[Top][All Lists]

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [PATCH 18/23] plugins: add an API to read registers

From: Alex Bennée
Subject: Re: [PATCH 18/23] plugins: add an API to read registers
Date: Fri, 23 Feb 2024 11:44:30 +0000
User-agent: mu4e 1.11.28; emacs 29.1

Akihiko Odaki <akihiko.odaki@daynix.com> writes:

> On 2024/02/23 2:27, Alex Bennée wrote:
>> Akihiko Odaki <akihiko.odaki@daynix.com> writes:
>>> On 2024/02/22 19:20, Alex Bennée wrote:
>>>> Akihiko Odaki <akihiko.odaki@daynix.com> writes:
>>>>> On 2024/02/21 23:14, Alex Bennée wrote:
>>>>>> Akihiko Odaki <akihiko.odaki@daynix.com> writes:
>>>>>>> On 2024/02/21 19:02, Alex Bennée wrote:
>>>>>>>> Akihiko Odaki <akihiko.odaki@daynix.com> writes:
>>>>>>>>> On 2024/02/20 23:14, Alex Bennée wrote:
>>>>>>>>>> Akihiko Odaki <akihiko.odaki@daynix.com> writes:
>>>>>>>>>>> On 2024/02/17 1:30, Alex Bennée wrote:
>>>>>>>>>>>> We can only request a list of registers once the vCPU has been
>>>>>>>>>>>> initialised so the user needs to use either call the get function 
>>>>>>>>>>>> on
>>>>>>>>>>>> vCPU initialisation or during the translation phase.
>>>>>>>>>>>> We don't expose the reg number to the plugin instead hiding it
>>>>>>>>>>>> behind
>>>>>>>>>>>> an opaque handle. This allows for a bit of future proofing should 
>>>>>>>>>>>> the
>>>>>>>>>>>> internals need to be changed while also being hashed against the
>>>>>>>>>>>> CPUClass so we can handle different register sets per-vCPU in
>>>>>>>>>>>> hetrogenous situations.
>>>>>>>>>>>> Having an internal state within the plugins also allows us to 
>>>>>>>>>>>> expand
>>>>>>>>>>>> the interface in future (for example providing callbacks on 
>>>>>>>>>>>> register
>>>>>>>>>>>> change if the translator can track changes).
>>>>>>>>>>>> Resolves: https://gitlab.com/qemu-project/qemu/-/issues/1706
>>>>>>>>>>>> Cc: Akihiko Odaki <akihiko.odaki@daynix.com>
>>>>>>>>>>>> Message-Id: <20240103173349.398526-39-alex.bennee@linaro.org>
>>>>>>>>>>>> Based-on: <20231025093128.33116-18-akihiko.odaki@daynix.com>
>>>>>>>>>>>> Signed-off-by: Alex Bennée <alex.bennee@linaro.org>
>>>>>>>>>>>> Reviewed-by: Pierrick Bouvier <pierrick.bouvier@linaro.org>
>>>>>>>>>> <snip>
>>>>>>>>>>>> +/*
>>>>>>>>>>>> + * Register handles
>>>>>>>>>>>> + *
>>>>>>>>>>>> + * The plugin infrastructure keeps hold of these internal data
>>>>>>>>>>>> + * structures which are presented to plugins as opaque handles. 
>>>>>>>>>>>> They
>>>>>>>>>>>> + * are global to the system and therefor additions to the hash 
>>>>>>>>>>>> table
>>>>>>>>>>>> + * must be protected by the @reg_handle_lock.
>>>>>>>>>>>> + *
>>>>>>>>>>>> + * In order to future proof for up-coming heterogeneous work we 
>>>>>>>>>>>> want
>>>>>>>>>>>> + * different entries for each CPU type while sharing them in the
>>>>>>>>>>>> + * common case of multiple cores of the same type.
>>>>>>>>>>>> + */
>>>>>>>>>>>> +
>>>>>>>>>>>> +static QemuMutex reg_handle_lock;
>>>>>>>>>>>> +
>>>>>>>>>>>> +struct qemu_plugin_register {
>>>>>>>>>>>> +    const char *name;
>>>>>>>>>>>> +    int gdb_reg_num;
>>>>>>>>>>>> +};
>>>>>>>>>>>> +
>>>>>>>>>>>> +static GHashTable *reg_handles; /* hash table of PluginReg */
>>>>>>>>>>>> +
>>>>>>>>>>>> +/* Generate a stable key - would xxhash be overkill? */
>>>>>>>>>>>> +static gpointer cpu_plus_reg_to_key(CPUState *cs, int gdb_regnum)
>>>>>>>>>>>> +{
>>>>>>>>>>>> +    uintptr_t key = (uintptr_t) cs->cc;
>>>>>>>>>>>> +    key ^= gdb_regnum;
>>>>>>>>>>>> +    return GUINT_TO_POINTER(key);
>>>>>>>>>>>> +}
>>>>>>>>>>> I have pointed out this is theoretically prone to collisions and
>>>>>>>>>>> unsafe.
>>>>>>>>>> How is it unsafe? The aim is to share handles for the same CPUClass
>>>>>>>>>> rather than having a unique handle per register/cpu combo.
>>>>>>>>> THe intention is legitimate, but the implementation is not safe. It
>>>>>>>>> assumes (uintptr)cs->cc ^ gdb_regnum is unique, but there is no such
>>>>>>>>> guarantee. The key of GHashTable must be unique; generating hashes of
>>>>>>>>> keys should be done with hash_func given to g_hash_table_new().
>>>>>>>> This isn't a hash its a non-unique key. It is however unique for
>>>>>>>> the same register on the same class of CPU so for each vCPU in a system
>>>>>>>> can share the same opaque handles.
>>>>>>>> The hashing is done internally by glib. We would assert if there was
>>>>>>>> a
>>>>>>>> duplicate key referring to a different register.
>>>>>>>> I'm unsure what you want here? Do you have a suggestion for the key
>>>>>>>> generation algorithm? As the comment notes I did consider a more 
