| Message ID | 20191113075335.31775-4-linus.walleij@linaro.org |
|---|---|
| State | New |
| Headers | show |
| Series | MMCI odd buffer alignment fixes | expand |
On Wed, Nov 13, 2019 at 8:53 AM Linus Walleij <linus.walleij@linaro.org> wrote: > + /* Pack the residue into a 32bit word */ > + for (i = 0; i < host->pio_write_residue_sz; i++) { > + val |= host->pio_write_residue[i]; > + val <<= 8; > + } > + /* Top up with new data */ > + for (i = 0; i < fill; i++) { > + val |= *ptr; > + val <<= 8; > + ptr++; > + remain--; > + } I'm worried that I might have gotten this wrong. iowrite32_rep() reads the data little-endian (native endianness) from memory does it not? Bytes [0 1 2 3] end up in the FIFO like [3 2 1 0]. So it will pack the first byte into the lowest 8 bits, second byte into bits 8-15 etc. So I should rewrite all the loops like this: for (i = 0; i < host->pio_write_residue_sz; i++) { val |= (host->pio_write_residue[i] << 24); val >>= 8; } So I shift the value down from the high bits instead of the other way around. This also gives a pretty plausible hint att what might be wrong with the DMA in non-divisible by 4. As suggested by Stephan in another context, I will try to set up my own test rig for this. Yours, Linus Walleij
On Wed, 13 Nov 2019 at 08:53, Linus Walleij <linus.walleij@linaro.org> wrote: > > The MMCI PIO write function contains a bug if using any buffers > with sglist miter contents that do not add upp to even 32bit > writes. The PIO write code currently does this: > > iowrite32_rep(base + MMCIFIFO, ptr, (count + 3) >> 2); > > This will make sure that a single buffer passed in gets written > into the FIFO even if it is odd, i.e. not evenly divisible > by 4. > > However it has the following problems: > > - It will read up to three bytes beyond the end of the buffer > and fill the FIFO with unpredictable junk and possibly cause > segmentation violations if the read passes a page boundary > such as at the end of an sglist buffer. > > - It will be fine if a single buffer is passed in, but the code > handles SG lists. There is a while (1) loop in mmci_pio_irq() > which repeatedly calls mmci_pio_write() as long as the FIFO > is not half-full, if it gets half-full it exits the IRQ and > waits for a new IRQ, also the while loop repeatedly calls > sg_miter_next() to consume bytes and this means that between > subsequent calls to mmci_pio_write() some random junk can be > inserted at the end of each call if the buffers passed in > do not contain an number of bytes evenly divisible by 4. Thanks for a well description to the problem! > > Fix this by simply doing this: > > iowrite32_rep(base + MMCIFIFO, ptr, count >> 2); > > But since the code will then just not write enough bytes if > the count is not evenly divisible by 4, we introduce a special > residue buffer and keep track of any odd bytes from 1..3 and > just write these padded out with new data in a separate > statement the next time we get a call of the mmci_pio_write(), > or, if there is for example only 1,2 or 3 bytes in the transfer, > or we end up with an odd number of bytes in the residue, > flushi it out when we end the current data transaction to > the host. > > Cc: Ludovic Barre <ludovic.barre@st.com> > Cc: Brian Masney <masneyb@onstation.org> > Cc: Stephan Gerhold <stephan@gerhold.net> > Cc: Niklas Cassel <niklas.cassel@linaro.org> > Cc: Russell King <rmk+kernel@armlinux.org.uk> > Signed-off-by: Linus Walleij <linus.walleij@linaro.org> > --- > ChangeLog v1->v3: > - New patch in v3 after discussion with Ulf > - I'm unable to test this because I cannot provoke > these odd reads/writes but