ata: ata_generic: Do not bind to devices that are not IDE controllers
Niklas Cassel
maintainerPATCH 1/1 ata: ata_generic: Do not bind to devices that are not IDE controllers
syzbot force-bound ata_generic to 0000:00:03.0 on a QEMU arm64 virt machine. That device is a virtio-blk-pci device holding the root file system, and it reports PCI class 0x010000, i.e. PCI_CLASS_STORAGE_SCSI.
QEMU gives the virtio-blk-pci device a legacy virtio I/O BAR0 and a 4 KiB MSI-X BAR1, so both resources are non-empty, the port is not discarded, and the device control register ends up in the middle of the MSI-X table.
The emulated device rejects the byte write, and arm64 reports the resulting bus error as a fatal synchronous external abort:
Internal error: synchronous external abort: 0000000096000050 [#1] SMP pc : ata_sff_freeze+0x7c/0x90 drivers/ata/libata-sff.c:1606 Call trace: ata_sff_freeze+0x7c/0x90 ata_eh_freeze_port+0x34/0x5c ata_host_start+0x13c/0x228 ata_pci_sff_activate_host+0x50/0x340 ata_pci_init_one+0x19c/0x1d8 ata_pci_bmdma_init_one+0x14/0x20 ata_generic_init_one+0xc4/0x1ac local_pci_probe+0x40/0xa8 pci_device_probe+0xd8/0x288 really_probe+0xbc/0x2bc device_driver_attach+0x48/0xb4 bind_store+0x7c/0xd8
Refuse devices which neither report the IDE class nor appear in our ID table. Table entries keep binding as before, because some of the listed controllers cannot be assumed to report the IDE class. A controller which needs ata_generic but does not report the IDE class should get an ID table entry, which is what the table is for.
Binding a driver to unrelated hardware requires root and is what driver_override is meant to do, so this does not fix a privilege boundary.
This change only stops ata_generic from binding to a PCI device which it has no reason to believe to be an IDE controller.
--- drivers/ata/ata_generic.c | 16 ++++++++++++++++ 1 file changed, 16 insertions(+)
drivers/ata/ata_generic.c +16 -0
@@ -150,6 +150,9 @@ static int is_intel_ider(struct pci_dev *dev)return 1 ; } +/* Forward declaration for the pci_match_id() call in ata_generic_init_one() */ +static const struct pci_device_id ata_generic[]; +/** * ata_generic_init_one - attach generic IDE * @dev: PCI device found
@@ -172,6 +175,19 @@ static int ata_generic_init_one(struct pci_dev *dev, const struct pci_device_id };const struct ata_port_info *ppi[] = { &info,NULL }; +/* +* A device matched through driver_override or through an ID added +* with new_id does not come from our ID table, so pci_match_device() +* hands us a synthetic ID with no driver_data and none of the checks +* below apply. Probing maps BAR0 and BAR1 as the ATA command and +* control blocks and writes to the device control register, so only +* continue for devices which report the IDE class or which we list +* ourselves. +*/ +if (!pci_match_id (ata_generic, dev) && + (dev->class >>8 ) !=PCI_CLASS_STORAGE_IDE ) +return -ENODEV ; +/* Don't use the generic entry unless instructed to do so */ if ((id->driver_data &ATA_GEN_CLASS_MATCH ) && all_generic_ide ==0 )return -ENODEV ;
Signature · 1 line
2.55.0
[email protected]
AI reviewThank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider:
Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider:
Pre-existing issues:
- [High] Preexisting Use-After-Free in generic_set_mode() when an IDE device is bound via sysfs new_id and the ID is later removed. --
--- Patch [1]: [PATCH] ata: ata_generic: Do not bind to devices that are not IDE controllers --- commit 6571983234fdb7767d7bb9ad7f55e962d8eec259
Author: Niklas Cassel <[email protected]>
ata: ata_generic: Do not bind to devices that are not IDE controllers
This commit prevents ata_generic from binding to devices that are not IDE controllers unless explicitly listed in the ID table. It resolves a fatal synchronous external abort caused by force-binding to a virtio-blk-pci device.
