| Message ID | 1463140760-7128-2-git-send-email-jszhang@marvell.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
On Friday 13 May 2016 19:59:19 Jisheng Zhang wrote: > /* port register accessors **************************************************/ > static inline u32 rdl(struct mv643xx_eth_private *mp, int offset) > { > - return readl(mp->shared->base + offset); > + return readl_relaxed(mp->shared->base + offset); > } > > static inline u32 rdlp(struct mv643xx_eth_private *mp, int offset) > { > - return readl(mp->base + offset); > + return readl_relaxed(mp->base + offset); > } I'd recommend not changing these in general, but introducing new 'rdl_relaxed()' and 'rdlp_relaxed()' etc variants that you use in the code that actually is performance sensitive, but use the normal non-relaxed versions by default. Then add a comment to each use of the relaxed accessors on how you know that it's safe for that caller. This usually is just needed for the xmit() function and for the interrupt handler. > @@ -2642,10 +2642,10 @@ mv643xx_eth_conf_mbus_windows(struct mv643xx_eth_shared_private *msp, > int i; > > for (i = 0; i < 6; i++) { > - writel(0, base + WINDOW_BASE(i)); > - writel(0, base + WINDOW_SIZE(i)); > + writel_relaxed(0, base + WINDOW_BASE(i)); > + writel_relaxed(0, base + WINDOW_SIZE(i)); > if (i < 4) > - writel(0, base + WINDOW_REMAP_HIGH(i)); > + writel_relaxed(0, base + WINDOW_REMAP_HIGH(i)); > } > > win_enable = 0x3f; Configuring the mbus for instance is clearly not an operation in which performance matters at all, so better not touch that. > @@ -2674,8 +2675,8 @@ static void infer_hw_params(struct mv643xx_eth_shared_private *msp) > * [21:8], or a 16-bit coal limit in bits [25,21:7] of the > * SDMA config register. > */ > - writel(0x02000000, msp->base + 0x0400 + SDMA_CONFIG); > - if (readl(msp->base + 0x0400 + SDMA_CONFIG) & 0x02000000) > + writel_relaxed(0x02000000, msp->base + 0x0400 + SDMA_CONFIG); > + if (readl_relaxed(msp->base + 0x0400 + SDMA_CONFIG) & 0x02000000) > msp->extended_rx_coal_limit = 1; > else > msp->extended_rx_coal_limit = 0; This also seems to be configuration, rather than in the packet rx/tx hotpath. Arnd
Dear Arnd, On Fri, 13 May 2016 14:09:43 +0200 Arnd Bergmann wrote: > On Friday 13 May 2016 19:59:19 Jisheng Zhang wrote: > > /* port register accessors **************************************************/ > > static inline u32 rdl(struct mv643xx_eth_private *mp, int offset) > > { > > - return readl(mp->shared->base + offset); > > + return readl_relaxed(mp->shared->base + offset); > > } > > > > static inline u32 rdlp(struct mv643xx_eth_private *mp, int offset) > > { > > - return readl(mp->base + offset); > > + return readl_relaxed(mp->base + offset); > > } > > I'd recommend not changing these in general, but introducing new 'rdl_relaxed()' > and 'rdlp_relaxed()' etc variants that you use in the code that actually > is performance sensitive, but use the normal non-relaxed versions by > default. > > Then add a comment to each use of the relaxed accessors on how you know > that it's safe for that caller. This usually is just needed for the xmit() > function and for the interrupt handler. Got your points and I do think it makes sense. But could we always use the relaxed version to