Message ID | 20221111144433.2421680-1-yangyingliang@huawei.com (mailing list archive) |
---|---|
State | Changes Requested |
Headers | show |
Series | [v2] scsi: scsi_transport_sas: fix error handling in sas_rphy_add() | expand |
On Fri, 2022-11-11 at 22:44 +0800, Yang Yingliang wrote: > In sas_rphy_add(), if transport_add_device() fails, the device > is not added, the return value is not checked, it won't goto > error path, when removing rphy in normal remove path, it causes > null-ptr-deref, because transport_remove_device() is called to > remove the device that was not added. > > Unable to handle kernel NULL pointer dereference at virtual address > 0000000000000108 > pc : device_del+0x54/0x3d0 > lr : device_del+0x37c/0x3d0 > Call trace: > device_del+0x54/0x3d0 > attribute_container_class_device_del+0x28/0x38 > transport_remove_classdev+0x6c/0x80 > attribute_container_device_trigger+0x108/0x110 > transport_remove_device+0x28/0x38 > sas_rphy_remove+0x50/0x78 [scsi_transport_sas] > sas_port_delete+0x30/0x148 [scsi_transport_sas] > do_sas_phy_delete+0x78/0x80 [scsi_transport_sas] > device_for_each_child+0x68/0xb0 > sas_remove_children+0x30/0x50 [scsi_transport_sas] > sas_rphy_remove+0x38/0x78 [scsi_transport_sas] > sas_port_delete+0x30/0x148 [scsi_transport_sas] > do_sas_phy_delete+0x78/0x80 [scsi_transport_sas] > device_for_each_child+0x68/0xb0 > sas_remove_children+0x30/0x50 [scsi_transport_sas] > sas_remove_host+0x20/0x38 [scsi_transport_sas] > scsih_remove+0xd8/0x420 [mpt3sas] > > Fix this by checking and handling return value of > transport_add_device() > in sas_rphy_add(). > > Fixes: c7ebbbce366c ("[SCSI] SAS transport class") > Signed-off-by: Yang Yingliang <yangyingliang@huawei.com> > --- > v1 -> v2: > Update commit message. > --- > drivers/scsi/scsi_transport_sas.c | 6 +++++- > 1 file changed, 5 insertions(+), 1 deletion(-) > > diff --git a/drivers/scsi/scsi_transport_sas.c > b/drivers/scsi/scsi_transport_sas.c > index 74b99f2b0b74..accc0afa8f77 100644 > --- a/drivers/scsi/scsi_transport_sas.c > +++ b/drivers/scsi/scsi_transport_sas.c > @@ -1526,7 +1526,11 @@ int sas_rphy_add(struct sas_rphy *rphy) > error = device_add(&rphy->dev); > if (error) > return error; > - transport_add_device(&rphy->dev); > + error = transport_add_device(&rphy->dev); > + if (error) { > + device_del(&rphy->dev); > + return error; > + } > transport_configure_device(&rphy->dev); > if (sas_bsg_initialize(shost, rphy)) > printk("fail to a bsg device %s\n", dev_name(&rphy- > >dev)); There is a slight problem with doing this in that if transport_device_add() ever fails it's likely because memory pressure caused the allocation of the internal_container to fail. What that means is that the visible sysfs attributes don't get added, but otherwise the rphy is fully functional as far as the driver sees it, so this condition doesn't have to be a fatal error which kills the device. There are two ways of handling this: 1. The above to move the condition from an ignored to a fatal error. It's so rare that we almost never see it in practice and if it ever happened, the machine is so low on memory that something else is bound to fail an allocation and kill the device anyway, so treating it as non-fatal likely serves no purpose. 2. Simply to make the assumption that transport_remove_device() is idempotent true by adding a flag in the internal_class to signify removal is required. This would preserve current behaviour and have the bonus that it only requires a single patch, not one patch per transport class object that has this problem. I'd probably prefer 2. since it's way less work, but others might have different opinions. James