Quoted 4 lines
drivers/ata/ata_generic.c +0
-0
[ ... ]
Quoted 7 lines
drivers/ata/ata_generic.c +3
-0@@ -172,6 +175,19 @@ static int ata_generic_init_one(struct pci_dev *dev, const struct pci_device_id };const struct ata_port_info *ppi[] = { &info,NULL }; +/* +* A device matched through driver_override or through an ID added +* with new_id does not come from our ID table, so pci_match_device()
drivers/ata/ata_generic.c static int ata_generic_init_one(struct pci_dev *dev, const struct pci_device_id · line 190
Quoted 10 lines
drivers/ata/ata_generic.c +10
-0+* hands us a synthetic ID with no driver_data and none of the checks +* below apply. Probing maps BAR0 and BAR1 as the ATA command and +* control blocks and writes to the device control register, so only +* continue for devices which report the IDE class or which we list +* ourselves. +*/ +if (!pci_match_id (ata_generic, dev) && + (dev->class >>8 ) !=PCI_CLASS_STORAGE_IDE ) +return -ENODEV ; +
[Severity: High]
This is a pre-existing issue, but does this bypass still leave the driver vulnerable to a use-after-free when a dynamic ID is removed?
If a device with PCI_CLASS_STORAGE_IDE is bound via the sysfs new_id mechanism, the PCI core passes a dynamically allocated struct pci_device_id pointer to ata_generic_init_one().
This bypasses the new -ENODEV check and eventually passes the dynamic ID pointer down to be stored in the host private data:
drivers/ata/ata_generic.c:ata_generic_init_one() {
...
return ata_pci_bmdma_init_one(dev, ppi, &generic_sht, (void *)id, 0);
}If root later writes to the remove_id sysfs attribute, the PCI core frees the dynamic ID without unbinding the device. Subsequently, if an ATA link reset occurs, the freed memory is dereferenced:
drivers/ata/ata_generic.c:generic_set_mode() {
...
const struct pci_device_id *id = ap->host->private_data;
...
if (id->driver_data & ATA_GEN_FORCE_DMA) {
...
}Could the driver avoid storing the synthetic ID in private data, or copy the ID contents to prevent accessing freed memory?
Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=1
Quoted 36 lines · Niklas Cassel
syzbot force-bound
ata_genericto 0000:00:03.0 on a QEMU arm64 virt machine. That device is a virtio-blk-pci device holding the root file system, and it reports PCI class0x010000, i.e.PCI_CLASS_STORAGE_SCSI.QEMU gives the virtio-blk-pci device a legacy virtio I/O BAR0 and a 4 KiB MSI-X BAR1, so both resources are non-empty, the port is not discarded, and the device control register ends up in the middle of the MSI-X table.
The emulated device rejects the byte write, and arm64 reports the resulting bus error as a fatal synchronous external abort:
Internal error: synchronous external abort: 0000000096000050 [#1] SMP pc : ata_sff_freeze+0x7c/0x90 drivers/ata/libata-sff.c:1606 Call trace: ata_sff_freeze+0x7c/0x90 ata_eh_freeze_port+0x34/0x5c ata_host_start+0x13c/0x228 ata_pci_sff_activate_host+0x50/0x340 ata_pci_init_one+0x19c/0x1d8 ata_pci_bmdma_init_one+0x14/0x20 ata_generic_init_one+0xc4/0x1ac local_pci_probe+0x40/0xa8 pci_device_probe+0xd8/0x288 really_probe+0xbc/0x2bc device_driver_attach+0x48/0xb4 bind_store+0x7c/0xd8Refuse devices which neither report the IDE class nor appear in our ID table. Table entries keep binding as before, because some of the listed controllers cannot be assumed to report the IDE class. A controller which needs
ata_genericbut does not report the IDE class should get an ID table entry, which is what the table is for.Binding a driver to unrelated hardware requires root and is what
driver_overrideis meant to do, so this does not fix a privilege boundary.
