Re: [PATCH v2 3/3] scsi: target: tcmu: Use address_space->invalidate_lock

[Bodo Stroesser <bostroesser@gmail.com>]

On 23.03.22 14:49, Xiaoguang Wang wrote:
> Currently tcmu_vma_fault() uses udev->cmdr_lock to avoid concurrent
> find_free_blocks(), which unmaps idle pages and truncates them. This
> work is really like many filesystem's truncate operations, but they
> use address_space->invalidate_lock to protect race.
> 
> This patch replaces cmdr_lock with address_space->invalidate_lock in
> tcmu fault procedure, which will also make page-fault have concurrency.
> 

Page fault happens only once for each allocated data page. After the
page fault was processed, the page is re-used for many cmds without
further page faults.

Of course, if find_free_blocks released a page and tcmu after that
allocates a new one, another page fault will occur. But that is a
very rare case.

So, what is the advantage of concurrency here?

OTOH, I don't like the switch from tcmu's cmdr_lock to
filemap_invalidate_* without a real technical reason.


> Signed-off-by: Xiaoguang Wang <xiaoguang.wang@linux.alibaba.com>
> ---
>   drivers/target/target_core_user.c | 14 ++++++++++----
>   1 file changed, 10 insertions(+), 4 deletions(-)
> 
> diff --git a/drivers/target/target_core_user.c b/drivers/target/target_core_user.c
> index 9196188504ec..81bfa553cc67 100644
> --- a/drivers/target/target_core_user.c
> +++ b/drivers/target/target_core_user.c
> @@ -1820,13 +1820,14 @@ static int tcmu_find_mem_index(struct vm_area_struct *vma)
>   
>   static struct page *tcmu_try_get_data_page(struct tcmu_dev *udev, uint32_t dpi)
>   {
> +	struct address_space *mapping = udev->inode->i_mapping;
>   	struct page *page;
>   
> -	mutex_lock(&udev->cmdr_lock);
> +	filemap_invalidate_lock_shared(mapping);
>   	page = xa_load(&udev->data_pages, dpi);
>   	if (likely(page)) {
>   		get_page(page);
> -		mutex_unlock(&udev->cmdr_lock);
> +		filemap_invalidate_unlock_shared(mapping);
>   		return page;
>   	}
>   
> @@ -1836,7 +1837,7 @@ static struct page *tcmu_try_get_data_page(struct tcmu_dev *udev, uint32_t dpi)
>   	 */
>   	pr_err("Invalid addr to data page mapping (dpi %u) on device %s\n",
>   	       dpi, udev->name);
> -	mutex_unlock(&udev->cmdr_lock);
> +	filemap_invalidate_unlock_shared(mapping);
>   
>   	return NULL;
>   }
> @@ -3116,6 +3117,7 @@ static void find_free_blocks(void)
>   	loff_t off;
>   	u32 pages_freed, total_pages_freed = 0;
>   	u32 start, end, block, total_blocks_freed = 0;
> +	struct address_space *mapping;
>   
>   	if (atomic_read(&global_page_count) <= tcmu_global_max_pages)
>   		return;
> @@ -3139,6 +3141,8 @@ static void find_free_blocks(void)
>   			continue;
>   		}
>   
> +		mapping = udev->inode->i_mapping;
> +		filemap_invalidate_lock(mapping);
>   		end = udev->dbi_max + 1;
>   		block = find_last_bit(udev->data_bitmap, end);
>   		if (block == udev->dbi_max) {
> @@ -3146,6 +3150,7 @@ static void find_free_blocks(void)
>   			 * The last bit is dbi_max, so it is not possible
>   			 * reclaim any blocks.
>   			 */
> +			filemap_invalidate_unlock(mapping);
>   			mutex_unlock(&udev->cmdr_lock);
>   			continue;
>   		} else if (block == end) {
> @@ -3159,10 +3164,11 @@ static void find_free_blocks(void)
>   
>   		/* Here will truncate the data area from off */
>   		off = udev->data_off + (loff_t)start * udev->data_blk_size;
> -		unmap_mapping_range(udev->inode->i_mapping, off, 0, 1);
> +		unmap_mapping_range(mapping, off, 0, 1);
>   
>   		/* Release the block pages */
>   		pages_freed = tcmu_blocks_release(udev, start, end - 1);
> +		filemap_invalidate_unlock(mapping);
>   		mutex_unlock(&udev->cmdr_lock);
>   
>   		total_pages_freed += pages_freed;