09:34PM +0800, Baolu Lu wrote: On 4/3/24 7:59 PM, Jason Gunthorpe wrote: On Wed, Apr 03, 2024 at 09:15:12AM +0800, Lu Baolu wrote: + /* A bond already exists, just take a reference`. */ + handle = iommu_attach_handle_get(group, iommu_mm->pasid); + if (handle) { + mutex_unlock(_sva_lock); + return handle; } At least in this context this is not enough we need to ensure that the domain on the PASID is actually an SVA domain and it was installed by this mechanism, not an iommufd domain for instance. ie you probably need a type field in the iommu_attach_handle to tell what the priv is. Otherwise this seems like a great idea! Yes, you are right. For the SVA case, I will add the following changes. The IOMMUFD path will also need such enhancement. I will update it in the next version. The only use for this is the PRI callbacks right? Maybe instead of adding a handle type let's just check domain->iopf_handler ? Ie SVA will pass _sva_iopf_handler as its "type" Sorry that I don't fully understand the proposal here. I was talking specifically about the type field you suggested adding to the handle struct. Instead of adding a type field check the domain->iopf_handler to determine the domain and thus handle type. The problem is that the context code (SVA, IOMMUFD, etc.) needs to make sure that the attach handle is really what it has installed during domain attachment. The context code needs some mechanism to include some kind of "owner cookie" in the attach handle, so that it could check against it later for valid use. Right, you have a derived struct for each user and you need a way to check if casting from the general handle struct to the derived struct is OK. I'm suggesting using domain->iopf_handle as the type key. Oh, I see. It works. Thanks! Best regards, baolu
Re: [PATCH v4 2/9] iommu: Replace sva_iommu with iommu_attach_handle
On Tue, Apr 09, 2024 at 10:11:28AM +0800, Baolu Lu wrote: > On 4/8/24 10:19 PM, Jason Gunthorpe wrote: > > On Sat, Apr 06, 2024 at 02:09:34PM +0800, Baolu Lu wrote: > > > On 4/3/24 7:59 PM, Jason Gunthorpe wrote: > > > > On Wed, Apr 03, 2024 at 09:15:12AM +0800, Lu Baolu wrote: > > > > > + /* A bond already exists, just take a reference`. */ > > > > > + handle = iommu_attach_handle_get(group, iommu_mm->pasid); > > > > > + if (handle) { > > > > > + mutex_unlock(_sva_lock); > > > > > + return handle; > > > > > } > > > > At least in this context this is not enough we need to ensure that the > > > > domain on the PASID is actually an SVA domain and it was installed by > > > > this mechanism, not an iommufd domain for instance. > > > > > > > > ie you probably need a type field in the iommu_attach_handle to tell > > > > what the priv is. > > > > > > > > Otherwise this seems like a great idea! > > > Yes, you are right. For the SVA case, I will add the following changes. > > > The IOMMUFD path will also need such enhancement. I will update it in > > > the next version. > > The only use for this is the PRI callbacks right? Maybe instead of > > adding a handle type let's just check domain->iopf_handler ? > > > > Ie SVA will pass _sva_iopf_handler as its "type" > > Sorry that I don't fully understand the proposal here. I was talking specifically about the type field you suggested adding to the handle struct. Instead of adding a type field check the domain->iopf_handler to determine the domain and thus handle type. > The problem is that the context code (SVA, IOMMUFD, etc.) needs to make > sure that the attach handle is really what it has installed during > domain attachment. The context code needs some mechanism to include some > kind of "owner cookie" in the attach handle, so that it could check > against it later for valid use. Right, you have a derived struct for each user and you need a way to check if casting from the general handle struct to the derived struct is OK. I'm suggesting using domain->iopf_handle as the type key. Jason
Re: [PATCH v4 2/9] iommu: Replace sva_iommu with iommu_attach_handle