Quoted 6 lines
This change only stops
ata_genericfrom binding to a PCI device which it has no reason to believe to be an IDE controller.Reported-by [email protected]Signed-off-by Niklas Cassel <[email protected]>
Looks sensible to me, and very surprising that this problem was not cought before...
Signature · 2 lines
Damien Le Moal Western Digital Research
Niklas Cassel
maintainerQuoted 40 lines · Damien Le Moal
Quoted 39 lines · Niklas Cassel
syzbot force-bound
ata_genericto 0000:00:03.0 on a QEMU arm64 virt machine. That device is a virtio-blk-pci device holding the root file system, and it reports PCI class0x010000, i.e.PCI_CLASS_STORAGE_SCSI.QEMU gives the virtio-blk-pci device a legacy virtio I/O BAR0 and a 4 KiB MSI-X BAR1, so both resources are non-empty, the port is not discarded, and the device control register ends up in the middle of the MSI-X table.
The emulated device rejects the byte write, and arm64 reports the resulting bus error as a fatal synchronous external abort:
Internal error: synchronous external abort: 0000000096000050 [#1] SMP pc : ata_sff_freeze+0x7c/0x90 drivers/ata/libata-sff.c:1606 Call trace: ata_sff_freeze+0x7c/0x90 ata_eh_freeze_port+0x34/0x5c ata_host_start+0x13c/0x228 ata_pci_sff_activate_host+0x50/0x340 ata_pci_init_one+0x19c/0x1d8 ata_pci_bmdma_init_one+0x14/0x20 ata_generic_init_one+0xc4/0x1ac local_pci_probe+0x40/0xa8 pci_device_probe+0xd8/0x288 really_probe+0xbc/0x2bc device_driver_attach+0x48/0xb4 bind_store+0x7c/0xd8Refuse devices which neither report the IDE class nor appear in our ID table. Table entries keep binding as before, because some of the listed controllers cannot be assumed to report the IDE class. A controller which needs
ata_genericbut does not report the IDE class should get an ID table entry, which is what the table is for.Binding a driver to unrelated hardware requires root and is what
driver_overrideis meant to do, so this does not fix a privilege boundary.This change only stops
ata_genericfrom binding to a PCI device which it has no reason to believe to be an IDE controller.
Quoted 7 lines
Quoted 3 lines
Reported-by [email protected]Signed-off-by Niklas Cassel <[email protected]>Looks sensible to me, and very surprising that this problem was not cought before...
Reviewed-by Damien Le Moal <[email protected]>
I mean, syzbot using driver_override like this is kind of a stupid test.
If you try to use the generic ATA driver for a PCI device that is something completely different, e.g. an Ethernet controller... you kind of get what you asked for :)
But it is such a small patch, so I think it makes sense, even if all it does it to stop us from getting more silly reports like this.
Tell me if you prefer the other patch, which checks BAR type instead. BAR type I/O is very rare, since it is old x86 legacy.
(But such a patch would still allow someone to bind a PCI device with BARs marked as type I/O, even if that PCI device does not have storage class IDE.)
I think the main argument for the patch in $subject... is:
if your PCI device is not class IDE... just add an entry to the match table. (I.e. we don't care that driver_override will not work with a device that does not have class IDE... If you really want the driver to work with such a device, you will have to add an explicit entry to the match table.)
Kind regards,
Niklas
Damien Le Moal
maintainerQuoted 59 lines · Niklas Cassel
Quoted 40 lines · Damien Le Moal
Quoted 39 lines · Niklas Cassel
syzbot force-bound
ata_genericto 0000:00:03.0 on a QEMU arm64 virt machine. That device is a virtio-blk-pci device holding the root file system, and it reports PCI class0x010000, i.e.PCI_CLASS_STORAGE_SCSI.QEMU gives the virtio-blk-pci device a legacy virtio I/O BAR0 and a 4 KiB MSI-X BAR1, so both resources are non-empty, the port is not discarded, and the device control register ends up in the middle of the MSI-X table.
