| Message ID | 20221223092703.61927-2-hch@lst.de (mailing list archive) |
|---|---|
| State | New |
| Headers | show |
| Series | [1/2] Revert "remoteproc: qcom_q6v5_mss: map/unmap metadata region before/after use" | expand |
On Fri, Dec 23, 2022 at 10:27:02AM +0100, Christoph Hellwig wrote: > This reverts commit fc156629b23a21181e473e60341e3a78af25a1d4. > > This commit manages to do three API violations at once: > > - dereference the return value of dma_alloc_attrs with the > DMA_ATTR_NO_KERNEL_MAPPING mapping, which is clearly forbidden and > will do the wrong thing on various dma mapping implementations. The > fact that dma-direct uses a struct page as a cookie is an undocumented > implementation detail > - include dma-map-ops.h and use pgprot_dmacoherent despite a clear > comment documenting that this is not acceptable > - use of the VM_DMA_COHERENT for something that is not the dma-mapping > code > - use of VM_FLUSH_RESET_PERMS for vmap, while it is only supported for > vmalloc > > Signed-off-by: Christoph Hellwig <hch@lst.de> Acked-by: Manivannan Sadhasivam <manivannan.sadhasivam@linaro.org> Thanks, Mani > --- > drivers/remoteproc/qcom_q6v5_mss.c | 38 +++++------------------------- > 1 file changed, 6 insertions(+), 32 deletions(-) > > diff --git a/drivers/remoteproc/qcom_q6v5_mss.c b/drivers/remoteproc/qcom_q6v5_mss.c > index fddb63cffee078..a8b141db4de63f 100644 > --- a/drivers/remoteproc/qcom_q6v5_mss.c > +++ b/drivers/remoteproc/qcom_q6v5_mss.c > @@ -10,7 +10,6 @@ > #include <linux/clk.h> > #include <linux/delay.h> > #include <linux/devcoredump.h> > -#include <linux/dma-map-ops.h> > #include <linux/dma-mapping.h> > #include <linux/interrupt.h> > #include <linux/kernel.h> > @@ -933,52 +932,27 @@ static void q6v5proc_halt_axi_port(struct q6v5 *qproc, > static int q6v5_mpss_init_image(struct q6v5 *qproc, const struct firmware *fw, > const char *fw_name) > { > - unsigned long dma_attrs = DMA_ATTR_FORCE_CONTIGUOUS | DMA_ATTR_NO_KERNEL_MAPPING; > - unsigned long flags = VM_DMA_COHERENT | VM_FLUSH_RESET_PERMS; > - struct page **pages; > - struct page *page; > + unsigned long dma_attrs = DMA_ATTR_FORCE_CONTIGUOUS; > dma_addr_t phys; > void *metadata; > int mdata_perm; > int xferop_ret; > size_t size; > - void *vaddr; > - int count; > + void *ptr; > int ret; > - int i; > > metadata = qcom_mdt_read_metadata(fw, &size, fw_name, qproc->dev); > if (IS_ERR(metadata)) > return PTR_ERR(metadata); > > - page = dma_alloc_attrs(qproc->dev, size, &phys, GFP_KERNEL, dma_attrs); > - if (!page) { > + ptr = dma_alloc_attrs(qproc->dev, size, &phys, GFP_KERNEL, dma_attrs); > + if (!ptr) { > kfree(metadata); > dev_err(qproc->dev, "failed to allocate mdt buffer\n"); > return -ENOMEM; > } > > - count = PAGE_ALIGN(size) >> PAGE_SHIFT; > - pages = kmalloc_array(count, sizeof(struct page *), GFP_KERNEL); > - if (!pages) { > - ret = -ENOMEM; > - goto free_dma_attrs; > - } > - > - for (i = 0; i < count; i++) > - pages[i] = nth_page(page, i); > - > - vaddr = vmap(pages, count, flags, pgprot_dmacoherent(PAGE_KERNEL)); > - kfree(pages); > - if (!vaddr) { > - dev_err(qproc->dev, "unable to map memory region: %pa+%zx\n", &phys, size); > - ret = -EBUSY; > - goto free_dma_attrs; > - } > - > - memcpy(vaddr, metadata, size); > - > - vunmap(vaddr); > + memcpy(ptr, metadata, size); > > /* Hypervisor mapping to access metadata by modem */ > mdata_perm = BIT(QCOM_SCM_VMID_HLOS); > @@ -1008,7 +982,7 @@ static int q6v5_mpss_init_image(struct q6v5 *qproc, const struct firmware *fw, > "mdt buffer not reclaimed system may become unstable\n"); > > free_dma_attrs: > - dma_free_attrs(qproc->dev, size, page, phys, dma_attrs); > + dma_free_attrs(qproc->dev, size, ptr, phys, dma_attrs); > kfree(metadata); > > return ret < 0 ? ret : 0; > -- > 2.35.1 >
On Fri, Dec 23, 2022 at 08:17:31PM +0530, Manivannan Sadhasivam wrote:
> Acked-by: Manivannan Sadhasivam <manivannan.sadhasivam@linaro.org>
Btw, if the hardware really does not like a kernel mapping, the
right way is to just keep using the normal dma allocator, but make
sure that there shared-dma-pool with the no-map property for the
device.
