On 2020-05-29 04:53, Dan Carpenter wrote: ...
What are the runtime implications of this patch? I'm still not clear on that honestly.
Instead of incrementing each page's refcount by 1 (with get_user_pages()), pin_user_pages*() will increment by GUP_PIN_COUNTING_BIAS, which is 1024. That by itself should not have any performance impact, of course, but there's a couple more things: For compound pages of more than 2 page size, it will also increment a separate struct page's field, via hpage_pincount_add(). And finally, it will update /proc/vmstat counters on pin and unpin, via the optimized mod_node_page_state() call. So it's expected to be very light. And, for DMA (as opposed to DIO) situations, the DMA setup time is inevitably much greater than any of the above overheads, so I expect that this patch will be completely invisible from a performance point of view. It would be a "nice to have", though, if anyone were able to do a performance comparison on the gasket driver for this patch, and/or basic runtime verification, since I'm sure it's a specialized setup. thanks, -- John Hubbard NVIDIA _______________________________________________ devel mailing list devel@xxxxxxxxxxxxxxxxxxxxxx http://driverdev.linuxdriverproject.org/mailman/listinfo/driverdev-devel
