Re: [Intel-gfx] refactor the i915 GVT support

["Wang, Zhi A" <zhi.a.wang@intel.com>]

On 9/28/21 2:00 PM, Luis Chamberlain wrote:
> On Tue, Sep 28, 2021 at 07:41:00AM +0000, Wang, Zhi A wrote:
>> Hey guys:
>>
>> After some investigation, I found the root cause this problem ("i915"
>> module loading will be stuck with Christoph's refactor patches), which
>> can be reproduced by building both i915 and kvmgt as kernel module and
>> the loading i915.
> Thanks for looking into this!
>
>> The root cause is: in Linux kernel loading, before a kernel module
>> loading is finished, its symbols can not be reached by other module when
>> resolving the symbols (even they can be found in /proc/kallsyms).
>> Because the status of the kernel module is MODULE_STATE_COMING and
>> resolve_symbol() from another kernel module will check this and return a
>> -EBUSY.
> Well, it would seem that way but...
>
>> In this case, before i915 loading is finished, the requested module
>> "kvmgt" cannot reach the symbols in module i915. Thus it kept waiting
>> and left message like below in the dmesg:
>>
>> [  644.152021] kvmgt: gave up waiting for init of module i915.
>> [  644.152039] kvmgt: Unknown symbol i915_gem_object_set_to_cpu_domain
>> (err -16)
>> [  674.871409] kvmgt: gave up waiting for init of module i915.
>> [  674.871427] kvmgt: Unknown symbol intel_ring_begin (err -16)
>> [  705.590586] kvmgt: gave up waiting for init of module i915.
>> [  705.590604] kvmgt: Unknown symbol i915_vma_move_to_active (err -16)
>> [  736.310230] kvmgt: gave up waiting for init of module i915.
>> [  736.310248] kvmgt: Unknown symbol shmem_unpin_map (err -16)
>> ...
>>
>> The error message is from execution path below:
>>
>> kernel/module.c:
>>
>> [i915 module loading] ->
>> request_module("kvmgt")->[modprobe]->init_module("kvmgt")->load_module()->simplify_symbols()->resolve_symbol_wait():
>>
>> static const struct kernel_symbol *
>> resolve_symbol_wait(struct module *mod,
>>               const struct load_info *info,
>>               const char *name)
>> {
>>       const struct kernel_symbol *ksym;
>>       char owner[MODULE_NAME_LEN];
>>
>>       if (wait_event_interruptible_timeout(module_wq,
>>               !IS_ERR(ksym = resolve_symbol(mod, info, name, owner))
>>               || PTR_ERR(ksym) != -EBUSY,
>>                            30 * HZ) <= 0) {
>>           pr_warn("%s: gave up waiting for init of module %s.\n",
>>               mod->name, owner);
>>
>> }
> Commit 9bea7f23952d5 ("module: fix bne2 "gave up waiting for init of
> module libcrc32c") is worth reviewing. It dealt with a similar issue,
> and in particular it addressed the issue with -EBUSY being returned
> by ref_module().
>
> And so, in theory that case should be dealt with in resolve_symbol_wait()
> already. And so can you try this just to verify something:
>
> diff --git a/kernel/module.c b/kernel/module.c
> index 40ec9a030eec..98f87cbb37de 100644
> --- a/kernel/module.c
> +++ b/kernel/module.c
> @@ -1459,7 +1459,7 @@ resolve_symbol_wait(struct module *mod,
>   	if (wait_event_interruptible_timeout(module_wq,
>   			!IS_ERR(ksym = resolve_symbol(mod, info, name, owner))
>   			|| PTR_ERR(ksym) != -EBUSY,
> -					     30 * HZ) <= 0) {
> +					     160 * HZ) <= 0) {
>   		pr_warn("%s: gave up waiting for init of module %s.\n",
>   			mod->name, owner);
>   	}
>
Hi Luis:

Thanks so much for the reply and patch.:)

I am afraid that this patch wouldn't work in this case as the 
request_module("kvmgt") happens in the init_module of i915. Before the 
initialization path is finished in i915, the i915 symbols are not 
available to be referenced. Unfortunately, It's matter of sequence, not 
of delay. :(

