Message-ID: <28ad3045-9238-2a77-d74d-9660a36aa4da@redhat.com>
Date:   Thu, 24 Sep 2020 20:12:51 +0200
From:   David Hildenbrand <david@...hat.com>
To:     Dan Williams <dan.j.williams@...el.com>
Cc:     Joao Martins <joao.m.martins@...cle.com>,
        Andrew Morton <akpm@...ux-foundation.org>,
        Vishal Verma <vishal.l.verma@...el.com>,
        Dave Hansen <dave.hansen@...ux.intel.com>,
        Pavel Tatashin <pasha.tatashin@...een.com>,
        Peter Zijlstra <peterz@...radead.org>,
        Ard Biesheuvel <ard.biesheuvel@...aro.org>,
        Linux MM <linux-mm@...ck.org>,
        linux-nvdimm <linux-nvdimm@...ts.01.org>,
        Linux Kernel Mailing List <linux-kernel@...r.kernel.org>,
        Linux ACPI <linux-acpi@...r.kernel.org>,
        Maling list - DRI developers 
        <dri-devel@...ts.freedesktop.org>
Subject: Re: [PATCH v4 11/23] device-dax: Kill dax_kmem_res

On 24.09.20 15:54, Dan Williams wrote:
> On Thu, Sep 24, 2020 at 12:26 AM David Hildenbrand <david@...hat.com> wrote:
>>
>> On 23.09.20 23:41, Dan Williams wrote:
>>> On Wed, Sep 23, 2020 at 1:04 AM David Hildenbrand <david@...hat.com> wrote:
>>>>
>>>> On 08.09.20 17:33, Joao Martins wrote:
>>>>> [Sorry for the late response]
>>>>>
>>>>> On 8/21/20 11:06 AM, David Hildenbrand wrote:
>>>>>> On 03.08.20 07:03, Dan Williams wrote:
>>>>>>> @@ -37,109 +45,94 @@ int dev_dax_kmem_probe(struct device *dev)
>>>>>>>      * could be mixed in a node with faster memory, causing
>>>>>>>      * unavoidable performance issues.
>>>>>>>      */
>>>>>>> -   numa_node = dev_dax->target_node;
>>>>>>>     if (numa_node < 0) {
>>>>>>>             dev_warn(dev, "rejecting DAX region with invalid node: %d\n",
>>>>>>>                             numa_node);
>>>>>>>             return -EINVAL;
>>>>>>>     }
>>>>>>>
>>>>>>> -   /* Hotplug starting at the beginning of the next block: */
>>>>>>> -   kmem_start = ALIGN(range->start, memory_block_size_bytes());
>>>>>>> -
>>>>>>> -   kmem_size = range_len(range);
>>>>>>> -   /* Adjust the size down to compensate for moving up kmem_start: */
>>>>>>> -   kmem_size -= kmem_start - range->start;
>>>>>>> -   /* Align the size down to cover only complete blocks: */
>>>>>>> -   kmem_size &= ~(memory_block_size_bytes() - 1);
>>>>>>> -   kmem_end = kmem_start + kmem_size;
>>>>>>> -
>>>>>>> -   new_res_name = kstrdup(dev_name(dev), GFP_KERNEL);
>>>>>>> -   if (!new_res_name)
>>>>>>> +   res_name = kstrdup(dev_name(dev), GFP_KERNEL);
>>>>>>> +   if (!res_name)
>>>>>>>             return -ENOMEM;
>>>>>>>
>>>>>>> -   /* Region is permanently reserved if hotremove fails. */
>>>>>>> -   new_res = request_mem_region(kmem_start, kmem_size, new_res_name);
>>>>>>> -   if (!new_res) {
>>>>>>> -           dev_warn(dev, "could not reserve region [%pa-%pa]\n",
>>>>>>> -                    &kmem_start, &kmem_end);
>>>>>>> -           kfree(new_res_name);
>>>>>>> +   res = request_mem_region(range.start, range_len(&range), res_name);
>>>>>>
>>>>>> I think our range could be empty after aligning. I assume
>>>>>> request_mem_region() would check that, but maybe we could report a
>>>>>> better error/warning in that case.
>>>>>>
>>>>> dax_kmem_range() already returns a memory-block-aligned @range but
>>>>> IIUC request_mem_region() isn't checking for that. Having said that
>>>>> the returned @res wouldn't be different from the passed range.start.
>>>>>
>>>>>>>     /*
>>>>>>>      * Ensure that future kexec'd kernels will not treat this as RAM
>>>>>>>      * automatically.
>>>>>>>      */
>>>>>>> -   rc = add_memory_driver_managed(numa_node, new_res->start,
>>>>>>> -                                  resource_size(new_res), kmem_name);
>>>>>>> +   rc = add_memory_driver_managed(numa_node, res->start,
>>>>>>> +                                  resource_size(res), kmem_name);
>>>>>>> +
>>>>>>> +   res->flags |= IORESOURCE_BUSY;
>>>>>>
>>>>>> Hm, I don't think that's correct. Any specific reason why to mark the
>>>>>> not-added, unaligned parts BUSY? E.g., walk_system_ram_range() could
>>>>>> suddenly stumble over it - and e.g., similarly kexec code when trying to
>>>>>> find memory for placing kexec images. I think we should leave this
>>>>>> !BUSY, just as it is right now.
>>>>>>
>>>>> Agreed.
>>>>>
>>>>>>>     if (rc) {
>>>>>>> -           release_resource(new_res);
>>>>>>> -           kfree(new_res);
>>>>>>> -           kfree(new_res_name);
>>>>>>> +           release_mem_region(range.start, range_len(&range));
>>>>>>> +           kfree(res_name);
>>>>>>>             return rc;
>>>>>>>     }
>>>>>>> -   dev_dax->dax_kmem_res = new_res;
>>>>>>> +
>>>>>>> +   dev_set_drvdata(dev, res_name);
>>>>>>>
>>>>>>>     return 0;
>>>>>>>  }
>>>>>>>
>>>>>>>  #ifdef CONFIG_MEMORY_HOTREMOVE
>>>>>>> -static int dev_dax_kmem_remove(struct device *dev)
>>>>>>> +static void dax_kmem_release(struct dev_dax *dev_dax)
>>>>>>>  {
>>>>>>> -   struct dev_dax *dev_dax = to_dev_dax(dev);
>>>>>>> -   struct resource *res = dev_dax->dax_kmem_res;
>>>>>>> -   resource_size_t kmem_start = res->start;