hoping for Stephan to take it > for a spin. > --- > drivers/mmc/host/mmci.c | 125 ++++++++++++++++++++++++++++++++++++---- > drivers/mmc/host/mmci.h | 2 + > 2 files changed, 117 insertions(+), 10 deletions(-) > > diff --git a/drivers/mmc/host/mmci.c b/drivers/mmc/host/mmci.c > index a08cd845dddc..0d01689016f0 100644 > --- a/drivers/mmc/host/mmci.c > +++ b/drivers/mmc/host/mmci.c > @@ -607,6 +607,7 @@ static void mmci_init_sg(struct mmci_host *host, struct mmc_data *data) > flags |= SG_MITER_FROM_SG; > > sg_miter_start(&host->sg_miter, data->sg, data->sg_len, flags); > + host->pio_write_residue_sz = 0; > } > > static u32 mmci_get_dctrl_cfg(struct mmci_host *host) > @@ -1422,30 +1423,111 @@ static int mmci_pio_write(struct mmci_host *host, char *buffer, unsigned int rem > struct variant_data *variant = host->variant; > void __iomem *base = host->base; > char *ptr = buffer; > + bool wrote_residue = false; > + int i; > + > + /* > + * This will normally not happen during block I/O, but can > + * happen during SDIO traffic, where odd byte chunks are > + * shoveled to the SDIO link. Fill up the residue buffer > + * and flush out. > + */ > + if (host->pio_write_residue_sz) { > + int fill = 4 - host->pio_write_residue_sz; > + u32 val = 0; > + > + /* > + * If we need more padding that what we have, just stash > + * some more in the residue then and continue. This only > + * happens if we're passed e.g. 1 or 2 bytes and there is > + * just 1 byte residue residue already: we can't fill up! > + */ > + if (fill > remain) { > + for (i = 0; i < fill; i++) { > + host->pio_write_residue[host->pio_write_residue_sz + i] = *ptr; > + host->pio_write_residue_sz++; > + ptr++; > + remain--; > + } > + return ptr - buffer; > + } > + > + /* Pack the residue into a 32bit word */ > + for (i = 0; i < host->pio_write_residue_sz; i++) { > + val |= host->pio_write_residue[i]; > + val <<= 8; > + } > + /* Top up with new data */ > + for (i = 0; i < fill; i++) { > + val |= *ptr; > + val <<= 8; > + ptr++; > + remain--; > + } > + iowrite32(val, base + MMCIFIFO); > + host->pio_write_residue_sz = 0; > + wrote_residue = true; > + } > + > + /* > + * Maybe the residue plus some few bytes was exactly filling a > + * 32bit word. > + */ > + if (!remain) > + return ptr - buffer; > > do { > - unsigned int count, maxcnt; > + unsigned int count, maxcnt, written, residue; > > + /* > + * If the FIFO is empty just stash it full with data else > + * just half-fill it. > + */ > maxcnt = status & MCI_TXFIFOEMPTY ? > - variant->fifosize : variant->fifohalfsize; > + variant->fifosize : variant->fifohalfsize; > + > + /* Watch it to not overfill the FIFO */ > + if (wrote_residue) > + maxcnt -= 4; > + > count = min(remain, maxcnt); > > /* > * SDIO especially may want to send something that is > * not divisible by 4 (as opposed to card sectors > - * etc), and the FIFO only accept full 32-bit writes. > - * So compensate by adding +3 on the count, a single > - * byte become a 32bit write, 7 bytes will be two > - * 32bit writes etc. > + * etc), but the FIFO only accepts full 32-bit writes so start by > + * just filling up quickly with as much as we can. > */ > - iowrite32_rep(base + MMCIFIFO, ptr, (count + 3) >> 2); > + iowrite32_rep(base + MMCIFIFO, ptr, count >> 2); > + residue = count & 3; > > - ptr += count; > - remain -= count; > + written = count - residue; > + ptr += written; > + remain -= written; > > - if (remain == 0) > + /* Handles the common case for block-IO */ > + if (!remain) > break; > > + /* > + * These were not written out since we only write 32bit > + * words to the FIFO. Since this function gets called > + * repeatedly