** Tags added: kernel-daily-bug -- You received this bug notification because you are subscribed to linux in Ubuntu. Matching subscriptions: Bgg, Bmail, Nb https://bugs.launchpad.net/bugs/2165806 Title: Ubuntu 26.04.1 Server ISO cannot install on MAXIO MAP1602 NVMe due to kernel 7.0.0-30 I/O regression; 7.0.0-31 works Status in linux package in Ubuntu: New Bug description: Ubuntu 26.04.1 Server cannot be installed on an ACEMAGIC M1A PRO+ with its factory HOGE H820 2 TB / MAXIO MAP1602 NVMe SSD. The Ubuntu 26.04.1 live-server ISO boots kernel 7.0.0-30. With this kernel, reproducible NVMe READ and WRITE failures occur during installation, and the installation eventually fails during grub-install. The same physical machine, SSD, M.2 slot, and SSD firmware operate normally with kernels 7.0.0-14, 6.8, 6.14, and 7.0.0-31. The regression is also independently reproducible on installed Ubuntu 24.04 and Ubuntu 26.04 systems by switching between 7.0.0-30 and 7.0.0-31. PRIMARY IMPACT The current Ubuntu 26.04.1 Server ISO cannot complete installation on this hardware. During installation, kernel 7.0.0-30 reports NVMe block I/O failures. The installer then eventually fails with: grub-install: error: cannot copy '/usr/lib/grub/x86_64-efi-signed/grubx64.efi.signed' to '/boot/efi/EFI/ubuntu/grubx64.efi': Invalid argument. The grub-install failure occurs in the same installation in which the kernel is reporting NVMe READ/WRITE failures. I believe the storage regression is likely contributing to the grub-install failure, but I have not independently proven that the NVMe errors are the direct cause of the grub-install error. HARDWARE System: ACEMAGIC M1A PRO+ AMD Ryzen AI Max+ 395 AMD Radeon 8060S 128 GB memory, 8000 MT/s BIOS: American Megatrends Project version: P10_F11_20_IEC0008_BI0010_AMI_120W Build date: 2026-01-13 NVMe: Model: HOGE H820 2TB Firmware: H240313a PCI vendor ID: 0x1e4b PCI device/controller: 0x1602 Controller family: MAXIO MAP1602 NVMe SMART data observed during troubleshooting: critical_warning: 0 available_spare: 100% percentage_used: 0% media_errors: 0 num_err_log_entries: 0 temperature: approximately 39 C UBUNTU 26.04.1 SERVER INSTALL FAILURE Installation method: - Ubuntu 26.04.1 live-server ISO - ISO written directly to USB using dd - UEFI boot - Secure Boot disabled - entire NVMe disk - no LVM - ext4 root filesystem - 1 GiB FAT EFI System Partition The installer successfully partitions the disk, creates the filesystems, and mounts the target. During installation, kernel 7.0.0-30 begins producing NVMe block I/O failures. Examples of WRITE failures observed: invalid error, dev nvme0n1, sector 11896 op 0x1:(WRITE) ... invalid error, dev nvme0n1, sector 2080 op 0x1:(WRITE) ... followed by messages including: Buffer I/O error on dev nvme0n1p1, logical block 32, lost async page write Buffer I/O error on dev nvme0n1p1, logical block 33, lost async page write Buffer I/O error on dev nvme0n1p1, logical block 34, lost async page write READ failures were also observed, including: invalid error, dev nvme0n1, sector 3104 op 0x0:(READ) ... invalid error, dev nvme0n1, sector 3144 op 0x0:(READ) ... invalid error, dev nvme0n1, sector 3184 op 0x0:(READ) ... ... invalid error, dev nvme0n1, sector 3480 op 0x0:(READ) ... The failing READ requests repeatedly showed multi-segment I/O, including phys_seg 33. Installation ultimately fails at grub-install with: grub-install: error: cannot copy '/usr/lib/grub/x86_64-efi-signed/grubx64.efi.signed' to '/boot/efi/EFI/ubuntu/grubx64.efi': Invalid argument. The Ubuntu installer offered to send an error report to Canonical after the failure, and I selected that option. Therefore an installer-side failure report from this machine should already exist in Canonical's error-reporting infrastructure, although I do not currently have an identifier for that report. REGRESSION TESTING All of the following tests used the same physical ACEMAGIC M1A PRO+, same HOGE H820 SSD, same M.2 slot, and same SSD firmware. Ubuntu 26.04 ------------ 1. Ubuntu 26.04 base installation, kernel 7.0.0-14 PASS Ubuntu 26.04 installs and boots successfully. No NVMe "invalid error", Buffer I/O errors, or related read/write failures are present. 2. Upgrade the installed Ubuntu 26.04 system to kernel 7.0.0-30 FAIL The same NVMe "invalid error" messages return. 