>>>>>>> -   resource_size_t kmem_size = resource_size(res);
>>>>>>> -   const char *res_name = res->name;
>>>>>>>     int rc;
>>>>>>> +   struct device *dev = &dev_dax->dev;
>>>>>>> +   const char *res_name = dev_get_drvdata(dev);
>>>>>>> +   struct range range = dax_kmem_range(dev_dax);
>>>>>>>
>>>>>>>     /*
>>>>>>>      * We have one shot for removing memory, if some memory blocks were not
>>>>>>>      * offline prior to calling this function remove_memory() will fail, and
>>>>>>>      * there is no way to hotremove this memory until reboot because device
>>>>>>> -    * unbind will succeed even if we return failure.
>>>>>>> +    * unbind will proceed regardless of the remove_memory result.
>>>>>>>      */
>>>>>>> -   rc = remove_memory(dev_dax->target_node, kmem_start, kmem_size);
>>>>>>> -   if (rc) {
>>>>>>> -           any_hotremove_failed = true;
>>>>>>> -           dev_err(dev,
>>>>>>> -                   "DAX region %pR cannot be hotremoved until the next reboot\n",
>>>>>>> -                   res);
>>>>>>> -           return rc;
>>>>>>> +   rc = remove_memory(dev_dax->target_node, range.start, range_len(&range));
>>>>>>> +   if (rc == 0) {
>>>>>>
>>>>>> if (!rc) ?
>>>>>>
>>>>> Better off would be to keep the old order:
>>>>>
>>>>>       if (rc) {
>>>>>               any_hotremove_failed = true;
>>>>>               dev_err(dev, "%#llx-%#llx cannot be hotremoved until the next reboot\n",
>>>>>                               range.start, range.end);
>>>>>               return;
>>>>>       }
>>>>>
>>>>>       release_mem_region(range.start, range_len(&range));
>>>>>       dev_set_drvdata(dev, NULL);
>>>>>       kfree(res_name);
>>>>>       return;
>>>>>
>>>>>
>>>>>>> +           release_mem_region(range.start, range_len(&range));
>>>>>>
>>>>>> remove_memory() does a release_mem_region_adjustable(). Don't you
>>>>>> actually want to release the *unaligned* region you requested?
>>>>>>
>>>>> Isn't it what we're doing here?
>>>>> (The release_mem_region_adjustable() is using the same
>>>>> dax_kmem-aligned range and there's no split/adjust)
>>>>>
>>>>> Meaning right now (+ parent marked as !BUSY), and if I am understanding
>>>>> this correctly:
>>>>>
>>>>> request_mem_region(range.start, range_len)
>>>>>    __request_region(iomem_res, range.start, range_len) -> alloc @parent
>>>>> add_memory_driver_managed(parent.start, resource_size(parent))
>>>>>    __request_region(parent.start, resource_size(parent)) -> alloc @child
>>>>>
>>>>> [...]
>>>>>
>>>>> remove_memory(range.start, range_len)
>>>>>  request_mem_region_adjustable(range.start, range_len)
>>>>>   __release_region(range.start, range_len) -> remove @child
>>>>>
>>>>> release_mem_region(range.start, range_len)
>>>>>   __release_region(range.start, range_len) -> doesn't remove @parent because !BUSY?
>>>>>
>>>>> The add/removal of this relies on !BUSY. But now I am wondering if the parent remaining
>>>>> unreleased is deliberate even on CONFIG_MEMORY_HOTREMOVE=y.
>>>>>
>>>>>       Joao
>>>>>
>>>>
>>>> Thinking about it, if we don't set the parent resource BUSY (which is
>>>> what I think is the right way of doing things), and don't want to store
>>>> the parent resource pointer, we could add something like
>>>> lookup_resource() - e.g., lookup_mem_resource() - , however, searching
>>>> properly in the whole hierarchy (instead of only the first level), and
>>>> traversing down to the last hierarchy. Then it would be as simple as
>>>>
>>>> remove_memory(range.start, range_len)
>>>> res = lookup_mem_resource(range.start);
>>>> release_resource(res);
>>>
>>> Another thought... I notice that you've taught
>>> register_memory_resource() a IORESOURCE_MEM_DRIVER_MANAGED special
>>> case. Lets just make the assumption of add_memory_driver_managed()
>>> that it is the driver's responsibility to mark the range busy before
>>> calling, and the driver's responsibility to release the region. I.e.
>>> validate (rather than request) that the range is busy in
>>> register_memory_resource(), and teach release_memory_resource() to
>>> skip releasing the region when the memory is marked driver managed.
>>> That would let dax_kmem drop its manipulation of the 'busy' flag which
>>> is a layering violation no matter how many comments we put around it.
>>
>> IIUC, that won't work for virtio-mem, whereby the parent resource spans
>> multiple possible (future) add_memory_driver_managed() calls and is
>> (just like for kmem) a pure indication to which device memory ranges belong.
>>
>> For example, when exposing 2GB via a virtio-mem device with max 4GB:
>>
>> (/proc/iomem)
>> 240000000-33fffffff : virtio0
>>   240000000-2bfffffff : System RAM (virtio_mem)
>>
>> And after hotplugging additional 2GB:
>>
>> 240000000-33fffffff : virtio0
>>   240000000-33fffffff : System RAM (virtio_mem)
>>
>> So marking "virtio0" always BUSY (especially right from the start) would
>> be wrong.
> 
> I'm not suggesting to busy the whole "virtio" range, just the portion
> that's about to be passed to add_memory_driver_managed().

