Re: [Qemu-devel] [PATCH RFC v2] s390/css: handle CCW_FLAG_SKIP

[Eric Farman <farman@linux.ibm.com>]

On 5/7/19 5:08 AM, Pierre Morel wrote:
> On 07/05/2019 10:12, Cornelia Huck wrote:
>> If a ccw has CCW_FLAG_SKIP set, and the command is of type
>> read, read backwards, or sense, no data should be written
>> to the guest for that command.
>>
>> Signed-off-by: Cornelia Huck <cohuck@redhat.com>
>> ---
>>
>> v1 -> v2: fixed checks for command type [Eric]
>>
>> Still only lightly tested (it boots); I don't think I have a tool
>> generating channel programs with the skip flag handy.

FWIW, my test program hits this code once (read of a single page, with 
SLI and SKIP flags set) before getting into the passthrough codepath.

>>
>> ---
>>   hw/s390x/css.c         | 22 ++++++++++++++++++----
>>   include/hw/s390x/css.h |  1 +
>>   2 files changed, 19 insertions(+), 4 deletions(-)
>>
>> diff --git a/hw/s390x/css.c b/hw/s390x/css.c
>> index 8fc9e35ba5d3..080ac7e5bc0b 100644
>> --- a/hw/s390x/css.c
>> +++ b/hw/s390x/css.c
>> @@ -830,8 +830,12 @@ static int ccw_dstream_rw_noflags(CcwDataStream 
>> *cds, void *buff, int len,
>>       if (op == CDS_OP_A) {
>>           goto incr;
>>       }
>> -    ret = address_space_rw(&address_space_memory, cds->cda,
>> -                           MEMTXATTRS_UNSPECIFIED, buff, len, op);
>> +    if (!cds->do_skip) {
>> +        ret = address_space_rw(&address_space_memory, cds->cda,
>> +                               MEMTXATTRS_UNSPECIFIED, buff, len, op);
>> +    } else {
>> +        ret = 0;
>> +    }
>>       if (ret != MEMTX_OK) {
>>           cds->flags |= CDS_F_STREAM_BROKEN;
>>           return -EINVAL;
>> @@ -928,8 +932,13 @@ static int ccw_dstream_rw_ida(CcwDataStream *cds, 
>> void *buff, int len,
>>       do {
>>           iter_len = MIN(len, cont_left);
>>           if (op != CDS_OP_A) {
>> -            ret = address_space_rw(&address_space_memory, cds->cda,
>> -                                   MEMTXATTRS_UNSPECIFIED, buff, 
>> iter_len, op);
>> +            if (!cds->do_skip) {
>> +                ret = address_space_rw(&address_space_memory, cds->cda,
>> +                                       MEMTXATTRS_UNSPECIFIED, buff, 
>> iter_len,
>> +                                       op);
>> +            } else {
>> +                ret = 0;
>> +            }
>>               if (ret != MEMTX_OK) {
>>                   /* assume inaccessible address */
>>                   ret = -EINVAL; /* channel program check */
>> @@ -968,6 +977,11 @@ void ccw_dstream_init(CcwDataStream *cds, CCW1 
>> const *ccw, ORB const *orb)
>>       cds->count = ccw->count;
>>       cds->cda_orig = ccw->cda;
>> +    /* skip is only effective for read, read backwards, or sense 
>> commands */
>> +    cds->do_skip = (ccw->flags & CCW_FLAG_SKIP) &&
>> +        ((ccw->cmd_code & CCW_CMD_BASIC_SENSE) == CCW_CMD_BASIC_SENSE ||
>> +         (ccw->cmd_code & 0x02) == 0x02 /* read */ ||
>> +         (ccw->cmd_code & 0x0c) == 0x0c /* read backwards */);
> 
> I think you should use masks like
> ((code & 0x3) == 2) => READ
> ((code & 0xf) == 0xc) => READ BACKWARD
> ((code & 0xf) == 0x4) => SENSE

I think Pierre is right.  In the v2 code, a control CCW (like NOP) would 
still be flagged as a READ.

> 
> Regards,
> Pierre
> 
> 
> 
> 
>>       ccw_dstream_rewind(cds);
>>       if (!(cds->flags & CDS_F_IDA)) {
>>           cds->op_handler = ccw_dstream_rw_noflags;
>> diff --git a/include/hw/s390x/css.h b/include/hw/s390x/css.h
>> index aae19c427229..7cc183ef4366 100644
>> --- a/include/hw/s390x/css.h
>> +++ b/include/hw/s390x/css.h
>> @@ -97,6 +97,7 @@ typedef struct CcwDataStream {
>>       int (*op_handler)(struct CcwDataStream *cds, void *buff, int len,
>>                         CcwDataStreamOp op);
>>       hwaddr cda;
>> +    bool do_skip;
>>   } CcwDataStream;
>>   /*
>>
> 
>