| Message ID | 20220919180949.388785-2-anirudh.venkataramanan@intel.com |
|---|---|
| State | Superseded |
| Delegated to: | Anthony Nguyen |
| Headers | show |
| Series | [RESEND,net-next,1/2] e1000: Remove unnecessary use of kmap_atomic() | expand |
On Mon, Sep 19, 2022 at 11:09:49AM -0700, Venkataramanan, Anirudh wrote: > alloc_rx_buf() allocates ps_page->page and buffer_info->page using either > GFP_ATOMIC or GFP_KERNEL. Memory allocated with GFP_KERNEL/GFP_ATOMIC can't > come from highmem and so there's no need to kmap() them. Just use > page_address(). > > I don't have access to a 32-bit system so did some limited testing on qemu > (qemu-system-i386 -m 4096 -smp 4 -device e1000e) with a 32-bit Debian 11.04 > image. > > Cc: Ira Weiny <ira.weiny@intel.com> > Cc: Fabio M. De Francesco <fmdefrancesco@gmail.com> > Cc: Jesse Brandeburg <jesse.brandeburg@intel.com> > Cc: Tony Nguyen <anthony.l.nguyen@intel.com> > Suggested-by: Ira Weiny <ira.weiny@intel.com> > Suggested-by: Fabio M. De Francesco <fmdefrancesco@gmail.com> > Signed-off-by: Anirudh Venkataramanan <anirudh.venkataramanan@intel.com> > --- > drivers/net/ethernet/intel/e1000e/netdev.c | 17 ++++------------- > 1 file changed, 4 insertions(+), 13 deletions(-) > > diff --git a/drivers/net/ethernet/intel/e1000e/netdev.c b/drivers/net/ethernet/intel/e1000e/netdev.c > index 321f2a9..05a59e5 100644 > --- a/drivers/net/ethernet/intel/e1000e/netdev.c > +++ b/drivers/net/ethernet/intel/e1000e/netdev.c > @@ -1393,21 +1393,14 @@ static bool e1000_clean_rx_irq_ps(struct e1000_ring *rx_ring, int *work_done, > */ The comment just above here says this: 1389 /* page alloc/put takes too long and effects small 1390 * packet throughput, so unsplit small packets and 1391 * save the alloc/put only valid in softirq (napi) 1392 * context to call kmap_* 1393 */ I'm unable to grok what that means exactly but I feel like the kmap part is no longer appropriate? Maybe just delete ... 'to call kmap_*'? The code itself looks ok. With some clean up to that comment: Reviewed-by: Ira Weiny <ira.weiny@intel.com> Ira > if (l1 && (l1 <= copybreak) && > ((length + l1) <= adapter->rx_ps_bsize0)) { > - u8 *vaddr; > - > ps_page = &buffer_info->ps_pages[0]; > > - /* there is no documentation about how to call > - * kmap_atomic, so we can't hold the mapping > - * very long > - */ > dma_sync_single_for_cpu(&pdev->dev, > ps_page->dma, > PAGE_SIZE, > DMA_FROM_DEVICE); > - vaddr = kmap_atomic(ps_page->page); > - memcpy(skb_tail_pointer(skb), vaddr, l1); > - kunmap_atomic(vaddr); > + memcpy(skb_tail_pointer(skb), > + page_address(ps_page->page), l1); > dma_sync_single_for_device(&pdev->dev, > ps_page->dma, > PAGE_SIZE, > @@ -1607,11 +1600,9 @@ static bool e1000_clean_jumbo_rx_irq(struct e1000_ring *rx_ring, int *work_done, > */ > if (length <= copybreak && > skb_tailroom(skb) >= length) { > - u8 *vaddr; > - vaddr = kmap_atomic(buffer_info->page); > - memcpy(skb_tail_pointer(skb), vaddr, > + memcpy(skb_tail_pointer(skb), > + page_address(buffer_info->page), > length); > - kunmap_atomic(vaddr); > /* re-use the page, so don't erase > * buffer_info->page > */ > -- > 2.37.2 >
On 9/29/2022 3:40 PM, Ira Weiny wrote: > On Mon, Sep 19, 2022 at 11:09:49AM -0700, Venkataramanan, Anirudh wrote: >> alloc_rx_buf() allocates ps_page->page and buffer_info->page using either >> GFP_ATOMIC or GFP_KERNEL. Memory allocated with GFP_KERNEL/GFP_ATOMIC can't >> come from highmem and so there's no need to kmap() them. Just use >> page_address(). >> >> I don't have access to a 32-bit system so did some limited testing on qemu >> (qemu-system-i386 -m 4096 -smp 4 -device e1000e) with a 32-bit Debian 11.04 >> image. >> >> Cc: Ira Weiny <ira.weiny@intel.com> >> Cc: Fabio M. De Francesco <fmdefrancesco@gmail.com> >> Cc: Jesse Brandeburg <jesse.brandeburg@intel.com> >> Cc: Tony Nguyen <anthony.l.nguyen@intel.com> >> Suggested-by: Ira Weiny <ira.weiny@intel.com> >> Suggested-by: Fabio M. De Francesco <fmdefrancesco@gmail.com> >> Signed-off-by: Anirudh