| Message ID | 1379606505-2529-2-git-send-email-ezequiel.garcia@free-electrons.com |
|---|---|
| State | Accepted |
| Commit | 62e8b851783138a11da63285be0fbf69530ff73d |
| Headers | show |
On Thu, Sep 19, 2013 at 01:01:25PM -0300, Ezequiel Garcia wrote: > This commit replaces the currently hardcoded buffer size, by a > dynamic detection scheme. First a small 256 bytes buffer is allocated > so the device can be detected (using READID and friends commands). > > After detection, this buffer is released and a new buffer is allocated > to acommodate the page size plus out-of-band size. > > Signed-off-by: Ezequiel Garcia <ezequiel.garcia@free-electrons.com> This is the patch that breaks Daniel's DMA setup, right? It looks a bit off. I'll wait to comment on it much until v2. BTW, there is a similar issue with at least one other driver (denali.c, maybe others) where the driver uses some hard assumptions about the maximum page/OOB sizes (NAND_MAX_PAGESIZE and NAND_MAX_OOBSIZE). This is kind of ugly and not very sustainable/maintainable, since these dimensions keep increasing. I appreciate your effort to avoid simply kludging in a larger MAX_BUFF_SIZE :) I had similar plans for the other drivers, but I don't know if we'll have much testing opportunities for the ancient ones... Also, it seems like your driver has a few potential leaks right now. If alloc_nand_resource() succeeds but pxa3xx_nand_probe() later fails (e.g., in pxa3xx_nand_scan()), you don't clean up after yourself. You should address this soon, even if not in this patch series. Brian
On Thu, Sep 26, 2013 at 01:10:32PM -0700, Brian Norris wrote: > On Thu, Sep 19, 2013 at 01:01:25PM -0300, Ezequiel Garcia wrote: > > This commit replaces the currently hardcoded buffer size, by a > > dynamic detection scheme. First a small 256 bytes buffer is allocated > > so the device can be detected (using READID and friends commands). > > > > After detection, this buffer is released and a new buffer is allocated > > to acommodate the page size plus out-of-band size. > > > > Signed-off-by: Ezequiel Garcia <ezequiel.garcia@free-electrons.com> > > This is the patch that breaks Daniel's DMA setup, right? It looks a bit > off. What do you mean by 'a bit off'? > I'll wait to comment on it much until v2. > Ok... let me re-work it then and prepare the v2. > BTW, there is a similar issue with at least one other driver (denali.c, > maybe others) where the driver uses some hard assumptions about the > maximum page/OOB sizes (NAND_MAX_PAGESIZE and NAND_MAX_OOBSIZE). This is > kind of ugly and not very sustainable/maintainable, since these > dimensions keep increasing. I appreciate your effort to avoid simply > kludging in a larger MAX_BUFF_SIZE :) I had similar plans for the other > drivers, but I don't know if we'll have much testing opportunities for > the ancient ones... > > Also, it seems like your driver has a few potential leaks right now. If > alloc_nand_resource() succeeds but pxa3xx_nand_probe() later fails > (e.g., in pxa3xx_nand_scan()), you don't clean up after yourself. You > should address this soon, even if not in this patch series. > Hm.. are you sure about that? AFAICS, there's no leak at all. If alloc_nand_resource() succeeds, the only leakable resources allocated are the ones allocated at pxa3xx_nand_init_buff() and the NAND base stuff. If pxa3xx_nand_probe() later fails to complete, it calls pxa3xx_nand_remove() in this part: if (!probe_success) { pxa3xx_nand_remove(pdev); return -ENODEV; } Which takes care of cleaning both the buffers and the NAND base stuff. Or am I missing something? Feel free to ask more questions and thanks a lot for the feedback.