On 4/8/24 10:19 PM, Jason Gunthorpe wrote: On Sat, Apr 06, 2024 at 02:09:34PM +0800, Baolu Lu wrote: On 4/3/24 7:59 PM, Jason Gunthorpe wrote: On Wed, Apr 03, 2024 at 09:15:12AM +0800, Lu Baolu wrote: + /* A bond already exists, just take a reference`. */ + handle = iommu_attach_handle_get(group, iommu_mm->pasid); + if (handle) { + mutex_unlock(_sva_lock); + return handle; } At least in this context this is not enough we need to ensure that the domain on the PASID is actually an SVA domain and it was installed by this mechanism, not an iommufd domain for instance. ie you probably need a type field in the iommu_attach_handle to tell what the priv is. Otherwise this seems like a great idea! Yes, you are right. For the SVA case, I will add the following changes. The IOMMUFD path will also need such enhancement. I will update it in the next version. The only use for this is the PRI callbacks right? Maybe instead of adding a handle type let's just check domain->iopf_handler ? Ie SVA will pass _sva_iopf_handler as its "type" Sorry that I don't fully understand the proposal here. We need to get the attach handle at least in below cases: 1. In the iommu_sva_bind_device() path so that the existing bind could be reused. 2. In the iommu_report_device_fault() path so that the context-specific data could be used in the fault handler. The problem is that the context code (SVA, IOMMUFD, etc.) needs to make sure that the attach handle is really what it has installed during domain attachment. The context code needs some mechanism to include some kind of "owner cookie" in the attach handle, so that it could check against it later for valid use. Best regards, baolu
Re: [PATCH v4 2/9] iommu: Replace sva_iommu with iommu_attach_handle
On Sat, Apr 06, 2024 at 02:09:34PM +0800, Baolu Lu wrote: > On 4/3/24 7:59 PM, Jason Gunthorpe wrote: > > On Wed, Apr 03, 2024 at 09:15:12AM +0800, Lu Baolu wrote: > > > + /* A bond already exists, just take a reference`. */ > > > + handle = iommu_attach_handle_get(group, iommu_mm->pasid); > > > + if (handle) { > > > + mutex_unlock(_sva_lock); > > > + return handle; > > > } > > At least in this context this is not enough we need to ensure that the > > domain on the PASID is actually an SVA domain and it was installed by > > this mechanism, not an iommufd domain for instance. > > > > ie you probably need a type field in the iommu_attach_handle to tell > > what the priv is. > > > > Otherwise this seems like a great idea! > > Yes, you are right. For the SVA case, I will add the following changes. > The IOMMUFD path will also need such enhancement. I will update it in > the next version. The only use for this is the PRI callbacks right? Maybe instead of adding a handle type let's just check domain->iopf_handler ? Ie SVA will pass _sva_iopf_handler as its "type" Jason
Re: [PATCH v4 2/9] iommu: Replace sva_iommu with iommu_attach_handle
On 4/3/24 7:59 PM, Jason Gunthorpe wrote: - iommu_detach_device_pasid(domain, dev, iommu_mm->pasid); - if (--domain->users == 0) { - list_del(>next); - iommu_domain_free(domain); + iommu_attach_handle_put(handle); + if (refcount_read(>users) == 1) { + iommu_detach_device_pasid(domain, dev, iommu_mm->pasid); + if (--domain->users == 0) { + list_del(>next); + iommu_domain_free(domain); + } } Though I'm not convinced the refcount should be elevated into the core structure. The prior patch I showed you where the caller can provide the memory for the handle and we don't have a priv would make it easy to put the refcount in a SVA dervied handle struct without more allocation. Then we don't need this weirdness. It's fine to move the refcount out of the core and allow the caller to specify and manage its own attach handler. The refcount would then be managed by the SVA code. For the IOMMUFD case, we've discussed that all outstanding iopf's should be automatically responded in the detach process. This ensures the attach handle won't be used once the detach process completes. Therefore, if this is true, there appears to be no need for a refcount for IOMMUFD. mutex_unlock(_sva_lock); - kfree(handle); Also do we need iommu_sva_lock here anymore? I wonder if the group mutex would be sufficient.. The iommu_sva_lock protects the whole process of a mm binding, from pasid allocation to domain attachment. While the group mutex only protects the data within it structure. I don't think we could replace iommu_sva_lock with group mutex in this patch. Or any misunderstanding? Best regards, baolu
