Re: [PULL v2 0/8] Ide patches

[Laszlo Ersek]

On 10/08/19 23:58, Laszlo Ersek wrote:
> On 10/07/19 19:55, Philippe Mathieu-Daudé wrote:
>> On 10/7/19 7:35 PM, John Snow wrote:
>>> On 10/7/19 8:33 AM, Peter Maydell wrote:
>>>> On Thu, 3 Oct 2019 at 20:33, John Snow <address@hidden> wrote:
>>>>>
>>>>> The following changes since commit
>>>>> 7f21573c822805a8e6be379d9bcf3ad9effef3dc:
>>>>>
>>>>>    Merge remote-tracking branch
>>>>> 'remotes/huth-gitlab/tags/pull-request-2019-10-01' into staging
>>>>> (2019-10-01 13:13:38 +0100)
>>>>>
>>>>> are available in the Git repository at:
>>>>>
>>>>>    https://github.com/jnsnow/qemu.git tags/ide-pull-request
>>>>>
>>>>> for you to fetch changes up to
>>>>> f6d61c9509c56eea3cdd2d23b40d285601b1c1ca:
>>>>>
>>>>>    hd-geo-test: Add tests for lchs override (2019-10-03 14:36:54 -0400)
>>>>>
>>>>> ----------------------------------------------------------------
>>>>> Pull request V2
>>>>>
>>>>> - Added signoff into the mirrored commits themselves (vs just the
>>>>> email)
>>>>> - Kudos to `stg-foreach stg edit --sign`
>>>>>
>>>>> ----------------------------------------------------------------
>>>>
>>>> Hi; the new tests in hd-geo-test seem to hang on
>>>> big-endian hosts (both s390x and ppc64 hung here):
>>>>
>>>> linux1@lxub05:~/qemu/build/all$ QTEST_QEMU_IMG=./qemu-img
>>>> QTEST_QEMU_BINARY=x86_64-softmmu/qemu-system-x86_64
>>>> ./tests/hd-geo-test
>>>> /x86_64/hd-geo/ide/none: OK
>>>> /x86_64/hd-geo/ide/drive/cd_0: OK
>>>> /x86_64/hd-geo/ide/drive/mbr/blank: OK
>>>> /x86_64/hd-geo/ide/drive/mbr/lba: OK
>>>> /x86_64/hd-geo/ide/drive/mbr/chs: OK
>>>> /x86_64/hd-geo/ide/device/mbr/blank: OK
>>>> /x86_64/hd-geo/ide/device/mbr/lba: OK
>>>> /x86_64/hd-geo/ide/device/mbr/chs: OK
>>>> /x86_64/hd-geo/ide/device/user/chs: OK
>>>> /x86_64/hd-geo/ide/device/user/chst: OK
>>>> /x86_64/hd-geo/override/ide:
>>>>
>>>
>>> :(
>>>
>>>>
>>>> thanks
>>>> -- PMM
>>>>
>>>
>>> Sam, can you investigate this?
>>
>> Not seeing my T-b tags makes me grumble because I don't remember which I
>> reviewed and need to go check on the list.
>>
>> If the error is a endianess bug related to fw_cfg, you can add the
>> "-trace fw_cfg*" in hd-geo-test::create_args() and rerun the tests on a
>> BE system, the bug should appear straightly on stdout.
>>
>> Are FWLCHSEntry fields little-endian? Shouldn't
>> get_boot_devices_lchs_list() use some le32_to_cpu() call for the LCHS
>> values?
>>
> 
> *One* problem is most likely in the find_fw_cfg_file() function, in patch 8.
> 
> +static uint16_t find_fw_cfg_file(QFWCFG *fw_cfg,
> +                                 const char *filename)
> +{
> +    struct QemuCfgFile qfile;
> +    uint32_t count, e;
> +    uint16_t select;
> +
> +    count = qfw_cfg_get_u32(fw_cfg, FW_CFG_FILE_DIR);
> +    count = be32_to_cpu(count);
> +    for (select = 0, e = 0; e < count; e++) {
> +        qfw_cfg_read_data(fw_cfg, &qfile, sizeof(qfile));
> +        if (!strcmp(filename, qfile.name)) {
> +            select = be16_to_cpu(qfile.select);
> +        }
> +    }
> +
> +    return select;
> +}
> 
> Note qfw_cfg_get_u32():
> 
> uint32_t qfw_cfg_get_u32(QFWCFG *fw_cfg, uint16_t key)
> {
>     uint32_t value;
>     qfw_cfg_get(fw_cfg, key, &value, sizeof(value));
>     return le32_to_cpu(value);
> }
> 
> This function assumes that the wire encoding of the value read is little
> endian. So, calling this function is wrong; and calling be32_to_cpu()
> afterwards does not help. Namely:
> 
> * On LE hosts, the find_fw_cfg_file() function happens to work, because:
> 
> - the le32_to_cpu() call in qfw_cfg_get_u32() does nothing (it's identity),
> - the subsequent be32_to_cpu() call in find_fw_cfg_file() corresponds to
> the *blob-specific* encoding of the "count" field, in the fw_cfg
> directory blob. (Which is BE) Therefore we perform the one byte-swap
> that we need.
> 
> * On BE hosts, stuff breaks, because:
> 
> - the le32_to_cpu() call in qfw_cfg_get_u32() swaps the byte-order,
> - the subsequent be32_to_cpu() call in find_fw_cfg_file() does nothing,
> - thus, ultimately we have byte-swapped the contents of the "count"
> field of the directory blob, even though the blob-specific wire format
> thereof is *already* BE (= host-endian). On a BE host, all in all, there
> should be zero byte swaps for consuming "count".

And the hang is probably due to the loop in find_fw_cfg_file() counting
up to cca. 0x2000_0000:

- The file directory currently has room for 0x20 files, and if you
byte-swap that as a uint32_t, you get 0x2000_0000. (You minimally get
0x0100_0000, if there's just one entry in the directory -- still
16,777,216 in decimal.)

- Additionally, the loop body does not contain a "break" statement for
when strcmp() matches; so even if there is a hit in the low numbers, the
loop continues to the limit.

Thanks
Laszlo

> Now, how to fix this: eliminate
> - QemuCfgFile,
> - find_fw_cfg_file(),
> - and read_fw_cfg_file()
> 
> altogether, and call qfw_cfg_get_file(), from "tests/libqos/fw_cfg.c".
> 
> Some other tests look up fw_cfg directory entries with that function
> already (see call sites in "tests/fw_cfg-test.c").
> 
> Thanks
> Laszlo
>