Message ID | 20250207-bam-read-fix-v1-1-027975cf1a04@oss.qualcomm.com |
---|---|
State | New |
Headers | show |
Series | dmaengine: qcom: bam_dma: Avoid accessing BAM_REVISION on remote BAM | expand |
Hello Bjorn, On Fri Feb 7, 2025 at 9:17 PM CET, Bjorn Andersson wrote: > Commit '57a7138d0627 ("dmaengine: qcom: bam_dma: Avoid writing > unavailable register")' made this read unconditional, in order to > identify if the instance is BAM-NDP or BAM-Lite. > But the BAM_REVISION register is not accessible on remotely managed BAM > instances and attempts to access it causes the system to crash. > > Move the access back to be conditional and expand the checks that was > introduced to restore the old behavior when no revision information is > available. > > Fixes: 57a7138d0627 ("dmaengine: qcom: bam_dma: Avoid writing unavailable register") > Reported-by: Georgi Djakov <djakov@kernel.org> > Closes: https://lore.kernel.org/lkml/9ef3daa8-cdb1-49f2-8d19-a72d6210ff3a@kernel.org/ > Signed-off-by: Bjorn Andersson <bjorn.andersson@oss.qualcomm.com> > --- > drivers/dma/qcom/bam_dma.c | 15 ++++++++------- > 1 file changed, 8 insertions(+), 7 deletions(-) > > diff --git a/drivers/dma/qcom/bam_dma.c b/drivers/dma/qcom/bam_dma.c > index c14557efd577..d42d913492a8 100644 > --- a/drivers/dma/qcom/bam_dma.c > +++ b/drivers/dma/qcom/bam_dma.c > @@ -445,8 +445,8 @@ static void bam_reset(struct bam_device *bdev) > writel_relaxed(val, bam_addr(bdev, 0, BAM_CTRL)); > > /* set descriptor threshold, start with 4 bytes */ > - if (in_range(bdev->bam_revision, BAM_NDP_REVISION_START, > - BAM_NDP_REVISION_END)) > + if (!bdev->bam_revision || > + in_range(bdev->bam_revision, BAM_NDP_REVISION_START, BAM_NDP_REVISION_END)) > writel_relaxed(DEFAULT_CNT_THRSHLD, > bam_addr(bdev, 0, BAM_DESC_CNT_TRSHLD)); > > @@ -1006,8 +1006,8 @@ static void bam_apply_new_config(struct bam_chan *bchan, > maxburst = bchan->slave.src_maxburst; > else > maxburst = bchan->slave.dst_maxburst; > - if (in_range(bdev->bam_revision, BAM_NDP_REVISION_START, > - BAM_NDP_REVISION_END)) > + if (!bdev->bam_revision || > + in_range(bdev->bam_revision, BAM_NDP_REVISION_START, BAM_NDP_REVISION_END)) > writel_relaxed(maxburst, > bam_addr(bdev, 0, BAM_DESC_CNT_TRSHLD)); > } > @@ -1199,11 +1199,12 @@ static int bam_init(struct bam_device *bdev) > u32 val; > > /* read revision and configuration information */ > - val = readl_relaxed(bam_addr(bdev, 0, BAM_REVISION)); > - if (!bdev->num_ees) > + if (!bdev->num_ees) { > + val = readl_relaxed(bam_addr(bdev, 0, BAM_REVISION)); > bdev->num_ees = (val >> NUM_EES_SHIFT) & NUM_EES_MASK; > > - bdev->bam_revision = val & REVISION_MASK; > + bdev->bam_revision = val & REVISION_MASK; > + } > > /* check that configured EE is within range */ > if (bdev->ee >= bdev->num_ees) > > --- > base-commit: 2014c95afecee3e76ca4a56956a936e23283f05b > change-id: 20250207-bam-read-fix-2b31297d3fa1 > > Best regards, Thank you, this fixes the early synchronous external abort in bam_dma_probe on my msm8916 device. Tested-by: Sam Day <me@samcday.com> Regards, -Sam
On 7.02.25 22:17, Bjorn Andersson wrote: > Commit '57a7138d0627 ("dmaengine: qcom: bam_dma: Avoid writing > unavailable register")' made this read unconditional, in order to > identify if the instance is BAM-NDP or BAM-Lite. > But the BAM_REVISION register is not accessible on remotely managed BAM > instances and attempts to access it causes the system to crash. > > Move the access back to be conditional and expand the checks that was > introduced to restore the old behavior when no revision information is > available. > > Fixes: 57a7138d0627 ("dmaengine: qcom: bam_dma: Avoid writing unavailable register") > Reported-by: Georgi Djakov <djakov@kernel.org> > Closes: https://lore.kernel.org/lkml/9ef3daa8-cdb1-49f2-8d19-a72d6210ff3a@kernel.org/ > Signed-off-by: Bjorn Andersson <bjorn.andersson@oss.qualcomm.com> Tested-by: Georgi Djakov <djakov@kernel.org> # db845c Thanks, Georgi