I'm afraid I don't get your point. For virtio-mem:

Before:

1. Create virtio0 container resource

2. (somewhen in the future) add_memory_driver_managed()
 - Create resource (System RAM (virtio_mem)), marking it busy/driver
   managed

After:

1. Create virtio0 container resource

2. (somewhen in the future) Create resource (System RAM (virtio_mem)),
   marking it busy/driver managed
3. add_memory_driver_managed()

Not helpful or simpler IMHO.

> 
>> The assumption is that anything that's IORESOURCE_SYSTEM_RAM
>> and IORESOUCE_BUSY is currently added to the system as system RAM (e.g.,
>> after add_memory() and friends, or during boot).
>>
>> I do agree that manually clearing the busy flag is ugly. What we most
>> probably want is request_mem_region() that performs similar checks (no
>> overlaps with existing BUSY resources), but doesn't set the region busy.
>>
> 
> I can't see that working without some way to export and hold the
> resource lock until some agent can atomically claim the range.

I don't think we have to care about races here. The "BUSY" checks is
really just a check for leftovers, e.g., after kexec or after driver
reloading. If somebody else would try to concurrently add System RAM
/something else within the range of your device, something else, very
weird, would be going on (let's call it a BUG, just like if somebody
would be removing system RAM in your device range ...).

For example, in case of virtio-mem, when you unload the driver, it
cannot remove the "virtio0" resource in case some system ram in the
range is still plugged (busy).

So when reloading the driver, it would try to re-create the virtio0
resource, detect that some system ram in the range is still BUSY, and
fail gracefully. This is how it works and how it's expected to work - at
least for virtio-mem.

I assume something similar can be observed with kmem, when trying to
reload the driver or similar - but races shouldn't be relevant here.

-- 
Thanks,

David / dhildenb

Powered by blists - more mailing lists