From 5b35746a0fdc73063a4c7fc6208b7abd644f9ef5 Mon Sep 17 00:00:00 2001 From: Linus Walleij Date: Fri, 27 Sep 2024 17:54:28 +0200 Subject: Revert "mmc: mvsdio: Use sg_miter for PIO" This reverts commit 2761822c00e8c271f10a10affdbd4917d900d7ea. When testing on real hardware the patch does not work. Revert, try to acquire real hardware, and retry. These systems typically don't have highmem anyway so the impact is likely zero. Cc: stable@vger.kernel.org Reported-by: Charlie Signed-off-by: Linus Walleij Link: https://lore.kernel.org/r/20240927-kirkwood-mmc-regression-v1-1-2e55bbbb7b19@linaro.org Signed-off-by: Ulf Hansson --- drivers/mmc/host/mvsdio.c | 71 ++++++++++++----------------------------------- 1 file changed, 18 insertions(+), 53 deletions(-) (limited to 'drivers/mmc/host') diff --git a/drivers/mmc/host/mvsdio.c b/drivers/mmc/host/mvsdio.c index af7f21888e27..ca01b7d204ba 100644 --- a/drivers/mmc/host/mvsdio.c +++ b/drivers/mmc/host/mvsdio.c @@ -38,9 +38,8 @@ struct mvsd_host { unsigned int xfer_mode; unsigned int intr_en; unsigned int ctrl; - bool use_pio; - struct sg_mapping_iter sg_miter; unsigned int pio_size; + void *pio_ptr; unsigned int sg_frags; unsigned int ns_per_clk; unsigned int clock; @@ -115,18 +114,11 @@ static int mvsd_setup_data(struct mvsd_host *host, struct mmc_data *data) * data when the buffer is not aligned on a 64 byte * boundary. */ - unsigned int miter_flags = SG_MITER_ATOMIC; /* Used from IRQ */ - - if (data->flags & MMC_DATA_READ) - miter_flags |= SG_MITER_TO_SG; - else - miter_flags |= SG_MITER_FROM_SG; - host->pio_size = data->blocks * data->blksz; - sg_miter_start(&host->sg_miter, data->sg, data->sg_len, miter_flags); + host->pio_ptr = sg_virt(data->sg); if (!nodma) - dev_dbg(host->dev, "fallback to PIO for data\n"); - host->use_pio = true; + dev_dbg(host->dev, "fallback to PIO for data at 0x%p size %d\n", + host->pio_ptr, host->pio_size); return 1; } else { dma_addr_t phys_addr; @@ -137,7 +129,6 @@ static int mvsd_setup_data(struct mvsd_host *host, struct mmc_data *data) phys_addr = sg_dma_address(data->sg); mvsd_write(MVSD_SYS_ADDR_LOW, (u32)phys_addr & 0xffff); mvsd_write(MVSD_SYS_ADDR_HI, (u32)phys_addr >> 16); - host->use_pio = false; return 0; } } @@ -297,8 +288,8 @@ static u32 mvsd_finish_data(struct mvsd_host *host, struct mmc_data *data, { void __iomem *iobase = host->base; - if (host->use_pio) { - sg_miter_stop(&host->sg_miter); + if (host->pio_ptr) { + host->pio_ptr = NULL; host->pio_size = 0; } else { dma_unmap_sg(mmc_dev(host->mmc), data->sg, host->sg_frags, @@ -353,12 +344,9 @@ static u32 mvsd_finish_data(struct mvsd_host *host, struct mmc_data *data, static irqreturn_t mvsd_irq(int irq, void *dev) { struct mvsd_host *host = dev; - struct sg_mapping_iter *sgm = &host->sg_miter; void __iomem *iobase = host->base; u32 intr_status, intr_done_mask; int irq_handled = 0; - u16 *p; - int s; intr_status = mvsd_read(MVSD_NOR_INTR_STATUS); dev_dbg(host->dev, "intr 0x%04x intr_en 0x%04x hw_state 