The emulated device rejects the byte write, and arm64 reports the resulting bus error as a fatal synchronous external abort:
Internal error: synchronous external abort: 0000000096000050 [#1] SMP pc : ata_sff_freeze+0x7c/0x90 drivers/ata/libata-sff.c:1606 Call trace: ata_sff_freeze+0x7c/0x90 ata_eh_freeze_port+0x34/0x5c ata_host_start+0x13c/0x228 ata_pci_sff_activate_host+0x50/0x340 ata_pci_init_one+0x19c/0x1d8 ata_pci_bmdma_init_one+0x14/0x20 ata_generic_init_one+0xc4/0x1ac local_pci_probe+0x40/0xa8 pci_device_probe+0xd8/0x288 really_probe+0xbc/0x2bc device_driver_attach+0x48/0xb4 bind_store+0x7c/0xd8Refuse devices which neither report the IDE class nor appear in our ID table. Table entries keep binding as before, because some of the listed controllers cannot be assumed to report the IDE class. A controller which needs
ata_genericbut does not report the IDE class should get an ID table entry, which is what the table is for.Binding a driver to unrelated hardware requires root and is what
driver_overrideis meant to do, so this does not fix a privilege boundary.This change only stops
ata_genericfrom binding to a PCI device which it has no reason to believe to be an IDE controller.Quoted 7 lines
Quoted 3 lines
Reported-by [email protected]Signed-off-by Niklas Cassel <[email protected]>Looks sensible to me, and very surprising that this problem was not cought before...
Reviewed-by Damien Le Moal <[email protected]>I mean, syzbot using
driver_overridelike this is kind of a stupid test.If you try to use the generic ATA driver for a PCI device that is something completely different, e.g. an Ethernet controller... you kind of get what you asked for :)
But it is such a small patch, so I think it makes sense, even if all it does it to stop us from getting more silly reports like this.
Quoted 5 lines
Tell me if you prefer the other patch, which checks BAR type instead. BAR type I/O is very rare, since it is old x86 legacy.
(But such a patch would still allow someone to bind a PCI device with BARs marked as type I/O, even if that PCI device does not have storage class IDE.)
I definitely prefer the patch as is, checking that the storage class is IDE, instead of doing guesswork with the BAR types.
Quoted 11 lines
I think the main argument for the patch in $subject... is:
if your PCI device is not class IDE... just add an entry to the match table. (I.e. we don't care thatdriver_overridewill not work with a device that does not have class IDE... If you really want the driver to work with such a device, you will have to add an explicit entry to the match table.)Kind regards,
Niklas
Signature · 2 lines
Damien Le Moal Western Digital Research
Niklas Cassel
maintainerQuoted 81 lines · Niklas Cassel
syzbot force-bound
ata_genericto 0000:00:03.0 on a QEMU arm64 virt machine. That device is a virtio-blk-pci device holding the root file system, and it reports PCI class0x010000, i.e.PCI_CLASS_STORAGE_SCSI.QEMU gives the virtio-blk-pci device a legacy virtio I/O BAR0 and a 4 KiB MSI-X BAR1, so both resources are non-empty, the port is not discarded, and the device control register ends up in the middle of the MSI-X table.
