Stefanha pointed out that I should try if JMicron driver can be chanloaded like: gPXE(jme driver)→gPXE(UNDI driver). Due to the UNDI interface would trigger .irq() function of net_device_operations.

First I spend a lot of time trying to figure out why chainloading with: syslinux→jme.lkrm→undionly.kpxe keeps crash(reboot) after some actions.

After a few tests, I found out that:

• syslinux→gPXE(jme.lkrm)→undionly.kkpxe
• Vendor PXE→gPXE(jme.pxe)→undionly.kkpxe
• Vendor PXE→gPXE(jme.kpxe)→undionly.kkpxe

All worked fine. So I guess that it's an issue with the chainloading processes, instead of in-correct driver implement.

Testing results can be found in this note: TCP Window Size and RX Csum Offload Benchmark

1. jme_process_receive() isn't using a read memory barrier.
2. Allocate/Free Ring buffers on open/close.
3. Using jme_free_?x_resources() for allocation fail recover.
4. Simply use struct io_buffer, instead of jme_buffer_info.
5. There is no need to discard pending TX buffers.
6. extra memset in probe.
7. Ring size tune, performance test.

Detailed modifications can reference the jme_driver_bak branch of my git tree.