Re: pagefaults and hang with 5.15.11

[John David Anglin <dave.anglin@bell.net>]

On 2022-01-25 11:54 a.m., Rolf Eike Beer wrote:
> Am Montag, 24. Januar 2022, 18:41:27 CET schrieb John David Anglin:
>> On 2022-01-24 12:24 p.m., John David Anglin wrote:
>>> Faulting instruction is "ldw 18(r3),r24".  Address in $r3 (and $r5) seems
>>> bad.  Think the sglist argument
>>> to sba_unmap_sg() is bad.  Don't have a clue as to why this might be.
>> Maybe try interchanging operands of following &&
>>
>>           while (sg_dma_len(sglist) && nents--) {
>>
>> so nents is checked first.
> But nents would be increased then even for the case that sg_dma_len() returns
> false, which may or may not be wanted.
You are correct.  The decrement of nents needs to be in loop so count in following DBG_RUN_SG is correct:

         while (sg_dma_len(sglist) && nents--) {

                 sba_unmap_page(dev, sg_dma_address(sglist), sg_dma_len(sglist),
                                 direction, 0);
#ifdef SBA_COLLECT_STATS
                 ioc->usg_pages += ((sg_dma_address(sglist) & ~IOVP_MASK) + sg_dma_len(sglist) + IOVP_SIZE - 1) >> PAGE_SHIFT;
                 ioc->usingle_calls--;   /* kluge since call is unmap_sg() */
#endif
                 ++sglist;
         }

         DBG_RUN_SG("%s() DONE (nents %d)\n", __func__,  nents);

However, nents still needs to be checked first.

What has happened is the sglist pointer has crossed a page boundary causing the TLB miss.  The offset of sg_dma_len is 0x18
and checking sg_dma_len(sglist) first causes the fault.

-- 
John David Anglin  dave.anglin@bell.net