and across page boundaries with new pointers > + * in *buffer we need to take special care to stash odd > + * bytes away and flush them out in next PIO round or at the > + * end of the whole data transfer. > + */ > + if (remain < 4) { > + WARN_ON(remain != residue); > + host->pio_write_residue_sz = residue; > + for (i = 0; i < residue; i++) { > + host->pio_write_residue[i] = *ptr; > + ptr++; > + remain--; > + } > + break; > + } > + > status = readl(base + MMCISTATUS); > } while (status & MCI_TXFIFOHALFEMPTY); > > @@ -1522,6 +1604,29 @@ static irqreturn_t mmci_pio_irq(int irq, void *dev_id) > if (host->size == 0) { > mmci_set_mask1(host, 0); > writel(readl(base + MMCIMASK0) | MCI_DATAENDMASK, base + MMCIMASK0); > + > + if ((status & MCI_TXACTIVE) && host->pio_write_residue_sz) { > + /* > + * If the last pass of the SG miter left some residue after the pio > + * write loop, push that out to the card. We are now at the end of > + * the transfer so it is OK to pad with zeroes. > + */ > + int fill = 4 - host->pio_write_residue_sz; > + u32 val = 0; > + int i; > + > + /* Pack the residue into a 32bit word */ > + for (i = 0; i < host->pio_write_residue_sz; i++) { > + val |= host->pio_write_residue[i]; > + val <<= 8; > + } > + /* Top up with zeroes */ > + for (i = 0; i < fill; i++) > + val <<= 8; > + > + iowrite32(val, base + MMCIFIFO); > + host->pio_write_residue_sz = 0; > + } Honestly, I am concerned about the above complex piece of code. And it also affects the hot path, etc, but please see my comment in the end around all this. However, I am open to apply something along the lines of this change, in case we can't come up with a better approach. > } > > return IRQ_HANDLED; > diff --git a/drivers/mmc/host/mmci.h b/drivers/mmc/host/mmci.h > index e20af17bb313..a7690b64d4cd 100644 > --- a/drivers/mmc/host/mmci.h > +++ b/drivers/mmc/host/mmci.h > @@ -425,6 +425,8 @@ struct mmci_host { > /* pio stuff */ > struct sg_mapping_iter sg_miter; > unsigned int size; > + u8 pio_write_residue[4]; > + u8 pio_write_residue_sz; > int (*get_rx_fifocnt)(struct mmci_host *h, u32 status, int remain); > > u8 use_dma:1; > -- > 2.21.0 > After reviewing this, I decided to also look deeper into the mmci_pio_read(). Unfortunately, it also doesn't cover all cases, which I believed it did. More precisely, mmc_pio_read() can't cope with a buffer being split across two pages, where the first element contains 2 bytes and the second contains 4 bytes (total size 6 bytes). The result is that we will end up filling the buffer with corrupt data. (While filling the first 2 bytes in the buffer, we read 4 bytes from the fifo, but throw away the last 2 bytes while copying to the buffer. Then when filling the last 4 bytes in the buffer, we read 4 bytes from the fifo, while only the first 2 bytes in fifo is actually valid data...) Now, why am I stating all this? Because I am wondering if it's really worth to support all kinds of different flavors of allocated buffers. It clearly introduces severe amount of complex code, that not only is hard to test, but also affects the "hot path" for "nicely" allocated buffers (whatever that means to performance). I guess what I am really questioning is, if it would be better to limit the driver to support only "nicely" allocated buffers. Of course, in the long run, that means we also need a way to inform upper layer drivers/code of what that means. I guess we need to introduce some new mmc host variables, along the lines of how host->max_seg_size and host->max_req_size is being used. What do you think, is it a terrible idea? :-) Kind regards Uffe
