[Deepin-Kernel-SIG] [linux 6.6.y] [Upstream] sync bugfixes from mainline

[opsiff]

Summary by Sourcery

Sync upstream mainline fixes across memory management, virtualization, architecture code, networking, wireless, cryptography, platform drivers, graphics, audio, and kernel self-tests.

New Features:

• Add support for newer Intel processor power-state residency events and expanded MediaTek wireless capabilities, including 320 MHz operation on supported hardware.
• Extend IAA compression handling with software fallbacks and support for fragmented buffers, and refine QAT fuse handling across hardware generations.
• Improve KVM memory-attribute invalidation, guest memory fault handling, and HugeTLB write-protection behavior.
• Add ChainDMA-aware audio delay tracking and improve SOF HDA lifecycle management.

Bug Fixes:

• Fix ARM64 KVM hypervisor state preservation, x86 perf event reassignment, Intel uncore PMU failure handling, and several DMA resource-management issues.
• Correct networking, wireless, PHY, firmware, audio, platform-device, and VFIO migration edge cases across multiple drivers.
• Fix transparent hugepage policy handling, IOVA bitmap range updates, TCP delayed-ACK bounds, ELF FDPIC auxiliary vector sizing, and radiotap antenna parsing.
• Correct KVM guest memory attribute filtering and strengthen related self-tests.

Enhancements:

• Refactor Intel platform driver module naming and improve synchronization, cleanup, and state tracking across kernel subsystems.
• Update Nouveau firmware allocation and synchronization to use noncoherent DMA APIs.
• Improve MediaTek WED/mt76 diagnostics and descriptor handling, Realtek MLO link identification, and Intel TPMI/RAPL debug and error reporting.

Documentation:

• Update transparent hugepage administration documentation and related API semantics.

Tests:

• Expand KVM memory-region validation and guest-memfd overlap coverage, and make pagemap tests use the system page size.

Chores:

• Sync a broad set of mainline fixes and hardware support updates into the 6.6 kernel tree.

[Copilot]

🟡 Changes recommended

The backport contains a compile failure and multiple correctness issues in crypto, memory tracking, audio, and DMA handling.

Once you've addressed the issues Copilot identified, you can request another Copilot review.

Review details

Suppressed comments (7)

drivers/crypto/intel/iaa/iaa_crypto_main.c:1659

• For a multi-entry source, this fallback still calls deflate_generic_decompress(), which maps only the first SG page but reads req->slen bytes as contiguous input. The entries need to be linearized first or passed to an SG-aware fallback; this is unsafe even when the total length is at most one page because the segments are not contiguous.

	if (sg_nents(req->src) > 1) {
		if (req->slen > PAGE_SIZE)
			return deflate_generic_decompress(req);

drivers/iommu/iommufd/iova_bitmap.c:416

• Pinning failures from iova_bitmap_advance_to() are silently discarded because this setter returns void, and iova_bitmap_for_each() now returns only the callback result. An ioctl can consequently report success while part of the dirty range was never written. Record and propagate the mapping error (or make the setter return it) so iteration cannot succeed after a failed pin.

		if (iova_bitmap_advance_to(bitmap, iova))
			return;

drivers/gpu/drm/nouveau/nvkm/core/firmware.c:229

• dma_alloc_noncoherent() may return NULL, but sg_init_one() is called before the common allocation check and dereferences/converts that NULL buffer. Check the allocation immediately, then initialize the scatterlist only for a valid buffer.

		sg_init_one(&fw->mem.sgl, fw->img, len);
		sg_dma_address(&fw->mem.sgl) = fw->phys;
		sg_dma_len(&fw->mem.sgl) = len;

sound/soc/sof/ipc4-pcm.c:1164

• The counters are masked with U32_MAX, which wraps them modulo 2^32, but this wrapped-distance formula uses a modulus of U32_MAX. Crossing from 0xffffffff to zero is therefore one frame short; include the missing wrap value.
  sound/soc/sof/intel/hda.c:1224
• Moving hda_init() into probe_early makes its HDA mapping and i915 state live before snd_sof_probe(). If the later hda_dsp_probe() fails, sof_init_environment() returns after only sof_ops_free() and never invokes remove_late, while the old HDA/i915 cleanup was removed from the probe error path. Ensure all post-early-probe failures call hda_dsp_remove_late() so these resources are released.
  include/linux/kvm_host.h:273
• attr_filter is assigned by the new HVA, guest_memfd, and memory-attribute paths but is never read by any range handler in this tree. Consequently private mappings are still processed by shared-only invalidations and the old-attribute filtering is ineffective. Backport the consumer-side KVM MMU changes that test this field before operating on a mapping.
  arch/x86/kvm/mmu/mmu.c:4430
• Correct the spelling of “preemptible” in this comment.

	 * Do the pre-check even for non-preemtible kernels, i.e. even if KVM

• Files reviewed: 80/80 changed files
• Comments generated: 9
• Review effort level: Balanced