The emulated device rejects the byte write, and arm64 reports the resulting bus error as a fatal synchronous external abort:
Internal error: synchronous external abort: 0000000096000050 [#1] SMP pc : ata_sff_freeze+0x7c/0x90 drivers/ata/libata-sff.c:1606 Call trace: ata_sff_freeze+0x7c/0x90 ata_eh_freeze_port+0x34/0x5c ata_host_start+0x13c/0x228 ata_pci_sff_activate_host+0x50/0x340 ata_pci_init_one+0x19c/0x1d8 ata_pci_bmdma_init_one+0x14/0x20 ata_generic_init_one+0xc4/0x1ac local_pci_probe+0x40/0xa8 pci_device_probe+0xd8/0x288 really_probe+0xbc/0x2bc device_driver_attach+0x48/0xb4 bind_store+0x7c/0xd8Refuse devices which neither report the IDE class nor appear in our ID table. Table entries keep binding as before, because some of the listed controllers cannot be assumed to report the IDE class. A controller which needs
ata_genericbut does not report the IDE class should get an ID table entry, which is what the table is for.Binding a driver to unrelated hardware requires root and is what
driver_overrideis meant to do, so this does not fix a privilege boundary.This change only stops
ata_genericfrom binding to a PCI device which it has no reason to believe to be an IDE controller.Reported-by [email protected]Signed-off-by Niklas Cassel <[email protected]>--- drivers/ata/ata_generic.c | 16 ++++++++++++++++ 1 file changed, 16 insertions(+)drivers/ata/ata_generic.c +16
-1@@ -150,6 +150,9 @@ static int is_intel_ider(struct pci_dev *dev)return 1 ; } +/* Forward declaration for the pci_match_id() call in ata_generic_init_one() */ +static const struct pci_device_id ata_generic[]; +/** * ata_generic_init_one - attach generic IDE * @dev: PCI device found @@ -172,6 +175,19 @@ static int ata_generic_init_one(struct pci_dev *dev, const struct pci_device_id };const struct ata_port_info *ppi[] = { &info,NULL }; +/* +* A device matched through driver_override or through an ID added +* with new_id does not come from our ID table, so pci_match_device() +* hands us a synthetic ID with no driver_data and none of the checks +* below apply. Probing maps BAR0 and BAR1 as the ATA command and +* control blocks and writes to the device control register, so only +* continue for devices which report the IDE class or which we list +* ourselves. +*/ +if (!pci_match_id (ata_generic, dev) && + (dev->class >>8 ) !=PCI_CLASS_STORAGE_IDE ) +return -ENODEV ; +/* Don't use the generic entry unless instructed to do so */ if ((id->driver_data &ATA_GEN_CLASS_MATCH ) && all_generic_ide ==0 )return -ENODEV ; --
Quoted 2 lines
2.55.0
Another suggested patch is to something like:
drivers/ata/libata-sff.c +0 -0
@@ -2115,6 +2115,14 @@ static bool ata_resources_present(struct pci_dev *pdev, int port)
/* Check the PCI resources for this channel are enabled */
port *= 2;
for (i = 0; i < 2; i++) {
+ /*
+ * The command and control blocks of a native mode PCI IDE
+ * controller are I/O resources, as are the legacy mode
+ * resources which the PCI core fabricates for compatibility
+ * mode controllers. A memory resource is not a taskfile.
+ */+ if (!(pci_resource_flags(pdev, port + i) & IORESOURCE_IO))
+ return false;
Since IDE controllers always have BAR0 and BAR1 as IORESOURCE_IO and not IORESOURCE_MEM.
Verified that:ata_generic, ata_piix, pata_rdc and the ~40 other pata_* PCI drivers, plus sata_nv, sata_sis, sata_uli, sata_via (vt6420/vt8251) all use I/O BARs for the register blocks.
That would still allow someone to use driver_override to bind a PCI device that has class code != PCI_CLASS_STORAGE_IDE, but would still fail e.g. a virtio-pci device, which has the BARs of type MEM and not type I/O.
Kind regards,
Niklas
Quoted 9 lines · Niklas Cassel
syzbot force-bound
ata_genericto 0000:00:03.0 on a QEMU arm64 virt machine. That device is a virtio-blk-pci device holding the root file system, and it reports PCI class0x010000, i.e.PCI_CLASS_STORAGE_SCSI.QEMU gives the virtio-blk-pci device a legacy virtio I/O BAR0 and a 4 KiB MSI-X BAR1, so both resources are non-empty, the port is not discarded, and the device control register ends up in the middle of the MSI-X table.
[...]
Applied to libata/linux.git (for-7.4), thanks!
[1/1] ata: ata_generic: Do not bind to devices that are not IDE controllers
https://git.kernel.org/libata/linux/c/29b86026
Kind regards,
Niklas