Re: [PATCH 02/24] ARM: l2c: Serialize device reads with cache maintenance when requested

[Navid Ghahremani <ghahramani.navid@gmail.com>]

Hi Arnd,

Thank you again for the suggestion to test arm,outer-sync-disable. I have now completed extensive hardware tests on the physical ZTE ZXHN H3600 across multiple bring-up orders and load scenarios, and unfortunately disabling outer-sync alone is not sufficient to prevent the interconnect deadlock.

Here is what I tested:
1. Added arm,outer-sync-disable to the l2cc node in the device tree (zx279128s.dtsi).
2. Rebuilt the kernel with CONFIG_CACHE_L2X0_IO_LOCK disabled and completely removed my custom read lock from both cache-l2x0 and mt76.
3. Netbooted the initramfs uImage over TFTP into RAM via U-Boot (zero flash modification).

The kernel booted cleanly and confirmed the configuration:
[    0.000000] L2C: disabling outer sync
[    0.000000] L2C-310 cache controller enabled, 16 ways, 256 kB

Both MT7915 cards enumerated over PCIe and initialized firmware. Here are the observed results across the different test scenarios:

1. Sequential Bring-Up with Long Delays:
When bringing up one radio and waiting several seconds until it was completely idle before bringing up the second radio, the system survived temporarily. Because idle beaconing generates very few register reads, the collision window was small enough that it avoided a collision purely by timing chance.

2. Simultaneous Bring-Up ('wifi up'):
When enabling both radios simultaneously in UCI and triggering 'wifi up' to bring up both 2.4 GHz and 5 GHz interfaces together, the interconnect froze solid immediately. The serial console went completely silent, Ethernet ping dropped to 100% loss with zero panic or oops output from the CPU, and the router required a physical power-cycle to recover.

3. Concurrent Wi-Fi Scans and Traffic:
Even when the radios were brought up sequentially, the moment concurrent operations were triggered across both PCIe endpoints (such as spectrum scan triggers via 'iw dev scan trigger' or active data traffic while DMA cache maintenance was running), the SoC interconnect locked up dead every time.

Looking into arch/arm/mm/cache-l2x0.c, arm,outer-sync-disable only suppresses standalone outer_sync() calls by setting fns.sync to NULL. It does not touch the line-by-line cache maintenance inside l2c210_clean_range(), l2c210_inv_range(), or l2c210_flush_range(). During active dual-band Wi-Fi operation, network DMA continuously invokes these range operations, which collide with the high frequency of PCIe MMIO register reads across the internal Sanechips AXI crossbar.

By contrast, my serialization lock approach (which serializes PCIe register reads with L2C maintenance operations) successfully passed 240+ continuous stress reload cycles and boots under heavy Wi-Fi traffic without a single freeze. Interestingly, examining ZTE's original stock kernel (Linux 4.1.25) reveals that ZTE's own chip engineers used this exact same solution, implementing a global spinlock around PCIe reads (fixed_pci_read_u32()) and L2C maintenance to work around this hardware erratum.

Given that this is a silicon crossbar flaw that requires software serialization, what would you recommend as the cleanest, most acceptable way to structure this for upstream? Would you prefer keeping the serialization lock as an SoC-specific quirk in cache-l2x0, or is there an alternative architectural mechanism you would suggest?

Best regards,
Navid