Re: [PATCH v4 2/9] iommu: Replace sva_iommu with iommu_attach_handle
On 4/3/24 7:59 PM, Jason Gunthorpe wrote: On Wed, Apr 03, 2024 at 09:15:12AM +0800, Lu Baolu wrote: + /* A bond already exists, just take a reference`. */ + handle = iommu_attach_handle_get(group, iommu_mm->pasid); + if (handle) { + mutex_unlock(_sva_lock); + return handle; } At least in this context this is not enough we need to ensure that the domain on the PASID is actually an SVA domain and it was installed by this mechanism, not an iommufd domain for instance. ie you probably need a type field in the iommu_attach_handle to tell what the priv is. Otherwise this seems like a great idea! Yes, you are right. For the SVA case, I will add the following changes. The IOMMUFD path will also need such enhancement. I will update it in the next version. diff --git a/drivers/iommu/iommu-priv.h b/drivers/iommu/iommu-priv.h index 08c0667cef54..9aee70f87a21 100644 --- a/drivers/iommu/iommu-priv.h +++ b/drivers/iommu/iommu-priv.h @@ -28,9 +28,22 @@ void iommu_device_unregister_bus(struct iommu_device *iommu, const struct bus_type *bus, struct notifier_block *nb); +enum attach_handle_type { + ATTACH_HANDLE_TYPE_DEFAULT = 0, + ATTACH_HANDLE_TYPE_SVA, + ATTACH_HANDLE_TYPE_IOMMUFD, +}; + struct iommu_attach_handle { struct iommu_domain *domain; refcount_t users; + + /* +* Set by the attach interface callers. The type field could be used +* by the caller to identify whether the priv field was installed by +* them. +*/ + enum attach_handle_type type; void*priv; }; diff --git a/drivers/iommu/iommu-sva.c b/drivers/iommu/iommu-sva.c index c66cf26137bf..3eb664cc3f3a 100644 --- a/drivers/iommu/iommu-sva.c +++ b/drivers/iommu/iommu-sva.c @@ -90,7 +90,11 @@ struct iommu_attach_handle *iommu_sva_bind_device(struct device *dev, struct mm_ handle = iommu_attach_handle_get(group, iommu_mm->pasid); if (handle) { mutex_unlock(_sva_lock); - return handle; + if (handle->type == ATTACH_HANDLE_TYPE_SVA) + return handle; + + iommu_attach_handle_put(handle); + return ERR_PTR(-EBUSY); } /* Search for an existing domain. */ @@ -118,6 +122,7 @@ struct iommu_attach_handle *iommu_sva_bind_device(struct device *dev, struct mm_ out: handle = iommu_attach_handle_get(group, iommu_mm->pasid); mutex_unlock(_sva_lock); + handle->type = ATTACH_HANDLE_TYPE_SVA; handle->priv = dev; return handle; Best regards, baolu
Re: [PATCH v4 2/9] iommu: Replace sva_iommu with iommu_attach_handle
On Wed, Apr 03, 2024 at 09:15:12AM +0800, Lu Baolu wrote: > + /* A bond already exists, just take a reference`. */ > + handle = iommu_attach_handle_get(group, iommu_mm->pasid); > + if (handle) { > + mutex_unlock(_sva_lock); > + return handle; > } At least in this context this is not enough we need to ensure that the domain on the PASID is actually an SVA domain and it was installed by this mechanism, not an iommufd domain for instance. ie you probably need a type field in the iommu_attach_handle to tell what the priv is. Otherwise this seems like a great idea! > - iommu_detach_device_pasid(domain, dev, iommu_mm->pasid); > - if (--domain->users == 0) { > - list_del(>next); > - iommu_domain_free(domain); > + iommu_attach_handle_put(handle); > + if (refcount_read(>users) == 1) { > + iommu_detach_device_pasid(domain, dev, iommu_mm->pasid); > + if (--domain->users == 0) { > + list_del(>next); > + iommu_domain_free(domain); > + } > } Though I'm not convinced the refcount should be elevated into the core structure. The prior patch I showed you where the caller can provide the memory for the handle and we don't have a priv would make it easy to put the refcount in a SVA dervied handle struct without more allocation. Then we don't need this weirdness. > mutex_unlock(_sva_lock); > - kfree(handle); Also do we need iommu_sva_lock here anymore? I wonder if the group mutex would be sufficient.. Jason