diff --git a/drivers/mmc/host/mmci.c b/drivers/mmc/host/mmci.c index a08cd845dddc..0d01689016f0 100644 --- a/drivers/mmc/host/mmci.c +++ b/drivers/mmc/host/mmci.c @@ -607,6 +607,7 @@ static void mmci_init_sg(struct mmci_host *host, struct mmc_data *data) flags |= SG_MITER_FROM_SG; sg_miter_start(&host->sg_miter, data->sg, data->sg_len, flags); + host->pio_write_residue_sz = 0; } static u32 mmci_get_dctrl_cfg(struct mmci_host *host) @@ -1422,30 +1423,111 @@ static int mmci_pio_write(struct mmci_host *host, char *buffer, unsigned int rem struct variant_data *variant = host->variant; void __iomem *base = host->base; char *ptr = buffer; + bool wrote_residue = false; + int i; + + /* + * This will normally not happen during block I/O, but can + * happen during SDIO traffic, where odd byte chunks are + * shoveled to the SDIO link. Fill up the residue buffer + * and flush out. + */ + if (host->pio_write_residue_sz) { + int fill = 4 - host->pio_write_residue_sz; + u32 val = 0; + + /* + * If we need more padding that what we have, just stash + * some more in the residue then and continue. This only + * happens if we're passed e.g. 1 or 2 bytes and there is + * just 1 byte residue residue already: we can't fill up! + */ + if (fill > remain) { + for (i = 0; i < fill; i++) { + host->pio_write_residue[host->pio_write_residue_sz + i] = *ptr; + host->pio_write_residue_sz++; + ptr++; + remain--; + } + return ptr - buffer; + } + + /* Pack the residue into a 32bit word */ + for (i = 0; i < host->pio_write_residue_sz; i++) { + val |= host->pio_write_residue[i]; + val <<= 8; + } + /* Top up with new data */ + for (i = 0; i < fill; i++) { + val |= *ptr; + val <<= 8; + ptr++; + remain--; + } + iowrite32(val, base + MMCIFIFO); + host->pio_write_residue_sz = 0; + wrote_residue = true; + } + + /* + * Maybe the residue plus some few bytes was exactly filling a + * 32bit word. + */ + if (!remain) + return ptr - buffer; do { - unsigned int count, maxcnt; + unsigned int count, maxcnt, written, residue; + /* + * If the FIFO is empty just stash it full with data else + * just half-fill it. + */ maxcnt = status & MCI_TXFIFOEMPTY ? - variant->fifosize : variant->fifohalfsize; + variant->fifosize : variant->fifohalfsize; + + /* Watch it to not overfill the FIFO */ + if (wrote_residue) + maxcnt -= 4; + count = min(remain, maxcnt); /* * SDIO especially may want to send something that is * not divisible by 4 (as opposed to card sectors - * etc), and the FIFO only accept full 32-bit writes. - * So compensate by adding +3 on the count, a single - * byte become a 32bit write, 7 bytes will be two - * 32bit writes etc. + * etc), but the FIFO only accepts full 32-bit writes so start by + * just filling up quickly with as much as we can. */ - iowrite32_rep(base + MMCIFIFO, ptr, (count + 3) >> 2); + iowrite32_rep(base + MMCIFIFO, ptr, count >> 2); + residue = count & 3; - ptr += count; - remain -= count; + written = count - residue; + ptr += written; + remain -= written; - if (remain == 0) + /* Handles the common case for block-IO */ + if (!remain) break; + /* + * These were not written out since we only write 32bit + * words to the FIFO. Since this function gets called + * repeatedly and across page boundaries with new pointers + * in *buffer we need to take special care to stash odd + * bytes away and flush them out in next PIO round or at the + * end of the whole data transfer. + */ + if (remain < 4) { + WARN_ON(remain != residue); + host->pio_write_residue_sz = residue; + for (i = 0; i < residue; i++) { + host->pio_write_residue[i] = *ptr; + ptr++; + remain--; + } + break; + } + status = readl(base + MMCISTATUS); } while (status & MCI_TXFIFOHALFEMPTY); @@ -1522,6 +1604,29 @@ static irqreturn_t mmci_pio_irq(int irq, void *dev_id) if (host->size == 0) { mmci_set_mask1(host, 0); writel(readl(base + MMCIMASK0) | MCI_DATAENDMASK, base + MMCIMASK0); + + if ((status & MCI_TXACTIVE) && host->pio_write_residue_sz) { + /* + * If the last pass of the SG miter left some residue after the pio + * write loop, push that out to the card. We are now at the end of + * the transfer so it is OK to pad with zeroes. + */ + int fill = 4 - host->pio_write_residue_sz; + u32 val = 0; + int i; + + /* Pack the residue into a 32bit word */ + for (i = 0; i < host->pio_write_residue_sz; i++) { + val |= host->pio_write_residue[i]; + val <<= 8; + } + /* Top up with zeroes */ + for (i = 0; i < fill; i++) + val <<= 8; + + iowrite32(val, base + MMCIFIFO); + host->pio_write_residue_sz = 0; + } } return IRQ_HANDLED; diff --git a/drivers/mmc/host/mmci.h b/drivers/mmc/host/mmci.h index e20af17bb313..a7690b64d4cd 100644 --- a/drivers/mmc/host/mmci.h +++ b/drivers/mmc/host/mmci.h @@ -425,6 +425,8 @@ struct mmci_host { /* pio stuff */ struct sg_mapping_iter sg_miter; unsigned int size; + u8 pio_write_residue[4]; + u8 pio_write_residue_sz; int (*get_rx_fifocnt)(struct mmci_host *h, u32 status, int remain); u8 use_dma:1;
The MMCI PIO write function contains a bug if using any buffers with sglist miter contents that do not add upp to even 32bit writes. The PIO write code currently does this: iowrite32_rep(base + MMCIFIFO, ptr, (count + 3) >> 2); This will make sure that a single buffer passed in gets written into the FIFO even if it is odd, i.e. not evenly divisible by 4. However it has the following problems: - It will read up to three bytes beyond the end of the buffer and fill the FIFO with unpredictable junk and possibly cause segmentation violations if the read passes a page boundary such as at the end of an sglist buffer. - It will be fine if a single buffer is passed in, but the code handles SG lists. There is a while (1) loop in mmci_pio_irq() which repeatedly calls mmci_pio_write() as long as the FIFO is not half-full, if it gets half-full it exits the IRQ and waits for a new IRQ, also the while loop repeatedly calls sg_miter_next() to consume bytes and this means that between subsequent calls to mmci_pio_write() some random junk can be inserted at the end of each call if the buffers passed in do not contain an number of bytes evenly divisible by 4. Fix this by simply doing this: iowrite32_rep(base + MMCIFIFO, ptr, count >> 2); But since the code will then just not write enough bytes if the count is not evenly divisible by 4, we introduce a special residue buffer and keep track of any odd bytes from 1..3 and just write these padded out with new data in a separate statement the next time we get a call of the mmci_pio_write(), or, if there is for example only 1,2 or 3 bytes in the transfer, or we end up with an odd number of bytes in the residue, flushi it out when we end the current data transaction to the host. Cc: Ludovic Barre <ludovic.barre@st.com> Cc: Brian Masney <masneyb@onstation.org> Cc: Stephan Gerhold <stephan@gerhold.net> Cc: Niklas Cassel <niklas.cassel@linaro.org> Cc: Russell King <rmk+kernel@armlinux.org.uk> Signed-off-by: Linus Walleij <linus.walleij@linaro.org> --- ChangeLog v1->v3: - New patch in v3 after discussion with Ulf - I'm unable to test this because I cannot provoke these odd reads/writes but hoping for Stephan to take it for a spin. --- drivers/mmc/host/mmci.c | 125 ++++++++++++++++++++++++++++++++++++---- drivers/mmc/host/mmci.h | 2 + 2 files changed, 117 insertions(+), 10 deletions(-) -- 2.21.0