arcmsr + archttp64 calls dma_free_coherent() with irqs disabled - dmesg filled with warnings

Thanks Dave,

I'm just updating this report with the upstream git commit id and summary for the patch to resolve this. Thanks.

commit 69e562c234440fb7410877b5b24f4b29ef8521d1
Author: Daniel Drake <email address hidden>
Date: Wed Feb 20 13:29:05 2008 +0000

    [SCSI] arcmsr: fix message allocation

    arcmsr_iop_message_xfer() is called from atomic context under the
    queuecommand scsi_host_template handler. James Bottomley pointed out
    that the current GFP_KERNEL|GFP_DMA flags are wrong: firstly we are in
    atomic context, secondly this memory is not used for DMA.
    Also removed some unneeded casts.

    Signed-off-by: Daniel Drake <email address hidden>
    Cc: Nick Cheng <email address hidden>
    Signed-off-by: James Bottomley <email address hidden>

I find that status "medium" on this bug does not do it justice... all Areca controllers, old and new, have become unusable in 2.6.24. This is a serious regression.

>have become unusable in 2.6.24.

Besides the annoying warning, what else goes wrong? (I recently reported duplicate bug #197982)

The flood of kernel warnings will hang your machine given a large enough I/O load... we currently need to kill klogd to keep our machines running for any longer period of time.

That's interesting; I only get the warning when I run the cli64 tool
to query drive status. Regular use of the SATA controller doesn't
trigger any warnings for me, no matter what the load. Probably the
difference is in how we are utilizing the Areca card; I'm just using
a simple passthrough mode and not actually invoking any of the
hardware RAID capabilities of the card.

Yeah I am still using the machines in question for pretty heavy I/O load (they are dedicated storage servers for an OpenVZ cluster). They are running RAID5 from the card. Can't remember the specs on it, if I could run the CLI tool I would tell you. But its a definite regression in that if the CLI tool doesn't work anymore, then we can't monitor the RAID for failures, and then don't know if a drive has failed. btw, just so everyone knows, there is a couple of nagios plugins to monitor this card using the CLI tool here:

http://www.nagiosexchange.org/Search_Projects.43.0.html?tx_netnagext_pi1%5Bphrase%5D=areca&tx_netnagext_pi1%5Bsubmit%5D=search&tx_netnagext_pi1%5Bsearch%5D=1

(sorry about the massive url).

I don't get any other errors out of the cards unless I run the tool. Of course areca may release a different version of CLI64 that makes the problem go away as well.
For my money, I would try and get the latest drive into the Hardy kernel though.

Applied back ported patch http://bugs.gentoo.org/attachment.cgi?id=143585

This bug was fixed in the package linux - 2.6.24-12.18

---------------
linux (2.6.24-12.18) hardy; urgency=low

  [Amit Kucheria]

  * Move Marvell 8686 and 8688 to LUM
  * Poulsbo: Sync patches with moblin/ume-hardy tree
  * Break if a patch fails to apply
  * SAUCE: implement smarter atime updates support
    - LP: #199427
  * Enable USB_PERSIST to allow devices with /root on usb to work with
    suspend
  * Enable USB_PERSIST across the board

  [Ben Collins]

  * build/config: Really fix ide on smp ppc configs
  * build/configs: Enable relatime config option for all flavors
  * build/abi: Ignore ide-core module for ppc, moved to built-in

  [Colin Ian King]

  * fix reversed logic for bbuild check leads to -j1 default
    - LP: #197040
  * Enable IDE_PMAC for powerpc-smp
    - LP: #196686
  * Disable CONFIG_USB_OHCI_HCD_SSB
    - LP: #182716
  * SAUCE: fix arcmsr + archttp64 calls dma_free_coherent() with irqs
    disabled - dmesg filled with warnings
    - LP: #194207

  [Jorge Boncompte [DTI2]]

  * Fix Messed multicast lists after dev_mc_sync/unsync
    - LP: #193468

  [Stefan Bader]

  * Add support for Apple Aluminium keyboards.
    - LP: #162083
  * SAUCE: Restore VT fonts on switch

  [Upstream Kernel Changes]

  * [NET]: Messed multicast lists after dev_mc_sync/unsync
  * KVM: x86 emulator: add support for group decoding
  * KVM: x86 emulator: group decoding for group 1A
  * KVM: x86 emulator: Group decoding for group 3
  * KVM: x86 emulator: Group decoding for groups 4 and 5
  * KVM: x86 emulator: add group 7 decoding
  * KVM: constify function pointer tables
  * KVM: Only x86 has pio
  * KVM: x86 emulator: group decoding for group 1 instructions
  * KVM: MMU: Decouple mmio from shadow page tables
  * KVM: Limit vcpu mmap size to one page on non-x86
  * KVM: VMX: Enable Virtual Processor Identification (VPID)
  * KVM: Use CONFIG_PREEMPT_NOTIFIERS around struct preempt_notifier
  * KVM: Disable pagefaults during copy_from_user_inatomic()
  * KVM: make EFER_RESERVED_BITS configurable for architecture code
  * KVM: align valid EFER bits with the features of the host system
  * KVM: allow access to EFER in 32bit KVM
  * kvm: i386 fix
  * KVM: export information about NPT to generic x86 code
  * KVM: MMU: make the __nonpaging_map function generic
  * KVM: export the load_pdptrs() function to modules
  * KVM: MMU: add TDP support to the KVM MMU
  * KVM: x86 emulator: Fix 'jmp abs'
  * KVM: x86 emulator: fix group 5 decoding
  * KVM: Fix kvm_arch_vcpu_ioctl_set_sregs so that set_cr0 works properly
  * KVM: Make the supported cpuid list a host property rather than a vm
    property
  * KVM: emulate access to MSR_IA32_MCG_CTL
  * KVM: remove the usage of the mmap_sem for the protection of the memory
    slots.
  * KVM: SVM: allocate the MSR permission map per VCPU
  * KVM: make MMU_DEBUG compile again
  * KVM: paravirtualized clocksource: host part
  * KVM: Add missing semicolon
  * KVM: x86 emulator: add ad_mask static inline
  * KVM: x86 emulator: make register_address, address_mask static inlines
  * KVM: x86 emulator: make register_address_increment and JMP_REL static
    inlines
  * KVM: Add API to retrieve the number of supported...