>> code:
>> https://github.com/intel/gvt-linux/blob/bd950a66c7919d7121d2530f30984351534a96dc/kernel/module.c#L1452
>>
>> In resolve_symbol_wait(), it calls resolve_symbol() to resolve the
>> symbols in "i915". In resolve_symbol() -> ref_module() ->
>> strong_try_module_get(), it will check the status of the module which
>> owns the symbol.
>>
>> static inline int strong_try_module_get(struct module *mod)
>> {
>>       BUG_ON(mod && mod->state == MODULE_STATE_UNFORMED);
>>       if (mod && mod->state == MODULE_STATE_COMING)
>>           return -EBUSY;
>>       if (try_module_get(mod))
>>           return 0;
>>       else
>>           return -ENOENT;
>> }
>>
>> code:https://github.com/intel/gvt-linux/blob/bd950a66c7919d7121d2530f30984351534a96dc/kernel/module.c#L318
>>
>> But unfortunately, this execution path begins in i915 module loading, at
>> this time, the status of kernel module "i915" is MODULE_STATE_COMING
>> until loading of "kvmgt" is finished. Thus a -EBUSY is always returned
>> when kernel is trying to resolve symbols for "kvmgt".
>>
>>
>> This patch below might need re-work:
> If the above test patch still fails, well.. that might be telling of
> another issue which is perhaps difficult to see at first glance. If
> resolve_symbol_wait() won't succeed until request_module("kvmgt")
> completes and if this means having kvmgt's init routine complete, that
> could end up in some longer chain or in the worst case a sort of
> circular dependency which is only implicated by module loading. It'd be
> really odd... but I cannot rule it out.
>
> This is one reason I hinted that you should strive to not do much on a
> module's init. If you can punt work off for later that's best.

Yes. I was thinking of the possibility of putting off some work later so 
that we don't need to make a lot of changes. GVT-g needs to take a 
snapshot of GPU registers as the initial virtual states for other vGPUs, 
which requires the initialization happens at a certain early time of 
initialization of i915. I was thinking maybe we can take other patches 
from Christoph like "de-virtualize*" except this one because currently 
we have to maintain a TEST-ONLY patch on our tree to prevent i915 built 
as kernel module.

Zhi.

>    Luis
>
>> Author: Christoph Hellwig <hch@lst.de>
>> Date:   Wed Jul 21 17:53:38 2021 +0200
>>
>>       drm/i915/gvt: move the gvt code into kvmgt.ko
>>
>>       Instead of having an option to build the gvt code into the main i915
>>       module, just move it into the kvmgt.ko module.  This only requires
>>       a new struct with three entries that the main i915 module needs to
>>       request before enabling VGPU passthrough operations.
>>
>>       This also conveniently streamlines the GVT initialization and avoids
>>       the need for the global device pointer.
>>
>>       Signed-off-by: Christoph Hellwig <hch@lst.de>
>>       Signed-off-by: Zhenyu Wang <zhenyuw@linux.intel.com>
>>       Link:
>> http://patchwork.freedesktop.org/patch/msgid/20210721155355.173183-5-hch@lst.de
>>       Acked-by: Zhenyu Wang <zhenyuw@linux.intel.com>
>>
>> On 8/26/21 6:12 AM, Zhenyu Wang wrote:
>>> On 2021.08.20 12:56:34 -0700, Luis Chamberlain wrote:
>>>> On Fri, Aug 20, 2021 at 04:17:24PM +0200, Christoph Hellwig wrote:
>>>>> On Thu, Aug 19, 2021 at 04:29:29PM +0800, Zhenyu Wang wrote:
>>>>>> I'm working on below patch to resolve this. But I met a weird issue in
>>>>>> case when building i915 as module and also kvmgt module, it caused
>>>>>> busy wait on request_module("kvmgt") when boot, it doesn't happen if
>>>>>> building i915 into kernel. I'm not sure what could be the reason?
>>>>> Luis, do you know if there is a problem with a request_module from
>>>>> a driver ->probe routine that is probably called by a module_init
>>>>> function itself?
>>>> Generally no, but you can easily foot yourself in the feet by creating
>>>> cross dependencies and not dealing with them properly. I'd make sure
>>>> to keep module initialization as simple as possible, and run whatever
>>>> takes more time asynchronously, then use a state machine to allow
>>>> you to verify where you are in the initialization phase or query it
>>>> or wait for a completion with a timeout.
>>>>
>>>> It seems the code in question is getting some spring cleaning, and its
>>>> unclear where the code is I can inspect. If there's a tree somewhere I
>>>> can take a peak I'd be happy to review possible oddities that may stick
>>>> out.
>>> I tried to put current patches under test here: https://github.com/intel/gvt-linux/tree/gvt-staging
>>> The issue can be produced with CONFIG_DRM_I915=m and CONFIG_DRM_I915_GVT_KVMGT=m.
>>>
>>>> My goto model for these sorts of problems is to abstract the issue
>>>> *outside* of the driver in question and implement new selftests to
>>>> try to reproduce. This serves two purposes, 1) helps with testing
>>>> 2) may allow you to see the problem more clearly.
>>>>
>>> I'll see if can abstract that.
>>>
>>> Thanks, Luis.
>>