lpfc sleeps while holding a spinlock

This is with 2.6.27-rc9; most recent non-janitorial commit to the lpfc directory is ff0f4cb5ea322dcc32d08bab2d758c050ba1ab07 (Update lpfc driver version to 8.2.7). The janitorial commits do not affect the code in question.

This bug gets reported by lockdep as a mismatch between holding the hbalock with interrupts on vs acquiring it in interrupt context, but actually the problem is a GFP_KERNEL allocation while holding a spinlock:

lpfc_sli_hbqbuf_fill_hbqs(struct lpfc_hba *phba, uint32_t hbqno, uint32_t count)
{
...

        spin_lock_irqsave(&phba->hbalock, flags);
...
        for (i = start; i < end; i++) {
                hbq_buffer = (phba->hbqs[hbqno].hbq_alloc_buffer)(phba);

where previously:
drivers/scsi/lpfc/lpfc_init.c: phba->hbqs[LPFC_ELS_HBQ].hbq_alloc_buffer = lpfc_els_hbq_alloc;

and:
lpfc_els_hbq_alloc(struct lpfc_hba *phba)

{
        struct hbq_dmabuf *hbqbp;
        hbqbp = kmalloc(sizeof(struct hbq_dmabuf), GFP_KERNEL);

Presumably these GFP_KERNEL should become GFP_ATOMIC, or the spinlock shouldn't be taken at this point, but I haven't looked deeply enough into the driver to determine which course of action is correct.

--
Matthew Wilcox Intel Open Source Technology Centre
"Bill, look, we understand that you're interested in selling us this operating system, but compare it to ours. We can't possibly take such a retrograde step."

We already fixed this in the 8.2.8 patches that were sent and are sitting in James's git pools. Specifically, see patch8:
http://marc.info/?l=linux-scsi&m=121962902731358&w=2