Each IOVA link call results in a page walk, which is relatively expensive. Reduce the number of IOVA link calls by collapsing contiguous physical pages into a single IOVA link operation.
Signed-off-by: Matthew Brost <[email protected]> --- drivers/gpu/drm/drm_pagemap.c | 49 ++++++++++++++++++++++++++--------- 1 file changed, 37 insertions(+), 12 deletions(-) diff --git a/drivers/gpu/drm/drm_pagemap.c b/drivers/gpu/drm/drm_pagemap.c index b928c89f4bd1..a167064ce41a 100644 --- a/drivers/gpu/drm/drm_pagemap.c +++ b/drivers/gpu/drm/drm_pagemap.c @@ -305,7 +305,8 @@ drm_pagemap_migrate_map_system_pages(struct device *dev, struct dma_iova_state *state) { struct page *dummy_page = NULL; - unsigned long i, psize; + unsigned long i, j, psize; + phys_addr_t phys = -1; bool try_alloc = false; for (i = 0; i < npages;) { @@ -339,26 +340,50 @@ drm_pagemap_migrate_map_system_pages(struct device *dev, if (dma_use_iova(state)) { bool found_dummy = page && !dummy_page; + bool phys_match; int err; - if (found_dummy) { - unsigned long j; + if (found_dummy && i) { + err = dma_iova_link(dev, state, + page_to_phys(page), + 0, i * PAGE_SIZE, + dir, 0); + if (err) + return err; + } + + if (phys == -1) { + phys = page_to_phys(page); + j = i; + phys_match = true; + } else { + phys_match = phys + (i - j) * PAGE_SIZE == + page_to_phys(page); + } - for (j = 0; j < i; ++j) { + if (psize != PAGE_SIZE || !phys_match || + (i + 1) == npages) { + err = dma_iova_link(dev, state, phys, + j * PAGE_SIZE, + psize * ((i - j) + + phys_match), + dir, 0); + if (err) + return err; + + if (!phys_match && (i + 1) == npages) { err = dma_iova_link(dev, state, page_to_phys(page), - j * PAGE_SIZE, - PAGE_SIZE, dir, 0); + i * PAGE_SIZE, + psize, + dir, 0); if (err) return err; } - } - err = dma_iova_link(dev, state, page_to_phys(page), - i * PAGE_SIZE, psize, - dir, 0); - if (err) - return err; + phys = page_to_phys(page); + j = i; + } if (page != dummy_page) dma_addr = state->addr + i * PAGE_SIZE; -- 2.34.1
