Skip to contentSkip to conversation

KVM: s390: pci: Fix handling of AIF enable without AISB

PATCH In review Matthew Rosato 3 participants 3 messages Jul 9

Subsystems (4)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 KVM: s390: pci: Fix handling of AIF enable without AISB

When a guest seeks to register IRQs without a summary bit specified, ensure that the associated GAITE then stores 0 for the guest AISB location instead of virt_to_phys(page_address(NULL)).

Fixes 3c5a1b6f0a18 ("KVM: s390: pci: provide routines for enabling/disabling interrupt forwarding")
---
 arch/s390/kvm/pci.c | 11 ++++++++---
 1 file changed, 8 insertions(+), 3 deletions(-)
arch/s390/kvm/pci.c +8 -3
index 5b075c38998e..dc748cde4703 100644
@@ -300,9 +300,14 @@ static int kvm_s390_pci_aif_enable(struct zpci_dev *zdev, struct zpci_fib *fib,
 
 	gaite->gisc = fib->fmt0.isc;
 	gaite->count++;
-	gaite->aisbo = fib->fmt0.aisbo;
-	gaite->aisb = virt_to_phys(page_address(aisb_page) + (fib->fmt0.aisb &
-							      ~PAGE_MASK));
+	if (fib->fmt0.sum == 1) {
+		gaite->aisbo = fib->fmt0.aisbo;
+		gaite->aisb = virt_to_phys(page_address(aisb_page) +
+					   (fib->fmt0.aisb & ~PAGE_MASK));
+	} else {
+		gaite->aisbo = 0;
+		gaite->aisb = 0;
+	}
 	aift->kzdev[zdev->aisb] = zdev->kzdev;
 	spin_unlock_irq(&aift->gait_lock);
 
Signature · 1 line
2.54.0
AI review Reviewed-by Farhan Ali on v1 1/1
#1 lore raw
Thank you for your contribution! Sashiko AI review found 6 potential issue(s) to consider:

Thank you for your contribution! Sashiko AI review found 6 potential issue(s) to consider:

Pre-existing issues:
- [Critical] Missing state check allows repeated KVM_S390_ZPCIOP_REG_AEN ioctls to leak resources and create dangling pointers.
- [Critical] Unvalidated guest-supplied offsets for the Adapter Interrupt Bit Vector (AIBV) can cause the host firmware to write beyond the pinned page boundary.
- [High] Error paths for account_mem and airq_iv_alloc_bit do not set an error code, leading to silent failures, and bypass memory accounting rollback.
- [High] Unchecked return value of airq_iv_create() leads to a kernel NULL pointer dereference on memory exhaustion.
- [High] The memory accounting functions account_mem() and unaccount_mem() leak a reference count to the user_struct on every call.
- [High] Failure in hardware configuration via kvm_zpci_set_airq bypasses local cleanup labels, leaving resources allocated.
--

--- Patch [1]: [PATCH] KVM: s390: pci: Fix handling of AIF enable without AISB --- commit 770cb99be23b8c546be19f495246867591501586
Author: Matthew Rosato <[email protected]>

KVM: s390: pci: Fix handling of AIF enable without AISB

When a guest seeks to register IRQs without a summary bit specified, ensure that the associated GAITE then stores 0 for the guest AISB location instead of virt_to_phys(page_address(NULL)).

Fixes 3c5a1b6f0a18 ("KVM: s390: pci: provide routines for enabling/disabling interrupt forwarding")
Quoted 12 lines
arch/s390/kvm/pci.c +1 -3
index 686113be05309..720bb58cabe2e 100644
@@ -300,9 +300,14 @@ static int kvm_s390_pci_aif_enable(struct zpci_dev *zdev, struct zpci_fib *fib,
 
 	gaite->gisc = fib->fmt0.isc;
 	gaite->count++;
-	gaite->aisbo = fib->fmt0.aisbo;
-	gaite->aisb = virt_to_phys(page_address(aisb_page) + (fib->fmt0.aisb &
-							      ~PAGE_MASK));
+	if (fib->fmt0.sum == 1) {

arch/s390/kvm/pci.c static int kvm_s390_pci_aif_enable(struct zpci_dev *zdev, struct zpci_fib *fib, · line 313

Quoted 10 lines
arch/s390/kvm/pci.c +7 -0
+		gaite->aisbo = fib->fmt0.aisbo;
+		gaite->aisb = virt_to_phys(page_address(aisb_page) +
+					   (fib->fmt0.aisb & ~PAGE_MASK));
+	} else {
+		gaite->aisbo = 0;
+		gaite->aisb = 0;
+	}
 	aift->kzdev[zdev->aisb] = zdev->kzdev;
 	spin_unlock_irq(&aift->gait_lock);
 

[Severity: High]
This is a pre-existing issue, but does this error path properly set an error code and roll back accounting?