+Cc: Greg Hi Greg, On 2022/11/11 23:51, James Bottomley wrote: > On Fri, 2022-11-11 at 22:44 +0800, Yang Yingliang wrote: >> In sas_rphy_add(), if transport_add_device() fails, the device >> is not added, the return value is not checked, it won't goto >> error path, when removing rphy in normal remove path, it causes >> null-ptr-deref, because transport_remove_device() is called to >> remove the device that was not added. >> >> Unable to handle kernel NULL pointer dereference at virtual address >> 0000000000000108 >> pc : device_del+0x54/0x3d0 >> lr : device_del+0x37c/0x3d0 >> Call trace: >> device_del+0x54/0x3d0 >> attribute_container_class_device_del+0x28/0x38 >> transport_remove_classdev+0x6c/0x80 >> attribute_container_device_trigger+0x108/0x110 >> transport_remove_device+0x28/0x38 >> sas_rphy_remove+0x50/0x78 [scsi_transport_sas] >> sas_port_delete+0x30/0x148 [scsi_transport_sas] >> do_sas_phy_delete+0x78/0x80 [scsi_transport_sas] >> device_for_each_child+0x68/0xb0 >> sas_remove_children+0x30/0x50 [scsi_transport_sas] >> sas_rphy_remove+0x38/0x78 [scsi_transport_sas] >> sas_port_delete+0x30/0x148 [scsi_transport_sas] >> do_sas_phy_delete+0x78/0x80 [scsi_transport_sas] >> device_for_each_child+0x68/0xb0 >> sas_remove_children+0x30/0x50 [scsi_transport_sas] >> sas_remove_host+0x20/0x38 [scsi_transport_sas] >> scsih_remove+0xd8/0x420 [mpt3sas] >> >> Fix this by checking and handling return value of >> transport_add_device() >> in sas_rphy_add(). >> >> Fixes: c7ebbbce366c ("[SCSI] SAS transport class") >> Signed-off-by: Yang Yingliang <yangyingliang@huawei.com> >> --- >> v1 -> v2: >> Update commit message. >> --- >> drivers/scsi/scsi_transport_sas.c | 6 +++++- >> 1 file changed, 5 insertions(+), 1 deletion(-) >> >> diff --git a/drivers/scsi/scsi_transport_sas.c >> b/drivers/scsi/scsi_transport_sas.c >> index 74b99f2b0b74..accc0afa8f77 100644 >> --- a/drivers/scsi/scsi_transport_sas.c >> +++ b/drivers/scsi/scsi_transport_sas.c >> @@ -1526,7 +1526,11 @@ int sas_rphy_add(struct sas_rphy *rphy) >> error = device_add(&rphy->dev); >> if (error) >> return error; >> - transport_add_device(&rphy->dev); >> + error = transport_add_device(&rphy->dev); >> + if (error) { >> + device_del(&rphy->dev); >> + return error; >> + } >> transport_configure_device(&rphy->dev); >> if (sas_bsg_initialize(shost, rphy)) >> printk("fail to a bsg device %s\n", dev_name(&rphy- >>> dev)); > There is a slight problem with doing this in that if > transport_device_add() ever fails it's likely because memory pressure > caused the allocation of the internal_container to fail. What that > means is that the visible sysfs attributes don't get added, but > otherwise the rphy is fully functional as far as the driver sees it, so > this condition doesn't have to be a fatal error which kills the device. > > There are two ways of handling this: > > 1. The above to move the condition from an ignored to a fatal error. > It's so rare that we almost never see it in practice and if it > ever happened, the machine is so low on memory that something > else is bound to fail an allocation and kill the device anyway, > so treating it as non-fatal likely serves no purpose. > 2. Simply to make the assumption that transport_remove_device() is > idempotent true by adding a flag in the internal_class to signify > removal is required. This would preserve current behaviour and > have the bonus that it only requires a single patch, not one > patch per transport class object that has this problem. > > I'd probably prefer 2. since it's way less work, but others might have > different opinions. Current some callers ignore the return value of transport_add_device(), if it fails, it will cause null-ptr-deref in transport_remove_device(). James suggested that add some check in transport_remove_device(), so all can be fix in one patch. Do you have any suggestion for this ? Thanks, Yang > > James > > .