On Fri, Feb 07, 2025 at 12:17:33PM -0800, Bjorn Andersson wrote: > Commit '57a7138d0627 ("dmaengine: qcom: bam_dma: Avoid writing > unavailable register")' made this read unconditional, in order to > identify if the instance is BAM-NDP or BAM-Lite. > But the BAM_REVISION register is not accessible on remotely managed BAM > instances and attempts to access it causes the system to crash. > > Move the access back to be conditional and expand the checks that was > introduced to restore the old behavior when no revision information is > available. > > Fixes: 57a7138d0627 ("dmaengine: qcom: bam_dma: Avoid writing unavailable register") > Reported-by: Georgi Djakov <djakov@kernel.org> > Closes: https://lore.kernel.org/lkml/9ef3daa8-cdb1-49f2-8d19-a72d6210ff3a@kernel.org/ > Signed-off-by: Bjorn Andersson <bjorn.andersson@oss.qualcomm.com> This patch fixes the most critical regression (the bus hang), but the in_range(..., BAM_NDP) checks are also wrong. They do not consider the plain "BAM" type where the register is apparently also available. The check should be !in_range(..., BAM_LITE) instead to fix this. I mentioned this twice to Md Sadre Alam [1, 2], but a fix was unfortunately never sent for that part of the regression. I think we should take Caleb's patch and revert the entire patch for the 6.14 cycle. There are several incorrect assumptions in the original patch, it will be easier to review a fixed version with the full diff, rather than looking at incremental fixups. On a somewhat related note, I'm working on a patch series for bam_dma to clean up the handling of remotely controlled BAMs. It will make it more clear when it's safe to access BAM registers and when not, and should allow reading the revision also for remotely controlled BAMs. This would avoid the need for all these if (!bdev->bam_revision) checks. Thanks, Stephan [1]: https://lore.kernel.org/linux-arm-msm/Z4D2jQNNW94qGIlv@linaro.org/ [2]: https://lore.kernel.org/linux-arm-msm/Z4_U19_QyH2RJvKW@linaro.org/ > --- > drivers/dma/qcom/bam_dma.c | 15 ++++++++------- > 1 file changed, 8 insertions(+), 7 deletions(-) > > diff --git a/drivers/dma/qcom/bam_dma.c b/drivers/dma/qcom/bam_dma.c > index c14557efd577..d42d913492a8 100644 > --- a/drivers/dma/qcom/bam_dma.c > +++ b/drivers/dma/qcom/bam_dma.c > @@ -445,8 +445,8 @@ static void bam_reset(struct bam_device *bdev) > writel_relaxed(val, bam_addr(bdev, 0, BAM_CTRL)); > > /* set descriptor threshold, start with 4 bytes */ > - if (in_range(bdev->bam_revision, BAM_NDP_REVISION_START, > - BAM_NDP_REVISION_END)) > + if (!bdev->bam_revision || > + in_range(bdev->bam_revision, BAM_NDP_REVISION_START, BAM_NDP_REVISION_END)) > writel_relaxed(DEFAULT_CNT_THRSHLD, > bam_addr(bdev, 0, BAM_DESC_CNT_TRSHLD)); > > @@ -1006,8 +1006,8 @@ static void bam_apply_new_config(struct bam_chan *bchan, > maxburst = bchan->slave.src_maxburst; > else > maxburst = bchan->slave.dst_maxburst; > - if (in_range(bdev->bam_revision, BAM_NDP_REVISION_START, > - BAM_NDP_REVISION_END)) > + if (!bdev->bam_revision || > + in_range(bdev->bam_revision, BAM_NDP_REVISION_START, BAM_NDP_REVISION_END)) > writel_relaxed(maxburst, > bam_addr(bdev, 0, BAM_DESC_CNT_TRSHLD)); > } > @@ -1199,11 +1199,12 @@ static int bam_init(struct bam_device *bdev) > u32 val; > > /* read revision and configuration information */ > - val = readl_relaxed(bam_addr(bdev, 0, BAM_REVISION)); > - if (!bdev->num_ees) > + if (!bdev->num_ees) { > + val = readl_relaxed(bam_addr(bdev, 0, BAM_REVISION)); > bdev->num_ees = (val >> NUM_EES_SHIFT) & NUM_EES_MASK; > > - bdev->bam_revision = val & REVISION_MASK; > + bdev->bam_revision = val & REVISION_MASK; > + } > > /* check that configured EE is within range */ > if (bdev->ee >= bdev->num_ees) > > --- > base-commit: 2014c95afecee3e76ca4a56956a936e23283f05b > change-id: 20250207-bam-read-fix-2b31297d3fa1 > > Best regards, > -- > Bjorn Andersson <bjorn.andersson@oss.qualcomm.com> >