On Mon, Sep 30, 2013 at 09:37:57AM -0300, Ezequiel Garcia wrote: > On Thu, Sep 26, 2013 at 01:10:32PM -0700, Brian Norris wrote: > > On Thu, Sep 19, 2013 at 01:01:25PM -0300, Ezequiel Garcia wrote: > > > This commit replaces the currently hardcoded buffer size, by a > > > dynamic detection scheme. First a small 256 bytes buffer is allocated > > > so the device can be detected (using READID and friends commands). > > > > > > After detection, this buffer is released and a new buffer is allocated > > > to acommodate the page size plus out-of-band size. > > > > > > Signed-off-by: Ezequiel Garcia <ezequiel.garcia@free-electrons.com> > > > > This is the patch that breaks Daniel's DMA setup, right? It looks a bit > > off. > > What do you mean by 'a bit off'? The existing code allocates data_buff with kmalloc() or dma_alloc_coherent() depending on ARCH_HAS_DMA, but your patch does the initial fixed-size allocation only with kmalloc(). It's just what you and Daniel already identified: that this causes problems for the DMA case. > > BTW, there is a similar issue with at least one other driver (denali.c, > > maybe others) where the driver uses some hard assumptions about the > > maximum page/OOB sizes (NAND_MAX_PAGESIZE and NAND_MAX_OOBSIZE). This is > > kind of ugly and not very sustainable/maintainable, since these > > dimensions keep increasing. I appreciate your effort to avoid simply > > kludging in a larger MAX_BUFF_SIZE :) I had similar plans for the other > > drivers, but I don't know if we'll have much testing opportunities for > > the ancient ones... > > > > Also, it seems like your driver has a few potential leaks right now. If > > alloc_nand_resource() succeeds but pxa3xx_nand_probe() later fails > > (e.g., in pxa3xx_nand_scan()), you don't clean up after yourself. You > > should address this soon, even if not in this patch series. > > > > Hm.. are you sure about that? AFAICS, there's no leak at all. No, I'm not sure :) > If alloc_nand_resource() succeeds, the only leakable resources allocated > are the ones allocated at pxa3xx_nand_init_buff() and the NAND base > stuff. > > If pxa3xx_nand_probe() later fails to complete, it calls pxa3xx_nand_remove() > in this part: > > if (!probe_success) { > pxa3xx_nand_remove(pdev); > return -ENODEV; > } > > Which takes care of cleaning both the buffers and the NAND base stuff. > > Or am I missing something? Nope, you were correct. I just missed this point when reading. Brian
diff --git a/drivers/mtd/nand/pxa3xx_nand.c b/drivers/mtd/nand/pxa3xx_nand.c index dd03dfd..b4f3784 100644 --- a/drivers/mtd/nand/pxa3xx_nand.c +++ b/drivers/mtd/nand/pxa3xx_nand.c @@ -39,6 +39,13 @@ #define NAND_STOP_DELAY (2 * HZ/50) #define PAGE_CHUNK_SIZE (2048) +/* + * Define a buffer size for the initial command that detects the flash device: + * STATUS, READID and PARAM. The largest of these is the PARAM command, + * needing 256 bytes. + */ +#define INIT_BUFFER_SIZE 256 + /* registers and bit definitions */ #define NDCR (0x00) /* Control register */ #define NDTR0CS0 (0x04) /* Timing Parameter 0 for CS0 */ @@ -164,6 +171,7 @@ struct pxa3xx_nand_info { unsigned int buf_start; unsigned int buf_count; + unsigned int buf_size; /* DMA information */ int drcmr_dat; @@ -912,26 +920,20 @@ static int pxa3xx_nand_detect_config(struct pxa3xx_nand_info *info) return 0; } -/* the maximum possible buffer size for large page with OOB data - * is: 2048 + 64 = 2112 bytes, allocate a page here for both the - * data buffer and the DMA descriptor - */ -#define MAX_BUFF_SIZE PAGE_SIZE - #ifdef ARCH_HAS_DMA static int pxa3xx_nand_init_buff(struct pxa3xx_nand_info *info) { struct platform_device *pdev = info->pdev; - int data_desc_offset = MAX_BUFF_SIZE - sizeof(struct pxa_dma_desc); + int data_desc_offset = info->buf_size - sizeof(struct pxa_dma_desc); if (use_dma == 0) { - info->data_buff = kmalloc(MAX_BUFF_SIZE, GFP_KERNEL); + info->data_buff = kmalloc(info->buf_size, GFP_KERNEL); if (info->data_buff == NULL) return -ENOMEM; return 0; } - info->data_buff = dma_alloc_coherent(&pdev->dev, MAX_BUFF_SIZE, + info->data_buff = dma_alloc_coherent(&pdev->dev, info->buf_size, &info->data_buff_phys, GFP_KERNEL); if (info->data_buff == NULL) { dev_err(&pdev->dev, "failed to allocate dma buffer\n"); @@ -945,7 +947,7 @@ static int pxa3xx_nand_init_buff(struct pxa3xx_nand_info *info) pxa3xx_nand_data_dma_irq, info); if (info->data_dma_ch < 0) { dev_err(&pdev->dev, "failed to request data dma\n"); - dma_free_coherent(&pdev->dev, MAX_BUFF_SIZE, + dma_free_coherent(&pdev->dev, info->buf_size, info->data_buff, info->data_buff_phys); return info->data_dma_ch; } @@ -958,7 +960,7 @@ static void pxa3xx_nand_free_buff(struct pxa3xx_nand_info *info) struct platform_device *pdev = info->pdev; if (use_dma) { pxa_free_dma(info->data_dma_ch); - dma_free_coherent(&pdev->dev, MAX_BUFF_SIZE, + dma_free_coherent(&pdev->dev, info->buf_size, info->data_buff, info->data_buff_phys); } else { kfree(info->data_buff); @@ -967,7 +969,7 @@ static void pxa3xx_nand_free_buff(struct pxa3xx_nand_info *info) #else static int pxa3xx_nand_init_buff(struct pxa3xx_nand_info *info) { - info->data_buff = kmalloc(MAX_BUFF_SIZE, GFP_KERNEL); + info->data_buff = kmalloc(info->buf_size, GFP_KERNEL); if (info->data_buff == NULL) return -ENOMEM; return 0; @@ -1081,7 +1083,16 @@ KEEP_CONFIG: else host->col_addr_cycles = 1; + /* release the initial buffer */ + kfree(info->data_buff); + + /* allocate the real data + oob buffer */ + info->buf_size = mtd->writesize + mtd->oobsize; + ret = pxa3xx_nand_init_buff(info); + if (ret) + return ret; info->oob_buff = info->data_buff + mtd->writesize; + if ((mtd->size >> chip->page_shift) > 65536) host->row_addr_cycles = 3; else @@ -1187,9 +1198,13 @@ static int alloc_nand_resource(struct platform_device *pdev) } info->mmio_phys = r->start; - ret = pxa3xx_nand_init_buff(info); - if (ret) + /* Allocate a buffer to allow flash detection */ + info->buf_size = INIT_BUFFER_SIZE; + info->data_buff = kmalloc(info->buf_size, GFP_KERNEL); + if (info->data_buff == NULL) { + ret = -ENOMEM; goto fail_disable_clk; + } /* initialize all interrupts to be disabled */ disable_int(info, NDSR_MASK); @@ -1207,7 +1222,7 @@ static int alloc_nand_resource(struct platform_device *pdev) fail_free_buf: free_irq(irq, info); - pxa3xx_nand_free_buff(info); + kfree(info->data_buff); fail_disable_clk: clk_disable_unprepare(info->clk); return ret;
This commit replaces the currently hardcoded buffer size, by a dynamic detection scheme. First a small 256 bytes buffer is allocated so the device can be detected (using READID and friends commands). After detection, this buffer is released and a new buffer is allocated to acommodate the page size plus out-of-band size. Signed-off-by: Ezequiel Garcia <ezequiel.garcia@free-electrons.com> --- drivers/mtd/nand/pxa3xx_nand.c | 45 ++++++++++++++++++++++++++++-------------- 1 file changed, 30 insertions(+), 15 deletions(-)