IB/hfi1: clear VM_MAYWRITE on read-only STATUS mmap
Subsystems (3)
PATCH 1/2 IB/hfi1: clear VM_MAYWRITE on read-only STATUS mmap
hfi1_file_mmap() rejects a writable mapping of the device status page, but leaves VM_MAYWRITE set in the VMA flags. Userspace can map the page read-only and then upgrade it to writable with mprotect(), bypassing the check and writing to a device-owned status page.
Clear VM_MAYWRITE on this path, as the RTAIL and SUBCTXT_EGRBUF cases in the same function already do.
---
drivers/infiniband/hw/hfi1/file_ops.c +1 -0
@@ -492,6 +492,7 @@ static int hfi1_file_mmap(struct file *fp, struct vm_area_struct *vma) ret = -EPERM ; goto done; } + flags &= ~VM_MAYWRITE ; memaddr =kvirt_to_phys ((void *)dd->status); memlen =PAGE_SIZE ;break ;
PATCH 2/2 IB/hfi1: reject writable SUBCTXT_EGRBUF mmap
hfi1_file_mmap() clears VM_MAYWRITE for the SUBCTXT_EGRBUF mapping, but never checks whether the caller requested VM_WRITE in the first place. A caller can pass PROT_WRITE to mmap() and get a writable mapping of the read-only shared eager receive buffers, unlike RCV_EGRBUF which rejects the request with -EPERM.
Add the missing VM_WRITE check, matching the other read-only mmap cases in this function.
---
drivers/infiniband/hw/hfi1/file_ops.c +4 -0
@@ -525,6 +525,10 @@ static int hfi1_file_mmap(struct file *fp, struct vm_area_struct *vma) is_vmalloc =1 ;break ;case SUBCTXT_EGRBUF : +if (flags &VM_WRITE ) { + ret = -EPERM ; + goto done; + } memvirt = uctxt->subctxt_rcvegrbuf; memlen = uctxt->egrbufs.size * uctxt->subctxt_cnt; flags &= ~VM_MAYWRITE ;
[email protected]
AI reviewRe: [PATCH 2/2] IB/hfi1: reject writable SUBCTXT_EGRBUF mmap
Thank you for your contribution! Sashiko AI review found 2 potential issues to consider: - [High] IB/hfi1: rejecting VM_WRITE on SUBCTXT_EGRBUF in hfi1_file_mmap() breaks subcontexts - [Low]…
Thank you for your contribution! Sashiko AI review found 2 potential issues to consider:
- [High] IB/hfi1: rejecting VM_WRITE on SUBCTXT_EGRBUF in hfi1_file_mmap() breaks subcontexts
- [Low] IB/hfi1: inaccurate Fixes trailer citing 12220267645c for hfi1_file_mmap()
--
commit 0342c207f57f73ef2cbfe9cff3f4ce685d284272
Author: Abdifatah Suruur <[email protected]>
Quoted 8 lines
IB/hfi1: reject writable
SUBCTXT_EGRBUFmmap
hfi1_file_mmap()clearsVM_MAYWRITEfor theSUBCTXT_EGRBUFmapping, but never checks whether the caller requestedVM_WRITEin the first place. A caller can passPROT_WRITEtommap()and get a writable mapping of the read-only shared eager receive buffers, unlikeRCV_EGRBUFwhich rejects the request with-EPERM.
Quoted 6 lines
Add the missing
VM_WRITEcheck, matching the other read-only mmap cases in this function.Fixes 12220267645c ("IB/hfi: Protect against writable mmap")Signed-off-by Abdifatah Suruur <[email protected]>
[Severity: Low]12220267645c ("IB/hfi: Protect against writable mmap") only added a VM_WRITE | VM_EXEC check to the STATUS case in hfi1_file_mmap(). The SUBCTXT_EGRBUF handling dates back to 7724105686e7 ("IB/hfi1: add driver files").
Quoted 5 lines
drivers/infiniband/hw/hfi1/file_ops.c +0
-0
drivers/infiniband/hw/hfi1/file_ops.c static int hfi1_file_mmap(struct file *fp, struct vm_area_struct *vma) · line 535
Quoted 10 lines
drivers/infiniband/hw/hfi1/file_ops.c +4
-0@@ -526,6 +526,10 @@ static int hfi1_file_mmap(struct file *fp, struct vm_area_struct *vma) is_vmalloc =1 ;break ;case SUBCTXT_EGRBUF : +if (flags &VM_WRITE ) { + ret = -EPERM ; + goto done; + } memvirt = uctxt->subctxt_rcvegrbuf; memlen = uctxt->egrbufs.size * uctxt->subctxt_cnt; flags &= ~VM_MAYWRITE ;
[Severity: High]
Does rejecting VM_WRITE on SUBCTXT_EGRBUF break userspace context sharing?
Unlike RCV_EGRBUF, which maps DMA buffers written by the HFI1 chip, uctxt->subctxt_rcvegrbuf is a vmalloc_user() buffer allocated in setup_subctxt() for userspace subcontexts sharing a hardware context, and neither hardware nor the kernel ever writes to it. The process polling the hardware receive queue forwards eager packets to sibling subcontexts by writing them into this shared mapping (alongside SUBCTXT_RCV_HDRQ and SUBCTXT_UREGS), so userspace maps SUBCTXT_EGRBUF with PROT_READ | PROT_WRITE. Rejecting VM_WRITE causes subcontext initialization to fail with -EPERM and leaves nothing able to populate uctxt->subctxt_rcvegrbuf.
Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=2
Re: [PATCH 2/2] IB/hfi1: reject writable SUBCTXT_EGRBUF mmap
In reply to Abdifatah SuruurHello,
Please drop patch 2/2 of this series ("IB/hfi1: reject writable SUBCTXT_EGRBUF mmap"). The automated review pointed out that uctxt->subctxt_rcvegrbuf is a vmalloc_user() buffer that userspace subcontexts sharing a hardware context intentionally map PROT_READ|PROT_WRITE: the process polling the hardware receive queue forwards eager packets into it. Rejecting VM_WRITE would break subcontext initialization, and the existing VM_MAYWRITE clear is deliberate policy. My patch was wrong; I am withdrawing it.
Patch 1/2 (STATUS VM_MAYWRITE) is unaffected and stands on its own.
Sorry for the noise,
Abdifatah Suruur
[email protected]
[email protected]
AI reviewSashiko has reviewed this patch and found no issues.
Quoted 11 lines
hfi1_file_mmap()rejects a writable mapping of the device status page, but leavesVM_MAYWRITEset in the VMA flags. Userspace can map the page read-only and then upgrade it to writable withmprotect(), bypassing the check and writing to a device-owned status page.Clear
VM_MAYWRITEon this path, as the RTAIL andSUBCTXT_EGRBUFcases in the same function already do.Fixes 12220267645c ("IB/hfi: Protect against writable mmap")Signed-off-by Abdifatah Suruur <[email protected]>
Sashiko has reviewed this patch and found no issues.
Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=1