save some LoCs, although I have no mv643xx_eth platform but I can confirm similar relaxed version changes in pxa168_eth is safe and this is what we do in product's kernel. Above is just my humble opinion, comments are welcome. Thanks, Jisheng > > > @@ -2642,10 +2642,10 @@ mv643xx_eth_conf_mbus_windows(struct mv643xx_eth_shared_private *msp, > > int i; > > > > for (i = 0; i < 6; i++) { > > - writel(0, base + WINDOW_BASE(i)); > > - writel(0, base + WINDOW_SIZE(i)); > > + writel_relaxed(0, base + WINDOW_BASE(i)); > > + writel_relaxed(0, base + WINDOW_SIZE(i)); > > if (i < 4) > > - writel(0, base + WINDOW_REMAP_HIGH(i)); > > + writel_relaxed(0, base + WINDOW_REMAP_HIGH(i)); > > } > > > > win_enable = 0x3f; > > Configuring the mbus for instance is clearly not an operation in which > performance matters at all, so better not touch that. > > > @@ -2674,8 +2675,8 @@ static void infer_hw_params(struct mv643xx_eth_shared_private *msp) > > * [21:8], or a 16-bit coal limit in bits [25,21:7] of the > > * SDMA config register. > > */ > > - writel(0x02000000, msp->base + 0x0400 + SDMA_CONFIG); > > - if (readl(msp->base + 0x0400 + SDMA_CONFIG) & 0x02000000) > > + writel_relaxed(0x02000000, msp->base + 0x0400 + SDMA_CONFIG); > > + if (readl_relaxed(msp->base + 0x0400 + SDMA_CONFIG) & 0x02000000) > > msp->extended_rx_coal_limit = 1; > > else > > msp->extended_rx_coal_limit = 0; > > > This also seems to be configuration, rather than in the packet rx/tx hotpath. > > Arnd
On Fri, May 13, 2016 at 08:19:55PM +0800, Jisheng Zhang wrote: > Dear Arnd, > > On Fri, 13 May 2016 14:09:43 +0200 Arnd Bergmann wrote: > > > On Friday 13 May 2016 19:59:19 Jisheng Zhang wrote: > > > /* port register accessors **************************************************/ > > > static inline u32 rdl(struct mv643xx_eth_private *mp, int offset) > > > { > > > - return readl(mp->shared->base + offset); > > > + return readl_relaxed(mp->shared->base + offset); > > > } > > > > > > static inline u32 rdlp(struct mv643xx_eth_private *mp, int offset) > > > { > > > - return readl(mp->base + offset); > > > + return readl_relaxed(mp->base + offset); > > > } > > > > I'd recommend not changing these in general, but introducing new 'rdl_relaxed()' > > and 'rdlp_relaxed()' etc variants that you use in the code that actually > > is performance sensitive, but use the normal non-relaxed versions by > > default. > > > > Then add a comment to each use of the relaxed accessors on how you know > > that it's safe for that caller. This usually is just needed for the xmit() > > function and for the interrupt handler. > > Got your points and I do think it makes sense. But could we always use the > relaxed version to save some LoCs. It is a trade off between lines of code and hard to find bugs. Getting this wrong can introduce subtle bugs. Best be paranoid and only touch the fast path, where this actually matters. Andrew
