Avalon-MM PCIe DMA Controller busy forever with some hosts

I've been using Cyclone V device that communicates with a host PC via

PCI Express. The communication logic is built using Qsys "Avalon-MM

Cyclone V Hard IP for PCI Express". On the host PC (x86-64),

CentOS 6.x (Linux kernel 2.6.x) is running. The system has been working fine

for more than a year without problem.

Recently, however, I noticed that the DMA write (memory write

initiated by the device) does not work properly on CentOS 7.x (kernel

3.10.x). Note that other functions including PCI configuration read/write and

the host initiated memory read/write are working fine. DMA read is not

implemented and thus not tested.

Looking inside the status register of Qsys DMA Controller, I found that the

bit1 remains high forever, which means DMA is started but never completes.

I am sure the following registers are properly set before starting the

DMA. The value of address 03h (length register) does not decrement at

all, even a few minutes after the kicking off of the DMA.

01h readaddress RW start address of a read.

02h writeaddress RW start address of a write.

03h length RW length in bytes. decrements to 0 as the transfer proceeds.

06h control RW

So my question is: what can make the DMA controller busy forever?

If a wrong value is set to 02h (writeaddress), could that

cause the DMA controller behave like this?

Please help. Any idea or suggestion would be greatly appreciated.