On 18/11/2022 03:11, Yang Yingliang wrote: >>>> ); >> There is a slight problem with doing this in that if >> transport_device_add() ever fails it's likely because memory pressure >> caused the allocation of the internal_container to fail. What that >> means is that the visible sysfs attributes don't get added, but >> otherwise the rphy is fully functional as far as the driver sees it, so >> this condition doesn't have to be a fatal error which kills the device. >> >> There are two ways of handling this: >> >> 1. The above to move the condition from an ignored to a fatal error. >> It's so rare that we almost never see it in practice and if it >> ever happened, the machine is so low on memory that something >> else is bound to fail an allocation and kill the device anyway, >> so treating it as non-fatal likely serves no purpose. >> 2. Simply to make the assumption that transport_remove_device() is >> idempotent true by adding a flag in the internal_class to signify >> removal is required. This would preserve current behaviour and >> have the bonus that it only requires a single patch, not one >> patch per transport class object that has this problem. >> >> I'd probably prefer 2. since it's way less work, but others might have >> different opinions. > Current some callers ignore the return value of transport_add_device(), > if it fails, > it will cause null-ptr-deref in transport_remove_device(). > > James suggested that add some check in transport_remove_device(), so all > can > be fix in one patch. > > Do you have any suggestion for this ? Personally I prefer 1. However did you develop a prototype patch for how 2. would look? And how many changes are still required for 1.? Thanks, John
On 2022/11/18 17:18, John Garry wrote: > On 18/11/2022 03:11, Yang Yingliang wrote: >>>>> ); >>> There is a slight problem with doing this in that if >>> transport_device_add() ever fails it's likely because memory pressure >>> caused the allocation of the internal_container to fail. What that >>> means is that the visible sysfs attributes don't get added, but >>> otherwise the rphy is fully functional as far as the driver sees it, so >>> this condition doesn't have to be a fatal error which kills the device. >>> >>> There are two ways of handling this: >>> >>> 1. The above to move the condition from an ignored to a fatal >>> error. >>> It's so rare that we almost never see it in practice and if it >>> ever happened, the machine is so low on memory that something >>> else is bound to fail an allocation and kill the device anyway, >>> so treating it as non-fatal likely serves no purpose. >>> 2. Simply to make the assumption that transport_remove_device() is >>> idempotent true by adding a flag in the internal_class to >>> signify >>> removal is required. This would preserve current behaviour and >>> have the bonus that it only requires a single patch, not one >>> patch per transport class object that has this problem. >>> >>> I'd probably prefer 2. since it's way less work, but others might have >>> different opinions. >> Current some callers ignore the return value of >> transport_add_device(), if it fails, >> it will cause null-ptr-deref in transport_remove_device(). >> >> James suggested that add some check in transport_remove_device(), so >> all can >> be fix in one patch. >> >> Do you have any suggestion for this ? > > Personally I prefer 1. However did you develop a prototype patch for > how 2. would look? And how many changes are still required for 1.? For 1, in total, there are 8 places need be checked in drivers/scsi/scsi_transport_sas.c, 2 places in drivers/scsi/scsi_sysfs.c, 3 places in drivers/scsi/scsi_transport_fc.c, 2 places in drivers/scsi/scsi_transport_srp.c, 1 place For 2, I think we can use device_is_registered() to check if add operation is successful, may be like this (not test yet): diff --git a/drivers/base/transport_class.c b/drivers/base/transport_class.c index ccc86206e508..ac41be7b724e 100644 --- a/drivers/base/transport_class.c +++ b/drivers/base/transport_class.c @@ -227,9 +227,11 @@ static int transport_remove_classdev(struct attribute_container *cont, tclass->remove(tcont, dev, classdev); if (tclass->remove != anon_transport_dummy_function) { - if (tcont->statistics) - sysfs_remove_group(&classdev->kobj, tcont->statistics); - attribute_container_class_device_del(classdev); + if (device_is_registered(classdev)) { + if (tcont->statistics) + sysfs_remove_group(&classdev->kobj, tcont->statistics); + attribute_container_class_device_del(classdev); + } } return 0; Thanks, Yang > > Thanks, > John > .