diff --git a/drivers/net/ethernet/marvell/mv643xx_eth.c b/drivers/net/ethernet/marvell/mv643xx_eth.c index 5583118..c6d8124 100644 --- a/drivers/net/ethernet/marvell/mv643xx_eth.c +++ b/drivers/net/ethernet/marvell/mv643xx_eth.c @@ -433,22 +433,22 @@ struct mv643xx_eth_private { /* port register accessors **************************************************/ static inline u32 rdl(struct mv643xx_eth_private *mp, int offset) { - return readl(mp->shared->base + offset); + return readl_relaxed(mp->shared->base + offset); } static inline u32 rdlp(struct mv643xx_eth_private *mp, int offset) { - return readl(mp->base + offset); + return readl_relaxed(mp->base + offset); } static inline void wrl(struct mv643xx_eth_private *mp, int offset, u32 data) { - writel(data, mp->shared->base + offset); + writel_relaxed(data, mp->shared->base + offset); } static inline void wrlp(struct mv643xx_eth_private *mp, int offset, u32 data) { - writel(data, mp->base + offset); + writel_relaxed(data, mp->base + offset); } @@ -2642,10 +2642,10 @@ mv643xx_eth_conf_mbus_windows(struct mv643xx_eth_shared_private *msp, int i; for (i = 0; i < 6; i++) { - writel(0, base + WINDOW_BASE(i)); - writel(0, base + WINDOW_SIZE(i)); + writel_relaxed(0, base + WINDOW_BASE(i)); + writel_relaxed(0, base + WINDOW_SIZE(i)); if (i < 4) - writel(0, base + WINDOW_REMAP_HIGH(i)); + writel_relaxed(0, base + WINDOW_REMAP_HIGH(i)); } win_enable = 0x3f; @@ -2654,16 +2654,17 @@ mv643xx_eth_conf_mbus_windows(struct mv643xx_eth_shared_private *msp, for (i = 0; i < dram->num_cs; i++) { const struct mbus_dram_window *cs = dram->cs + i; - writel((cs->base & 0xffff0000) | + writel_relaxed((cs->base & 0xffff0000) | (cs->mbus_attr << 8) | dram->mbus_dram_target_id, base + WINDOW_BASE(i)); - writel((cs->size - 1) & 0xffff0000, base + WINDOW_SIZE(i)); + writel_relaxed((cs->size - 1) & 0xffff0000, + base + WINDOW_SIZE(i)); win_enable &= ~(1 << i); win_protect |= 3 << (2 * i); } - writel(win_enable, base + WINDOW_BAR_ENABLE); + writel_relaxed(win_enable, base + WINDOW_BAR_ENABLE); msp->win_protect = win_protect; } @@ -2674,8 +2675,8 @@ static void infer_hw_params(struct mv643xx_eth_shared_private *msp) * [21:8], or a 16-bit coal limit in bits [25,21:7] of the * SDMA config register. */ - writel(0x02000000, msp->base + 0x0400 + SDMA_CONFIG); - if (readl(msp->base + 0x0400 + SDMA_CONFIG) & 0x02000000) + writel_relaxed(0x02000000, msp->base + 0x0400 + SDMA_CONFIG); + if (readl_relaxed(msp->base + 0x0400 + SDMA_CONFIG) & 0x02000000) msp->extended_rx_coal_limit = 1; else msp->extended_rx_coal_limit = 0; @@ -2685,12 +2686,12 @@ static void infer_hw_params(struct mv643xx_eth_shared_private *msp) * yes, whether its associated registers are in the old or * the new place. */ - writel(1, msp->base + 0x0400 + TX_BW_MTU_MOVED); - if (readl(msp->base + 0x0400 + TX_BW_MTU_MOVED) & 1) { + writel_relaxed(1, msp->base + 0x0400 + TX_BW_MTU_MOVED); + if (readl_relaxed(msp->base + 0x0400 + TX_BW_MTU_MOVED) & 1) { msp->tx_bw_control = TX_BW_CONTROL_NEW_LAYOUT; } else { - writel(7, msp->base + 0x0400 + TX_BW_RATE); - if (readl(msp->base + 0x0400 + TX_BW_RATE) & 7) + writel_relaxed(7, msp->base + 0x0400 + TX_BW_RATE); + if (readl_relaxed(msp->base + 0x0400 + TX_BW_RATE) & 7) msp->tx_bw_control = TX_BW_CONTROL_OLD_LAYOUT; else msp->tx_bw_control = TX_BW_CONTROL_ABSENT;
Since appropriate memory barriers are already there, use the relaxed version to improve performance a bit. Signed-off-by: Jisheng Zhang <jszhang@marvell.com> --- drivers/net/ethernet/marvell/mv643xx_eth.c | 33 +++++++++++++++--------------- 1 file changed, 17 insertions(+), 16 deletions(-)