Venkataramanan <anirudh.venkataramanan@intel.com> >> --- >> drivers/net/ethernet/intel/e1000e/netdev.c | 17 ++++------------- >> 1 file changed, 4 insertions(+), 13 deletions(-) >> >> diff --git a/drivers/net/ethernet/intel/e1000e/netdev.c b/drivers/net/ethernet/intel/e1000e/netdev.c >> index 321f2a9..05a59e5 100644 >> --- a/drivers/net/ethernet/intel/e1000e/netdev.c >> +++ b/drivers/net/ethernet/intel/e1000e/netdev.c >> @@ -1393,21 +1393,14 @@ static bool e1000_clean_rx_irq_ps(struct e1000_ring *rx_ring, int *work_done, >> */ > > The comment just above here says this: > > 1389 /* page alloc/put takes too long and effects small > 1390 * packet throughput, so unsplit small packets and > 1391 * save the alloc/put only valid in softirq (napi) > 1392 * context to call kmap_* > 1393 */ I did see this when I was making the code change but it evaporated from my working memory. Thanks for catching it. > > I'm unable to grok what that means exactly but I feel like the kmap part is no > longer appropriate? > > Maybe just delete ... 'to call kmap_*'? Okay, so did some git archeology and found that the original comment just said this: /* page alloc/put takes too long and effects small packet * throughput, so unsplit small packets and save the alloc/put*/ ... which is what I think we need to revert to. Tony, do you want me to send a v2? Or is making an inline edit to the patch you've already applied to dev-tree less work for you? I am okay either way. Just tell me which one makes life easier for you. Thanks! Ani
On 9/30/2022 11:20 AM, Anirudh Venkataramanan wrote: > Tony, do you want me to send a v2? Or is making an inline edit to the > patch you've already applied to dev-tree less work for you? > > I am okay either way. Just tell me which one makes life easier for you. v2 would be great. Thanks, Tony
diff --git a/drivers/net/ethernet/intel/e1000e/netdev.c b/drivers/net/ethernet/intel/e1000e/netdev.c index 321f2a9..05a59e5 100644 --- a/drivers/net/ethernet/intel/e1000e/netdev.c +++ b/drivers/net/ethernet/intel/e1000e/netdev.c @@ -1393,21 +1393,14 @@ static bool e1000_clean_rx_irq_ps(struct e1000_ring *rx_ring, int *work_done, */ if (l1 && (l1 <= copybreak) && ((length + l1) <= adapter->rx_ps_bsize0)) { - u8 *vaddr; - ps_page = &buffer_info->ps_pages[0]; - /* there is no documentation about how to call - * kmap_atomic, so we can't hold the mapping - * very long - */ dma_sync_single_for_cpu(&pdev->dev, ps_page->dma, PAGE_SIZE, DMA_FROM_DEVICE); - vaddr = kmap_atomic(ps_page->page); - memcpy(skb_tail_pointer(skb), vaddr, l1); - kunmap_atomic(vaddr); + memcpy(skb_tail_pointer(skb), + page_address(ps_page->page), l1); dma_sync_single_for_device(&pdev->dev, ps_page->dma, PAGE_SIZE, @@ -1607,11 +1600,9 @@ static bool e1000_clean_jumbo_rx_irq(struct e1000_ring *rx_ring, int *work_done, */ if (length <= copybreak && skb_tailroom(skb) >= length) { - u8 *vaddr; - vaddr = kmap_atomic(buffer_info->page); - memcpy(skb_tail_pointer(skb), vaddr, + memcpy(skb_tail_pointer(skb), + page_address(buffer_info->page), length); - kunmap_atomic(vaddr); /* re-use the page, so don't erase * buffer_info->page */
alloc_rx_buf() allocates ps_page->page and buffer_info->page using either GFP_ATOMIC or GFP_KERNEL. Memory allocated with GFP_KERNEL/GFP_ATOMIC can't come from highmem and so there's no need to kmap() them. Just use page_address(). I don't have access to a 32-bit system so did some limited testing on qemu (qemu-system-i386 -m 4096 -smp 4 -device e1000e) with a 32-bit Debian 11.04 image. Cc: Ira Weiny <ira.weiny@intel.com> Cc: Fabio M. De Francesco <fmdefrancesco@gmail.com> Cc: Jesse Brandeburg <jesse.brandeburg@intel.com> Cc: Tony Nguyen <anthony.l.nguyen@intel.com> Suggested-by: Ira Weiny <ira.weiny@intel.com> Suggested-by: Fabio M. De Francesco <fmdefrancesco@gmail.com> Signed-off-by: Anirudh Venkataramanan <anirudh.venkataramanan@intel.com> --- drivers/net/ethernet/intel/e1000e/netdev.c | 17 ++++------------- 1 file changed, 4 insertions(+), 13 deletions(-)