| Message ID | 20230217-kmap-qib-v1-1-e5a6fde167e0@intel.com (mailing list archive) |
|---|---|
| State | Accepted |
| Headers | show |
| Series | rdma/qib: Remove deprecated kmap() call | expand |
On 3/7/23 12:15 PM, Ira Weiny wrote: > kmap() has been deprecated in favor of the kmap_local_page() call. > kmap_local_page() is thread local. > > In the sdma coalesce case the page allocated is potentially free'ed in a > different context through qib_sdma_get_complete() -> > qib_user_sdma_make_progress(). The use of kmap_local_page() is > inappropriate in this call path. However, the page is allocated using > GFP_KERNEL and will never be from highmem. > > Remove the use of kmap calls and use page_address() in this case. > > Signed-off-by: Ira Weiny <ira.weiny@intel.com> Acked-by: Dennis Dalessandro <dennis.dalessandro@cornelisnetworks.com>
On Tue, 07 Mar 2023 09:15:31 -0800, Ira Weiny wrote: > kmap() has been deprecated in favor of the kmap_local_page() call. > kmap_local_page() is thread local. > > In the sdma coalesce case the page allocated is potentially free'ed in a > different context through qib_sdma_get_complete() -> > qib_user_sdma_make_progress(). The use of kmap_local_page() is > inappropriate in this call path. However, the page is allocated using > GFP_KERNEL and will never be from highmem. > > [...] Applied, thanks! [1/1] rdma/qib: Remove deprecated kmap() call https://git.kernel.org/rdma/rdma/c/4f00848e490c59 Best regards,
diff --git a/drivers/infiniband/hw/qib/qib_user_sdma.c b/drivers/infiniband/hw/qib/qib_user_sdma.c index 9fe03d6ffac1..336eb15a721f 100644 --- a/drivers/infiniband/hw/qib/qib_user_sdma.c +++ b/drivers/infiniband/hw/qib/qib_user_sdma.c @@ -320,7 +320,6 @@ static int qib_user_sdma_page_to_frags(const struct qib_devdata *dd, unpin_user_page(page); } else { /* coalesce case */ - kunmap(page); __free_page(page); } ret = -ENOMEM; @@ -572,7 +571,7 @@ static int qib_user_sdma_coalesce(const struct qib_devdata *dd, goto done; } - mpage = kmap(page); + mpage = page_address(page); mpage_save = mpage; for (i = 0; i < niov; i++) { int cfur; @@ -581,7 +580,7 @@ static int qib_user_sdma_coalesce(const struct qib_devdata *dd, iov[i].iov_base, iov[i].iov_len); if (cfur) { ret = -EFAULT; - goto free_unmap; + goto page_free; } mpage += iov[i].iov_len; @@ -592,8 +591,7 @@ static int qib_user_sdma_coalesce(const struct qib_devdata *dd, page, 0, 0, len, mpage_save); goto done; -free_unmap: - kunmap(page); +page_free: __free_page(page); done: return ret; @@ -627,9 +625,6 @@ static void qib_user_sdma_free_pkt_frag(struct device *dev, pkt->addr[i].dma_length, DMA_TO_DEVICE); - if (pkt->addr[i].kvaddr) - kunmap(pkt->addr[i].page); - if (pkt->addr[i].put_page) unpin_user_page(pkt->addr[i].page); else
kmap() has been deprecated in favor of the kmap_local_page() call. kmap_local_page() is thread local. In the sdma coalesce case the page allocated is potentially free'ed in a different context through qib_sdma_get_complete() -> qib_user_sdma_make_progress(). The use of kmap_local_page() is inappropriate in this call path. However, the page is allocated using GFP_KERNEL and will never be from highmem. Remove the use of kmap calls and use page_address() in this case. Signed-off-by: Ira Weiny <ira.weiny@intel.com> --- drivers/infiniband/hw/qib/qib_user_sdma.c | 11 +++-------- 1 file changed, 3 insertions(+), 8 deletions(-) --- base-commit: 8ca09d5fa3549d142c2080a72a4c70ce389163cd change-id: 20230217-kmap-qib-27a563231f85 Best regards,