On 2/10/2025 2:14 PM, Stephan Gerhold wrote: > On Fri, Feb 07, 2025 at 12:17:33PM -0800, Bjorn Andersson wrote: >> Commit '57a7138d0627 ("dmaengine: qcom: bam_dma: Avoid writing >> unavailable register")' made this read unconditional, in order to >> identify if the instance is BAM-NDP or BAM-Lite. >> But the BAM_REVISION register is not accessible on remotely managed BAM >> instances and attempts to access it causes the system to crash. >> >> Move the access back to be conditional and expand the checks that was >> introduced to restore the old behavior when no revision information is >> available. >> >> Fixes: 57a7138d0627 ("dmaengine: qcom: bam_dma: Avoid writing unavailable register") >> Reported-by: Georgi Djakov <djakov@kernel.org> >> Closes: https://lore.kernel.org/lkml/9ef3daa8-cdb1-49f2-8d19-a72d6210ff3a@kernel.org/ >> Signed-off-by: Bjorn Andersson <bjorn.andersson@oss.qualcomm.com> > > This patch fixes the most critical regression (the bus hang), but the > in_range(..., BAM_NDP) checks are also wrong. They do not consider the > plain "BAM" type where the register is apparently also available. The > check should be !in_range(..., BAM_LITE) instead to fix this. > > I mentioned this twice to Md Sadre Alam [1, 2], but a fix was > unfortunately never sent for that part of the regression. I apologize for the delay. I was attending to a family member's medical emergency and couldn't address this sooner. I will test and post a new revision as soon as possible. Thanks Alam > > I think we should take Caleb's patch and revert the entire patch for the > 6.14 cycle. There are several incorrect assumptions in the original > patch, it will be easier to review a fixed version with the full diff, > rather than looking at incremental fixups. > > On a somewhat related note, I'm working on a patch series for bam_dma to > clean up the handling of remotely controlled BAMs. It will make it more > clear when it's safe to access BAM registers and when not, and should > allow reading the revision also for remotely controlled BAMs. This would > avoid the need for all these if (!bdev->bam_revision) checks. > > Thanks, > Stephan > > [1]: https://lore.kernel.org/linux-arm-msm/Z4D2jQNNW94qGIlv@linaro.org/ > [2]: https://lore.kernel.org/linux-arm-msm/Z4_U19_QyH2RJvKW@linaro.org/ > >> --- >> drivers/dma/qcom/bam_dma.c | 15 ++++++++------- >> 1 file changed, 8 insertions(+), 7 deletions(-) >> >> diff --git a/drivers/dma/qcom/bam_dma.c b/drivers/dma/qcom/bam_dma.c >> index c14557efd577..d42d913492a8 100644 >> --- a/drivers/dma/qcom/bam_dma.c >> +++ b/drivers/dma/qcom/bam_dma.c >> @@ -445,8 +445,8 @@ static void bam_reset(struct bam_device *bdev) >> writel_relaxed(val, bam_addr(bdev, 0, BAM_CTRL)); >> >> /* set descriptor threshold, start with 4 bytes */ >> - if (in_range(bdev->bam_revision, BAM_NDP_REVISION_START, >> - BAM_NDP_REVISION_END)) >> + if (!bdev->bam_revision || >> + in_range(bdev->bam_revision, BAM_NDP_REVISION_START, BAM_NDP_REVISION_END)) >> writel_relaxed(DEFAULT_CNT_THRSHLD, >> bam_addr(bdev, 0, BAM_DESC_CNT_TRSHLD)); >> >> @@ -1006,8 +1006,8 @@ static void bam_apply_new_config(struct