{"id":"CVE-2026-98216","published":"2026-10-06T09:18:08.260","lastModified":"2026-10-07T07:17:04.653","description":"In the Linux kernel, the following vulnerability has been resolved:\n\nIB/hfi1: Fix the PIO_CRED credit-return mmap\n\nhfi1_file_mmap()'s PIO_CRED case must hand user space the single\ncredit-return page that holds this context's entry.  That page is the\nsecond or third page of the per-node credit-return allocation once the\nhardware send context index reaches 64 or 128, so the failure below is\nintermittent: when the entry lands on the first page the offset is zero\nand everything works.\n\nTwo things are wrong.\n\nFirst, cr_page_offset is a byte offset but .va is a struct\ncredit_return *, so adding it is pointer arithmetic and scales the offset\nby sizeof(struct credit_return) == 64.  memvirt then lands 256 KiB or\n512 KiB past a 10240-byte allocation.  With an IOMMU translating, that\naddress is inside the vmalloc range but in no vm_area, so\ndma_mmap_coherent() -> iommu_dma_mmap() finds no pages, vmalloc_to_pfn()\nreturns page_to_pfn(NULL), and remap_pfn_range() installs a frame above\nMAXPHYADDR.  The first user read then takes:\n\n  psm2_ep_open_pr: Corrupted page table at address 7a14d007e000\n  PGD 800000013886a067 P4D 800000013886a067 PUD 13886b067 PMD 13886c067\n                                            PTE 800049168e911235\n  Oops: Bad pagetable: 000d [#1] SMP PTI\n\nSecond, and still wrong once the arithmetic is corrected,\ndma_mmap_coherent() describes a whole coherent buffer and selects the\npage within it with vma->vm_pgoff.  Offsetting cpu_addr has no effect:\nfor a vmap'd allocation iommu_dma_mmap() uses cpu_addr only to locate the\nvm_area and then maps pages[vm_pgoff], which hfi1_file_mmap() has just\nset to 0.  User space therefore always receives the first credit-return\npage, every credit read is for the wrong context, and send PIO stalls\nforever.\n\nUse the DMA API as intended: pass the base of the allocation with its\nfull length and select the page with vm_pgoff.  A separate length is\nneeded because memlen must keep describing the VMA for the existing size\ncheck.  The dma-direct path stays correct as well, since dma_direct_mmap()\nadds the same vm_pgoff to the base pfn.\n\nTested on a Dell T7610 (Xeon E5-2650 v2, Intel IOMMU in DMA-FQ mode)\nagainst a Threadripper PRO 3995WX peer, both Omni-Path 100.  Before this\nchange psm2_ep_open() Oopses the kernel; with only the arithmetic\ncorrected psm2_ep_open() succeeds but any transfer that uses send PIO\nhangs, PSM2_SDMA=2 (send PIO disabled) completing normally while\nPSM2_SDMA=0 (send PIO only) hangs every time.  With this change send PIO,\nsend DMA and the default mixed mode all work.","cvssScore":7.1,"cvssSeverity":"HIGH","cvssVector":"CVSS:3.1/AV:L/AC:L/PR:L/UI:N/S:U/C:H/I:N/A:H","cwes":[],"vendors":[],"products":[],"references":[{"url":"https://git.kernel.org/stable/c/180752deb7270ad37394ab6ef7cf4978fad040b3","tags":[]},{"url":"https://git.kernel.org/stable/c/535530bb2ea5254e1e9f55280143d262dd065204","tags":[]},{"url":"https://git.kernel.org/stable/c/62f0f34fbd2b2d5653d33d3b9d42fdcabb1c0101","tags":[]},{"url":"https://git.kernel.org/stable/c/bafeac9ce5d1ce5256bcf7e5702e831e9aaf419b","tags":[]},{"url":"https://git.kernel.org/stable/c/dcebe0b0bb080a25fe08011fd6b6e741f7912f01","tags":[]}],"exploitRefs":[],"hasPoc":false,"ai":null}