+++ This bug was initially created as a clone of Bug #363961 +++

-- Additional comment from dzickus@redhat.com on 2008-01-14 15:05 EST --
in 2.6.18-68.el5
You can download this test kernel from http://people.redhat.com/dzickus/el5

-- Additional comment from nick.cheng@areca.com.tw on 2008-01-14 21:10 EST --
Hi Tomas,
About Comment #39, I found if I use pci_alloc_consistent and then open Areca AP,
the memory usage will increase gradually. But if I use the kmalloc, it will not.
This was ever reported by Falconstor and Pogolinux.
Especially for Falconstor, if their AP is open, it will use up the memory
overnight and the system will hang. 
We though it only happens on specific version and the  pci_alloc_consistent()
was also proposed by the kernel org fellow as well then there is no need to
change, but to my surprise it happens again.

I check the kernel source for this piece of code of pci_alloc_consistent but
found it has been unchanged for a long time, therefore we guest it malfunctions
on the memory management or somewhere in the kernel code.

This is why this time I decide to upstream to RedHat and kernel org.
The most important is it will hang the system for long-term running.


 

-- Additional comment from nick.cheng@areca.com.tw on 2008-01-14 23:38 EST --
Tomas,
Thanks for your patience.
I would explain one by one and let you decide which is acceptable and which is not.
(1). fix the portability problems
[Discription]
This fixs the endian issue. This will hang the system, which defines big-endian,
while I/O.

(2). fix the iomem release on type B
[Discription]
This fixs the io-memory allocation and release issue on Type B. It could cause
memory leakage on the heavy  load system and leads the system unstable. 

(3). add return -ENOMEM in case of ioremap() failing
[Discription]
It fixs the system unstable issue while iomemory allocation fails.

(4). modify acb->devstate[i][j] as ARECA_RAID_GONE in the initial stage instead
of ARECA_RAID_GOOD in arcmsr_alloc_ccb_pool()
[Discription]
This fixs the exiting volumes' initial states. The wrong setting could lead to
system hang. 

(5). fix the assignment of arcmsr_cdb->Context as (unsigned long)arcmsr_cdb
[Discription]
This fixs the scsi command allocation to a predefined structure. It could lead
to system unstable while the AP accesses this address.

(6). add the checking state of (outbound_intstatus &
ARCMSR_MU_OUTBOUND_HANDLE_INT) == 0 in arcmsr_handle_hba_isr()
[Discription]
This fixs the interrupt routine to handle the controller abnormal interrupts. If
not, it could not handle share IRQ  and leads to system crash. 

(7). fix the scsi error handling in arcmsr_polling_hbb_ccbdone()
[Discription]
It fixs the host scsi error handling callback in case of I/O on failed volumes.
If not, the continual I/O on failed volume could hang the system. 


-- Additional comment from thenzl@redhat.com on 2008-01-15 04:47 EST --
Nick,
thanks for the explanation. (No more questions now.)

Please do not forget to create bugzilla for 4.7 even without having the patch
right now.




-- Additional comment from thenzl@redhat.com on 2008-01-15 11:06 EST --
Nick,
I've made minor changes to your previous patches, it would help me if you could
create the patch against our latest sources.I'm going to send the files to you
separately asap. 
Btw. I could swear that I've seen this part somewhere, but now it magically 
vanished from the patch
-		ver_addr = pci_alloc_consistent(acb->pdev, 1032, &buf_handle);
-		if (!ver_addr) {
+		tmp = kmalloc(1032, GFP_KERNEL|GFP_DMA);
+		ver_addr = (unsigned long *)tmp;