bam_chan *bchan, >> maxburst = bchan->slave.src_maxburst; >> else >> maxburst = bchan->slave.dst_maxburst; >> - if (in_range(bdev->bam_revision, BAM_NDP_REVISION_START, >> - BAM_NDP_REVISION_END)) >> + if (!bdev->bam_revision || >> + in_range(bdev->bam_revision, BAM_NDP_REVISION_START, BAM_NDP_REVISION_END)) >> writel_relaxed(maxburst, >> bam_addr(bdev, 0, BAM_DESC_CNT_TRSHLD)); >> } >> @@ -1199,11 +1199,12 @@ static int bam_init(struct bam_device *bdev) >> u32 val; >> >> /* read revision and configuration information */ >> - val = readl_relaxed(bam_addr(bdev, 0, BAM_REVISION)); >> - if (!bdev->num_ees) >> + if (!bdev->num_ees) { >> + val = readl_relaxed(bam_addr(bdev, 0, BAM_REVISION)); >> bdev->num_ees = (val >> NUM_EES_SHIFT) & NUM_EES_MASK; >> >> - bdev->bam_revision = val & REVISION_MASK; >> + bdev->bam_revision = val & REVISION_MASK; >> + } >> >> /* check that configured EE is within range */ >> if (bdev->ee >= bdev->num_ees) >> >> --- >> base-commit: 2014c95afecee3e76ca4a56956a936e23283f05b >> change-id: 20250207-bam-read-fix-2b31297d3fa1 >> >> Best regards, >> -- >> Bjorn Andersson <bjorn.andersson@oss.qualcomm.com> >>
diff --git a/drivers/dma/qcom/bam_dma.c b/drivers/dma/qcom/bam_dma.c index c14557efd577..d42d913492a8 100644 --- a/drivers/dma/qcom/bam_dma.c +++ b/drivers/dma/qcom/bam_dma.c @@ -445,8 +445,8 @@ static void bam_reset(struct bam_device *bdev) writel_relaxed(val, bam_addr(bdev, 0, BAM_CTRL)); /* set descriptor threshold, start with 4 bytes */ - if (in_range(bdev->bam_revision, BAM_NDP_REVISION_START, - BAM_NDP_REVISION_END)) + if (!bdev->bam_revision || + in_range(bdev->bam_revision, BAM_NDP_REVISION_START, BAM_NDP_REVISION_END)) writel_relaxed(DEFAULT_CNT_THRSHLD, bam_addr(bdev, 0, BAM_DESC_CNT_TRSHLD)); @@ -1006,8 +1006,8 @@ static void bam_apply_new_config(struct bam_chan *bchan, maxburst = bchan->slave.src_maxburst; else maxburst = bchan->slave.dst_maxburst; - if (in_range(bdev->bam_revision, BAM_NDP_REVISION_START, - BAM_NDP_REVISION_END)) + if (!bdev->bam_revision || + in_range(bdev->bam_revision, BAM_NDP_REVISION_START, BAM_NDP_REVISION_END)) writel_relaxed(maxburst, bam_addr(bdev, 0, BAM_DESC_CNT_TRSHLD)); } @@ -1199,11 +1199,12 @@ static int bam_init(struct bam_device *bdev) u32 val; /* read revision and configuration information */ - val = readl_relaxed(bam_addr(bdev, 0, BAM_REVISION)); - if (!bdev->num_ees) + if (!bdev->num_ees) { + val = readl_relaxed(bam_addr(bdev, 0, BAM_REVISION)); bdev->num_ees = (val >> NUM_EES_SHIFT) & NUM_EES_MASK; - bdev->bam_revision = val & REVISION_MASK; + bdev->bam_revision = val & REVISION_MASK; + } /* check that configured EE is within range */ if (bdev->ee >= bdev->num_ees)
Commit '57a7138d0627 ("dmaengine: qcom: bam_dma: Avoid writing unavailable register")' made this read unconditional, in order to identify if the instance is BAM-NDP or BAM-Lite. But the BAM_REVISION register is not accessible on remotely managed BAM instances and attempts to access it causes the system to crash. Move the access back to be conditional and expand the checks that was introduced to restore the old behavior when no revision information is available. Fixes: 57a7138d0627 ("dmaengine: qcom: bam_dma: Avoid writing unavailable register") Reported-by: Georgi Djakov <djakov@kernel.org> Closes: https://lore.kernel.org/lkml/9ef3daa8-cdb1-49f2-8d19-a72d6210ff3a@kernel.org/ Signed-off-by: Bjorn Andersson <bjorn.andersson@oss.qualcomm.com> --- drivers/dma/qcom/bam_dma.c | 15 ++++++++------- 1 file changed, 8 insertions(+), 7 deletions(-) --- base-commit: 2014c95afecee3e76ca4a56956a936e23283f05b change-id: 20250207-bam-read-fix-2b31297d3fa1 Best regards,