In kvm_s390_pci_aif_enable():

    if (account_mem(pcount))
        goto unpin2;

    /* AISB must be allocated before we can fill in GAITE */
    mutex_lock(&aift->aift_lock);
    bit = airq_iv_alloc_bit(aift->sbv);
    if (bit == -1UL)
        goto unlock;

If account_mem() fails, rc appears to remain 0.

If airq_iv_alloc_bit() fails, the code jumps to unlock and falls through to unpin2. This successfully unpins the pages but bypasses unaccount_mem().

Does this permanently leak the user's memory limits?

[Severity: High]
This isn't a bug introduced by this patch, but does the return value of airq_iv_create() need validation here?

In kvm_s390_pci_aif_enable():

    zdev->aibv = airq_iv_create(msi_vecs, AIRQ_IV_DATA |
                                AIRQ_IV_BITLOCK |
                                AIRQ_IV_GUESTVEC,
                                phys_to_virt(fib->fmt0.aibv));

If airq_iv_create() returns NULL under memory pressure, zdev->aibv is assigned NULL.

Immediately after, kvm_zpci_set_airq() is called, which appears to unconditionally dereference it:

kvm_zpci_set_airq() {
    ...
    fib.fmt0.noi = airq_iv_end(zdev->aibv);
    ...

Could this lead to a NULL pointer dereference?

[Severity: Critical]
This is a pre-existing issue, but does the ioctl handler check if AEN is already enabled before allowing another registration?

Looking at kvm_s390_pci_zpci_op():

    r = kvm_s390_pci_zpci_reg_aen(zdev, args);

If a guest repeatedly calls the KVM_S390_ZPCIOP_REG_AEN ioctl, it unconditionally invokes kvm_s390_pci_aif_enable().

Inside kvm_s390_pci_aif_enable():

    zdev->aisb = bit;
    zdev->aibv = airq_iv_create(...);

Is it possible this overwrites previous bit indexes and pointers without cleaning up the old resources? This looks like it might leave dangling pointers in the global aift->kzdev array for the overwritten bits.

[Severity: Critical]
This isn't a bug introduced by this patch, but does this code restrict the guest-provided offset within the single pinned page?

In kvm_s390_pci_aif_enable():

    npages = pin_user_pages_fast(hva, 1, FOLL_WRITE | FOLL_LONGTERM, pages);
    ...
    gaddr = page_to_phys(aibv_page) + (fib->fmt0.aibv & ~PAGE_MASK);

Since only 1 page is pinned, if the guest offset plus the required interrupt vector size exceeds the page boundary, can the hardware adapter end up writing interrupt bits into unpinned host memory?

[Severity: High]
This is a pre-existing issue, but do these accounting functions leak a reference count?

In account_mem():

    struct user_struct *user = get_uid(current_user());

And in unaccount_mem():

    struct user_struct *user = get_uid(current_user());

Both functions call get_uid(), which increments the reference count, but they do not appear to call free_uid() to release it.

Additionally, unaccount_mem() uses current_user(), which might not be the same user who originally accounted the memory if AEN is disabled by a different process. Are we at risk of mismatched locked_vm accounting?

[Severity: High]
This isn't a bug introduced by this patch, but does an error from kvm_zpci_set_airq() bypass the cleanup paths?

In kvm_s390_pci_aif_enable():

    /* Issue the clp to setup the irq now */
    rc = kvm_zpci_set_airq(zdev);
    return rc;

unlock:

    mutex_unlock(&aift->aift_lock);

unpin2:
...

If kvm_zpci_set_airq() fails, the function returns directly. Should the function jump to the cleanup labels instead to prevent leaving pages pinned and structures like airq_iv permanently allocated?

Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=1
#2 lore raw
In reply to Matthew Rosato

Am 09.07.26 um 15:54 schrieb Matthew Rosato:

Quoted 8 lines

When a guest seeks to register IRQs without a summary bit specified, ensure that the associated GAITE then stores 0 for the guest AISB location instead of virt_to_phys(page_address(NULL)).

Fixes 3c5a1b6f0a18 ("KVM: s390: pci: provide routines for enabling/disabling interrupt forwarding")

applied and queued for master.
Thank you