Hi Shimoda-san,
> -----Original Message-----
> From: Yoshihiro Shimoda <yoshihiro.shimoda.uh@xxxxxxxxxxx>
> Sent: 03 April 2020 10:34
> To: Prabhakar Mahadev Lad <prabhakar.mahadev-lad.rj@xxxxxxxxxxxxxx>; Bjorn Helgaas <bhelgaas@xxxxxxxxxx>; Rob Herring
> <robh+dt@xxxxxxxxxx>; Mark Rutland <mark.rutland@xxxxxxx>; Geert Uytterhoeven <geert+renesas@xxxxxxxxx>; Magnus Damm
> <magnus.damm@xxxxxxxxx>; Kishon Vijay Abraham I <kishon@xxxxxx>; Lorenzo Pieralisi <lorenzo.pieralisi@xxxxxxx>; Marek Vasut
> <marek.vasut+renesas@xxxxxxxxx>; linux-pci@xxxxxxxxxxxxxxx
> Cc: Catalin Marinas <catalin.marinas@xxxxxxx>; Will Deacon <will@xxxxxxxxxx>; Arnd Bergmann <arnd@xxxxxxxx>; Greg Kroah-Hartman
> <gregkh@xxxxxxxxxxxxxxxxxxx>; Andrew Murray <andrew.murray@xxxxxxx>; devicetree@xxxxxxxxxxxxxxx; linux-kernel@xxxxxxxxxxxxxxx;
> linux-arm-kernel@xxxxxxxxxxxxxxxxxxx; linux-renesas-soc@xxxxxxxxxxxxxxx; Chris Paterson <Chris.Paterson2@xxxxxxxxxxx>; Frank Rowand
> <frowand.list@xxxxxxxxx>; Gustavo Pimentel <gustavo.pimentel@xxxxxxxxxxxx>; Jingoo Han <jingoohan1@xxxxxxxxx>; Simon Horman
> <horms@xxxxxxxxxxxx>; Shawn Lin <shawn.lin@xxxxxxxxxxxxxx>; Tom Joseph <tjoseph@xxxxxxxxxxx>; Heiko Stuebner
> <heiko@xxxxxxxxx>; linux-rockchip@xxxxxxxxxxxxxxxxxxx; Lad Prabhakar <prabhakar.csengg@xxxxxxxxx>
> Subject: RE: [PATCH v6 06/11] PCI: endpoint: Add support to handle multiple base for mapping outbound memory
>
> Hi Prabhakar-san,
>
> > From: Prabhakar Mahadev Lad, Sent: Friday, April 3, 2020 6:12 PM
> <snip>
> > > > @@ -122,31 +167,56 @@ EXPORT_SYMBOL_GPL(pci_epc_mem_exit);
> > > > void __iomem *pci_epc_mem_alloc_addr(struct pci_epc *epc,
> > > > phys_addr_t *phys_addr, size_t size)
> > > > {
> > > > -int pageno;
> > > > void __iomem *virt_addr = NULL;
> > > > -struct pci_epc_mem *mem = epc->mem;
> > > > -unsigned int page_shift = ilog2(mem->page_size);
> > > > +struct pci_epc_mem *mem;
> > > > +unsigned int page_shift;
> > > > +int pageno = -EINVAL;
> > > > int order;
> > > > +int i;
> > > >
> > > > -size = ALIGN(size, mem->page_size);
> > > > -order = pci_epc_mem_get_order(mem, size);
> > > > -
> > > > -mutex_lock(&mem->lock);
> > > > -pageno = bitmap_find_free_region(mem->bitmap, mem->pages, order);
> > > > -if (pageno < 0)
> > > > -goto ret;
> > > > +for (i = 0; i < epc->num_windows; i++) {
> > > > +mem = epc->windows[i];
> > > > +mutex_lock(&mem->lock);
> > >
> > > This is my feeling though, calling mutex_lock() in the loop seems
> > > to cause overhead. And, if we call mutex_lock() at out-of the loop,
> > > I think we can write single mutex_unlock() calling.
> > >
> > But the mutex is for each window, are you suggesting to add a global mutex ?
>
> Oops, that's right. So, I'd like to recall.
>
> > > > +size = ALIGN(size, mem->window.page_size);
> > >
> > > I'm sorry I should have realized this in the previous review,
> > > but overwriting this size is possible to cause an issue at second time or more loops.
> > > So, the first argument of ALIGN should be kept for the loop.
> > >
> > Could you please elaborate on this.
>
> My concern is the following.
>
> For example, the size of argument of pci_epc_mem_alloc_addr() is 4096.
> epc->windows[0].window.page_size = 8192
> --> then the size will be changed to 0.
>
> epc->windows[1].window.page_size = 4096
> --> since the size was changed to 0 on the first loop, the result is 0.
> But, this should be 4096.
>
> Does such a case never happen?
> (Or, is my understanding incorrect?)
>
Good catch, yes that needs fixing probably by having a local variable for size.
