Message-ID: <f45c26ae-865b-b7b6-0f2a-5906d9c0a349@redhat.com>
Date:   Thu, 16 Mar 2023 10:59:39 +0100
MIME-Version: 1.0
User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:102.0) Gecko/20100101
 Thunderbird/102.7.1
Subject: Re: [PATCH v3 6/8] platform/x86/intel/ifs: Implement Array BIST test
Content-Language: en-US, nl
To:     "Joseph, Jithu" <jithu.joseph@intel.com>, markgross@kernel.org
Cc:     tglx@linutronix.de, mingo@redhat.com, bp@alien8.de,
        dave.hansen@linux.intel.com, x86@kernel.org, hpa@zytor.com,
        gregkh@linuxfoundation.org, rostedt@goodmis.org,
        ashok.raj@intel.com, tony.luck@intel.com,
        linux-kernel@vger.kernel.org, platform-driver-x86@vger.kernel.org,
        patches@lists.linux.dev, ravi.v.shankar@intel.com,
        thiago.macieira@intel.com, athenas.jimenez.gonzalez@intel.com,
        sohil.mehta@intel.com
References: <20230214234426.344960-1-jithu.joseph@intel.com>
 <20230301015942.462799-1-jithu.joseph@intel.com>
 <20230301015942.462799-7-jithu.joseph@intel.com>
 <628e092c-793e-96ce-71f5-80392ad69569@redhat.com>
 <9dde72aa-a28d-2b0e-7b90-6b1996dbf202@intel.com>
From:   Hans de Goede <hdegoede@redhat.com>
In-Reply-To: <9dde72aa-a28d-2b0e-7b90-6b1996dbf202@intel.com>
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 7bit
Precedence: bulk

Hi,

On 3/13/23 17:37, Joseph, Jithu wrote:
> Hans,
> 
> Thank-you very much for the review.
> 
> On 3/13/2023 9:24 AM, Hans de Goede wrote:
> 
> 
>>>  
>>> +#define SPINUNIT 100 /* 100 nsec */
>>> +static atomic_t array_cpus_out;
>>
>> This variable is only inc-ed + read, it is never reset to 0
>> so the "while (atomic_read(t) < all_cpus)"
>> check only works for the first test run.
>>
> 
> It is reset to zero as annotated below. Let me know if this doesn't address your concern.

Ah, I somehow missed that despite looking for it twice, yes that is fine.

>> Also even static atomic_t variables must be initialized, you cannot
>> assume that using using zeroed mem is a valid value for an atomic_t.
>>
>> And this is also shared between all smt pairs, so if 2 "real"
>> CPU cores with both 2 sibblings are asked to run IFS tests at
>> the same time, then array_cpus_out will get increased 4 times
>> in total, breaking the wait_for_sibbling loop as soon as
>> the counter reaches 2, so before the tests are done.
> 
> Only one IFS test is allowed at a time. This is done using "ifs_sem" defined in sysfs.c

Ah I see.

After taking a closer look I do see one unrelated issue with this patch
sysfs.c: run_test_store() does:

        if (!ifsd->loaded)
                rc = -EPERM;
        else
                rc = do_core_test(cpu, dev);

But AFAICT the loaded check really only applies to the first (intel_ifs_0 device) test type and the 
Array BIST test should work fine when loaded is false.

So I think that the if (!ifsd->loaded) error check should be moved to 
ifs_test_core() ?

Regards,

Hans


> 
> ...
> 
>>> +static void ifs_array_test_core(int cpu, struct device *dev)
>>> +{
>>> +	union ifs_array command = {};
>>> +	bool timed_out = false;
>>> +	struct ifs_data *ifsd;
>>> +	unsigned long timeout;
>>> +
>>> +	ifsd = ifs_get_data(dev);
>>> +
>>> +	command.array_bitmask = ~0U;
>>> +	timeout = jiffies + HZ / 2;
>>> +
>>> +	do {
>>> +		if (time_after(jiffies, timeout)) {
>>> +			timed_out = true;
>>> +			break;
>>> +		}
>>> +		atomic_set(&array_cpus_out, 0);
> 
> The above line is where the zero initialization happens before every test.
> 
>>> +		stop_core_cpuslocked(cpu, do_array_test, &command);
>>> +
>>> +		if (command.ctrl_result)
>>> +			break;
>>> +	} while (command.array_bitmask);
>>> +
>>> +	ifsd->scan_details = command.data;
>>> +
>>> +	if (command.ctrl_result)
>>> +		ifsd->status = SCAN_TEST_FAIL;
>>> +	else if (timed_out || command.array_bitmask)
>>> +		ifsd->status = SCAN_NOT_TESTED;
>>> +	else
>>> +		ifsd->status = SCAN_TEST_PASS;
>>> +}
> 
> Jithu
>