On 19/11/2022 08:58, Yang Yingliang wrote: >> Personally I prefer 1. However did you develop a prototype patch for >> how 2. would look? And how many changes are still required for 1.? > For 1, in total, there are 8 places need be checked > in drivers/scsi/scsi_transport_sas.c, 2 places > in drivers/scsi/scsi_sysfs.c, 3 places > in drivers/scsi/scsi_transport_fc.c, 2 places > in drivers/scsi/scsi_transport_srp.c, 1 place and in linux-next there are 4x places which do already check the return code... Not sure what's best to do. I'll leave it to James' wisdom. However, we do seem to have a common pattern: error = device_add(dev); if (error) return error; transport_add_device(dev); transport_configure_device(dev); Could we make an "add" (which does as above pattern) and "remove" helper? It might simplify things such that we not only fixing the possible crash but also reducing code. Thanks, John > > For 2, I think we can use device_is_registered() to check if add > operation is successful, may be like this (not test yet): > > diff --git a/drivers/base/transport_class.c > b/drivers/base/transport_class.c > index ccc86206e508..ac41be7b724e 100644 > --- a/drivers/base/transport_class.c > +++ b/drivers/base/transport_class.c > @@ -227,9 +227,11 @@ static int transport_remove_classdev(struct > attribute_container *cont, > tclass->remove(tcont, dev, classdev); > > if (tclass->remove != anon_transport_dummy_function) { > - if (tcont->statistics) > - sysfs_remove_group(&classdev->kobj, tcont->statistics); > - attribute_container_class_device_del(classdev); > + if (device_is_registered(classdev)) { > + if (tcont->statistics) > + sysfs_remove_group(&classdev->kobj, tcont->statistics); > + attribute_container_class_device_del(classdev); > + } > } > > return 0;
diff --git a/drivers/scsi/scsi_transport_sas.c b/drivers/scsi/scsi_transport_sas.c index 74b99f2b0b74..accc0afa8f77 100644 --- a/drivers/scsi/scsi_transport_sas.c +++ b/drivers/scsi/scsi_transport_sas.c @@ -1526,7 +1526,11 @@ int sas_rphy_add(struct sas_rphy *rphy) error = device_add(&rphy->dev); if (error) return error; - transport_add_device(&rphy->dev); + error = transport_add_device(&rphy->dev); + if (error) { + device_del(&rphy->dev); + return error; + } transport_configure_device(&rphy->dev); if (sas_bsg_initialize(shost, rphy)) printk("fail to a bsg device %s\n", dev_name(&rphy->dev));
In sas_rphy_add(), if transport_add_device() fails, the device is not added, the return value is not checked, it won't goto error path, when removing rphy in normal remove path, it causes null-ptr-deref, because transport_remove_device() is called to remove the device that was not added. Unable to handle kernel NULL pointer dereference at virtual address 0000000000000108 pc : device_del+0x54/0x3d0 lr : device_del+0x37c/0x3d0 Call trace: device_del+0x54/0x3d0 attribute_container_class_device_del+0x28/0x38 transport_remove_classdev+0x6c/0x80 attribute_container_device_trigger+0x108/0x110 transport_remove_device+0x28/0x38 sas_rphy_remove+0x50/0x78 [scsi_transport_sas] sas_port_delete+0x30/0x148 [scsi_transport_sas] do_sas_phy_delete+0x78/0x80 [scsi_transport_sas] device_for_each_child+0x68/0xb0 sas_remove_children+0x30/0x50 [scsi_transport_sas] sas_rphy_remove+0x38/0x78 [scsi_transport_sas] sas_port_delete+0x30/0x148 [scsi_transport_sas] do_sas_phy_delete+0x78/0x80 [scsi_transport_sas] device_for_each_child+0x68/0xb0 sas_remove_children+0x30/0x50 [scsi_transport_sas] sas_remove_host+0x20/0x38 [scsi_transport_sas] scsih_remove+0xd8/0x420 [mpt3sas] Fix this by checking and handling return value of transport_add_device() in sas_rphy_add(). Fixes: c7ebbbce366c ("[SCSI] SAS transport class") Signed-off-by: Yang Yingliang <yangyingliang@huawei.com> --- v1 -> v2: Update commit message. --- drivers/scsi/scsi_transport_sas.c | 6 +++++- 1 file changed, 5 insertions(+), 1 deletion(-)