0x%04x\n", @@ -382,36 +370,15 @@ static irqreturn_t mvsd_irq(int irq, void *dev) spin_lock(&host->lock); /* PIO handling, if needed. Messy business... */ - if (host->use_pio) { - /* - * As we set sgm->consumed this always gives a valid buffer - * position. - */ - if (!sg_miter_next(sgm)) { - /* This should not happen */ - dev_err(host->dev, "ran out of scatter segments\n"); - spin_unlock(&host->lock); - host->intr_en &= - ~(MVSD_NOR_RX_READY | MVSD_NOR_RX_FIFO_8W | - MVSD_NOR_TX_AVAIL | MVSD_NOR_TX_FIFO_8W); - mvsd_write(MVSD_NOR_INTR_EN, host->intr_en); - return IRQ_HANDLED; - } - p = sgm->addr; - s = sgm->length; - if (s > host->pio_size) - s = host->pio_size; - } - - if (host->use_pio && + if (host->pio_size && (intr_status & host->intr_en & (MVSD_NOR_RX_READY | MVSD_NOR_RX_FIFO_8W))) { - + u16 *p = host->pio_ptr; + int s = host->pio_size; while (s >= 32 && (intr_status & MVSD_NOR_RX_FIFO_8W)) { readsw(iobase + MVSD_FIFO, p, 16); p += 16; s -= 32; - sgm->consumed += 32; intr_status = mvsd_read(MVSD_NOR_INTR_STATUS); } /* @@ -424,7 +391,6 @@ static irqreturn_t mvsd_irq(int irq, void *dev) put_unaligned(mvsd_read(MVSD_FIFO), p++); put_unaligned(mvsd_read(MVSD_FIFO), p++); s -= 4; - sgm->consumed += 4; intr_status = mvsd_read(MVSD_NOR_INTR_STATUS); } if (s && s < 4 && (intr_status & MVSD_NOR_RX_READY)) { @@ -432,13 +398,10 @@ static irqreturn_t mvsd_irq(int irq, void *dev) val[0] = mvsd_read(MVSD_FIFO); val[1] = mvsd_read(MVSD_FIFO); memcpy(p, ((void *)&val) + 4 - s, s); - sgm->consumed += s; s = 0; intr_status = mvsd_read(MVSD_NOR_INTR_STATUS); } - /* PIO transfer done */ - host->pio_size -= sgm->consumed; - if (host->pio_size == 0) { + if (s == 0) { host->intr_en &= ~(MVSD_NOR_RX_READY | MVSD_NOR_RX_FIFO_8W); mvsd_write(MVSD_NOR_INTR_EN, host->intr_en); @@ -450,10 +413,14 @@ static irqreturn_t mvsd_irq(int irq, void *dev) } dev_dbg(host->dev, "pio %d intr 0x%04x hw_state 0x%04x\n", s, intr_status, mvsd_read(MVSD_HW_STATE)); + host->pio_ptr = p; + host->pio_size = s; irq_handled = 1; - } else if (host->use_pio && + } else if (host->pio_size && (intr_status & host->intr_en & (MVSD_NOR_TX_AVAIL | MVSD_NOR_TX_FIFO_8W))) { + u16 *p = host->pio_ptr; + int s = host->pio_size; /* * The TX_FIFO_8W bit is unreliable. When set, bursting * 16 halfwords all at once in the FIFO drops data. Actually @@ -464,7 +431,6 @@ static irqreturn_t mvsd_irq(int irq, void *dev) mvsd_write(MVSD_FIFO, get_unaligned(p++)); mvsd_write(MVSD_FIFO, get_unaligned(p++)); s -= 4; - sgm->consumed += 4; intr_status = mvsd_read(MVSD_NOR_INTR_STATUS); } if (s < 4) { @@ -473,13 +439,10 @@ static irqreturn_t mvsd_irq(int irq, void *dev) memcpy(((void *)&val) + 4 - s, p, s); mvsd_write(MVSD_FIFO, val[0]); mvsd_write(MVSD_FIFO, val[1]); - sgm->consumed += s; s = 0; intr_status = mvsd_read(MVSD_NOR_INTR_STATUS); } - /* PIO transfer done */ - host->pio_size -= sgm->consumed; - if (host->pio_size == 0) { + if (s == 0) { host->intr_en &= ~(MVSD_NOR_TX_AVAIL | MVSD_NOR_TX_FIFO_8W); mvsd_write(MVSD_NOR_INTR_EN, host->intr_en); @@ -487,6 +450,8 @@ static irqreturn_t mvsd_irq(int irq, void *dev) } dev_dbg(host->dev, "pio %d intr 0x%04x hw_state 0x%04x\n", s, intr_status, mvsd_read(MVSD_HW_STATE)); + host->pio_ptr = p; + host->pio_size = s; irq_handled = 1; } -- cgit v1.2.3 From 27e8fe0da3b75520edfba9cee0030aeb5aef1505 Mon Sep 17 00:00:00 2001 From: Michal Wilczynski Date: Tue, 8 Oct 2024 12:03:27 +0200 Subject: mmc: sdhci-of-dwcmshc: Prevent stale command interrupt handling While working with the T-Head 1520 LicheePi4A SoC, certain conditions arose that allowed me to reproduce a race issue in the sdhci code. To reproduce the bug, you need to enable the sdio1 controller in the device tree file `arch/riscv/boot/dts/thead/th1520-lichee-module-4a.dtsi` as follows: &sdio1 { bus-width = <4>; max-frequency = <100000000>; no-sd; no-mmc; broken-cd; cap-sd-highspeed; post-power-on-delay-ms = <50>; status = "okay"; wakeup-source; keep-power-in-suspend; }; When resetting the SoC using the reset button, the following messages appear in the dmesg log: [ 8.164898] mmc2: Got command interrupt 0x00000001 even though no command operation was in progress. [ 8.174054] mmc2: sdhci: ============ SDHCI REGISTER DUMP =========== [ 8.180503] mmc2: sdhci: Sys addr: 0x00000000 | Version: 0x00000005 [ 8.186950] mmc2: sdhci: Blk size: 0x00000000 | Blk cnt: 0x00000000 [ 8.193395] mmc2: sdhci: Argument: 0x00000000 | Trn mode: 0x00000000 [ 8.199841] mmc2: sdhci: Present: 0x03da0000 | Host ctl: 0x00000000 [ 8.206287] mmc2: sdhci: Power: 0x0000000f | Blk gap: 0x00000000 [ 8.212733] mmc2: sdhci: Wake-up: 0x00000000 | Clock: 0x0000decf [ 8.219178] mmc2: sdhci: Timeout: 0x00000000 | Int stat: 0x00000000 [ 8.225622] mmc2: sdhci: Int enab: 0x00ff1003 | Sig enab: 0x00ff1003 [ 8.232068] mmc2: sdhci: ACmd stat: 0x00000000 | Slot int: 0x00000000 [ 8.238513] mmc2: sdhci: Caps: 0x3f69c881 | Caps_1: 0x08008177 [ 8.244959] mmc2: sdhci: Cmd: 0x00000502 | Max curr: 0x00191919 [ 8.254115] mmc2: sdhci: Resp[0]: 0x00001009 | Resp[1]: 0x00000000 [ 8.260561] mmc2: sdhci: Resp[2]: 0x00000000 | Resp[3]: 0x00000000 [ 8.267005] mmc2: sdhci: Host ctl2: 0x00001000 [ 8.271453] mmc2: sdhci: ADMA Err: 0x00000000 | ADMA Ptr: 0x0000000000000000 [ 8.278594] mmc2: sdhci: ============================================ I also enabled some traces to better understand the problem: kworker/3:1-62 [003] ..... 8.163538: mmc_request_start: mmc2: start struct mmc_request[000000000d30cc0c]: cmd_opcode=5 cmd_arg=0x0 cmd_flags=0x2e1 cmd_retries=0 stop_opcode=0 stop_arg=0x0 stop_flags=0x0 stop_retries=0 sbc_opcode=0 sbc_arg=0x0 sbc_flags=0x0 sbc_retires=0 blocks=0 block_size=0 blk_addr=0 data_flags=0x0 tag=0 can_retune=0 doing_retune=0 retune_now=0 need_retune=0 hold_retune=1 retune_period=0 -0 [000] d.h2. 8.164816: sdhci_cmd_irq: hw_name=ffe70a0000.mmc quirks=0x2008008 quirks2=0x8 intmask=0x10000 intmask_p=0x18000 irq/24-mmc2-96 [000] ..... 8.164840: sdhci_thread_irq: msg= irq/24-mmc2-96 [000] d.h2. 