3. Upgrade the same Ubuntu 26.04 installation from 7.0.0-30 to 7.0.0-31 from ppa:canonical-kernel-team/ppa PASS The NVMe errors disappear again. Ubuntu 26.04.1 live-server ISO ------------------------------ 4. Ubuntu 26.04.1 live-server ISO, kernel 7.0.0-30 FAIL The same NVMe READ/WRITE errors occur during installation. Unlike the Ubuntu 24.04.4 installer case below, the failures are severe enough here that grub-install fails and Ubuntu 26.04.1 Server cannot be installed. Ubuntu 24.04 ------------ 5. Ubuntu 24.04.4 installation using kernel 7.0.0-30 KERNEL REGRESSION PRESENT, BUT INSTALLATION COMPLETES Ubuntu 24.04.4 installs successfully. After boot, the same NVMe "invalid error" messages are present in the kernel logs. Thus the storage regression is present with 7.0.0-30 on Ubuntu 24.04 as well, even though that installer happens to complete. 6. Downgrade the same Ubuntu 24.04 system to kernel 6.8 PASS The NVMe errors disappear. Sustained fio testing completes without NVMe errors. The Radeon 8060S does not initialize correctly on this older kernel, which is unrelated to the NVMe regression. 7. Upgrade the same Ubuntu 24.04 system to kernel 6.14 PASS NVMe remains clean. Radeon 8060S / amdgpu initializes successfully. Sustained fio testing completes without NVMe errors. 8. Upgrade the same Ubuntu 24.04 system to kernel 7.0.0-31 from ppa:canonical-kernel-team/ppa PASS NVMe errors remain absent. A 10-minute mixed direct-I/O fio workload using libaio, iodepth=32, numjobs=4, and direct=1 completes without NVMe errors. SUMMARY Ubuntu 26.04 + 7.0.0-14 -> PASS Ubuntu 26.04 + 7.0.0-30 -> FAIL Ubuntu 26.04 + 7.0.0-31 -> PASS Ubuntu 26.04.1 ISO + -30 -> FAIL, installation cannot complete Ubuntu 24.04.4 + 7.0.0-30 -> regression present, install completes Ubuntu 24.04 + 6.8 -> PASS Ubuntu 24.04 + 6.14 -> PASS Ubuntu 24.04 + 7.0.0-31 -> PASS The -30 -> -31 comparison has been reproduced on both Ubuntu 24.04 and Ubuntu 26.04. This strongly isolates the problem to the kernel code present in 7.0.0-30 and absent/fixed in 7.0.0-31, rather than to the Ubuntu userspace release or SSD hardware. APST TESTING NVMe APST was tested as a possible cause. The Ubuntu 26.04.1 installer was booted with: nvme_core.default_ps_max_latency_us=0 The runtime value was verified as: /sys/module/nvme_core/parameters/default_ps_max_latency_us = 0 The same NVMe WRITE / Buffer I/O errors still occurred. Disabling NVMe APST therefore does not resolve the failure. STORAGE VALIDATION The SSD does not appear to contain defective media at the sector range where errors were reported. On a working kernel, a direct read spanning the affected region completed successfully: dd if=/dev/nvme0n1 of=/dev/null bs=512 skip=3000 count=600 iflag=direct 600+0 records in 600+0 records out No new NVMe errors were produced. The same SSD also completed substantial mixed read/write fio testing on working kernels without media errors or kernel block-I/O failures. This included sustained testing on: - 6.8 - 6.14 - 7.0.0-31 LIKELY UPSTREAM REGRESSION / FIX I have not performed a source-level git bisect, so the following is a suspected cause rather than independently proven. Linux 7.0.11 included: iommupt: Avoid rewalking during map upstream commit: d6c65b0fd6218bd21ed0be7a8d3218e8f6dc91de Linux 7.0.13 later included: iommu/dma: Do not try to iommu_map a 0 length region in swiotlb upstream commit: 6ec91df8aff77e2e8fe3179c1f3fc15b43a40ba3 The latter fix describes an IOMMU DMA mapping path in which an unaligned mapping can produce a zero-length middle region. iommu_map() then rejects the zero-length mapping as illegal. The upstream description explicitly notes NVMe as a frequent trigger because NVMe can issue unusually aligned buffers in some paths. Ubuntu kernel 7.0.0-31 contains the Linux 7.0.13 and 7.0.14 upstream stable updates, and 7.0.0-31 is the first tested Ubuntu 7.0 kernel in this environment where the problem disappears. Given the observed: 7.0.0-30 -> FAIL 7.0.0-31 -> PASS and the I/O errors being reported as "invalid error", this IOMMU/DMA fix appears to be a strong candidate for the fix. Please confirm whether this system is hitting that regression, or another change incorporated between Ubuntu 7.0.0-30 and 7.0.0-31. EXPECTED RESULT Ubuntu 26.04.1 Server should install successfully on this hardware. The HOGE H820 / MAXIO MAP1602 NVMe controller should operate without kernel block-I/O failures. IMPACT The practical issue is not merely log noise. The kernel shipped on the Ubuntu 26.04.1 Server ISO causes storage errors during installation and prevents the current 26.04.1 live-server image from completing installation on this system. The same regression can also be reproduced after installation by booting 7.0.0-30 on both Ubuntu 24.04 and Ubuntu 26.04. Kernel 7.0.0-31 from ppa:canonical-kernel-team/ppa does not reproduce the problem on either release. To manage notifications about this bug go to: https://bugs.launchpad.net/ubuntu/+source/linux/+bug/2165806/+subscriptions
Комментариев нет:
Отправить комментарий