Cheers,
--Prabhakar
> Best regards,
> Yoshihiro Shimoda
>
>
> > > > +order = pci_epc_mem_get_order(mem, size);
> > > >
> > > > -*phys_addr = mem->phys_base + ((phys_addr_t)pageno << page_shift);
> > > > -virt_addr = ioremap(*phys_addr, size);
> > > > -if (!virt_addr)
> > > > -bitmap_release_region(mem->bitmap, pageno, order);
> > > > +pageno = bitmap_find_free_region(mem->bitmap, mem->pages,
> > > > + order);
> > > > +if (pageno >= 0) {
> > > > +page_shift = ilog2(mem->window.page_size);
> > > > +*phys_addr = mem->window.phys_base +
> > > > +((phys_addr_t)pageno << page_shift);
> > > > +virt_addr = ioremap(*phys_addr, size);
> > > > +if (!virt_addr)
> > > > +bitmap_release_region(mem->bitmap,
> > > > + pageno, order);
> > > > +mutex_unlock(&mem->lock);
> > > > +return virt_addr;
> > >
> > > As I mentioned above, if mutex_lock() is called at out-of-loop,
> > > we can use "goto ret;" here like the original code,
> > >
> > > > +}
> > > > +mutex_unlock(&mem->lock);
> > >
> > > and we can remove this.
> > >
> > > > +}
> > > >
> > > > -ret:
> > > > -mutex_unlock(&mem->lock);
> > > > return virt_addr;
> > > > }
> > > > EXPORT_SYMBOL_GPL(pci_epc_mem_alloc_addr);
> > > >
> > > > +struct pci_epc_mem *pci_epc_get_matching_window(struct pci_epc *epc,
> > > > +phys_addr_t phys_addr)
> > > > +{
> > > > +struct pci_epc_mem *mem;
> > > > +int i;
> > > > +
> > > > +for (i = 0; i < epc->num_windows; i++) {
> > > > +mem = epc->windows[i];
> > > > +
> > > > +if (phys_addr >= mem->window.phys_base &&
> > > > + phys_addr < (mem->window.phys_base + mem->window.size))
> > > > +return mem;
> > > > +}
> > > > +
> > > > +return NULL;
> > > > +}
> > > > +
> > > > /**
> > > > * pci_epc_mem_free_addr() - free the allocated memory address
> > > > * @epc: the EPC device on which memory was allocated
> > > > @@ -159,14 +229,23 @@ EXPORT_SYMBOL_GPL(pci_epc_mem_alloc_addr);
> > > > void pci_epc_mem_free_addr(struct pci_epc *epc, phys_addr_t phys_addr,
> > > > void __iomem *virt_addr, size_t size)
> > > > {
> > > > +struct pci_epc_mem *mem;
> > > > +unsigned int page_shift;
> > > > +size_t page_size;
> > > > int pageno;
> > > > -struct pci_epc_mem *mem = epc->mem;
> > > > -unsigned int page_shift = ilog2(mem->page_size);
> > > > int order;
> > > >
> > > > +mem = pci_epc_get_matching_window(epc, phys_addr);
> > > > +if (!mem) {
> > > > +pr_err("failed to get matching window\n");
> > > > +return;
> > > > +}
> > > > +
> > > > +page_size = mem->window.page_size;
> > > > +page_shift = ilog2(page_size);
> > > > iounmap(virt_addr);
> > > > -pageno = (phys_addr - mem->phys_base) >> page_shift;
> > > > -size = ALIGN(size, mem->page_size);
> > > > +pageno = (phys_addr - mem->window.phys_base) >> page_shift;
> > > > +size = ALIGN(size, page_size);
> > > > order = pci_epc_mem_get_order(mem, size);
> > > > mutex_lock(&mem->lock);
> > > > bitmap_release_region(mem->bitmap, pageno, order);
> > > > diff --git a/include/linux/pci-epc.h b/include/linux/pci-epc.h
> > > > index e0ed9d01f6e5..d5da11cf0f2a 100644
> > > > --- a/include/linux/pci-epc.h
> > > > +++ b/include/linux/pci-epc.h
> > > > @@ -65,20 +65,28 @@ struct pci_epc_ops {
> > > > struct module *owner;
> > > > };
> > > >
> > > > +/**
> > > > + * struct pci_epc_mem_window - address window of the endpoint controller
> > > > + * @phys_base: physical base address of the PCI address window
> > > > + * @size: the size of the PCI address window
> > > > + * @page_size: size of each page
> > > > + */
> > > > +struct pci_epc_mem_window {
> > > > +phys_addr_tphys_base;
> > > > +size_tsize;
> > > > +size_tpage_size;
> > > > +};
> > > > +
> > > > /**
> > > > * struct pci_epc_mem - address space of the endpoint controller
> > > > - * @phys_base: physical base address of the PCI address space
> > > > - * @size: the size of the PCI address space
> > > > + * @window: address window of the endpoint controller
> > > > * @bitmap: bitmap to manage the PCI address space
> > > > - * @pages: number of bits representing the address region
> > > > - * @page_size: size of each page
> > > > * @lock: mutex to protect bitmap
> > > > + * @pages: number of bits representing the address region
> > >
> > > Perhaps, we should not change the "@pages" line.
> > >
> > OK will drop this change.
> >
> > Cheers,
> > --Prabhakar
> >
> > > Best regards,
> > > Yoshihiro Shimoda
Renesas Electronics Europe GmbH, Geschaeftsfuehrer/President: Carsten Jauch, Sitz der Gesellschaft/Registered office: Duesseldorf, Arcadiastrasse 10, 40472 Duesseldorf, Germany, Handelsregister/Commercial Register: Duesseldorf, HRB 3708 USt-IDNr./Tax identification no.: DE 119353406 WEEE-Reg.-Nr./WEEE reg. no.: DE 14978647