8.164896: sdhci_cmd_irq: hw_name=ffe70a0000.mmc quirks=0x2008008 quirks2=0x8 intmask=0x1 intmask_p=0x1 irq/24-mmc2-96 [000] ..... 8.285142: mmc_request_done: mmc2: end struct mmc_request[000000000d30cc0c]: cmd_opcode=5 cmd_err=-110 cmd_resp=0x0 0x0 0x0 0x0 cmd_retries=0 stop_opcode=0 stop_err=0 stop_resp=0x0 0x0 0x0 0x0 stop_retries=0 sbc_opcode=0 sbc_err=0 sbc_resp=0x0 0x0 0x0 0x0 sbc_retries=0 bytes_xfered=0 data_err=0 tag=0 can_retune=0 doing_retune=0 retune_now=0 need_retune=0 hold_retune=1 retune_period=0 Here's what happens: the __mmc_start_request function is called with opcode 5. Since the power to the Wi-Fi card, which resides on this SDIO bus, is initially off after the reset, an interrupt SDHCI_INT_TIMEOUT is triggered. Immediately after that, a second interrupt SDHCI_INT_RESPONSE is triggered. Depending on the exact timing, these conditions can trigger the following race problem: 1) The sdhci_cmd_irq top half handles the command as an error. It sets host->cmd to NULL and host->pending_reset to true. 2) The sdhci_thread_irq bottom half is scheduled next and executes faster than the second interrupt handler for SDHCI_INT_RESPONSE. It clears host->pending_reset before the SDHCI_INT_RESPONSE handler runs. 3) The pending interrupt SDHCI_INT_RESPONSE handler gets called, triggering a code path that prints: "mmc2: Got command interrupt 0x00000001 even though no command operation was in progress." To solve this issue, we need to clear pending interrupts when resetting host->pending_reset. This ensures that after sdhci_threaded_irq restores interrupts, there are no pending stale interrupts. The behavior observed here is non-compliant with the SDHCI standard. Place the code in the sdhci-of-dwcmshc driver to account for a hardware-specific quirk instead of the core SDHCI code. Signed-off-by: Michal Wilczynski Acked-by: Adrian Hunter Fixes: 43658a542ebf ("mmc: sdhci-of-dwcmshc: Add support for T-Head TH1520") Cc: stable@vger.kernel.org Link: https://lore.kernel.org/r/20241008100327.4108895-1-m.wilczynski@samsung.com Signed-off-by: Ulf Hansson --- drivers/mmc/host/sdhci-of-dwcmshc.c | 8 ++++++++ 1 file changed, 8 insertions(+) (limited to 'drivers/mmc/host') diff --git a/drivers/mmc/host/sdhci-of-dwcmshc.c b/drivers/mmc/host/sdhci-of-dwcmshc.c index 8999b97263af..8fd80dac11bf 100644 --- a/drivers/mmc/host/sdhci-of-dwcmshc.c +++ b/drivers/mmc/host/sdhci-of-dwcmshc.c @@ -852,6 +852,14 @@ static void th1520_sdhci_reset(struct sdhci_host *host, u8 mask) sdhci_reset(host, mask); + /* The T-Head 1520 SoC does not comply with the SDHCI specification + * regarding the "Software Reset for CMD line should clear 'Command + * Complete' in the Normal Interrupt Status Register." Clear the bit + * here to compensate for this quirk. + */ + if (mask & SDHCI_RESET_CMD) + sdhci_writel(host, SDHCI_INT_RESPONSE, SDHCI_INT_STATUS); + if (priv->flags & FLAG_IO_FIXED_1V8) { ctrl_2 = sdhci_readw(host, SDHCI_HOST_CONTROL2); if (!(ctrl_2 & SDHCI_CTRL_VDD_180)) { -- cgit v1.2.3