>>>>>>>> complex
>>>>>>>> mixing algorithm using xxhash but that wouldn't guarantee no clash
>>>>>>>> either.
>>>>>>> I suggest using a struct that holds both of cs->cc and gdb_regnum, and
>>>>>>> pass g_direct_equal() and g_direct_hash() to g_hash_table_new().
>>>>>> We already do:
>>>>>>            if (!reg_handles) {
>>>>>>                reg_handles = g_hash_table_new(g_direct_hash, 
>>>>>> g_direct_equal);
>>>>>>            }
>>>>>> But we can't use g_direct_equal with something that exceeds the
>>>>>> width of
>>>>>> gpointer as it is a straight equality test of the key. What you are
>>>>>> suggesting requires allocating memory for each key and de-referencing
>>>>>> with a custom GEqualFunc.
>>>>> My bad. I wrongly remembered g_direct_equal() and g_direct_hash(). It
>>>>> indeed seems to need a more complicated solution.
>>>>> It is possible to write a GEqualFunc and a GHashFunc that consumes a
>>>>> struct but it is a chore. How about having a two-level GHashTable?
>>>>> reg_handles will be a GHashTable keyed with cs->cc, and another
>>>>> GHashTable will be keyed with gdb_regnum.
>>>> That still seems overkill for a clash that can't happen. What do you
>>>> think about the following:
>>>>     /*
>>>>      * Generate a stable key shared across CPUs of the same class
>>>>      *
>>>>      * In order to future proof for up-coming heterogeneous work we want
>>>>      * different entries for each CPU type while sharing them in the
>>>>      * common case of multiple cores of the same type. This makes the
>>>>      * assumption you won't see two CPUClass pointers that are similar
>>>>      * enough that the low bits mixed with different registers numbers
>>>>      * will give you the same key.
>>>>      *
>>>>      * The build time assert will fire if CPUClass goes on a sudden diet
>>>>      * and we assert further down if we detect two keys representing
>>>>      * different regnums. In practice allocations of CPUClass are much
>>>>      * farther apart making clashes practically impossible.
>>>>      */
>>>>     static gpointer cpu_plus_reg_to_key(CPUState *cs, int gdb_regnum)
>>>>     {
>>>>         uintptr_t key = (uintptr_t) cs->cc;
>>>>         /* this protects some of the assumptions above */
>>>>         qemu_build_assert(sizeof(*cs->cc) >= 256);
>>>>         key ^= gdb_regnum;
>>>>         return GUINT_TO_POINTER(key);
>>>>     }
>>> I think the assertion and comments are overkill. Doesn't having a
>>> nested GHashTable save some words you have to wrote for the
>>> assumption?
>> A nested hash table for a single entry is overkill.
> You mean that the first level will be indexed by only one CPUClass (or
> few if we support a heterogeneous system).
> I think it's still OK though. It's not like we will need more code
> when having few entries.
>>> I'm also not quite convinced that the comments and assertions are
>>> enough to say this hack is safe; what if some sort of pointer
>>> authentication is added and shuffles bits? Will this hack be
>>> compatible with static and dynamic checkers we may have in the future?
>> We are not using the value as a pointer so that should be irrelevant
>> although generally those bits tend to be at the top of pointers so they
>> can be masked off.
>> I'm not sure what we are trying to achieve here. I've got something
>> that
>> works, doesn't fail any tests and has some guards in for potential
>> future problems. At the same time I'm not prepared to over-engineer the
>> solution for a theoretical future problem we haven't got yet.
>> What about if I just key based of gdb_regnum and we accept that that
>> might break the one heterogeneous system we model today?
> That's the best option in my opinion. gdbstub won't work well with
> such a system anyway, and fixing it will need something similar to
> GHashTable. But if I would fix gdbstub for a heterogeneous system, I
> would add a field to CPUClass instead of having a GHashTable keyed
> with tuples of CPUClass pointers and register numbers. It should be
> fine considering that CPUState already has gdbstub-specific fields
> like gdb_regs.

It would be nice to move all register code into CPUClass to avoid
repeating ourselves but I suspect that is quite an invasive change for a
later series. Currently all the CPUClass values are set on init and
shouldn't really be changed after that otherwise we'll have to start
messing with locking. 

Alex Bennée
Virtualisation Tech Lead @ Linaro

reply via email to

[Prev in Thread] Current Thread [Next in Thread]