On Fri, Dec 23, 2022 at 03:57:22PM +0100, Christoph Hellwig wrote: > On Fri, Dec 23, 2022 at 08:17:31PM +0530, Manivannan Sadhasivam wrote: > > Acked-by: Manivannan Sadhasivam <manivannan.sadhasivam@linaro.org> > > Btw, if the hardware really does not like a kernel mapping, the > right way is to just keep using the normal dma allocator, but make > sure that there shared-dma-pool with the no-map property for the > device. Sibi posted a series that uses a separate no-map carveout for this usecase: https://lore.kernel.org/lkml/20221213140724.8612-1-quic_sibis@quicinc.com/ But that doesn't use dma allocator with shared-dma-pool. Thanks, Mani
On Fri, Dec 23, 2022 at 09:10:39PM +0530, Manivannan Sadhasivam wrote: > On Fri, Dec 23, 2022 at 03:57:22PM +0100, Christoph Hellwig wrote: > > On Fri, Dec 23, 2022 at 08:17:31PM +0530, Manivannan Sadhasivam wrote: > > > Acked-by: Manivannan Sadhasivam <manivannan.sadhasivam@linaro.org> > > > > Btw, if the hardware really does not like a kernel mapping, the > > right way is to just keep using the normal dma allocator, but make > > sure that there shared-dma-pool with the no-map property for the > > device. > > Sibi posted a series that uses a separate no-map carveout for this usecase: > https://lore.kernel.org/lkml/20221213140724.8612-1-quic_sibis@quicinc.com/ Oh, I've missed that entire thread. I actually stumbled over this today while finding it during a vmap audit..
diff --git a/drivers/remoteproc/qcom_q6v5_mss.c b/drivers/remoteproc/qcom_q6v5_mss.c index fddb63cffee078..a8b141db4de63f 100644 --- a/drivers/remoteproc/qcom_q6v5_mss.c +++ b/drivers/remoteproc/qcom_q6v5_mss.c @@ -10,7 +10,6 @@ #include <linux/clk.h> #include <linux/delay.h> #include <linux/devcoredump.h> -#include <linux/dma-map-ops.h> #include <linux/dma-mapping.h> #include <linux/interrupt.h> #include <linux/kernel.h> @@ -933,52 +932,27 @@ static void q6v5proc_halt_axi_port(struct q6v5 *qproc, static int q6v5_mpss_init_image(struct q6v5 *qproc, const struct firmware *fw, const char *fw_name) { - unsigned long dma_attrs = DMA_ATTR_FORCE_CONTIGUOUS | DMA_ATTR_NO_KERNEL_MAPPING; - unsigned long flags = VM_DMA_COHERENT | VM_FLUSH_RESET_PERMS; - struct page **pages; - struct page *page; + unsigned long dma_attrs = DMA_ATTR_FORCE_CONTIGUOUS; dma_addr_t phys; void *metadata; int mdata_perm; int xferop_ret; size_t size; - void *vaddr; - int count; + void *ptr; int ret; - int i; metadata = qcom_mdt_read_metadata(fw, &size, fw_name, qproc->dev); if (IS_ERR(metadata)) return PTR_ERR(metadata); - page = dma_alloc_attrs(qproc->dev, size, &phys, GFP_KERNEL, dma_attrs); - if (!page) { + ptr = dma_alloc_attrs(qproc->dev, size, &phys, GFP_KERNEL, dma_attrs); + if (!ptr) { kfree(metadata); dev_err(qproc->dev, "failed to allocate mdt buffer\n"); return -ENOMEM; } - count = PAGE_ALIGN(size) >> PAGE_SHIFT; - pages = kmalloc_array(count, sizeof(struct page *), GFP_KERNEL); - if (!pages) { - ret = -ENOMEM; - goto free_dma_attrs; - } - - for (i = 0; i < count; i++) - pages[i] = nth_page(page, i); - - vaddr = vmap(pages, count, flags, pgprot_dmacoherent(PAGE_KERNEL)); - kfree(pages); - if (!vaddr) { - dev_err(qproc->dev, "unable to map memory region: %pa+%zx\n", &phys, size); - ret = -EBUSY; - goto free_dma_attrs; - } - - memcpy(vaddr, metadata, size); - - vunmap(vaddr); + memcpy(ptr, metadata, size); /* Hypervisor mapping to access metadata by modem */ mdata_perm = BIT(QCOM_SCM_VMID_HLOS); @@ -1008,7 +982,7 @@ static int q6v5_mpss_init_image(struct q6v5 *qproc, const struct firmware *fw, "mdt buffer not reclaimed system may become unstable\n"); free_dma_attrs: - dma_free_attrs(qproc->dev, size, page, phys, dma_attrs); + dma_free_attrs(qproc->dev, size, ptr, phys, dma_attrs); kfree(metadata); return ret < 0 ? ret : 0;
This reverts commit fc156629b23a21181e473e60341e3a78af25a1d4. This commit manages to do three API violations at once: - dereference the return value of dma_alloc_attrs with the DMA_ATTR_NO_KERNEL_MAPPING mapping, which is clearly forbidden and will do the wrong thing on various dma mapping implementations. The fact that dma-direct uses a struct page as a cookie is an undocumented implementation detail - include dma-map-ops.h and use pgprot_dmacoherent despite a clear comment documenting that this is not acceptable - use of the VM_DMA_COHERENT for something that is not the dma-mapping code - use of VM_FLUSH_RESET_PERMS for vmap, while it is only supported for vmalloc Signed-off-by: Christoph Hellwig <hch@lst.de> --- drivers/remoteproc/qcom_q6v5_mss.c | 38 +++++------------------------- 1 file changed, 6 insertions(+), 32 deletions(-)